Subscribe now

It’s been a busy quarter since I wrote about Sid’s incredible journey to cure his own cancer. This story has gone completely viral and driven an overwhelming amount of interest from patients, regulators, scientists, and founders. How can we scale what Sid has accomplished?

My wife and I also welcomed Rosalind to our family. Amplify Bio has been growing, too. Exciting times all around. You can check out a short essay I wrote about our investment in Infinitopes, an Oxford spinout working to develop the next generation of breakthrough cancer vaccines.

Today, we’re going to explore the case for optimism about AI-powered drug discovery.

Enjoy! 🧬

Typically, a steep cost decline for a foundational technology leads to a Cambrian explosion of new applications and products.

Moore’s Law gave rise to personal computers, portable computers, smart phones, smart watches, and even more questionable devices like smart fridges.

recently analyzed in exhaustive detail how plummeting costs in the “electric stack” (batteries, magnets, power electronics, etc.) led to a similar proliferation of new engineering marvels.

This pattern hasn’t held for biomedicine. DNA sequencing has decreased exponentially in cost, alongside a whole slew of rapidly improving drug discovery technologies and therapeutic modalities. Despite that, there hasn’t been a meaningful increase in the number of new drugs approved each year, which hovers around ~50.1

And the cost to bring a new drug to market keeps rising. In fact, exponentially so.

In 2012, an equity analyst named Jack Scannell and his colleagues published an article showing that “the number of new drugs approved per billion US dollars spent on R&D has halved roughly every 9 years since 1950.”

Recently, Scannell showed how this phenomenon, coined “Eroom’s Law” (the inverse of Moore’s Law), has steadily continued even with the birth of biotech, the Human Genome Project, and the rise of AI drug discovery.

Based on these observations, Scannell and others have voiced skepticism about expecting any different for the next wave of AI breakthroughs in the life sciences. Every time somebody posts a hyperbolic message about AI revolutionizing pharma, Scannell quips, “At last I understand why it costs 100x more to discover a drug today than in 1950. Our computers have slowed down.” (Other good ones here and here.)

Recently, , a fellow geneticist and writer, has published several great articles outlining why the cost of clinical development—not pre-clinical drug discovery—is the major driver of Eroom’s Law. And she has detailed why policy rather than technology is likely the biggest lever to solve this problem.

Teslo is now a Fellow at Renaissance Philanthropy working to make her vision for Clinical Trial Abundance a reality. Her research here is excellent, and if you haven’t checked out her personal newsletter yet, you should.

Several founders have made the bet that technology can meaningfully reduce the time and cost of trials. I’ve written about how Vial aims to do this. Similarly, Formation Bio is focused on developing their own AI-powered clinical infrastructure.

When explaining the rationale for this focus, Ben Liu, the CEO of Formation, outlines a similar view that discovery is no longer the bottleneck. He tells a story of pitching an exciting academic drug candidate to pharma, only to be told, “We already have more good drugs discovered than we can afford to prosecute.” He goes on to say that “if there’s one competitive advantage that you want to have as a pharma company, ironically it isn’t in discovery. It’s actually in the drug development process.”

I agree with a lot of this.

Scannell’s meta-analysis is foundational work. Teslo’s efforts on regulatory innovation and clinical trial abundance are sorely needed. Vial and Formation are tackling real problems and have the potential to be generational businesses.

The rising cost of clinical development is the primary driver of Eroom’s Law.

At the same time, I spend most of my efforts as an early-stage biotech investor on discovery, with a focus on new technologies and approaches. I still think this has the chance to deliver real impact and great returns.

I want to make the case for both points here.

I think that AI will help make more and better medicines.

I also think that there are a lot of misconceptions about biotech investing, especially in new platforms, let alone AI platforms. Similar to Bruce Booth’s classic “defense of life sciences investing,” I want to unpack why this is an attractive investment.

Here, I’ll outline three arguments:

1. Pharma R&D efficiency is not the same thing as biotech R&D efficiency.

2. Better trial regulation and infrastructure would be a big win for little biotech.

3. Discussions centered on efficiency may miss the point when it comes to AI-enabled drugs.

To start, let’s examine the nature of the relationship between biotech and pharma.

On the Financialization of Pharma

Many of the largest pharma companies have been around for a long time and have wacky origin stories. I previously wrote about how some of these 100+ year old businesses started out making dyestuffs for textiles before using their chemistry infrastructure to make pharmaceutical products. Obviously, these businesses have evolved considerably. And the world around them has, too.

Skipping over some very interesting history, some of the most critical business model evolution for the pharma industry started in the 1990s. Three major trends collided to restructure how drugs are discovered, developed, and sold.

The first trend was the increasing financialization of the global economy, especially in developed countries. This term has been defined in many ways, but the core idea is simple. As markets mature, there is a trend towards the finance industry accounting for a larger percentage of economic activity over time.

Consider the U.S. economy. From 1980 onwards, the share in GDP of the financial sector has roughly doubled on a percentage basis.

Investment products have increased in complexity and investment firms have scaled massively. Pharma has not been immune from this evolution. The shares of the largest pharma companies are now primarily owned by the world’s largest institutional investment firms.

This change in ownership was accompanied by different incentives. Pharma companies came under greater pressure to increase efficiency and maximize value for shareholders. As a result, these businesses began to spend much more of their earnings on dividends and share buybacks. The percentage of pharma executive compensation that comes in the form of share options and share awards is now also much higher than most other industries.

This trend is generally referred to as the financialization of the pharmaceutical industry.

In parallel, the trend of exponentially declining R&D efficiency (since the 1950s!) showed no signs of slowing up. Pharma executives were forced to answer hard questions about what to do.

Faced with declining R&D efficiency, looming patent cliffs for some of the first blockbuster drugs, and shareholder demands for increased profits, several large pharma companies decided to consolidate.

Mergers combined revenue, cash, and resources. Circling the wagons, so to speak. Scale has advantages. But the decline in R&D efficiency didn’t stop. Another solution was needed.

This brings us to our third trend: the rise and maturation of the biotech industry. Breakthroughs in genetic engineering led to the first wave of biotechs such as Genentech, Biogen, and Amgen. These businesses focused exclusively on R&D. Their entire raison d'être was to bring a wholly new category of medicines to the market. By definition, these startups didn’t have the clinical infrastructure or commercial scale of big pharma.

A natural symbiotic relationship quickly emerged. Pharma companies began to hunt for innovation externally. And biotechs leaned into these partnerships for early financial stability and the existing channels to bring their products to market.

Over time, two important things happened. First, the aperture of biotech expanded far beyond recombinant DNA, encompassing a wide range of approaches to early-stage drug discovery. Second, pharma started divesting their own internal R&D and manufacturing efforts and focused more heavily on acquisition for new products.2

As one analyst wrote, “The large, complex organizations of the originators [pharma] are unsuited to fostering innovation. An ecosystem of venture capital has proven much more effective in selecting early-stage biomedical research opportunities. Essentially, venture capitalists today pre-finance the early-stage development for pharma companies.”

As my friend and fellow investor describes it, we are basically the “crazy guys” that stepped in to underwrite early-stage drug discovery risk. When stuff works, pharma can just buy it, rather than wasting a lot of money on R&D themselves.3

It’s worth underscoring how big this trend is. At this point, roughly two out of three Phase 1 trial submissions come from “emerging biopharma” companies with less than $200M in R&D expenditures and annual sales of less than $500M.

In her research on this topic, Joanna Shepherd, a legal scholar at Emory, wrote, “Whereas a few decades ago almost all drug discovery took place inside traditional pharmaceutical companies, today most drug innovation is externally-sourced from biotech companies and smaller firms. Internal R&D is no longer the primary source, or even an important source, of drug innovation.”

So, let’s consider the following take:

Pharma may have historical backlogs of molecules, but they are absolutely hungry for promising new experimental medicines. Over $1 trillion dollars have gone towards pharma M&A since 2010. This is how biotech works.

With this structure in mind, let’s turn our attention to the cost of drug development.

It turns out that modeling the cost of R&D is a contentious topic with estimates that vary wildly, ranging from $161 million to $4.54 billion per approved drug. The reason for the huge spread is that people don’t agree on what costs should be included. Methodologies are varied.4

A recent research report on this topic broke down the three components of R&D cost for a new medicine that makes it to market:

1. Out-of-pocket (OOP) costs of R&D incurred year by year.

2. The cost of capital combined with the number of years it takes to bring

   a new medicine from discovery to being sold on the market for use with

   patients.

3. The attrition rate, i.e. the percentage of successful medicines relative to

   the number researched at the outset.

When people say things like, “It only costs ~$200M to get a drug approved,” this is likely a claim about the OOP costs. When people say things like, “It costs ~$4B to get a drug approved,” this is likely a claim about R&D costs, plus the opportunity cost of not investing the money elsewhere, multiplied by an assumed clinical probability of success.5

For simplicity, let’s focus on OOP costs. We’ll consider the numbers from a report prepared for the HHS in 2024:

Focusing on the numbers across all therapeutic areas, non-clinical R&D was estimated to account for only 7% of spend. The authors of the report wrote, “The clinical phases of drug development (Phase 1, 2, and 3) are the largest contributor to total out of pocket development costs, comprising around 68 percent of total costs inclusive of post-approval studies.”

Looking at this picture, discovery looks like a drop in the bucket. One could even argue it makes more sense to invest in reducing the cost of post-approval studies, which cost more than three times that of non-clinical R&D!6

But in reality, that’s not quite how things work.

As we just outlined, there are normally two parties that split these R&D costs:

1. The small biotechs that fund discovery and early-stage proof-of-concept, and

2. The big pharma companies that buy assets (or partner) and fund late-stage development and post-approval studies.

Considering this fact, the relative percentages for a small biotech are quite different. The ~$90M for a phase 3 study (let alone multiple phase 3 studies) and the ~$40M for post-approval studies are often not their problem. On a relative basis, the cost of non-clinical development is actually a bottleneck.

The discovery phase can cost more than the first trial. In aggregate, non-clinical costs are ~30% of the total spend through Phase 2, which is typically the steepest point of value inflection and de-risking. And critically, the discovery phase is what dictates whether or not there will be any return at all! The quality of the science and the molecule determines whether or not anybody will show up to fund the proof-of-concept study.

And this ignores one of the most costly factors of all: time! When tackling exceptionally difficult problems, discovery campaigns can drag on for years, consuming more resources than early clinical studies. Numbers can soar past the $11.8M estimate for non-clinical work.7

Which means that more efficient discovery can have a big impact on biotech returns!

Biotech R&D efficiency and pharma R&D efficiency are not the same thing. To Ben’s point, development efficiency is a great competitive advantage for a pharma. For a biotech, the name of the game is efficient and differentiated discovery.

Remember, we are the “crazy guys” doing the discovery work that pharma largely divested! But the definition of insanity is doing the same thing over and over and expecting different results. If we can find new ways to attack this problem and more reliably generate value, it is a very attractive opportunity.

Thankfully, there are many examples of greatly improved biotech efficiency even while the macro trend of Eroom’s Law has continued.

One example is the use of human genetic evidence to develop drugs with much higher success rates. Drug mechanisms supported by genetic data are ~2-3x more likely to succeed. This is a clear example where a new measurement technology can greatly improve R&D success rates.

BridgeBio, a hub-and-spoke biotech that is exclusively focused on this opportunity set, has success rates far higher than industry averages.

It’s also possible to invent new drug modalities that have a higher clinical success rate. Alnylam is a canonical example. Their RNAi platform makes it possible to drug genetically validated targets with astonishing accuracy.8

In aggregate, biologics have consistently been estimated to have roughly double the success rate of small molecule drugs. If we think back to the three factors of R&D costs, halving attrition rates in our industry should make a dent on the economics of our industry in the long run.

New technologies can and do make an impact on R&D productivity.

This picture leads to a slightly more hopeful mindset when it comes to AI for drug discovery. Rather than assuming all prior R&D innovation has been for naught, the question becomes: can AI have as big of an impact as human genetics or biologics?

“AI for drug discovery” is a broad enough term to encompass many things. There are many efforts to use AI for both target discovery and drug design—across basically every drug modality you can imagine. And new ideas are constantly being tried.

Some of the first movers, like Recursion Pharmaceuticals and Insilico Medicine, emphasized the use of robotics, large-scale data capture, and compute to systematize drug discovery. These businesses moved the industry forward and are advancing internal pipelines of new medicines.

The subsequent generation of founders learned from the first wave and put forward new ideas.

Consider Viswa Colluru and Ron Alfa, two members of the “excursion” mafia. Viswa founded Enveda, which uses AI to explore the chemical space produced by living systems (natural products). In only seven years, Enveda has raced into the clinic with a large pipeline of new molecules, several of which are already showing strong results. Ron founded Noetik, which uses AI to get as much information as possible from valuable primary human tissue samples.

Both ideas are centered around acquiring the data that will deliver the most predictive validity for new AI models. Already, self-reported “AI biotechs” are proving they can make drugs faster and with greater clinical success. With each iteration, founders are testing new theses about what will drive the greatest improvement on both fronts.

Again, this will not be the biggest driver of cost reduction in late-stage clinical development. We should expect to see the impact first show up in improved returns for biotech investment, rather than the fully loaded costs for a big pharma company to get a new drug approved.

Let’s assume this happens. Biotech investment becomes much more attractive. More exciting new drugs make it through proof of concept faster.

Will we still be rate-limited by the number of drugs that big pharma can ingest? In theory, yes. Which is why little biotech may benefit the most from regulatory innovation for clinical development.

Building to Endure, Not to Sell

Let’s recap. Most early-stage innovation comes from biotechs. But the primary exit strategy for biotechs is to sell to pharma, rather than to incur the costs of late-stage development and commercialization themselves.

This is arguably a good setup for both parties. Biotechs can have a laser focus on R&D. And there are less duplicative efforts to build massive sales forces for every company. Pharma can amortize these costs, putting “multiple products in the bag of sales reps” in a way that single-product drug companies cannot.

But there are consequences. One major issue is that the preferences and risk tolerance of big pharma can bound the set of ideas that biotech companies pursue. In other words, taste can be top down rather than bottom up.

Consider the case of GLP-1 drugs. Over the last few decades, biotechs have increasingly crowded their efforts around the targets that have been de-risked and prioritized by pharma. Ultimately, it was two pharma companies, Eli Lilly and Novo Nordisk, that pushed forward the original breakthrough products in obesity and weight loss.

Yang wrote a great post about this dynamic, saying, “Investors rely heavily on pharma for exits, and there was little pharma M&A interest in obesity.” Once the first big products validated the market opportunity, early-stage investors financed the companies making the next generation of drugs in this space. We’ve already seen a string of big acquisitions rewarding this activity.

Relying on pharma as top-down tastemakers can stifle the emergence of new therapeutic ideas from startups.

This is not a minor issue. Biotech founders and investors overlooked the biggest product opportunity in the history of our industry because of this bias!

But there is also just a simpler numerical problem: there is a finite set of companies with the scale to consistently pursue late-stage clinical development. Each company can pursue a finite number of new products.

If this doesn’t change, why should we expect an increase in new drug approvals per year?

So let’s run a thought experiment.

We wave a wand. Poof! Clinical trials cost 10x less. Swish! Hundreds of new surrogate endpoints suddenly make many trials 5x faster. Pow! The cost and time to file a biologics license application (BLA) or new drug application (NDA) with the FDA goes from ~$4.68M and ~1-1.5 years to ~$300k and ~3 months.9

One possible outcome is that pharma companies could acquire and commercialize more drugs for the same total cost.

Another possible outcome—which I find quite exciting—is that we could see the birth of a new generation of enduring biotechs commercializing their own products. If it becomes more financially viable to launch your own product, why sell?

George Yancopoulos, the co-founder and CSO of Regeneron, has spoken openly about the dramatic increase in trial costs since the company’s inception. Founded in 1988, Regeneron is one of the rare biotechs to scale from a startup into a $100B+ company. In an interview with Matt Herper of STAT News, he said that it cost $10,000 per patient to conduct a clinical trial when the business first started. Now, it can cost as much as $500,000.

With this increase in costs, would it be possible to build Regeneron again in 2026?

In the conditions of our thought experiment, it would be much more likely.10

Recently, we have seen an uptick in the number of “first-time launchers” making the transition from R&D organization to commercial enterprise.

But this growth has primarily been a function of greater capital availability, not greater capital efficiency. In the last market cycle, low interest rates and strong pharma M&A activity drove more dollars into biotech. Valuations increased. At a certain point, the price pharma would pay was lower than what investors would pay. So more companies decided to go it their own and become commercial organizations.

Now let’s imagine the capital efficiency scenario: companies are able to get to product approval faster and at much lower cost. This could drive a similar dynamic where biotechs can finance at more attractive terms than what pharma can afford to pay.

If we make trials faster and cheaper, could we see an even bigger wave of first-time launchers?

What if biotech efficiency (early-stage R&D) and pharma efficiency (clinical development and commercialization) both improve?

Hyper-efficient discovery firms are able to vertically integrate and become full-stack drug companies.

Could we go from ~50 drugs a year to ~500?

The points I’m making here are admittedly a bit nuanced.

I’m arguing:

1. Better discovery matters. The market financing early-stage discovery is its own little universe that is less directly impacted by Eroom’s Law.

2. If we can make clinical development more efficient, the biggest impact could be on the early-stage companies that would be able to swim out of their local estuaries and compete in the open sea with their own products, like Regeneron once did.11

If discovery becomes more efficient, more money should flow into biotech. If development becomes more efficient, more biotechs should become enduring, independent companies. These companies have the potential to get very big, further reinforcing the flywheel of more capital into biotech.

Patients would get more medicines, faster.

One way to talk about investments is to outline the “bear case,” the “base case,” and the “bull case” scenarios. The bear case for AI biotech is that it underperforms more traditional discovery approaches. The base case is that AI-enabled companies do at least as well as their traditional counterparts, and some do much better. The bull case is that a huge tailwind of regulatory innovation enables some of the most successful players to break out and become generational drug discovery companies.12

But the ultimate impact of AI on biomedicine may have nothing to do with efficiency.

Digital Biologics

My friend , who runs research at Dimension, gets “annoyed at the concept of an AI drug.”

His point is valid. Our industry has developed a genuinely annoying habit of arguing about whether or not any “AI drugs” have been approved yet. Normally, people use this term to mean conventional drugs that have been developed using AI discovery tools. This opens the door to an elaborate semantic argument about which drugs do or don’t meet this bar. It isn’t very productive.

But there is a growing set of true AI drugs—those that consist of both algorithms and biology, bits and atoms. If Digital Biology is the discipline at the intersection of computation and the life sciences, let’s call this new category of products Digital Biologics.

Consider the following examples.

Moderna has developed a personalized mRNA cancer vaccine called intismeran autogene. Merck and Moderna partnered to test this vaccine in combination with Keytruda in clinical trials. So far, they’ve seen a 49% reduction in the risk of cancer recurrence or death after five years of follow-up in Melanoma patients.

How does the vaccine work?

First, a patient’s tumor is sequenced. Next, a machine learning algorithm sifts through the sequencing data and ranks the most likely cancer-specific antigens. Then, “Upon administration into the body, the algorithmically derived and RNA-encoded neoantigen sequences are endogenously translated and undergo natural cellular antigen processing and presentation, a key step in adaptive immunity.”

I don’t know about you, but that sentence gives me chills. Am I quoting a science fiction novel? Nope, a Merck press release. What a time to be alive.

What is the drug? The RNA molecules? But you don’t get those without the algorithm. So is it the algorithm? As I’ve argued, I think the answer is that it is both things. The drug is the composite process of computational design and molecular instantiation.

Another example is sasineprocel, a new cell therapy product for Parkinson’s disease being developed by Aspen Neuroscience. In a Phase 1/2a study, all eight patients treated with the experimental therapy showed improvement. Some responses were dramatic:

Videos of one patient, recorded before treatment and at the one-year mark, were particularly striking to the Aspen team. Before receiving sasineprocel, the patient struggled to stand from a chair, open a door and walk down the hallway, with a low voice and a muted facial expression, McDevitt said. One year after the treatment, he is “like a different person,” rising from the chair and walking straight out the door with a smooth gait.

“It’s night and day,” McDevitt said, with the video making many at the company emotional. “There were a lot of tears.”

Sasineprocel is a complex product that combines cutting-edge cell engineering and bioinformatics. The first step is to collect a small biopsy of the patient’s skin. The skin cells are then transformed into pluripotent stem cells (iPSCs), which are differentiated into dopaminergic neuronal pre-cursor cells (DANPCs). When injected, the aim is for the engineered cells to “restore dopamine function and rebuild neural networks.”

This intricate transformation process is carefully controlled by a suite of four machine learning algorithms powered by genomic data: PluriTest, NeuriTest, GraftTest, and DopaTest. Without a scalable and data-driven approach, it’s hard to imagine that this therapeutic strategy would be viable. As their Director of Bioinformatics said, these tools are “the final check point for making autologous cell replacement therapy a personalized medicine reality for people with Parkinson’s.”

Again, what is sasineprocel? It’s definitely not an invariant chemical structure with discrete IP. But is it just the neuronal pre-cursor cells that are administered? Or is it the composite process required to administer the cells, algorithms and all? It feels like the latter.

Intismeran and sasineprocel are both Digital Biologics. My prediction is that there will be more products like this over time, not less.

Other Digital Biologics may have a more distinct separation between platform and product. We may be able to front-load more of the AI to make off-the-shelf cancer vaccines. Exquisitely complex AI-designed cell therapies may be manufactured like other cell therapies.

But these drugs will start to feel like the “Move 37” moment where AlphaGo played a 1 in 10,000 move to pull ahead of the best human Go player. AI could help enable sophisticated multi-specific antibodies that were previously out of reach of even the best antibody engineers. The “AI drug” distinction will start to feel less arbitrary.

Bending the arc of Eroom’s Law is super important for our industry.

One of the most impactful levers of change is likely to be regulatory innovation. If America wants to remain competitive in clinical development, this is sorely needed. Hopefully it’s clear that I’m in complete agreement with the brilliant people working to make this happen.

My arguments here are threefold.

First, innovation in discovery is still critical. Biotech companies live and die by their ability to make breakthrough medicines. For the startups that deliver most of the new products in our industry, this is still a bigger bottleneck than the ballooning costs of late-stage clinical development.

Second, if we succeed in reducing trial costs, it could fundamentally alter the relationship between little biotechs and big pharma companies. We could see a new wave of full-stack drug companies like Regeneron. This could be how we see an order of magnitude more drugs get to market every year.

Third, efficiency is not the only story for AI drug discovery. Cancer patients are living longer and Parkinson’s patients are walking because of a new wave of Digital Biologics. These medicines aren’t exciting because of a quantitative shift in discovery timelines. They represent a qualitative shift in what is possible.

Faster and cheaper trials would enable more of this.

Should we focus on accelerating discovery or development?

The answer is yes.

Subscribe now

Today, we’re going to study Sid Sijbrandij’s incredible journey to cure his own cancer.

Enjoy! 🧬

Sid Sijbrandij is an information maximalist.

On October 14, 2021, GitLab, the business he had started in his upstairs home office in the Netherlands only ten years earlier, went public. GitLab started as an open-source collaboration tool for software developers. Sid led its transformation into a massive application that tracks every stream of information in the software development lifecycle.

GitLab makes a great product. Businesses of all sizes use it. But GitLab is arguably best known in the business community for something else: it is one of the largest fully remote companies in the world. With over 2,500 employees and a $6.4B market capitalization, they still don’t operate a single in-person office.

Achieving this required doing things differently. One example is a cultural norm that Sid refers to as “radical transparency.” Consider the GitLab Handbook. Now totaling over 3,000 pages, the handbook is a central repository (updated and maintained with GitLab’s own version control system) for every piece of company information you could imagine. Not only is it available to all employees, it is publicly hosted on the World Wide Web for all to see.

Sid developed a complex information processing system (GitLab’s unique operating culture) to scale the development of a complex information product (GitLab) built for creating and maintaining complex information products (software).

It’s safe to say that Sid really likes information. But on November 18, 2022, Sid got information that absolutely nobody wants.

He had cancer.

Sid had a lot to lose. At this point, he was a self-made billionaire entrepreneur who was happily married to his life partner of over 25 years. Suddenly, the six centimeter mass growing out of his upper spine threatened to end all of that.

Throughout 2023, Sid dutifully underwent a brutal care regimen that he can only describe in retrospect as “devastating.” His tumorous vertebrae was surgically removed and his spine was fused with a titanium frame. He underwent rounds of radiation and chemotherapy so intense that four blood transfusions were required to keep him alive.

Despite all of this, his cancer resurfaced in 2024. Sid says the message he got from his physician was basically, “You’re done with standard of care, maybe there is a trial somewhere, good luck!” But that wasn’t going to cut it for Sid. He wanted to live.

So he decided to go founder mode on his cancer.

Over the last two years, Sid has assembled a veritable SWAT team to navigate—and in many cases create—his care journey.

Many of the ingredients resemble GitLab.

The top layer of his care stack is a system of intensive information gathering and documentation that is not dissimilar from the GitLab Handbook. In a massive Google Doc entitled “Sid Health Notes,” he and his team compile detailed entries for every medical interaction or meeting with a cancer researcher or oncology company they take. The document has grown to over 1,000 pages for just 2025.

Hyperlinked within this Doc is the next part of the stack, which Sid refers to as “maximal diagnostics.” The raw data for every lab test, scan, and genomic sequencing result is meticulously stored. And there are a lot of results. “I’m doing every diagnostic I can get my hands on, and doing them often,” he says.

The final part of Sid’s care stack is the actual set of therapies he can take. After exhausting his options with the standard of care, he set out to create his own “therapeutic ladder” of possible treatments. Some rungs on this ladder are existing medicines that could potentially be repurposed from other cancers based on information gained from diagnostics. Others are individualized therapies designed in collaboration with researchers or companies specifically for Sid.

Sid developed a complex information management system (his care notes) to manage streams of complex information (his diagnostics) that guide the use and creation of complex information products (drugs).

Cancer is a disease of information. Loss of genomic information threatens to destroy the self. It may have met its match with Sid.

So far, the results of this regimen have been nothing short of extraordinary. Sid’s cancer is in remission. He has enough energy that he decided to start another software company, called Kilo Code, all while running a venture capital fund, a charitable foundation, and traveling the world with his wife.

Sid is very clearly an anomalous person. The approach to his care and the results are equally anomalous. A simplistic takeaway from this story could be, “Wow! A brilliant billionaire seemingly cured his cancer. Good for him!”

But as I’ve gotten to know Sid, it’s become abundantly clear to me that there is more to the story than that.

First off, his story serves as an existence proof of what is possible.

There hasn’t always been a better tier of healthcare available to the wealthy. As an extreme example, consider the British royal family, who for centuries experienced dramatically higher infant mortality rates than ordinary people living in the rural countryside. Greater access to healthcare was a bad thing.

But sometimes, better care is available. One example is Magic Johnson’s story. When he announced that he had contracted HIV at the height of his legendary basketball career, people mourned like he had just told the world he was going to die. Because at the time, HIV was practically a death sentence. But through a combination of early detection, new breakthrough treatments, and a strict diet and fitness regimen, Johnson is alive and thriving. His HIV is now undetectable.

This proved to be an early indicator of a dramatically improved treatment paradigm for HIV. While he likely benefited from exceptional care and medical access at the time, these treatments are now much more widely available. It was a case where “the future is here, it’s just not evenly distributed,” as William Gibson famously put it.

Sid’s story also reveals lessons about the barriers that limit our ability to reimagine cancer care.

Even with extraordinary resources and motivation, Sid has faced major obstacles at every turn. He still struggles to gain access to his own tissues and order diagnostics. When advocating for more aggressive treatment, he has had to contend with a system that equates the Hippocratic Oath of “first do no harm” with a lethargy that can feel like “first do nothing,” in his view.

But a rigid healthcare system hasn’t been the only bottleneck. There are also structural challenges with how drugs are discovered and developed. And in some cases, even with how drugs are defined.

In recent years, China has adopted a dramatically different posture towards testing experimental medicines that now threatens the future competitiveness of the American biotech sector. This has reopened the debate within the biotech industry about what evidence and standards should be required to test and approve a new experimental medicine. It currently costs far too much.

Most of this conversation centers around a scientific and regulatory paradigm where medicines have one chemical recipe that is given to all eligible patients. But with new technologies, it is increasingly possible to consider making bespoke therapies tailored for an individual patient. Researchers and patients have pushed the FDA to reconsider its regulatory stance towards these types of personalized therapies.

As Sid puts it, “It costs $1B to get a drug approved. But it costs $1M to dose a single person [with a personalized therapy]. That discrepancy is the highest it’s ever been. And it keeps growing because it keeps getting easier and easier to make new medicines while the cost of the Phase 3 keeps going up.”

Drug discovery has gotten less efficient over time despite extraordinary technological progress. There is now intense pressure to do something about that.

Considering these massive tailwinds, Sid’s story may be a harbinger of what’s to come rather than a curious human-interest story about one billionaire’s medical luck.

That’s the story we’re going to explore here.

Chapter One: “It Became My Own Job to Keep Myself Alive”

Sid has always loved to build things. After studying engineering physics and management science at the University of Twente, he took a job in 2003 at a small Dutch firm called U-Boat Worx that specialized in building personal recreational submarines.

He worked closely with the founder, Bert Houtman, who had taken an interest in “private submersibles” after a successful career as a software entrepreneur. While Sid loved the engineering challenges, it became clear that the growth potential for this industry was limited. So he started to spend his free time teaching himself the Ruby programming language and scrolling Hacker News.

This website gave Sid a chance to imbibe some of the early Web and startup culture that was forming on the other side of the Atlantic. Paul Graham and his co-founders had just launched Y Combinator (YC) in Cambridge, Massachusetts. YC created Hacker News to curate and discuss interesting links and ideas about technology and business.

After working as a programmer and starting his first startup, Sid stumbled on a Hacker News link that caught his attention. It was about a project called GitLab. Started by a Ukrainian programmer named Dmitriy Zaporozhets, GitLab began as an open-source application for programmers collaborating using the Git version control tool.

According to the GitLab Handbook, “He created GitLab from his house in Ukraine. It was a house without running water but Dmitriy perceived not having a great collaboration tool as a bigger problem than his daily trip to the communal well.”

Sid was excited about the project for a few reasons. First, he thought the Ruby implementation was beautiful code. He also thought it made sense for the source code of a code collaboration tool to be open, meaning that coders could collaborate on making the tool itself better. This was in contrast to other related applications like GitHub, whose source code was proprietary.

But the project was only a repository. If a programmer wanted to use it, they’d need to download the source code and host it themselves. Sid had an idea. He’d create a hosted version of the application called Gitlab.io to solve this problem.

That evening, while making some pancakes for dinner with his girlfriend Karen, he posted a link to sign up for a beta version of this product on Hacker News. While they were cooking, he’d occasionally dash up to his office to check if it was getting any traction. At first, it wasn’t.

And then one time, he didn’t come back downstairs. Karen brought up the finished pancakes and found him absorbed in the comment section of the post. It had made the front page of Hacker News. In the first three hours, over 150 people had signed up for early access. Sid could see the first faint glimmer of product-market-fit.

Next, he sent an email to Dmitriy telling him he was going to make this product. Reflecting a true open-source ethos, Dmitriy’s response was basically, “Great. Thanks for doing that!” And he asked for nothing in return. It was an open project. People could use it as they pleased.

Over time though, GitLab kept gaining traction and Dmitriy decided he wanted to work on the project full-time. Sid jumped at the opportunity and found a way to sponsor him with the proceeds of his tiny bootstrapped company. Eventually, Dmitriy became the co-founder and CTO of the company.

Things started to get more serious. More companies signed up for the service and they iterated on the right commercial offerings to complement the open-source project. And they were accepted into the winter 2015 (W15) batch of YC.

In the following years, GitLab’s growth was dramatic. By 2016, the product had millions of users and the company had raised a $20M Series B. Companies and institutions like IBM, Macy’s, ING, NASA and VMWare were paying enterprise customers.

At the risk of stating the obvious, this was well before the COVID pandemic had normalized remote work. In Silicon Valley, the expectation from investors was that startups needed to work in person. At every stage of growth, Sid heard this opinion.

But Sid prefers to make these types of decisions based on empirical data. He wanted to hire the best people. And the best people weren’t always in the valley. After all, his co-founder Dmitriy was still living in a small village in Ukraine.

Despite the early team’s geographic diversity, they all seemed to work together effectively. In fact, the product they were building was designed to make that possible. Short experiments in establishing an office presence didn’t seem to impact productivity. So they leaned into making the best possible fully remote culture.

Great companies reflect the preferences of their founders. As Guy Kawasaki, Apple’s former Chief Evangelist, put it, “Apple is Steve Jobs with ten thousand lives.” One major preference is how a company communicates. Amazon and Stripe are famous for their internal writing cultures. NVIDIA has a flat organization structure with a large number of executives constantly reporting directly to Jensen.

For Sid and the early GitLab team, it was natural to maintain company information in large wikis much like how open-source projects are typically structured. This became the GitLab Handbook, which is edited and maintained using GitLab itself.

There are simple rules for its use. If any employee has a question about GitLab, they should first consult the Handbook. If there is no answer, they need to ask around to figure it out and then contribute what they’ve learned to the Handbook. It is a collaborative, living document. Over the years, the page count—which is also programmatically tracked and shared within the Handbook—has grown to over 3,000.

As GitLab has grown, they have trended towards even greater transparency and information sharing. The GitLab Unfiltered YouTube channel has now posted over 13,000 videos of internal meetings. Fairly banal recordings of code reviews or explanations of GitLab’s architecture are posted for all to see. There is seemingly no unit of information too small to record and share.

When the world was forced to go remote during COVID, all of GitLab’s geeky quirks and processes became a fascination of the business community. Multiple Harvard case studies were written about the business. Sid became a regular guest on TV shows and podcasts to talk about how to manage distributed teams.

By 2022, Sid was the CEO of a fast-growing public software company. His net worth had surpassed $1B and he was a widely studied and respected expert on remote work. And his long-term girlfriend Karen was now his wife. Things were going pretty well.

And then one day, during his typical workout, he felt a sharp pain in his chest. This wasn’t the first time he’d felt it, so he wasn’t too concerned. Typically, it went away. But this time, it got worse. Two weeks later, he found himself laying awake at four in the morning, still in pain. He decided to go to the emergency room.

After a workup and an X-ray, he was sent home. Later that morning, his doctor gave him a call and asked, “Do you know how to meditate?” His blood pressure was severely elevated from the pain. His doctor was concerned he might be suffering from an aneurysm.

Back at the hospital, there was good news and bad news. The good news was that he didn’t have an aneurysm. But the bad news was that he did have a six centimeter mass extending from his T5 vertebrae.

Sid was diagnosed with osteosarcoma, a form of bone cancer that is rare for an otherwise healthy 45 year old to have. The first order of business was to get the cancer out. A surgeon successfully removed the cancerous vertebrae and fused his spine with a titanium frame.

With the tumor removed, Sid underwent an aggressive version of the standard of care for his cancer type. He did stereotactic body radiotherapy (SBRT), several rounds of intensive chemotherapy, and added proton beam therapy.

The impact was brutal. The chemotherapy was so intense that he needed four separate blood transfusions. Like all chemo patients, he lost his hair. For weeks, he could barely get out of bed to go to the toilet. “It destroyed me,” Sid told me.

And these were just the temporary effects. His heart is less flexible now and he is anemic. Sid says that he is “stupider now,” referring to the toxic impact of systemic chemotherapy on his brain. (I’d note that his intellect is still clearly far out of distribution relative to most people.)

In his first round of treatment, Sid deviated only once from more traditional care. José M. Mejía Oneto had founded an oncology startup called Shasqi and gone through the same YC batch as Sid in 2015. Well before he was ever a patient himself, Sid had befriended José and invested in his company.

In 2020, Shasqi became the first YC company to start a clinical trial. They were testing new chemical strategies to make chemotherapy more targeted. But the trial wasn’t designed for Sid’s cancer type. As Sid struggled through systemic chemotherapy, he worked with Shasqi to file an investigational new drug (IND) application with the FDA for a trial where he would be the only patient. The application was approved and he was able to use the technology.

For two years, regular checkups confirmed that Sid’s cancer remained in remission. But during a checkup in 2024, imaging showed evidence of local recurrence. The cancer was back.

Sid had reached the end of the road for the standard of care. He’d had surgery. Radiation. The most aggressive chemotherapy possible. And the cancer was back.

To Sid’s surprise, his medical team didn’t have much else to say. It wasn’t unusual for a cancer to outsmart therapy and recur. It was the norm. But in Sid’s programmer brain, this didn’t compute. He wanted to live.

Because the diagnosis with osteosarcoma at his age is rare, he didn’t meet the inclusion criteria for any clinical trials. It dawned on Sid that without taking extreme measures, he would die. “It became my own job to keep myself alive. Nobody else was going to do it for me at this point,” he said.

At the end of 2024, Sid transitioned from CEO to Executive Chair of GitLab, saying, “I want more time to focus on my cancer treatment and health.” He was beginning to reorient his life to be able to go “all out” to cure his cancer.

Just a few months earlier, Paul Graham had written an influential essay about what makes founder-led companies so special. Entitled Founder Mode, Graham reflected on the fact that great founders often embed themselves more deeply in the details than hired professional managers. He wrote, “Whatever founder mode consists of, it's pretty clear that it's going to break the principle that the CEO should engage with the company only via his or her direct reports.” Many crucial decisions lurk below that level of abstraction.

Sid was leading GitLab in founder mode and his cancer care in manager mode. It was time to flip that before it was too late.

Chapter Two: “I’ll Talk to Anyone, I’ll Go Anywhere, and I Can Be There Anytime”

Programmers and oncologists have overlapping terminology. The decision trees for treatment regimens based on a patient’s biomarkers are referred to as algorithms.

These algorithms are often depicted in flowcharts. Here’s an example of what the current standard of care looks like in esophageal and gastric cancer.

Based on clinical staging (e.g. cT1, cT2) determined by doctors using scans or other diagnostics and pathologic staging (e.g. pT1, pT2) determined by pathologists using tissues, different combinations of treatment are applied. If a cancer progresses or metastasizes, more sophisticated biomarkers can guide more sophisticated therapies.

Sid had exhausted the treatment algorithm for his osteosarcoma. Thinking about the problem from first principles, he knew that he needed more advanced diagnostics to make decisions about what to do next. And he needed to find or develop new therapies to take based on that information.

He established some principles for his next phase of care:

1. Maximal Diagnostics. The goal is simple: do every diagnostic available as often as possible. Just like GitLab’s operating culture, no unit of information is too small to document.

2. Make 10+ Personalized Treatments. Since no more treatments were available, Sid scaled up his engagement with companies and academic researchers to tailor new drugs based on what he learned about his cancer.

3. Treatments in Parallel, Not Serial. Typically, a single therapy is used until it’s clear there is a need to change course. Instead, Sid aims to iterate faster, testing multiple therapeutic hypotheses in quick succession and empirically measuring the response with diagnostics, “rather than waiting for my cancer to progress.”

Each principle radically differs from how oncologists typically approach care. So he began to assemble a team willing to embark on this experiment with him. The first pillar of his new team included private concierge care practices and services—such as Private Medical, Private Health Management and Pathfinder Oncology—that could help him scale up and manage his diagnostic regimen.

Sid’s new suite of diagnostics can be divided into five primary categories.

Every cutting-edge genomic tool is deployed.

He uses single-cell sequencing instruments developed by 10x Genomics to measure cell-type specific expression of genes that could potentially be targeted by treatments. This technology also resolves the exact T-cell receptors (TCRs) of the T cells that have infiltrated his tumor, potentially enabling highly targeted immunotherapies.

This is paired with “bulk” DNA and RNA sequencing where his cancer cells are measured in aggregate to assess the broader mutational landscape of his tumor.

But that’s not where sequencing stops. He routinely uses minimal residual disease (MRD) tests from multiple providers, which scan for trace amounts of tumor DNA that can end up circulating in the blood. These tests have varying sensitivity and specificity but can be used as an early signal for disease recurrence, especially if multiple tests are positive multiple times.

Another more experimental tool in the toolbox is the use of assays for modeling the potential response to a drug in the lab. Sid has initiated collaborations with various academic labs developing different organoid systems for this purpose. He has even embarked on making organoids using his own cancer cells.

The last major pillar is the use of pathology stains to confirm the most promising genomic hypotheses in tissue samples from his tumor. Blood samples are much easier to collect than tissue samples, which are a more finite resource. So this step is done less frequently and with more consideration.

As Sid began blazing this trail, he reconnected with Jacob Stern. The two had been connected by José at Shasqi earlier, but they now shared an obsession: single-cell sequencing. Sid was a personal power user of 10x Genomics and Jacob was a director at the business.

When Jacob heard about how Sid was using their single-cell technology, he invited him to visit 10x and speak about it. As Jacob helped prep the talk, Sid showed him his health notes. Jacob was blown away. “I realized that this guy was living thirty years in the future,” Jacob said. It was the most exciting application of 10x technology he’d ever seen. Sid was making the mission of “mastering biology to advance human health” a reality for himself.

Sid recruited Jacob to join him full-time, effectively becoming the CEO of his care. Jacob now sequences and orchestrates the complex symphony of Sid’s diagnostic and therapeutic regimen. He liaises with hospitals and physicians. Meetings with researchers and startup founders pack the remainder of his calendar. No stone is left unturned in pursuit of nuggets of information that could lead to a new treatment.

Throughout 2025, Sid and Jacob have moved extraordinarily fast. With the support of the broader team, which consists of concierge medical services, a clinical advisory board, and a scientific advisory board, Sid’s armamentarium of treatment options has expanded from zero to dozens.

With all of this activity, it can be challenging to get a sense for what decisions actually look like in practice. But zooming in, one diagnostic discovery helped credential a therapeutic decision that made a particularly outsized impact.

When expert computational biologists analyzed Sid’s single-cell data, they made an important observation: many of the genes with the greatest differential expression relative to healthy cells were canonical markers of fibroblasts.

Fibroblast cells are widely distributed throughout healthy tissues and play a critical role in wound healing. But tumors can warp the wound healing pathway to gain mass and scaffold a protective matrix within the tumor microenvironment, acting like a wound that never heals. Sid’s cancer appeared to have learned this trick.

In parallel, one of the concierge medical services had surfaced an experimental therapy in Germany that directly targets FAP, one of the fibroblast-related genes that was shown to be up-regulated. Soon, Sid was en route to Germany. As Sid says, “I’ll talk to anyone, I’ll go anywhere, and I can be there anytime.”

The experimental treatment being developed in Germany was a radioligand therapy. These medicines consist of two parts. The first is a targeting compound called a ligand. Attached to the ligand is a radioisotope, which is an unstable atom that undergoes decay, emitting radiation. The goal is for the ligand to precisely deliver this infinitesimally small nuclear bomb to cancer cells, triggering their death.

A critical feature of radiotherapy is that the targeting ligand can be used as both a diagnostic and therapeutic tool—which is why it is sometimes called a theranostic. First, the ligand can shepherd a “cold” isotope that can be used for imaging. Sid’s tumor lit up. In conjunction with the single-cell analysis, the localized imaging response gave Sid confidence to move forward with the therapy.

Next came the “hot” payload. In Sid’s case, this was Lutetium-177 (Lu-177), the same warhead used in Pluvicto, one of the first breakthrough radiotherapies. Sid was dosed with the medicine and then quarantined for two days. He was given a small handheld monitor to track his radioactivity levels. For two weeks after the treatment, his urine was also carefully collected and discarded as his body flushed out the isotopes.

While radiotherapy can sound extreme, Sid found it dramatically more tolerable than chemotherapy. The experiences were as different as night and day. Rather than spreading throughout his entire body and wreaking havoc, the radiotherapy primarily concentrated at the site of his tumor. After getting out of quarantine, Sid got busy planning a wine tasting to celebrate.

And in the coming months, there would in fact be a lot to celebrate. The radiotherapy worked. Sid’s cancer shrank enough to become surgically operable again.

With fresh tissue from the surgery, they reanalyzed the immune cells within the tumor. It was a chance to empirically measure the response to the immunotherapy drugs that Sid had been taking in parallel.

The results were striking.

One measurement is what percentage of infiltrating cells are T cells. These cells are critical members of the adaptive immune system, which can learn to recognize persistent threats and establish immune memory. Most immunotherapies—including the checkpoint inhibitor that Sid had taken—work by boosting the anti-cancer T cell response.

When Sid’s cancer had returned, only 19% of the infiltrating immune cells measured by single-cell sequencing were T cells. After the surgery following radiotherapy, 89% of the cells measured were T cells. It seemed that the checkpoint inhibitor, a neoantigen peptide vaccine, an oncolytic virus, and the radiotherapy itself, had set Sid’s anti-cancer immune response into hyperdrive.

Now Sid and his care team are playing a different game. Like Magic Johnson’s HIV, Sid’s cancer is no longer detectable. But with a motto of “Stay Paranoid,” they are continuing to build out infrastructure even with negative test results.

Sid’s next course of treatment is an mRNA-based personalized neoantigen vaccine with the goal of sustaining the immune response they have already elicited. Of course, this treatment will be coupled with an in-depth monthly blood review to assess its impact.

As a backup to the backup plan, they are also developing personalized cell-based therapies in collaboration with leading academic groups. Based on more diagnostic data, the plan is to equip these cells with cutting-edge genetic logic gates that trigger cell killing in response to multiple signals rather than just one—another example of oncology and computer science blending together. But the goal is to never need these “nuclear options” if they can prevent recurrence with the less aggressive tools in their arsenal.

All of this is nothing short of remarkable. “Sid’s is a unique case,” Jacob said. Few people in the world have the resources or sophistication to pull even a fraction of this off. Sid has also made it clear that he would rather die from treatment than disease, which is an aggressive and personal stance.

But extreme cases like this often carry a residue from the future—especially where technology is concerned. Consider the computer revolution. The first computers filled entire rooms. They were only accessible to a select clergy of experts capable of operating them. Now we all carry more computational power in our pockets than NASA engineers had access to when landing men on the moon.

Sid’s case reveals the power of frontier cancer therapies. Is it possible that the average cancer journey of the future comes to look more like Sid’s? If so, what are the current bottlenecks that could hinder that transition?

Chapter Three: “I’m the Kool-Aid Man Breaking Through the Wall”

Sid and Jacob are both technologists. They can’t help but think of the ways that lessons from Sid’s journey could be scaled. It’s in their very nature.

When I first met Jacob, we went for a walk. As we wandered along The Embarcadero, spanning San Francisco’s eastern waterfront, we talked about Sid’s case. I asked him how much of this he thought could be repeated. He told me a story.

Before Jacob joined 10x, he’d spent time at Palantir earlier in his career. Although Palantir’s market capitalization is approaching $500B, few people have a great sense for what the company actually does. In her article, What Does Palantir Actually Do?, Caroline Haskins argued that this may be because “unlike consumer-facing startups that need to clearly explain their products to everyday users, Palantir’s main audience is sprawling government agencies and Fortune 500 companies.”

Palantir builds software infrastructure for these incredibly complex organizations. A big part of doing this is figuring out what these organizations need. Once a solution is engineered, it needs to be installed. To help make both steps easier, Palantir pioneered a role called the Forward Deployed Software Engineer (FDSE) that embeds within an organization.

FDSEs learn the shape of the problem and bolt on the first solution as quickly as possible. Over time, the solution can be refined, automated, and scaled. Many of Palantir’s features originate from lessons learned in “deployment.” Jacob told me about this philosophy. First, learn the problem. Next, solve it. Finally, harden the solution and scale it.

Similarly, Sid and Jacob and the broader team have accumulated immense firsthand experience navigating the Leviathan that is the American healthcare system. They now have a map of the territory. In order to solve problems for Sid, they’ve had to craft many rough solutions. In the future, these solutions could be scaled.

To understand how to scale a radically different approach to oncology, it’s worth examining the problems they faced when the rubber met the road.

First off, oncologists and the hospitals that employ them are not accustomed to patients going founder mode. They are used to manager mode, where patients delegate responsibility and autonomy to physicians. Oncologists determine what tests are ordered and what treatment is administered. Deviating from the standard path is met with resistance.

Consider diagnostics. Sid’s “maximal diagnostics” strategy is unusual. Achieving this ethos has required two big differences from standard clinical practice.

First, some of the more involved assays require cryopreserved tissue samples. Typically, hospitals only collect formalin-fixed paraffin embedded (FFPE) samples, which are helpful for conventional pathology reports, but destroy most of the nucleic acid information within the tissue. Second, Sid wants access to these tissue samples in order to generate data and store it.

Sid and his team were astonished by the difficulty of achieving both tasks.

Pathology departments prefer to run standard protocols. Cryopreservation is not standard. Accommodating this introduced novelty and risk. What if they made a mistake? Would they be liable?

And hospitals are very familiar with collecting samples and running tests. They are much less accustomed to collecting tissue samples and handing them off to patients. In many cases, there are simply no procedures to do this. Again, what is the liability associated with doing this?

Sid described the “incredible struggle at every hospital” to get access to his own tissue. Doing so has required “forward deployed tissue extractors,” who navigate hospital bureaucracy and don’t take no for an answer until they get sign-off and transfer. It is still a major struggle to access and move around historical samples.

Gathering genomic data has been another challenge. After the completion of the Human Genome Project in 2003, genomic technologies underwent a cost reduction more extreme than Moore’s Law. What cost billions of dollars can now be accomplished for less than $1,000. Still, whole-genome sequencing hasn’t become a fixture in the standard of care in oncology, a disease of the genome.

Jacob describes it as “shockingly hard” to obtain sequencing data beyond standard reports. For research labs, companies like Plasmidsaurus offer overnight plasmid sequencing. But using this type of service in a clinical setting requires jumping through hoops and cutting red tape.1

Part of this is arguably for good reason. There is a real fear that our sequencing capabilities have outstripped our ability to make sense of the resulting information. Even if we could, it’s unclear that there will be actionable treatment decisions available. It’s more pragmatic to stay under the lamp post where we can see—and to use the most well understood treatments.

But imagine trying to land this argument with Sid, the information maximalist. He obviously wants the information. And just like the GitLab Handbook, he expects that if there isn’t an answer already, they can figure it out and write their own.

When considering this problem, a natural set of questions emerges. Why is the Oncology Handbook so sparsely populated? Shouldn’t the genomics revolution have resulted in a Cambrian explosion of new treatment options? Why has measurement outstripped treatment to such a large degree? As Ruxandra Teslo asked, Why haven’t biologists cured cancer?

Part of the answer is that cancer is an extraordinarily complicated disease. Cancers are not static targets. They can evolve to evade even the most promising therapies.

But after his care experience, Sid has joined the growing group of people asking hard questions about the efficiency—or lack thereof—of biopharma innovation. As I frequently cite, in 2012, Jack Scannell and his colleagues observed that “the number of new drugs approved per billion US dollars spent on R&D has halved roughly every 9 years since 1950, falling around 80-fold in inflation-adjusted terms.”

This phenomenon, now widely known as “Eroom’s Law,” leads to an upside down picture where the exponential biotech revolution coincides with an exponential decline in the efficiency of pharmaceutical innovation.

Scannell and many others have attempted to outline the causes.

Much like literature, drugmakers need to compete against the entire canon of work that has come before them. And they do so in a much less favorable regulatory environment. The Eroom’s Law authors wrote, “Each real or perceived sin by the industry, or genuine drug misfortune, leads to a tightening of the regulatory ratchet, and the ratchet is rarely loosened, even if it seems as though this could be achieved without causing significant risk to drug safety.”

But let’s set aside discussion of the causes for now. The fact of the matter is that it’s become exorbitantly expensive to develop new drugs—especially for oncology. A recent analysis estimated that between 2017 and 2020, the average cost to develop a new oncology medicine was $4.4 billion.

This warps the incentives of drugmakers. One needs to see a path to astronomical peak yearly sales—on the order of billions of dollars—to justify this type of R&D expenditure. Anything less than a blockbuster hit doesn’t move the needle.

What this means in practice is that it’s become “so expensive to generate evidence that we don’t generate evidence for very much,” as Sid puts it. Several times now, Sid’s team has discovered promising experimental medicines that they’ve had to rescue from companies on the verge of bankruptcy. Drugs that may mean the difference between life and death for one patient can be discarded because of underwhelming financial estimates of their broader market potential.

And it’s hard to argue with this logic. In a paradigm of scarcity, we should be using our resources to advance the assets with the greatest chance to help as many people as possible. But there might be another way.

What if we could scale the development of medicines tailored specifically for a given patient?

There are two broad buckets to consider for this paradigm.

The first is drug repurposing. Sid has had a lot of success with this approach already. Guided by maximal diagnostics, Sid’s team has identified several drugs under development for other cancer indications that seemed likely to also improve his disease. And there are surprisingly flexible regulatory pathways already in place to accommodate this.

So far, using the Individual Patient Expanded Access - Investigational New Drug Application (IND), known as Form 3926, Sid has gained access to five experimental medicines where no trial would have otherwise been available to him. In each case, the FDA has accepted his application within 48 hours. “The FDA wants me to live,” Sid says.2

Surprisingly, hospital institutional review boards (IRBs) have been more challenging to navigate than the FDA. Sid describes it as a “vetocracy” where one member of the board can block treatment based on even the smallest concern. Large hospital systems, which are concerned with adhering to procedures and avoiding liability, have strong organizational antibodies to mitigate any threat of risk. Still, Sid has found drug repurposing to be a powerful tool in his arsenal.

Beyond repurposing, there is the vision of designing a drug from scratch based on a molecular understanding of one’s disease. This is typically referred to as personalized medicine. So far, Sid has initiated the development of multiple new experimental medicines in collaboration with academic research groups and startup companies.

But there are other examples where it’s possible to see a path to more scalable personalized medicine. One example is the personalized use of CRISPR for genetic disease. In June of 2024, I wrote about the vision, passionately spearheaded by Fyodor Urnov, to use CRISPR’s inherent programmability to design custom therapies for individual patients. In parallel to the scientific and clinical advances, the FDA proposed a new regulatory framework to more flexibly accommodate this type of programmable platform technology.

In May of 2025, this all came together in spectacular fashion. A baby boy in Philadelphia suffering from a rare metabolic disease was dosed with the first ever personalized CRISPR therapy.3 Baby KJ, as he is called, went viral. His story was picked up by every major news outlet. In response, leaders at the FDA proposed a new “plausible mechanism pathway” to enable more treatments like what KJ received.

Oncology is undergoing a similar renaissance. CAR-T therapy, which is one of the biggest breakthroughs in cancer care in decades, is not a drug with a universally defined chemical structure or biological sequence. Instead, it is a process for genetically engineering a patient’s own immune cells to become sentinels on the hunt for cancer. This also required regulatory flexibility.

Similarly, the type of personalized neoantigen vaccine that Sid is going to take is another striking example of a custom tailored molecular medicine. Let’s consider the workflow of this therapy, which I’ve written about before.

A patient’s cancer is sequenced. This data is piped into a bioinformatics pipeline with complex machine learning algorithms that predict which antigens should be used in the vaccine. These designs are then physically instantiated in personalized lots and packed into lipid nanoparticles—much like the type of medicine that baby KJ received.

In December of 2023, Moderna published incredible clinical results showing that their version of this process, in combination with a checkpoint inhibitor, cut the risk of cancer recurrence or death nearly in half for melanoma patients.

While these results are remarkable, there is still enormous room for improvement. Much like early computers, these therapies are still too expensive and slow for broader adoption. And they still aren’t delivering the types of complete responses that CAR-T therapies have now repeatedly achieved.

Decade, a company we partnered with at Amplify, is on the case. While they are still in stealth, the long-term vision set forth by the founders, Devin Scannell and Brian Naughton, is to enable “programmable radiotherapy.” The idea is to use many of the tools pioneered by the vaccine developers to make custom radiotherapies like what Sid received in Germany much more commonplace. If they succeed and the process is as potent as we hope, it’s the type of treatment I’d want for any member of my family.

Zooming out, all of these new personalized platform technologies represent a major shift. Remember: a real concern with collecting rich genomic data for a cancer patient is that we won’t be able to do anything with it. Now, there are approaches being developed to close the loop. Platform technologies are being developed to ingest exactly this type of information and to spit out custom drugs.

One of Sid’s biggest problems is that he is grasping at a future that hasn’t quite come into fruition yet. Or as he likes to say, “I’m the Kool-Aid Man breaking through the wall.”

Analyzing his data has required hiring computational biologists with graduate degrees. Repurposing medicines has required interactions with the FDA and hospital IRBs. Developing a personalized therapy has required sweating every little detail about the manufacturing process. He couldn’t have done any of this without his resources and team.

But let’s imagine what this could look like, in ten or twenty years, at scale.

An individual spots an abnormality on a consumer cancer diagnostic test, from one of many providers such as Prenuvo or Function Health.4 They decide to sign up for a consumer concierge oncology platform, which charges $1,499 for an initial workup.

Once they are onboarded and tissues are transferred to their personal biobank, the oncology service fires up some AI agents. After processing the individual’s entire medical history, the diagnostic agent orders a few more tests. Once returned, the results are passed off to the bioinformatics agent who generates a PhD’s worth of analysis in a matter of hours before distilling down the core details into a succinct five page report.

At this point, it’s clear the individual has a small cancer in their pancreas. But it’s early enough that surgery isn’t required. Based on the specifics of their tumor microenvironment and the exact antigens that are present, the drug hunter agent selects a checkpoint inhibitor (now generic, it costs less than $1,000) and an off-the-shelf cancer vaccine to rev up the immune system. Supercharged T cells hunt down any rogue cancer cells straying from the pancreas and kill them.

No chemotherapy is ever used. A custom radiotherapy is used to obliterate the physical tumor. Of course, imaging is used to confirm safety and tumor targeting before the cancer-killing warhead is ever delivered.

The total bill from start to cure is $175,000. Part of this is covered by the pool of money for discretionary health spending in their CHOICE account. The remainder is picked up by their health insurance provider—who is happy to save a considerable amount relative to the prior average cost of treating a pancreatic cancer patient, which was ~$250,000. Everybody wins.

Let’s be clear about a few things.

First off, despite starting my career as a cancer researcher, I am not a trained oncologist. And neither is Sid. Therefore, none of this should be construed as medical advice.

For example, Sid’s choice to test multiple drugs at lower doses in parallel is a personal one. While he felt it greatly reduced the side effects of any one drug, this is not how oncologists currently deliver care. The same can be said for all of his diagnostic testing.

This is also not an elaborate conspiracy about doctors or pharmaceutical companies hiding a secret cure for cancer behind closed doors. Reality is far more nuanced than that.

This is a case study. Just like business school professors devoted themselves to studying the operational practices that Sid established at GitLab, we should study the personalized protocols that he implemented during his extraordinary care journey.

Sid’s story has left me with a visceral feeling of William Gibson’s quote that we considered earlier: “The future is here, it’s just not evenly distributed.”

Widely considered the father of the cyberpunk genre, Gibson is one of the most prolific and influential living science fiction authors. His body of work sometimes reads like a detailed examination of this one central idea. Gibson’s cyberpunk cities are disorienting landscapes where people use cutting-edge virtual reality headsets while living in decrepit project apartments. Even sci-fi space habitats contain rooms filled with ancient relics.

The future always crawls clumsily out of the past.

This is what American healthcare feels like. I get my own medical care through a large health maintenance organization (HMO) in the Bay Area. Bugs in the janky Web portal frequently make it impossible to book an appointment or access my records.

But when I am seen, the care is often remarkable. My wife and I are expecting our first child, who we’ve already seen in stunning detail with futuristic imaging technology that has become totally routine. The experience was especially surreal because the machines were operated in tiny rooms full of charts and furniture that look like they haven’t changed since the time my mother would have been visiting for her own prenatal appointments.

Sid’s story is the most extreme example of this. He is living in the future. Many of the technologies used in his care represent the frontier of medical science. But navigating our existing labyrinth of a system to gain access to them has required herculean effort and persistence.

And that’s where things are. Oncologists want to cure their patients. Researchers and companies are making big strides in inventing the tools to help them do that. And regulators want to implement guidelines and pathways that keep pace with innovation.

The challenge is to line up all of these vectors of progress in order to achieve large-scale structural change. It can feel like every biomedical breakthrough is absorbed by a black hole that renders it invisible to any metric of progress. We’ve come to accept measurements of clinical progress in oncology on the scale of months, rather than years—let alone lifetimes.

Maybe the biggest lesson from Sid’s story is to be hopeful that the future is actually closer than it can sometimes feel. Things have a surprising way of happening gradually and then suddenly.

Thanks for reading this essay about Sid’s care journey. This wouldn’t have been possible to write without Sid’s “radical transparency” about his experiences. I’ve learned a lot from both Sid and Jacob, so I hope this essay proves to be of value for you, too. If you want to learn more, it’s worth checking out a website, sytse.com/cancer, that Sid and his team have put together. There are more incredible scientific vignettes and data resources than I was able to cover here.

If you don’t want to miss upcoming essays, you should consider subscribing for free to have them delivered to your inbox:

Subscribe now

Until next time! 🧬

Subscribe now

Today, we’re going to consider why it might now be possible to build big AI infrastructure businesses in biology.

Enjoy! 🧬

Mike Tyson famously said that “everyone has a plan until they get punched in the face.” As the COVID bubble burst, we learned that “every biotech has a plan until the market collapses.”

The pandemic spurred enormous enthusiasm for biotech investment. Stocks soared and many new companies were listed. I’m not going to belabor the point, but this catalyst was temporary. High-flying companies found themselves in a world of hurt.

We are still living through the aftermath right now. While the S&P 500 has more than doubled over the past five years, the XBI has failed to fully regain its value over the same time frame.

But the biotech industry has fundamentals on its side. Patent cliffs are still looming and the world’s armamentarium is still far from complete. Nothing about the current market cycle changes the fact that we are living through a renaissance in biology and technology. Our tools for understanding and engineering life are compounding.

And few industries offer a more direct way to produce financial returns by advancing scientific and human progress. In other words, it’s still a beautiful way of doing well by doing good.

Investors had to figure out how to get back on the horse. The return of capital to the sector happened in fits and starts. In 2023, there was a big contraction in investment relative to the hyperactive cycle of 2020-2022. Less deals happened.

When dealmaking returned, it happened unevenly. In 2024, we saw a spree of $100M+ “mega-rounds” happen. Large syndicates formed to insulate companies from market volatility. Basically, a small set of anointed companies raised a lot of money. Investors seemingly sought safety in numbers.

Many of these bets reflected viewpoints on where the next big blockbuster will be. Immunology, neuroscience, and oncology were major focus areas. Next-generation modalities like in vivo cell therapies and T-cell engagers were also palatable.

Another major theme—which is also one of the major themes of this newsletter—has been AI for biotech. Massive new investments are being made.

In April of 2024, we saw a $1B “Seed” round go to Xaira, a business aiming to reimagine many layers of drug discovery R&D with AI. This was one of the largest initial financings in the history of our industry.

Xaira was not alone. In June of 2024, Formation Bio raised $372M for their own thesis on accelerating biotech with AI. In March of this year, Isomorphic Labs raised $600M to continue scaffolding a new pharma organization around the research lab that produced the AlphaFold breakthrough.

As an early-stage investor, part of my job involves talking to a lot of different people across academia, biotech, and pharma. Broadly speaking, people don’t find these investments that crazy. We’ve been desensitized to mega-rounds over the years. Drug discovery is expensive! And it’s becoming consensus that AI will play a meaningful role in the future of biotech R&D.

But another set of deals is drawing more initial skepticism. Businesses such as EvolutionaryScale, Chai Discovery, Cradle Bio, Achira, Latent Labs, and Axiom have also recently raised large sums of capital to build AI solutions for biotech. A major difference between this set of companies and the likes of Xaira, Formation, and Isomorphic is that they seem to be pursuing value creation by selling technology rather than drugs.

This triggers strong reactions. Schrödinger is constantly cited as an example of how hard this business model is. After 35 years in business selling some of the leading drug discovery software solutions, their market cap is ~$1.5B. This is nothing to sneeze at, but it pales in comparison to the value created from single drugs developed using their technology, such as the sale of Nimbus’s TYK2 program for more than $4B.

The overarching sentiment is that drugs create far more value than discovery tools and that nothing about AI will change that. If you have a great new discovery technology, the most logical thing to do is to use it to make new drugs. The subtext is, “What are all these founders and investors thinking?!”

There is a kernel of truth to this. I spend a lot of my time searching for new platform companies with technologies that can make otherwise impossible drugs. One of the evaluation criteria for AI in biotech should be whether it leads to the development of fantastic new medicines that improve the human condition.1

But I don’t think that means there can’t be massive, objectively investable infrastructure businesses that create a lot of value for the industry. In tech, it’s completely obvious that infrastructure businesses providing tools rather than end products can be huge. Consider Datadog, a >$50B public company that provides cloud observability solutions—in other words, infrastructure for managing your infrastructure.

The most extreme example is NVIDIA. The biggest company in the history of the world is an infrastructure provider.

So, how might AI infrastructure companies actually create and capture value in biotech?

To explore this, I want to tell you three possible stories about the technological and structural evolution of our industry. Each story has parameters, which are numerical placeholders whose values you can personally estimate based on your own beliefs about the future.

Compute as a Reagent

Twenty five years ago, the biotech industry spent exactly $0 on high-throughput DNA and RNA sequencing. That’s an obvious statement because the technology hadn’t been invented yet. And it certainly hadn’t dramatically decreased in cost.

Last year, Illumina, the largest next-generation sequencing (NGS) company, generated $4.4B in revenue.

Factoring in the rest of the players in the market, the total annual spend on NGS is around ~$10B—which is the number most of the market research firms converge on.2

How did this happen? What drove the increase from $0 to ~$10B?

First, NGS became a ubiquitous tool in the life sciences research market. Illumina describes this market by saying, “This includes laboratories associated with universities, research centers, and government institutions, along with biotechnology and pharmaceutical companies.”

Labs around the world were able to answer new questions. Initially, researchers used NGS to sequence and compare genomes. Use cases evolved. As functional genomics emerged, massive gene expression experiments became drivers of continued growth and adoption. In 2022, we saw the first genome-wide CRISPR perturbation experiment, which involved sequencing over two million single cells.

This scale of experiment—let alone the CRISPR technology—would be basically unimaginable to scientists even a decade ago. The crucial point is, as NGS declined in cost, it proved to be elastic, meaning that scientists kept figuring out ways to use even more of it.

In addition to the research market, NGS is now used in the clinic. Two of the largest clinical uses are noninvasive prenatal testing (NIPT) and cancer screening (both for diagnostics and prognostics to help predict treatment responses and guide care). These are both valuable applications with huge markets that are projected to keep growing.

Rare disease diagnosis and infectious disease monitoring are also big drivers of clinical NGS growth. And just like the growth of the research market, it seems likely that we’ll keep discovering new clinical applications of sequencing. For example, T-cell receptor and B-cell repertoire sequencing could become an integral part of autoimmune diagnosis in the future.

It’s worth noting what NGS didn’t do. Despite the optimism around deciphering “the genetic blueprint for human beings,” as President Clinton stated, genomics didn’t solve biology. And it didn’t solve drug discovery or disease diagnosis.

But it did prove to be a foundational technology that justified converting a growing amount of capital into biological information. NGS was a technological breakthrough for biology. It continues to deliver real value to researchers and clinicians around the world.

The NGS market has consistently sustained a double digit (~11%) compound annual growth rate and doesn’t seem to be slowing down.3 It’s not absurd to imagine the market growing to >$30B over the next decade.

So, that’s value creation. What about value capture?

The primary way that NGS companies make money is by selling consumables—which are the reagents and materials needed for an experiment—at a healthy margin. In 2024, Illumina made 12% of their revenue from selling instruments and 72% of their revenue from consumables.4 The gross margins on their core business were 67.1%.

As NGS became a ubiquitous technology, people built businesses on top of it. For example, 10X Genomics sells instruments and consumables to prepare samples for single-cell sequencing experiments. In 2024, they generated $610M of revenue, with $493M of that coming from consumables.

Clinical NGS companies make money by selling tests. Some of these businesses are growing very quickly. In 2022, Natera generated ~$800M in revenue. By 2024, this had more than doubled to ~$1.7B, with gross margins increasing from 44.4% to 60.3% in the same time period.

Interestingly, because clinical NGS is growing so quickly, Natera is now a bigger company than Illumina—who is the provider of the actual underlying sequencing technology! As of writing, Natera’s market cap is ~$31B and Illumina’s is ~$18B.5

What Powers give these businesses the ability to command such strong margins?

I’ve already taken a crack at trying to explain this for Illumina. In large part, the primary Powers are two-fold. They have Cornered Resources in the form of a complex thicket of patents protecting their proprietary sequencing chemistry. Illumina also emerged as a definitive research standard, which introduced Switching Costs as well as Network Economies. People benefit from having the sequencing machine that most assay and software developers are developing new solutions for.

Similarly, Natera has their own IP portfolio for their tests. There is likely also the benefit of being an early mover to gain reimbursement and adoption by clinicians. But Natera will have to continue to fend off competitors, including Illumina and BillionToOne, who just had a fantastic IPO.

So, what does any of this have to do with AI?

Just as NGS rapidly permeated the research market, scientists around the world are integrating new AI tools into their work. Arguably, the adoption is happening faster and is more widespread.

And just as NGS led to growth in spend on chemistry reagents, AI is driving growth for another type of research reagent: compute.

Taken together, these two facts could lead to greater research productivity and the growth of large businesses.

Let’s make this more concrete.

Google Deepmind released AlphaFold 2, which was the tipping point for protein structure prediction, in November of 2020. In July of the following year, they launched the AlphaFold Protein Structure Database in collaboration with the European Bioinformatics Institute. In just over four years, DeepMind says the database “has over two million users in 190 countries.”

Think about that number for a second. How many biologists even are there in the world? In 2021, UNESCO estimated there were 8.854 million researchers worldwide, inclusive of every scientific discipline. Let’s assume somewhere between 20-30% of those researchers are doing something related to biology. That’s somewhere between 1.8 million and 2.6 million researchers.

That would mean that in less than five years, we went from basically no biologists using AI structure prediction models to basically every biologist using them.

This squares with my subjective experience. I don’t know any researchers who aren’t at least playing around with these types of databases and doing basic searches.

To be clear, this type of usage represents a small fraction of any given research project. My former PhD advisor Michael Fischbach has a good framework for thinking about this type of tool. In his course on problem choice and decision trees in science and engineering, he compares and contrasts how widely used a tool is versus the criticality of a tool for a given experiment.

Some tools, like complex and expensive microscopes, can be used to answer certain research questions all on their own. But their use is limited to a small number of specialist labs. Other tools, like CRISPR, are equivalently powerful, but much more widely adopted. Others, like the BLAST software program, are very widely used but are not as mission critical.

The AlphaFold Protein Structure Database resembles BLAST. This is just one example of this type of ambient AI usage for biology.

Consider this anecdote from a recent article on Amgen’s approach to AI adoption:

Traditionally, Amgen’s scientists would manufacture a few hundred milligrams of any protein of interest to assess viscosity. Today, Amgen runs an algorithm predicting viscosity from the sequence of a protein, before it’s ever made in the lab. If the prediction is not promising, it’s not made.

“We now do it every day on every molecule,” said Alan Russell, head of R&D technology and innovation at Amgen. “There are literally 20 or 30 other predictive models that we use to correct all of the therapeutic molecular attributes that could cause a problem.”

Like NGS, these types of models aren’t solving biology. But they are delivering sufficient value to justify converting a growing amount of capital into biological information.

It also seems like these models can serve as mission critical tools for research projects just like microscopy or CRISPR.

Last year, I read an interesting research paper about the possible evolutionary trajectories of the SARS-CoV-2 virus. The primary goal was to better understand how the virus’s spike protein might mutate and change over time.

In the first version of the preprint that was posted, they performed a comprehensive deep mutational scanning experiment of the protein sequence. They also ran some experiments to assess epistasis and look for signals of evolutionary selection.

But none of these experiments were done in the lab. Instead, they were all in silico experiments carried out using a protein language model (pLM). Using a single protein sequence, a pLM, and a little bit of existing experimental data, this group produced a novel research paper.6

Of course, computational biologists publish “dry” papers with no experiments in journals every day. But this type of project felt like a glimpse into a world where these models are used as a new form of experimental apparatus.

If model progress paused today, we would still have years worth of fruitful interpretability research to tease out new biology from what these models have already learned.

Considering both the ambient usage of AI in biology and its use as a new form of experimental tool, we should expect greater conversion of capital into biological information via compute over time.

There are some leading indicators of this being true. Lilly recently announced they are building a giant supercomputer with NVIDIA. Several other pharmaceutical companies including Novo Nordisk, GSK, and AstraZeneca have already invested in similar compute buildouts.

This introduces an interesting question: will other businesses capture any of this value besides NVIDIA?

In Illumina’s case, they owned the proprietary chemistry reagent. Here, NVIDIA owns the compute reagent.

I think the answer is yes. In tech, people talk about “platform shifts,” where a new foundational technology enables all sorts of new applications and businesses. For example, cloud and mobile were both big platform shifts. It sure seems like AI is the next big platform shift.

In this type of transition, it’s incredibly hard for one platform company to offer every solution needed to build the next wave of applications. AWS doesn’t ship everything that developers need to make full use of the cloud. There are tons of massive companies that command healthy margins by solving hard infrastructure problems on top of this underlying platform. I happen to work at Amplify, which is a firm that has done phenomenally well investing in these types of infrastructure businesses.

Joe Lonsdale recently wrote a nice piece on the different layers of opportunities within the AI stack being built out.7

We’re seeing pharmaceutical companies invest heavily in chips, which means there could be a growing life science enterprise market for solutions higher up in the stack.

Beyond the enterprise, I think we could be dramatically under-calling the size of the broader life sciences research market across smaller labs, institutions, and startups around the world. Just like NGS, it seems like researchers will keep figuring out new ways to use these tools to address all sorts of biological problems—especially as costs keep coming down.

There are also clinical use cases for AI. For example, a wave of “digital oncology” companies like Artera and Valar Labs are popping up. Their goal is to use AI to create new biomarkers from existing clinical data—such as pathology slides—to help guide cancer care. Over time, these models could compete head-to-head with clinical NGS offerings.

Now let’s turn this story into a concrete prediction like forecasters love to do.

Ten years from now, $X billion will be spent on AI compute in biology, and Y% of that spend will be captured by new startups.

You pick the parameters. When doing so, it’s worth having a bit of humility. Bets on market creation are often non-obvious. When Uber got started, many people passed on investing because the taxi market didn’t seem that big. This missed the fact that the mobile ride share market was much bigger. Predicting the future is hard.

“From the Wet Lab to the GPU Cluster”

What is AI, really?

Let’s peel back one layer of the onion and instead ask: what is deep learning? This is the set of algorithmic techniques that is now synonymous with modern AI.

Deep learning methods are fundamentally concerned with approximating functions. At the highest level, a function is a process that maps a set of inputs to a set of outputs. Lots of processes, like applying a label to an image, translating text between languages, or folding an amino acid sequence into a 3D protein structure can be described using the language of functions.

Deep learning has been incredibly successful at solving these types of problems. As Michal Bronstein, a prolific deep learning researcher, has previously argued, much of this success boils down to two core principles:

1. “The notion of representation or feature learning, whereby adapted, often hierarchical, features capture the appropriate notion of regularity for each task”

2. “Learning by gradient descent-type optimization, typically implemented as back-propagation.”

Deep learning provides a set of architectures, which are computational primitives for detecting the key features of an input that dictate what the corresponding output will be when it is passed through a function.

Next, optimization algorithms like back-propagation make it possible to learn approximations of a function from data.

Let’s consider a concrete example. In the early days of computer vision, a specific architecture called a convolutional neural network made it possible to extract the hierarchical features in an image that were predictive of what the correct label was. This type of mapping is then optimized with massive amounts of image data using the back-propagation algorithm—meaning the neural network is constantly tuned to have higher accuracy.

Bronstein had another key insight about what led to the success of deep learning. He wrote, “While learning generic functions in high dimensions is a cursed estimation problem, most tasks of interest are not generic and come with essential predefined regularities arising from the underlying low-dimensionality and structure of the physical world.”

Basically, many complex functions seem impossible to learn. But it turns out that they aren’t. AI can learn them. This is because the physical world actually has a simpler underlying structure that can be learned. This is why science works in the first place.

Physicists have had a lot of success in deriving some of these functions with lots of math and mental horse power. But other functions have resisted a similar treatment. Deep learning offers a set of tools for learning these thornier functions from data.

Let’s turn this general intuition towards the question of AI’s utility in biology.

Biology is chock full of problems that have resisted mathematical description. It’s incredibly hard to figure out the underlying function that dictates how an amino acid sequence will fold into a protein structure, what drives the differentiation from one cell type to another, or why some patients respond to specific drugs and others don’t.

But we can collect data by performing experiments. There are dozens of experimental techniques for resolving a protein’s structure. For decades, researchers around the world deposited structural data into the Protein Data Bank. Google DeepMind trained the AlphaFold models using this data.

So in a very real sense, the function AlphaFold is approximating is an experiment. It’s computationally deriving what we would otherwise measure in the physical world.

Right now, structure prediction models still have lower accuracy and resolution than physical experiments. The major difference is in speed, cost, and complexity. Producing a crystal structure is very hard, but basically any scientist can now ask structural questions as part of their project at the speed of inference.

Now let’s take this to its logical extreme. Given the rate of progress in molecular machine learning, is it possible that AI surpasses experimental accuracy?

It seems possible. Researchers have already made the argument that AlphaFold 2 predictions “are usually more accurate” than structures produced using nuclear magnetic resonance (NMR) techniques. That’s pretty crazy.

Compared to experiments, AI models have a few advantages:

1. They can average out background experimental noise,

2. incorporate priors from massive data sets like the PDB, and

3. interpolate between nearby data points.

Assume we keep figuring out clever representations (like CNNs for images) for these problems. And we keep pouring more high-quality data into these models. Could inference be more accurate than measurement?

Already, for proteins without existing structures, “the use of AlphaFold models in structure-determination protocols has been shown to reduce the time and effort required relative to ab initio model building.” The starting point is a computational prediction, which can then be refined with a smaller set of experimental data points.

Scientists used to generate data and then fit models. Now, in many cases, scientists are using models to generate predictions and then refine them with data. This is an interesting inversion.

In our first story, we compared AI models to the rise of NGS. The idea was that AI can be thought of as a new form of information generation whose usage will grow. This is a story of category creation.

Here, the idea is that AI can potentially replace existing experiments.

Consider Amgen’s AI usage again:

Traditionally, Amgen’s scientists would manufacture a few hundred milligrams of any protein of interest to assess viscosity. Today, Amgen runs an algorithm predicting viscosity from the sequence of a protein, before it’s ever made in the lab. If the prediction is not promising, it’s not made.

They are predicting the outcome of an experiment before they do it. I’m assuming they have amassed a lot of this data running R&D at Amgen over the last 35 years. Seemingly, enough data to learn an approximation of the mapping between a protein’s sequence and this physical property. And they have enough confidence in this approximation to let it determine what experiments they do.

At that point, the actual physical experiment becomes a verification step. Until verification, R&D can move at the speed of inference. More ideas can be tested.

Let’s consider another example.

Recently, I wrote about the growth in research on “virtual cell” models. For decades, researchers have attempted to build computational models of cells. But right now, this term is mainly being used to describe a very specific thing.

The main thrust of virtual cell research is predicting single-cell perturbations.

As NGS scaled, it became possible to measure gene expression in millions of individual cells.8 With the advent of CRISPR and improvements in DNA synthesis, our ability to perturb—basically, poke or manipulate—the genetics of individual cells also scaled dramatically.

The Arc Institute’s Virtual Cell Challenge is centered around predicting the result of this type of experiment. Using the massive data sets that have been generated using these types of technologies, the goal is to predict things like the gene expression profile in response to a new perturbation.9

Again, predicting experiments. At the moment, these virtual cell models are far less performant than protein models. But they may already be surprisingly close to replacing some experiments.

This is because the number of possible experiments is massive. In a 2023 talk on AI for drug discovery, Aviv Regev said, “When you talk about new medicines, the numbers become really, really big.”

There are roughly 20,000 genes in the human genome, meaning there are around ten trillion possible gene combinations to consider.10 The human body has thousands of different cell types, each of which can adopt different states in response to signals. We can’t possibly test all of these combinations and states in the lab.

For some of the leading researchers working on virtual cells, the goal is simply to make a model that is good enough that researchers start testing ideas in silico before proceeding with experiments. This would make it possible to run much larger digital experiments before doing smaller sets of physical experiments.

If these models get much better, there is a chance the ratio of digital and physical experiments could change dramatically. Stephen Quake envisions a future where “cell biology goes from being 90% experimental and 10% computational to the other way around.” Imagine a world where the next Yamanaka factors are discovered computationally.

Ron Alfa, the CEO of Noetik, described this shift by saying, “We’re moving biology from the wet lab to the gpu cluster.”

Now let’s return to our earlier logic for value creation and value capture.

The global pharma industry spends ~$250B annually on R&D in pursuit of new medicines. Historically, this research happened inside of Big Pharma companies. But as the industry matured, they divested more of this work to external organizations.

And smaller biotechs, who now collectively represent a large fraction of early-stage R&D, also benefit from outsourcing portions of their work instead of setting up every single assay from scratch internally.

These external organizations are called contract research organizations (CROs) and contract development and manufacturing organizations (CDMOs). Outsourced R&D spend that goes to CROs and CDMOs is growing faster than internal R&D spend.

Assume roughly a third of the outsourced R&D spend goes to CROs. That’s ~$30B annually. How much of that will ultimately go to computational experiments instead of lab experiments?

Imagine a CRO that offers what Amgen already does internally as a product. They provide a suite of tools for computational drug design and ~20-30 (or more) different experimental simulators to test the designs. Maybe they even experimentally verify the final predictions with a robotic lab or a passthrough contract to an experimental CRO.

This doesn’t seem that far-fetched. Just like NGS, there could be elasticity. If a computational experiment was a tenth the price, my bet is that researchers would find ten times the number of experiments to do—or more.

This would help alleviate the bottleneck of physical experimentation on a researcher’s creativity.

The global R&D market is already really big. The main question is if new AI companies will be able to scale and compete for a bigger slice of the pie over time.

Framing it as a prediction:

Ten years from now, X% of experiments in the R&D market will be computational, and Y% of those experiments will be run by new startups.

If this shift happens, there will be another big question. What will drive pricing power between companies in this market? In other words, why won’t it be a race to the bottom?

This could come from differentiated model research, sticky workflows and integrations, or simply access to compute. But time will tell. First, we’ll need to see scientists shift more work from the bench to the cluster.

AI for Program Decisions

I distinctly remember the first time I met Brandon White. At the time, he was the Head of Product at Spring Discovery, a startup selling AI-powered image analysis tools into pharma. He was trying to poach me from Stanford to come work as an engineer there.

We met at Verve Coffee Roasters on University Avenue in Palo Alto, right off of campus. I was there early, so I grabbed a coffee and sat down. A few minutes later, Brandon strode in wearing a crisp pair of sneakers that happened to be the largest I’d ever seen.

As we started talking, ideas about technology, AI, product design, and business history seemed to just stream out of this guy. I realized his LinkedIn header might not be an overstatement. There was a good chance I was talking to “The Best Seven Foot Engineer in the World.”

After becoming the Best Seven Foot World of Warcraft Player in the World (he and his team were also ranked #1 in North America, irrespective of height) in his late teens, Brandon’s obsession shifted from video games to machine learning, which led to a job as a ML engineer at Uber in 2014.

But ultimately, he had a deeper obsession: biology. Dealing with the challenges of moving through the world with a stutter, Brandon has an acute awareness for the pain of pathology and the frustration that comes from our inability to restore physiology.

So he jumped from tech to biotech, joining a cancer detection startup called Freenome as one of the first employees. As the company scaled rapidly, Brandon moved to Spring to help scale their platform and refine their product strategy.

Spring presented an interesting set of challenges. The business had built an exciting AI-powered platform for analyzing high-throughput images that were captured at scale in their lab. While they originally started with a thesis to develop longevity drugs, it became increasingly clear their strengths were in platform and technology development.

So Spring decided to pivot “from running its own discovery programs to collaborating with the larger scientific community through partnerships and software.” Brandon was tasked with helping to productize the platform. But in the Fall of 2023, he decided to leave and rethink what selling AI to biopharma might look like starting from a blank slate.11

That Fall, Brandon wrote an epic blog post on this topic. It was part reflection on past experience, part catharsis from his struggles, and part manifesto for how things could go differently in the future.

He made the following observations:

There are three fundamental truths you need to understand if you want to use AI to create huge value in drug discovery:

1. The further you are from the drug program decisions, the less value you create

2. The further the decisions you influence are from the commercial + clinical stages of drug discovery, the less value you create

3. The decisions that create the most value are choosing the right target/phenotype, reducing toxicity, and predicting drug response + choosing the right target population

Brandon derived these tenets from firsthand experience. Drug programs at the very start of their journey are worth astonishingly little. In fact, for companies with late-stage assets, public investors can assign negative enterprise value to early programs because of the risk and capital requirements.

But as risk is retired through clinical development, value increases dramatically. In a toy example, Richard Murphey from Bay Bridge Bio showed how a program worth $3M at the outset can grow to over $1B after a successful Phase 2 study demonstrating efficacy.

Selling tools and tech for the long tail on the left is hard. In theory, selling tools and tech that help guide decisions later in development would be much better. Based on this logic, Brandon developed a thesis for a new business.

Out of the high-value problems he outlined, reducing toxicity risk felt like the best opportunity. Convincing people to take on novel target risk is hard. And predicting drug efficacy is even harder—it’s the Holy Grail of drug discovery! But it felt like there would be strong pull for a product that flagged toxicity risks in molecules people were already developing.

This vision became Axiom. Roughly half of the drugs tested in clinical trials fail due to toxicity. Axiom wants to change that.

Amplify, the VC firm I ultimately joined, led Axiom’s first financing. And I wrote an angel check into the round out of my own personal savings. It felt like the right choice to bet on The Best Seven Foot Engineer in the world. Toxicity didn’t stand a chance.

Axiom’s strategy is unique. Brandon co-founded Axiom with Alex Beatson, a talented computer scientist with experience leading ML teams at platform biotech companies. A lot of what they do rhymes with how people have built platform companies in the past.

They generate massive amounts of proprietary data. So far, over 100,000 molecules have been screened in primary human liver cells, which required correctly identifying and labeling nearly 400 million cells with a computer vision pipeline.12 This lab data is integrated with another data set of over 10,000 molecules annotated with toxicity outcomes from human clinical trials. Both of these data sets grow every day.

But they don’t use this data platform to make a drug pipeline or strike traditional discovery deals. They use it to ship a product. Axiom integrates their AI predictions into a beautiful, modern web application. Additionally, they offer computational and experimental services for clinical risk assessment. Multiple leading pharma companies are now using the platform to screen their drugs.

Amgen has internal models that guide their R&D decisions. Axiom is building a suite of models to guide clinical decisions. Already, Axiom’s models outperform many of the current experimental toxicity assays used in discovery.

Here, Illumina isn’t the right analogy. Rather, this business could resemble Cadence Design Systems, which provides electronic design automation software to the semiconductor industry. Cadence’s product is so critical that they generate over $3B in annual revenue from software and IP licensing.

Imagine a future where Axiom’s AI is so good that it would be a fiduciary liability to make a decision to advance a drug program without using it. Cadence has already achieved this status in the semiconductor industry. Owning this type of platform for biopharma would be pretty valuable.

Our first story was about compute as a new reagent for biological discovery. Next, we considered the value of replacing existing experiments with computational simulation. Here, we are talking about the development of AI models that can produce information no existing experiment can rival. In other words, a net new capability.

Axiom is working hard to tackle toxicity. But the other problems Brandon outlined are still big bottlenecks. Discovering new drug targets is hard. Drugs still routinely fail in Phase 2 trials because they lack efficacy. We still struggle to identify the right patient populations for clinical trials. The list of problems like this is long.

So let’s generalize this story into another tangible prediction about how big of an opportunity this is for new startups.

Ten years from now, $X million will be spent per drug program on AI tools for making critical decisions, and Y% of that spend will be captured by new startups developing these models.

Historically, biotech has been synonymous with biopharma. The overwhelming majority of the venture dollars pointed at this market flow into businesses making new drug products. The central source of value is clinical assets.

While there have been large technology businesses like Illumina, they are somewhat anomalous. Selling software solutions into pharma is hard. And despite the growth of the contract research market, the expected returns from investing in CROs are a match for private equity, not venture. The revenue for most CROs grows linearly with headcount, not exponentially like a software or therapeutics business. (Sort of like consulting.)

Because of this, the growth of venture investment into AI biotechs selling technology rather than drugs is met with skepticism. Something about this technology platform shift has to be different for this to be a good idea. I think there is a chance this time is different.

I just told you three stories about how this could be true. I use the word story intentionally. These are narratives about how our industry could evolve, not deterministic models. Nobody knows the answer.

Frankly, we should hope that infrastructure for our industry becomes more investable. We are in continual pursuit of beautiful medicines. That is rightfully the North Star. But our relative underinvestment in shared infrastructure requires constant reinvention of the wheel, which is highly inefficient.

The resources we do share are not great. People like working with CROs about as much as they like going to the dentist. (Sorry, dentists.) We lack the market competition that leads to beautiful product experiences.

The tech industry has invested heavily in infrastructure. As a result, software developers get to use tools like Modal and Vercel, which are genuinely delightful products. I want a future where biologists have delightful tools, too.

Crucially, better infrastructure could expand our ambitions for biotech. Right now, R&D is so expensive and slow that it’s hard to actually generate returns without focusing on high-value drugs. We are still in the Tesla Roadster era.

But with better shared tools, we can move faster towards a future of sustainable abundance where our world runs on biology.

Thanks for reading this essay on AI infrastructure for biology.

If you don’t want to miss upcoming essays, you should consider subscribing for free to have them delivered to your inbox:

Subscribe now

Until next time! 🧬

Subscribe now

Today’s essay is about Manifold Bio, a promising Boston-based startup that spun out of the Church Lab at Harvard. In many ways, their story is a continuation of the key themes of this newsletter. Alongside Dyno Therapeutics, a company I’ve previously written about, the founders of Manifold are some of the world experts in combining cutting-edge DNA technologies with machine learning. They’ve also thought deeply about the best platform strategy to realize the full potential of their inventions.

We’re going to dive deep into the history of large-scale DNA libraries, the potential of multiplexed protein measurement in living organisms, and the business decisions Manifold has made in pursuit of breakthrough medicines.

Enjoy! 🧬

In biology, great technologists have a way of measuring the unmeasurable. And Manifold had done just that.

They had fifty antibody medicines, each carefully designed to traffic across the blood-brain barrier and dissolve the plaques thought to drive the progression of Alzheimer’s disease.

But testing each antibody would require over five hundred monkeys. Even if they could procure this many animals, the cost—both financially and ethically—would be enormous.

Instead, Manifold ran the study with a molecular sleight of hand. Using their protein barcoding system, they resolved the distribution of each drug in the body in parallel, requiring an order of magnitude less animals. And what they saw was incredible.

But in order to explain how they got here, I’d like to tell you a story.

Innovation rarely follows a linear trajectory. When one bottleneck is addressed, it often reveals another one not too far behind it. For this reason, progress is shaped more like a staircase, with discontinuous steps upwards that compound over time.

Over the last fifty years since the biotech industry started, many of our foundational capabilities have compounded like this.

The first “killer app” for Genentech was simple in hindsight: producing an existing protein in microbes instead of animals using recombinant DNA technology. Gradually, dozens of advances stacked on top of each other and expanded the scope of what we could build with DNA. Monoclonal antibodies emerged as a full-fledged app store, with over one hundred drug approvals solidifying biologics as a new cornerstone of medicine.

Along with the app store came an entire suite of tools for app developers. Multiple generations of antibody discovery technologies were invented, each building on previous efforts. Now antibody developers have a smorgasbord of approaches to choose from to tackle their specific problem.

Like any other technology, these techniques are quickly maturing and becoming commoditized. Biotech companies can choose from a large number of different vendors that offer antibody discovery services. We are getting really good at identifying proteins that bind to other proteins—binders.

At this point, an antibody discovery campaign rarely fails at the stage of finding binders. If that weren’t enough, new AI technologies are making it possible to computationally design protein binders in a matter of minutes. Imagine a future where every scientist with a laptop has this capability.

But a binder is not a drug.

This simple fact helps to explain one of the most surprising disconnects in our industry.

As Jack Scannell observed in 2012, despite all of these extraordinary developer tools, “the number of new drugs approved per billion US dollars spent on R&D has halved roughly every 9 years since 1950, falling around 80-fold in inflation-adjusted terms.”

In one of his “diagnoses” for this decline in efficiency, Scannell points to “The ‘basic research–brute force’ bias.” We have a tendency to over-invest in the discovery technologies that can be rapidly scaled and industrialized. But the ability to be scaled doesn’t necessarily correlate with a tool’s ability to help successfully predict a drug outcome.

So, we’ve made huge strides in finding new starting points for drug programs. But our downstream techniques for discerning their quality haven’t improved at the same rate.

We can now screen hundreds of billions of possible antibodies to find hits. Once it’s time to see how they behave in a living system, we still test them one by one in mice.

Manifold Bio has a vision to change that.

Founded in 2019 during the onset of the COVID pandemic, Manifold got started with nothing more than a small team and an idea. What if we could point our exponentially improving DNA technologies at the bottlenecks that come after we find drug starting points? What if we could test massive numbers of drug candidates in a single animal experiment?

Fast forwarding to the present moment, this is no longer just theory.

Using sophisticated molecular engineering, Manifold has built a drug discovery platform capable of screening over 100,000 antibodies in the type of in vivo experiments in organisms that would typically test one antibody. Narrower screens can resolve over 1,000 antibodies in the body at a resolution so precise it would be like detecting a single grain of sugar in over 120 Olympic-size swimming pools.

Along the way, Manifold has raised over $70M, recruited word-class drug discovery talent, and developed a pipeline of promising drug candidates that aim to precisely target one of the most challenging organs to develop biologics for: the brain.

Just like the industry’s evolution, Manifold’s story involves a number of crucial—and often serendipitous—insights and pivots stacking on top of each other to arrive at their current position.

In fact, if it weren’t for a throwaway elective at the end of Gleb Kuznetsov’s time at MIT, Manifold probably wouldn’t exist.

ML²

During his undergrad at MIT, Kuznetsov didn’t study biology. To the contrary, he actively avoided it. Biology didn’t feel like an engineering discipline. It wasn’t until the last credit of his Master’s degree in computer science that one course changed his tune.

In a class with Manolis Kellis, a Greek computational biologist who can preach about his discipline with the fervor of a church pastor, Kuznetsov saw the future of biology painted in a very different light. With rich data streaming out of modern lab instruments and DNA sequencing machines, the field started to feel like an open frontier where somebody with his computational skillset would soon become the protagonist.

Looking back on the experience years later, he said, “Taking that class was like a moment in a video game where you walk into the correct room, and the music starts getting louder, so you know you’re doing the right thing.”

After a brief stint as a software engineer at Google, the music led him to George Church.

Church is a walking embodiment of the biotech revolution that Kellis outlined. For decades, a seemingly endless stream of advances has flowed out of his research lab at Harvard.

He played a key role in inventing modern high-throughput DNA sequencing and synthesis tools, programmable gene-editing technologies, longevity therapies, “de-extinction” strategies, genetically recoded organisms, molecular data storage, AI-based drug design approaches, and many, many more projects that read like science fiction.

Dwarkesh Patel recently introduced him by saying, “it would honestly be easier to list out the major breakthroughs in biology over the last few decades that you haven’t been involved in.”

Kuznetsov pitched Church on a simple idea. He’d join the lab as a programmer and help build any software tools that were needed. In exchange, he’d get the chance to learn how to design and run experiments. Church happily agreed. “I was flattered that I could possibly lure somebody away from Google with a much lower salary,” he told me.

Ultimately, Kuznetsov went so far down the research rabbit hole that he enrolled in Harvard’s biophysics PhD program to pursue a graduate thesis with Church.

At the time, the Church Lab had an absolutely extraordinary concentration of talent. Kuznetsov overlapped with scientists like Sri Kosuri, Dan Goodman, Eric Kelsic, Surge Biswas, Marc Lajoie, Lexi Rovner, and Adam Marblestone, who have all gone on to become rising stars in science, philanthropy, and business.

What attracted many of these brilliant people—aside from the general appeal of training with a legendary researcher—was the unique environment Church cultivated. Unlike other labs where graduate students mainly focused on their own projects, collaboration and cross-pollination, especially between disciplines, was actively encouraged. And technology development wasn’t treated like a second class citizen to basic research.

Tinkering with the latest technologies and exploring their boundaries was considered to be a normal activity, not an unwelcome distraction.

One of the most exciting technological frontiers in this era was DNA synthesis. After the Human Genome Project had completed in 2003, it was clear that “DNA reading” (sequencing) approaches vastly exceeded our “DNA writing” (synthesis) capabilities.

But since the 1990s, companies like Affymetrix had been developing systems called DNA microchips. Using techniques from the semiconductor industry, they invented approaches for synthesizing billions of short DNA polymers called oligos (short for oligonucleotides) on miniature glass wafers.

Initially, these chips were used as an analysis tool. Distinct nucleic acid sequences in a sample could be detected when they bound to the oligos on the chip.

Beyond detecting other DNA molecules, the amount of DNA on the surface of these chips was so minuscule that it was unclear it would be useful for anything else. But the Church Lab started tinkering.

In 2004, just one year after the Human Genome Project, they found an extremely clever use case.

They came up with a way to cleave the DNA off of the chip, effectively harvesting a “pool” of these tiny oligo sequences. To get around the small volume of DNA, they would flank the sequences they wanted larger quantities of with PCR primers that they called “adaptors.”

By including these sequences, they could initiate a molecular chain reaction that would create billions of DNA copies.

Finally, they would cut off the adaptors using restriction enzymes—the same type of enzyme used to produce recombinant DNA that kicked off the biotech revolution in the first place.

It was nanotechnology in action. By “programming” the microchip with stepwise reactions, they could produce large quantities of short chunks of DNA that could be used for a wide variety of applications. One obvious use case was to stitch them together into complete gene sequences.

They published a demonstration of this in Nature, synthesizing 21 distinct genes for one experiment. It was an extraordinary feat at the time.

The genius was in the skillful synthesis of three powerful DNA technologies.

DNA microchips + PCR + restriction enzymes = magic.

Once the paper was out in such a widely read journal, Church told me he thought gaining large-scale adoption was “going to be a piece of cake.” It felt like an obviously important technology for large-scale testing of all the hypotheses being developed by the new field of genomics.

To his surprise, few people knew what to make of it. For over a decade, his lab had “this entire incredible field to ourselves,” he said. And so they kept tinkering away, exploring application after application with little competition.

In 2009, they published a landmark paper describing the use of complex pools of synthetic DNA for genome engineering. The technology, called Multiplex Automated Genome Engineering (MAGE), made it possible to introduce billions of genetic variants per day to rapidly evolve organisms toward desired traits—such as optimizing the rate of production of a desired molecule.

At a high level, Church refers to this as Super Evolution. “In principle, evolution might incorporate a few base pair changes in a million years. Now we can make billions of changes in an afternoon,” he said recently.

This was the massive playground that Kuznetsov and his brilliant lab mates freely explored.

In his first few years in the lab, Kuznetsov worked with a small team to publish a string of papers using these tools to genetically recode microbes—creating or deleting entire codons—and optimize their genomes. They wrote, “by recoding bacterial genomes, it is possible to create organisms that can potentially synthesize products not commonly found in nature.”

These were ambitious, pie-in-the-sky research projects to make synthetic organisms.

Over time, though, his interests evolved. Two of his neighbors at the bench, Eric Kelsic and Pierce Ogden, had more translationally relevant ideas in mind. They wanted to use these DNA technologies to engineer therapeutically relevant molecules.

Kelsic had originally studied physics but was drawn to the potential of engineering biology. By the time he arrived at the Church Lab, he was interested in pursuing a project with near-term commercial relevance.

Ogden, who grew up in the Bay Area, originally studied at the University of San Francisco, a small Jesuit college tucked next to Golden Gate Park. After peppering the TA of a graduate-level bioinformatics course with questions, she told him to seek out research opportunities in U.C. San Francisco’s world-class labs on the other side of the city.

There, Ogden became deeply obsessed with technology development. He interned in a number of labs before becoming a technician in the Small Molecule Discovery Center run by Jim Wells. For two and a half years, Ogden spent his time programming, building databases, learning the ins and outs of lab robots, and optimizing drug discovery assays.

And then one book changed his trajectory.

Regenesis: How Synthetic Biology Will Reinvent Nature and Ourselves, is Church’s manifesto on the future of bioengineering. Narrating the advances in DNA sequencing and synthesis, he paints a picture where writing “genetic software” is the most important technology of the future.

After reading the book, Ogden had a simple conclusion: he needed to go work with this guy.

So, he moved across the country to start graduate school and join the lab.1

Together, Kelsic and Ogden decided to work on engineering better viral vectors for gene therapy. Using the programmable DNA synthesis techniques, they could create “multiplexed libraries” of the vectors, meaning that a large number of genetic designs (library) were created and tested in parallel (multiplexed).

In 2019, they published impressive results from this approach applied to the adeno-associated virus (AAV) genome. This line of research ultimately became the basis of a company called Dyno Therapeutics, which Kelsic still runs.

Rather than joining Dyno, Ogden pursued a different technology direction: protein engineering.

Compared to proteins, gene therapies have a major engineering advantage: their payloads consist of DNA. This meant that it was possible to naturally apply all of the DNA technology tricks the Church Lab had refined over the years.

Proteins, which are chains of amino acids, lack this immediate advantage. But given their centrality in biology—as well as their enormous market share in both the research tools and therapeutics markets—Ogden thought it was worth figuring out a way to use multiplex libraries to engineer them.

In parallel to the AAV project, Ogden, Kuznetsov, and Surge Biswas formed a small team to explore this direction. To start, they worked on engineering green fluorescent protein (GFP), an important research tool that, well, glows green! Many efforts had been made to make it glow even brighter, but GFP has a complex structure that made progress difficult. They wanted to see if they could do any better.

So they took a Super Evolution approach.

Combining gene synthesis technology and error-prone PCR, which introduces a wide range of “mutations” while it makes copies of the DNA, they produced roughly 300,000 variants of GFP. Using yet another technology invented in the Church Lab called FlowSeq, they sorted the variants by their brightness and sequenced the DNA that produced them.

This created a data set of genetic variants matched with protein brightness.

Next came the question: what was the best way to analyze this data set and predict new designs?

It was a large data set…

With lots of paired samples of inputs and outputs…

Which made it a perfect fit for machine learning!

This was not by accident. Over the years, a number of Church Lab members had become ML early adopters. It was another technology to explore—and to mix and match with other tools.

Sam Sinai and Eric Kelsic had prototyped early generative models for biological sequences, and a major theme of both the AAV and protein engineering projects was to use ML for “machine-guided design.”

What does this mean?

Consider how Evolution works. An organism has a genome that encodes a set of traits. These traits impact fitness, which is the likelihood of reproduction. At any given time, it’s highly unlikely that an organism has the genome with the greatest possible fitness. Over time, random mutations provide the substrate for Evolution to effectively sample from the underlying landscape of possible genome sequences adjacent to an organism’s current genome.

This fitness landscape is often visualized as a three-dimensional space with peaks and valleys. Evolution is gradually selecting the sequences at higher and higher points of fitness.

For genome or protein engineering, we get to pick what sequence we are optimizing and what fitness is. For example, optimizing the GFP gene sequence to encode a brighter protein.

Technologies like directed evolution already harness this type of approach to engineering.

But the major conceptual advance for machine-guided design is that we can use machine learning models to predict the fitness impact of mutations. By doing that, the models need to learn an approximation of the underlying fitness landscape of that sequence.2

Having access to this approximation makes it possible to sample much farther out from the original starting point of the sequence.

It’s a technology to learn the rules of Evolution—and how to steer it!

Or, ML² (multiplex libraries x machine learning), as Church calls it.

These incredibly powerful technologies proved to be quite complementary.

Using this approach, they identified advantageous combinations of mutations far outside the “local neighborhoods” of possible sequences that other technologies sampled. They could produce much brighter GFP—steering the Evolution of a complex and important protein.

With these results in hand, they proceeded to completely bury the lede.

In June of 2018, they published a short, highly technical, ten page preprint on bioRxiv entitled Toward machine-guided design of proteins. Nobody ever bothered to shepherd the study all the way through the peer-review process into a prestigious journal. Years later, another group’s ML-guided evolution of brighter GFP was published in Science.

But the publication strategy was an afterthought. They had already convinced themselves that ML² for protein engineering had huge potential. Unlike other academics, they were beginning to ask themselves if this could form the basis of a company, not a CNS paper.

mCodes

With over 400 publications that have been collectively cited over 200,000 times, it’s safe to say that George Church has had a prolific scientific career. His involvement in the biotech industry is equally extreme. Church maintains a spreadsheet of his commercial involvements over the years. It now has well over 700 rows.

He loves seeing technologies translated out of the lab and into the world. So when his trainees started pursuing entrepreneurship rather than academic careers, it didn’t bother him in the slightest.

One of the early successes was AbVitro, an oncology startup co-founded in 2012 by Francois Vigneault, a French-Canadian postdoc in his lab. After only three and a half years, the company was snapped up by Juno Therapeutics for $125M, delivering a 40x return on investment in a short time period.3

Soon after AbVitro was started, Luhan Yang raised $38M to found eGenesis, a company with a sci-fi vision to genetically engineer pig organs for human transplantation.4

These early spin-outs led by trainees quickly created an “infectious entrepreneurial culture,” according to Church. More and more lab members started to successfully raise money to commercialize their research.

As Kelsic and several others left to start companies, Kuznetsov, Ogden, and Biswas wondered: “why not us?” They felt that they had great technology and could figure out what it took to successfully pitch their idea and build a business around it.

Kuznetsov started to make moves to figure the latter part out. While they kept working on the protein engineering project, he defended his PhD thesis and transitioned into a role on the Advanced Technology Team at the Wyss Institute. It was part technology development, part entrepreneurial information gathering. Between coding and pipetting, Kuznetsov talked to as many industry experts and potential customers as he could.

Early on, the commercial ideas were broad and unspecific. At one point, Kuznetsov, Ogden, and Biswas pitched a VC on a vision to build a protein engineering company that could help make “proteins for therapeutics, genome editing, materials, and reagents.” Name a major protein market, they would solve it.

After an unsuccessful pitch, the team realized they needed to refine their focus. Biswas decided to double down on machine learning methods development with another group of researchers in the Church Lab.5

Kuznetsov and Ogden realized they needed to decide on one market to clarify their pitch. One of the central ideas was that they could use ML² to “learn the binder code” that dictated how antibodies recognized their protein targets.

A conversation with Dane Wittrup, a co-founder of Adimab (one of the most successful antibody design businesses), gave them another wakeup call. He poured some cold water on their binder design ideas, pointing out how competitive this market had become with a litany of different screening technologies.6

While it might be tough to break through the noise with a new screening system for finding binders, he pointed out the greenfield space for technologies that helped people advance their antibodies through the rest of pre-clinical development.

In particular, in vivo experiments, which test drug candidates in animals, had remained basically unchanged since the start of the biotech revolution. More conversations helped to strengthen their conviction that solving this problem would deliver real value.

A new idea clicked into place for Ogden.

His earlier project with Kelsic had done exactly that—but for AAVs, not antibodies. They were able to screen massive libraries of possible AAV designs in a single animal experiment.

What if they could achieve the same result for antibodies?

Doing this would require solving a crucial problem: figuring out how to actually keep track of where the different antibodies were going and what they were doing.

Unlike antibodies, AAV naturally shuttles DNA into the nucleus of cells. In order to easily keep track of different designs, they introduced DNA barcodes—short stretches of nucleotides that could be quickly discriminated by DNA sequencing—into the genetic payloads of the viruses.

These synthetic DNA signatures let them quantify which AAV mutations were driving more efficient production or delivery of therapeutic payloads.

Unfortunately, stitching DNA barcodes onto antibodies was a non-starter. The reaction for joining antibodies with unique DNA sequences is much less scalable. And even if they could do it, adding such a large and negatively charged molecule would confound their measurements.

They devised a plan to skirt around this issue.

Using ML², they’d screen a massive number of antibodies against a massive number of antigens in a giant all-by-all experiment. By screening every antibody against every antigen, they could find antibody-antigen pairs that were highly specific to each other.

It would be like building a big set of lock-and-key pairs.

Now, what if they could discretely put the key (antigen) on an antibody they wanted to track?

And what if they attached DNA to the lock (antibody) that would bind it?

This would convert antibody measurement into a DNA barcode detection problem just like for AAV.

Simply put, they aimed to “turn counting proteins into DNA sequencing.”7

If they could develop these protein barcodes—which they called mCodes—they could track proteins much more easily. In theory, they could multiplex in vivo experiments, testing hundreds, or even thousands, of drug designs every animal experiment instead of just one.

While compelling, all of these ideas were just that—ideas.

To change that, they incorporated a business named Manifold Bio in March of 2019, alluding to the enormous mathematical space of possible protein sequences they navigated with ML². And in January of 2020, they started their first dedicated fundraising process.

It turned out that something else was getting started right around the same time: the COVID pandemic. As they pushed to get their new business off the ground, the entire world—including the entire global economy—ground to a halt around them. Many investors kept their hands in their pockets and waited for things to stabilize.

A few months in, Seth Bannon at Fifty Years wrote the first ~$1M check that would help catalyze the rest of the round. After dozens of Zoom diligence calls, Jory Bell at Playground Global stepped up to lead a Seed round totaling $5.4M.

Around the time the Seed round was coming together, they were joined by another co-conspirator: Shane Lofgren. After connecting at a friend’s party in Boston, it became clear that Kuznetsov, Ogden, and Lofgren shared a certain “crazy idea energy.” He also was highly motivated to have the autonomy to refine these ideas and put them into practice.

Lofgren brought a distinct set of experiences and skills to the trio. After working as a scientist at Stanford for several years, he transitioned into a role in the university’s technology transfer office. There, he pored over thousands of patents, helped research commercial opportunities, refine pitches, and broker connections to his growing VC network. When the opportunity came along, he rolled up his sleeves and worked with Atlas Venture to incubate and co-found a venture creation effort in oncology.

Lofgren met Kuznetsov and Ogden while he was building this company. At the time, he had his own set of ideas for new ventures to pursue next. But he saw two things. The science underlying Manifold seemed incredibly powerful. And the commercial ideas were incredibly raw.

Over a quickly ballooning email thread, Lofgren poked a litany of holes in Kuznetsov’s early commercial strategy. In over 2,000 words, Lofgren argued that their proposed drug programs and partnering plans were unfocused and underutilized the true advantages of the platform.

Unperturbed, Kuznetsov responded with an email to the effect of: “Well, what do you think we should focus on?” Another 2,000 words quickly landed in his inbox detailing a more focused and differentiated business strategy. Kuznetsov got excited.

Realizing that Lofgren provided a valuable commercial conduit to his and Ogden’s platform vision, he recruited him to join Manifold as a co-founder and help refine and run their business development. Lofgren quickly accepted the offer. He loved the science. And he was ready for a role with greater agency and ownership.

With a company, early team, and Seed funding in place, Manifold was ready to turn their vision for mCodes and in vivo multiplexing for biologics into reality.

Platform Evolution

For the rest of their first year in operation, the Manifold team laid the foundation for their platform. They needed to hire great scientists, build custom libraries, assays, and sequencing protocols. All of this took time. Progress was far from immediate.

But in the following year, it became clear that their mCode strategy had legs. Quarter after quarter, the number of mCodes in their toolbox kept growing. By the end of 2021, they had over 150 mCodes, effectively increasing the number of antibody designs they could track in vivo by two orders of magnitude relative to the rest of the industry.

In parallel to their engineering efforts, they pursued two commercial directions. The most central objective was to use their platform to produce highly differentiated internal drug programs. Additionally, they would engage with Big Pharma companies to license and access their mCode technology.

The team thought long and hard about platform-product fit: what type of drug discovery problem would be most uniquely advantaged by the technology they had developed? T-cell engagers (TCEs) seemed like an ideal product to focus on.

TCEs are a fascinating type of biological drug.

One of the biggest breakthroughs in oncology in recent history has been the advent of immunotherapy. In particular, cell-based therapies, such as CAR-T therapy, have delivered curative results for blood cancers. These medicines work by engineering a patient’s own T-cells with a synthetic receptor to target and kill cancer cells.

While the clinical results have been extraordinary, the immense time and cost associated with personalized cell engineering have limited widespread adoption of CAR-T therapy.

TCEs aim to drive the same cell-based mechanism with the dosing simplicity of a traditional antibody. One part of the antibody binds to molecules on the surface of a T-cell. Another part binds to a tumor-associated antigen (TAAs) on the surface of a cancer cell. It’s like a sort of cellular adaptor that brings immune cells and cancer cells into physical proximity, triggering an immune response.

One of the major challenges for developing the next generation of TCEs is figuring out what TAA—or combination of TAAs—will enable precise targeting to a tumor. If the TAA is unspecific, meaning it is also on healthy cells, it can cause serious immune toxicity.

It felt like a perfect problem for Manifold’s platform to solve. No other screening approach could rapidly test hundreds of possible designs to figure out the best possible tumor-specific targeting inside of an organism. So the Manifold team started to design new TCE formats and test them at scale in vivo.

By 2022, everything was tracking positively to move forward and raise a Series A. The goal was to raise enough money to keep scaling the platform and nominate their first clinical development candidates (DCs).

But things got complicated. Just like their Seed round at the start of COVID, they found themselves in one of the most difficult fundraising markets in recent history.

Right after their Seed, the biotech market had gone on a historic bull run driven by the enormous demand for vaccines and diagnostics, coupled with incredibly low interest rates to stimulate the economy. In February of 2021, the XBI, a widely tracked barometer of biotech stock performance, hit an all-time high of $174.

Towards the end of 2021 and start of 2022, cracks started to show. Public investors started to rotate out of “pandemic trades” and assess biotech companies with far greater scrutiny. Once the Fed started to increase rates in March of 2022, the bottom fell out. The public biotech market effectively collapsed. By June of 2022, the XBI had lost two thirds of its value.

Despite the bleak macro environment, Manifold managed to push forward and close an oversubscribed $40M Series A led by Jeff Huber at Triatomic Capital that same month. To use Kuznetsov’s video game analogy, Manifold had been able to close two fundraising rounds while playing on hard mode.

With enough capital to weather the upcoming biotech winter, Manifold kept cranking on their platform and drug programs. While things were headed in the right direction, they continued to reevaluate and tighten their strategy.

As Lofgren advanced the company’s BD and partnering discussions, they started to realize a concerning pattern. At the outset, their plan was to land small pilot projects with pharma offering access to their mCode screening technology. Over time, the idea was that these pilots could serve as a toehold for landing larger discovery collaborations.

After a few attempts, problems with this plan started to become more apparent. Timelines for negotiation were long and contracts were small. Even worse, it proved to be challenging for pharma partners to adopt their tools in a piecemeal fashion.

Sure, they could provide mCodes. But what good would that do without an approach to systematically design and synthesize a complex library of antibodies to measure in the first place?

In addition to their set of mCodes, Manifold had built an in-house suite of AI tools and custom-designed antibody discovery assays specifically for this purpose. Each piece of the stack was designed to be perfectly compatible with their in vivo multiplexing technology.

Leveraging his lab robotics chops, Ogden had helped design and build an automated rig that could execute large chunks of these complex experimental workflows with the push of a button.

The mCodes were one piece of a suite of vertically integrated technologies that embodied a distinct drug discovery philosophy. At Manifold, new campaigns start with in vivo experiments that inform iterative cycles of design and testing.

Pharma partners wouldn’t get the full benefit of the platform with in vivo multiplexing alone. Manifold realized they would need to deprioritize services contracts and prove to partners that they could be trusted as full-fledged discovery partners as soon as possible.

Ideally, Manifold would work with a potential pharma partner to scope out a drug program they could help deliver with their platform. From there, they could negotiate a collaboration agreement to deliver results in the form of new IP for chemical matter—rather than technology or services that would just be a means to an end.8

To help structure and land these larger deals, Manifold made a crucial addition to their team, recruiting the one and only Steve Holtzman to join as their Executive Chairman. Over decades on both the sell-side and buy-side of countless biotech BD deals, Holtzman had developed a coherent theory of platform value creation.

In order to drive larger partnerships over the line, Manifold knew they would need to keep advancing their internal programs to prove what their platform was capable of.

Once they had IP of their own, everything would get easier. As Steve Holtzman told me, for biotech platforms, “We don’t sell labor. We sell rights to certain applications of our intellectual property.”

Throughout 2022 and the first half of 2023, it looked like their internal TCE program would be their main value driver. But given the breadth of possible applications, they continued to survey conferences and results in other therapeutic areas.

One of these areas was neuroscience. Since the start of the company, the team had considered developing approaches to solve the challenge of shuttling biologics across the blood-brain barrier (BBB). While the clinical relevance was clear, it felt like a daunting technical challenge. And the time and cost associated with proving efficacy for a complex neurological disease felt intractable for an early-stage startup.

So the idea was relegated to the “blue skies opportunities” list. They ran a few screens, but it was a lower priority project.

In the Fall of 2023, this all started to change

In October of that year, Lofgren headed to the Logan Airport. He was on his way to the Oligonucleotide Therapeutics Society (OTS) conference in Barcelona. At the airport, Lofgren called Kuznetsov. He wasn’t sure it was worth the trip to Europe when the Clinical Trials in Alzheimer’s Disease (CTAD) conference was taking place in their backyard in Boston.

Ultimately, he decided to go. In Barcelona, Lofgren saw extraordinary data. It felt like the oligonucleotide therapy field was in the midst of a neuroscience renaissance. Multiple companies demonstrated the ability for these genetic medicines to precisely knock down a whole slew of challenging neurological drug targets deep in the brain.

In the clinical session, Kirk Brown presented some of the early Phase 1 data for Alnylam’s ALN-APP program, which aims to wipe out the precursor protein that accumulates into amyloid beta (Aβ) plaques in Alzheimer’s disease.

But all of this exciting progress felt rate-limited by a central bottleneck: the ability to deliver the oligonucleotides to the brain. Dosing required intrathecal administration, which involves a lumbar puncture to deliver the medicine directly into the cerebrospinal fluid, bypassing the BBB altogether.

Denali Therapeutics presented early pre-clinical data on their approach to solve this problem. By targeting the transferrin receptor (TfR), they showed an early proof-of-concept that oligonucleotides might be able to be delivered into the bloodstream before being shuttled across the BBB.

While the result was exciting, the technology was still very raw. Lofgren got excited by the prospect that Manifold could do much better. If they could, it would unlock all of these exciting neurological applications of oligonucleotides.

Back in Boston, Lofgren caught up on the results from CTAD. It turned out that Roche had delivered a major BBB advance of their own. Using a similar TfR delivery mechanism, their Aβ-targeting antibody called trontinemab had been able to make its way across the BBB.

Crucially, the design of the clinical study was uniquely tractable. In an interim analysis of only 44 patients, Roche had demonstrated impressive clearance of Aβ plaques using a simple PET scan. As Kuznetsov discussed the trontinemab results with people in other pharma organizations, they told him that their “jaws had dropped” upon seeing the data.

Suddenly, it felt that their BBB program was no longer just a “blue skies opportunity,” it was a chance to quickly become leaders on a new frontier.

If they could use their platform to produce better and more specific BBB shuttles, it would be possible for them to run a trial to prove that it worked. And compared to the white hot TCE space, there were far fewer competitors. The Manifold team quickly assembled the case to pivot their internal focus in this direction and presented it to their board.

Assuming the equity markets might not recover, Manifold realized that they might not be able to raise more money to keep building technology. This would mean they may have to “collapse their platform” and make a “bet the company” decision on advancing a single lead program with the highest possible value. It seemed like a BBB shuttle was the best bet to make. Their board agreed.

Even in the worst case scenario, their platform provided an incredible advantage. A massive wave of opportunity for brain shuttle technology was starting to crest. And they were in the water with a powerful surf board that could be pointed in that direction.

Manifold started to paddle furiously. Over the next year, they iterated through tens of thousands of shuttle designs, testing each molecule in vivo. Each time, they stared at a cross section of the brain, expectantly looking for a signal.

And the signal got brighter and brighter.

Several months in, they already had impressive results in mice. Some of their designs displayed an overwhelming ability to permeate the brain.

Crucially, these findings also held up in non-human primate studies—which were of course multiplexed as well. They were well on their way to having their own brain shuttles!

Towards the end of 2024, it became increasingly clear that their BBB pivot was the right bet. In October that year, AbbVie acquired Aliada Therapeutics, a startup developing brain shuttles against TfR and CD98, for $1.4B. The company hadn’t even completed their Phase 1 trial.

The data and commercial opportunity were compelling enough for an expert to vote with her feet. After two decades leading neuroscience discovery projects across Genentech, Denali, and Cajal Neuroscience, Kimberly Scearce-Levie has seen a lot of promising early findings. But Manifold’s data package stood out.

Having been involved in the development of the first TfR shuttles at Denali, she was particularly impressed by how rapidly Manifold had validated totally new shuttle targets beyond TfR. It seemed like they were really on to something.

Like Lofgren, Scearce-Levie evaluated the opportunity in front of her. Both the team and the platform seemed exceptional. And they could certainly benefit from her clinical development experience. So she agreed to join Manifold as their CSO at the start of 2025.

With wind in their sails as the BBB delivery space heated up, Manifold continued advancing their BD and partnering discussions. While the equity markets remained challenging, pharma deals were still happening. And it was becoming clear that any organization with a meaningful neuroscience pipeline would want to gain access to the new wave of brain shuttles being developed.

In lock step with Holtzman, they developed a framework for the types of deals they wanted to strike. Being an engineer, Kuznetsov describes the space of possible partnerships as an n-by-n array. Along one dimension, there are Manifold’s distinct delivery shuttles. On the other dimension, there are the possible payloads being delivered.

The entirety of the array represents the possible value that Manifold can create. As a platform company, they have two primary options to fuel that value creation. They can sell equity to invest in internal development for a given cell in the array. This comes at the cost of equity dilution. Alternatively, they can sell the future rights to part of the array to a pharma partner in exchange for non-dilutive capital and resources. But this comes at the cost of asset dilution.

This is the fundamental platform dilemma. Selling too much equity can collapse future value. But selling too much future asset value does the same. The two need to be carefully balanced and titrated—like hands at ten and two steering a wheel—in order to realize the greatest potential.

Consider Manifold’s options. Assume they have developed a fantastic brain shuttle. One option would be to raise a large amount of equity capital to advance this asset as far as possible with one payload. Let’s say it works. Maybe they get bought. This could be an enormous win for patients, founders, and investors.

But what about the rest of the array? Or all of the technology that was developed?

Probably never realized.

Alternatively, imagine that they license their brain shuttle IP to a pharma partner. After negotiating, they sell >90% of the future value for use of the shuttle with any payload in any indication. In exchange, they get tens of millions of dollars in cash, some milestones, and maybe some royalties if the drug makes it to market in the future. Success would be decoupled from one program and they could reinvest in the platform. But the company would be a lot less valuable as soon as the deal closed.

This deal structure would be the equivalent to selling an entire column in Manifold’s value array.

Instead of doing this, Manifold pursues deals that transact on cells in the array.

They aim to partner with pharma companies on the use of their shuttle for specific therapeutic payloads, ideally for specific indications as well (which would be a third dimension in the array). This makes for a much better deal. Pharma partners still get access to the shuttles that Manifold has developed. And Manifold still gets non-dilutive capital to reinvest in their platform while retaining the ability to use their shuttle for their own internal programs.

As they advance these conversations, Manifold can see a future where their full platform vision remains intact. Rather than betting everything on their first brain shuttles, they could soon have the resources to start meaningfully advancing shuttles for other tissues.

One exciting opportunity with compelling early data is targeted delivery of medicines to fat cells (adipose tissue). After the advent of GLP-1s, the field of metabolic drug discovery has come roaring to life. This could be another massive wave that Manifold is uniquely prepared to surf with their platform.

Over time, Manifold aims to strategically balance deal making with equity fundraising, driving forward high conviction internal and partnered drug programs using their growing suite of tissue-specific delivery shuttles.

One question in the back of Kuznetsov’s mind is if there is another possible type of deal on the table for Manifold. Historically, Big Pharma has been the sole party interested in transacting with early-stage biotech startups. But as the worlds of AI and biology converge, Big Tech is now in the mix.

For example, NVIDIA, which is now larger than all of the top twenty pharma companies combined, has made major investments in digital biology. As the technology hyperscalers aim to offer new digital biology products, Manifold’s exponentially growing proprietary in vivo dataset could gain more value in its own right.

Historically, biological data has had a short half life. New measurement tools are constantly being developed and scaled. As Holtzman has argued, “The history of data in the biopharmaceutical industry is the history of its commoditization.” Just think about the latest bids for 23andMe. Regeneron’s original winning bid in bankruptcy court for a genetic data set of three million people was $256M. To produce this data set, 23andMe had raised roughly a billion dollars and was once valued at over $6B.

But in an AI-centric world, will this change? Meta’s $14.8B quasi-acquisition of Scale AI shows the premium afforded to businesses with large corpuses of data specifically designed and captured for the purpose of model training.

In this new world order, Manifold is exploring data-centric collaborations with Big Tech in addition to asset-centric collaborations with Big Pharma.

Being Contrarian and Right

A lot of bets that Manifold has made have paid off.

Riding the tailwind of modern genomics and synthetic biology has enabled them to produce enormous data sets.

Riding the tailwind of molecular machine learning has enabled them to generate new drug designs with incredible accuracy.9

Their bet on protein multiplexing has also paid off in spades. What started as a back-of-napkin idea has scaled from tens, to hundreds, to thousands, to tens of thousands, to over one hundred thousand proteins that can be detected in parallel in vivo.

ML² sits at the intersection of two exponentially improving technologies that people believe strongly in.

In other words, it is becoming consensus.

But arguably the core bet of Manifold, which is to point these technologies at in vivo drug testing, is not.

Let’s revisit the very beginning of our story. Rather than becoming more efficient with new technologies, drug discovery is becoming more expensive—and exponentially so. Understandably, people want to figure out why.

Mice seem like a potential culprit. It’s well known that many drugs that look promising in mouse experiments end up failing in human clinical trials. The best laid plans of mice and men often go awry.

And the methodology just feels outdated. As Alex Telford traced in an essay on the origins of lab mice, the initial choice of this model organism and its subsequent proliferation throughout biomedical research had many points of random path dependency.

Telford writes, “This raises a question: what is the purpose of biomedical science? Because if it’s to efficiently discover new drugs, our investment in the enterprise of mouse models leaves much to be desired.”

A growing body of research aims to develop new approach methods (NAMs) and non-animal technologies (NATs) to ultimately replace animal testing. The year is 2025. Shouldn’t we be using AI or bioengineered models of human organs?

In the chaos of the new FDA, this viewpoint has escalated to the policy level, too. Earlier this year, the administration published a news release stating “a groundbreaking step to advance public health by replacing animal testing in the development of monoclonal antibody therapies and other drugs with more effective, human-relevant methods.”

To be clear, animal research is still the gold standard in drug discovery. Part of Telford’s argument is against mice as a model rather than animal studies altogether. But it’s a question of where people are laying down investment for the future. After the FDA’s announcement, public market investors punished Charles River Labs—one of the largest providers of animal testing services—based on a belief that animal studies will become a smaller part of biomedical research over time.

I asked Kuznetsov about all of this. He reframed it with another question. Peter Thiel, one of Silicon Valley’s most successful investors, believes that the biggest ideas are often contrarian before they gain acceptance. One of his famous interview questions is: “What important truth do very few people agree with you on?”

Kuznetsov firmly believes that biotech investors—having been burned many times by translational failures—have become overly pessimistic about the value of pre-clinical models.

The problem isn’t the animals. It’s that the models are treated as a gating filter at the very end of discovery. But what if you are able to pull these models earlier in discovery and iterate based on data from living systems? This can lead to a fundamental step change in the translational quality of the molecule—especially if this iteration is based on more advanced models like non-human primates.

In Scannell’s original paper on declining R&D efficiency, he made an interesting observation: “In the 1950s and 1960s, initial screening was typically performed in animals, not in vitro or in silico, and drug candidates were given in early stages of the development process to a range of physicians.”

In the heyday of drug discovery, screening started in animals—much like Manifold does now.10 But Manifold’s thesis isn’t just predicated on a historical observation. It’s also about picking the strategy with the highest predictive validity for the problem they are trying to solve.

It’s very possible that certain pre-clinical animal models should be augmented or ultimately replaced by new technologies. At Amplify, we strongly believe that Axiom’s approach to model human liver toxicity with AI will help reduce trial failures. And there are likely many other problems with this shape.

But for designing tissue-targeting medicines in the discovery phase, animal studies can be extraordinarily useful.

The story of Trontinemab—Roche’s brain-shuttling antibody—provides a compelling example of this. Early research was done in mice and dose-finding studies were done in primates. Rather than failing to translate, the plaque clearance observed at even the lowest doses vastly exceeded expectations. The drug worked even better in humans than it did in animals!

Manifold’s bet is that like the brain, our most predictive model for finding new tissue-targeting medicines will be mouse and monkey studies. You simply need to be able to see where a molecule traffics in a living organism. But with ML², we can increase the amount learned from each experiment by several orders of magnitude.

This will lead to much better molecules and also reduce the number and size of animal studies required to make them.

Of course, Manifold doesn’t just want to build a better mouse trap, so to speak. The driving question the founders ask themselves is: “What model has the greatest predictive validity for our molecule?” At the outset, they scaled their ML² platform with mice. At this point, non-human primates are becoming their primary model for their multiplexed experiments.

In the future, the founders wonder if they can integrate their compounding in vivo data set with other layers of information that can be generated with NAMs and NATs. As their “in vivo world model” continues to become more predictive, this could be the type of AI that helps realize a future where the behavior of a drug in the body can be accurately simulated.

So, that’s the thesis. Returning to Thiel’s belief in contrarianism, it’s not enough just to believe something others don’t. You also have to be right.

Let’s imagine what the future looks like if Manifold is right.

Their first brain shuttle is a major success. Paired with a thoughtful partnering strategy, they are able to capture value while reinvesting in the next wave of shuttles for additional tissues. It becomes increasingly clear that they have built a world-class platform for targeted delivery of both biologics and nucleic acid medicines.

This would be huge. Alnylam, the iconic biotech that’s helped unlock the siRNA modality, has built an enduring $60B business on the ability to precisely target a single organ: the liver.

What would the future of drug discovery look like if we could precisely engineer the delivery of these programmable medicines to any tissue in the body?

Of course, there is a lot of work to do for Manifold where the rubber meets the road today. They’ll need to demonstrate success for their first brain shuttles in human trials and land their first major partnerships. Achieving one breakthrough—let alone repeating the process—will be no small feat.

The Manifold team knows that. Kuznetsov draws inspiration from Pixar’s story. Like the famous studio that brought leaders in computer graphics together with talented story tellers, he aspires for Manifold to be a place for the best DNA technologists, protein engineers, and drug hunters to realize the full potential of ML².

If they succeed in tackling the bottleneck of tissue-targeted delivery, they could cement another foundational step in the biotech industry’s compounding staircase.

Thanks for reading this essay on Manifold.

I’d like to thank Gleb Kuznetsov, Pierce Ogden, Shane Lofgren, George Church and Steve Holtzman for sharing their lessons and ideas with me. My hope is that this essay helps the next generation of technical founders.

If you don’t want to miss upcoming essays, you should consider subscribing for free to have them delivered to your inbox:

Subscribe now

Until next time! 🧬

Subscribe now

Before jumping into today’s essay, I’ve got an update to share.

I’ve joined Amplify Partners to help build a $200M early-stage venture capital fund dedicated to technical founders in bio. It’s pretty obvious that this has been my obsession for years—I’ve got receipts! We are building a firm in service of the next generation of technical founders like Neil Kumar at BridgeBio, Eric Kelsic at Dyno Therapeutics, Simon Burns at Vial, Viswa Colluru at Enveda, and more. There is so much left to build…

For more, you can read our Manifesto or check out the news coverage.

This is also a good reminder: I’m not a journalist! My goal is not to record events from a detached and objective vantage point—like the incredible cameras that hover over the action in modern sports stadiums. I’ve got opinions. Even worse, I’ve got positions in companies, and relationships with many of the people in my essays.

In other words, I’m inextricably linked to the subject—accelerating biotech is my life’s work, after all. Just as before, I’ll specifically highlight when I’m writing about companies that we’ve decided to partner with.

Lastly, I’ve got a message for : thank you, thank you, thank you. I learned an extraordinary amount working at Not Boring. About writing and investing, sure. But also, life. Packy is one of the smartest and most generous people I’ve ever met. Despite being a New Yorker, Packy is the clearest embodiment of Silicon Valley positive sum thinking. It feels like he is hardwired to use his platform and gifts as a writer to boost others. I describe Packy as “chaotic good.” He’s a prolific writer, absurdly fast marathoner, brother, husband, dad, and friend. I’m not sure where he finds the energy to do it all, but I know that I’ll be a lifelong reader and Packy fan. Thank you.

Alright, let’s jump in! 🧬

On August 4th, 2017, Andrej Karpathy told us something profound about where the world was going in a cryptic tweet: “Gradient descent can write code better than you. I’m sorry.”

He was talking about the magic of deep learning—a subject he knew a lot about. At the time, Karpathy was leading Tesla’s autopilot efforts. By then, he’d spent over a decade obsessed with this subfield of AI, beginning his exploration well before its string of sweeping breakthroughs and Nobel Prizes. When he started out, the field was a backwater academic discipline.

As an undergrad at the University of Toronto, he studied under Geoff Hinton—now considered “the godfather of AI”—and attended his small reading groups on the subject. I can only imagine Hinton, a tireless, rail thin professor who claims to have literally never sat down since 2005, proselytizing to a young Karpathy in a practically empty university class room about a secret that few people yet understood.

The secret was that artificial neural networks—especially deep neural networks—are algorithms that can learn complex computer programs from data.

Consider an image classifier. How would you implement a program to determine if an image contained a kitten? Maybe you’d start by looking for patterns of pixels that looked like the outline of a face with pointy ears. But this would be brittle and would require thousands of lines of code accounting for different orientations and edge cases. Extending beyond kittens would quickly get out of hand.

Instead of writing code for an image classifier, we can train neural networks to do this by providing pairs of images and their corresponding labels. With sufficient data, the learned classifiers can even surpass human performance on the same task. The deep kernel of truth in Karpathy’s tweet was that gradient descent—the optimization algorithm responsible for tuning the networks to learn this type of mapping—could produce programs that no human could replicate.

In a subsequent blog post entitled Software 2.0, he produced a simple diagram describing this concept that has been etched into my own neural network ever since.

Software 1.0 consists of the set of programs—like terminals, word processors, or Web browsers—that humans can write by hand. Software 2.0 is the term Karpathy uses to describe the expanded set of programs that can be learned by neural networks with optimization algorithms (such as gradient descent).

Karpathy continued, “It turns out that a large portion of real-world problems have the property that it is significantly easier to collect the data (or more generally, identify a desirable behavior) than to explicitly write the program.”

Because of this, deep learning—which is now basically synonymous with AI—has steadily marched across many disciplines of computer science delivering world-class results. Even for problems that seem very much like traditional software 1.0, such as sorting and hashing, algorithms are now being rewritten by AI.

And of course, AI can deliver product experiences that simply aren’t possible to implement with traditional engineering. ChatGPT lives very far outside of the program space accessible via human coding. This paradigm shift has already generated trillions of dollars of market value.

Karpathy was right.

Gradient descent can write code better than you. I’m sorry.

Few disciplines feel this more acutely than biology.

A longstanding joke about biology has been that if you enjoy science but are bad at math, it’s a great major. Molecular biology is not like physics. The complexity—and nonlinearity—of living systems has required a brute force approach to knowledge creation and dissemination. Biology textbooks are full of facts and cartoons, not equations.

AI is starting to change this in a big way. After decades of research—including lots of attempted math and software 1.0—the problem of 3D protein structure prediction was effectively solved by deep learning.1 As Karpathy said, many real-world problems are easier to solve by collecting data rather than attempting to write programs. This is definitely one of those problems.

This breakthrough—and the resulting progress on the inverse problem of protein design—earned David Baker, Demis Hassabis, and John Jumper the 2024 Nobel Prize in Chemistry.

It’s not too hard to understand the logic behind this. Few macromolecules are as important as proteins. Delivering a step change in our ability to resolve and design their structures is a big deal.

With all the excitement and progress, protein modeling is now one of the fastest moving areas of biological research. As Brian Naughton has chronicled, our ability to design protein binders is rapidly improving. One way to think about this is that the binding interaction between two molecules—either proteins or ligands—is a logical next step for molecular machine learning.

But what about entire complexes of molecules, or even interactions between those complexes?

In 2023, I speculated about this, saying, “We already have powerful models of specific parts. Now we’re learning models of parts composed together into circuits. What about models of combinations of circuits—getting us closer to predictive models of whole cells? Or combinations of cells—getting us closer to models of tissues? Can we compose the outputs of our BioML models together like a set of Matryoshka dolls?”

Even broken clocks are right twice a day. This does seem to be where research efforts are now headed. There has been a palpable uptick in focus on a longstanding Holy Grail of biology: effectively simulating the behavior of cells.

Top labs and companies around the world are now racing to build the “Virtual Cell,” a concept that has been around since the turn of the century. With new tools, the big bet is that this vision is now within reach.

This concept is like an onion. At the highest level, “virtual cell” is a catch-all term for garnering excitement—and funding—that is abstract enough to mean many things to many different people. Basically, a buzz word.

Going one layer deeper, researchers are producing perspective articles—the scientific equivalent of manifestos—laying out the shape of more specific research programs. A particularly important example is the recent paper How to build the virtual cell with artificial intelligence: Priorities and opportunities, which we’ll soon explore in more detail.

At the core of the onion, of course, are the actual research results. We’re seeing the first wave of architecture innovation—and importantly, benchmarking—and several concerted efforts to scale data generation.

But before we start peeling back the onion on AI virtual cells, let’s consider the alternative approach. What it might look like to directly write down math to describe cells?

“A Leap of Faith”

Coming of age in the Bay Area in the 1980s, Markus Covert remembers a special blend of anarchy and excellence all around him. There were few rules and none of the extreme pressures around crafting the perfect set of “extracurriculars” that kids face today. But intellectual stimulation was easy to find. He’d end a day of aimless wandering by stopping over to visit a friend whose family who had just bought the first Macintosh. The sleepy suburbs of Silicon Valley were evolving into the tech mecca of the world.

In university, Covert decided to study chemical engineering, partially because of his longstanding interest in science, partially because it was considered to be the hardest major on campus. It lived up to its reputation: it was really hard. But the payoff wasn’t quite worth it. Modeling petrochemical plants wasn’t particularly awe-inspiring.

Underwhelmed by chemical engineering, Covert searched for the small set of graduate programs in biological engineering that were offered at the time. He found his way to U.C. San Diego, which had launched the first bioengineering department in the entire U.C. system just a few years earlier in 1994. This proved to be a wise choice.

At the time, the field of genomics was just starting to coalesce. One of the centers of excellence—and plenty of anarchy—was San Diego. Craig Venter had recently established The Institute for Genome Research, a private research center that was tackling seemingly insane projects.

Venter, an extraordinarily forward-thinking and competitive scientist who would later become famous for racing against the public Human Genome Project, assembled a small group of scientists who matched his intensity and ambition.

In successive landmark papers, they produced the first genomic map of a free-living organism, the genome of the world’s smallest organism (Mycoplasma genitalium), and the genome of a famous gut pathogen.2

But the vision extended far beyond genome sequencing. Venter and his team pushed forward a research agenda to simulate entire cells on a computer and ultimately even synthesize their genomes. Sequencing. Synthesis. Simulation. These ideas hung together as part of a mission to engineer life in a way that was fundamentally new.

Covert remembers this work blowing his mind. “To me and other young biologists in the late 1990s, this gang was Led Zeppelin: iconoclastic, larger-than-life personalities playing music we had never heard before.”

In his own research, Covert worked under Bernhard Palsson, one of the early pioneers of computer modeling in biology. It was a perfect intellectual fit. He became obsessed with the beauty of generating verifiable predictions. For his main project, Covert produced an important paper connecting Palsson’s techniques for modeling cellular metabolism with approaches for representing gene regulation.

After an initial academic faculty search with little success, Covert convinced David Baltimore—another giant of modern biology—to take him on as a postdoc. After earning the Nobel Prize at age 37 for his work on reverse transcription—which required revising the Central Dogma of molecular biology—Baltimore became a scientific leader. He helped establish the Whitehead Institute before serving as the president of Rockefeller University and, ultimately, the president of Caltech.

Initially, Baltimore was skeptical of Covert’s Venter-inspired “systems biology” ideas. Covert remembers chasing after Baltimore—who was whipping a Segway prototype around the Caltech campus—as he advocated for model-driven biological research. Finally, Baltimore relented, saying “We’ve never had somebody with your type of background in the lab before, so maybe that’s reason enough.”

At Caltech, Covert fell in love with experimentation. He learned how to combine his modeling techniques with meticulous experiments tracking individual transcription factors in actual cells. Crucially, he was exposed to the early prototypes for live-cell imaging technology. In order to learn, Covert volunteered to monitor the microscopes overnight, holding together a makeshift rig consisting of a chicken egg incubator, cardboard, and duct tape, making sure the temperature remained a consistent 37 degrees to keep the cells alive and happy.

Finally, Covert got the chance to move back to the Bay to start his own lab at Stanford. He continued his work on computer modeling and live-cell imaging. But there was a deeper question below the surface.

What would it take to simulate a cell?

In 1984, a Yale biophysicist named Harold Morowitz had laid out a plan of attack in an article entitled The completeness of molecular biology.

First, sequence the smallest organism, a Mycosplasma bacterium. Covert’s former neighbors in San Diego at the Venter Institute had already done this. Check.

Next, simulate the behavior of each of the ~600 genes estimated to be in the genome. Morowitz reasoned, “At 600 steps, a computer model is feasible, and every experiment that can be carried out in the laboratory can also be carried out on the computer. The extent to which these match measures the completeness of the paradigm of molecular biology.”

In practice, this step proved to be much harder than expected. Venter and Clyde Hutchinson’s early prototypes represented only 127 genes in the M. genitalium genome and the simulation results barely resembled experiments. The paradigm of molecular biology felt far from complete.

Biking home on Valentine’s Day in 2008, Covert mulled this over. Suddenly, discrete ideas collected throughout all of his training started clicking together.

At Caltech, he had stared down the microscope at individual cells. A whole-cell simulation should aim to approximate one cell rather than a population of cells. And similar to his thesis work, it would require integrating distinct mathematical representations for different cellular processes into one model. Thinking back to his chemical engineering days, he thought of HYSYS, a chemical process simulator composed of discrete modules.

He immediately got to work sketching out modules for each cellular process in a single M. genitalium cell. Things started to snowball. Covert recruited two graduate students, Jonathan Karr and Jayodita Sanghvi, who were ambitious (and maybe, crazy) enough to sign up for this new undertaking.

This was not exactly glamorous work. Karr, Sanghvi, and Covert spent two years visiting multiple research libraries to physically scan nearly a thousand research papers. They were looking for any crumb of molecular information about M. genitalium they could get their hands on. All of these data points, along with theoretical assumptions and measurements from other species, were stored in a MySQL database.

As they expanded and refined their modeling efforts, the team grew. Covert recruited researchers with a variety of different backgrounds. One crucial addition was Jared Jacobs, a childhood friend of Covert who had gone on to become a talented software engineer at Google. Before joining a new company, Jacobs took a leave of absence to help out for a few months.3

With this influence, the team adopted test-driven development and moved towards an object-oriented programming model. Each discrete cellular process and molecule was represented as its own object to help grapple with the complexity of the problem.4

Another essential modeling assumption was that “even though all these biological processes occur simultaneously in a living cell, their actions are effectively independent over periods of less than a second.” This meant that each module—of which there were 28 in total—could be executed independently for each one second interval. Information that needed to flow between modules could be exchanged between these discrete time steps forward.

Years into this project, there was no empirical evidence that this crucial assumption—or many others that were baked into the model—would actually yield results. As Covert recalls, “it was a leap of faith.”

It’s the type of modeling problem that “would drive mathematicians insane,” as Covert puts it. There is no bounded solution. “But that’s where engineers shine.”

The model started out with abysmal performance. Simulations had little to no correspondence with reality. But after over a year of hacking, tinkering, and tuning, experimental and digital results started to converge. It was gradient descent by hand.

The final product, published in Cell in 2012—four years after Covert’s initial spark of inspiration—was extraordinary.5 Every single annotated gene was accounted for. Distinct mathematical tools were matched with each cellular process.6

While far from perfect, the whole-cell model approximated cell growth and division, produced values within an order of magnitude of a wide range of metabolic data, and correlated with experimental gene expression data.

It was a big step forward. A central design element was the incorporation of “sensors” that read out cell state at each step of the simulation. With this digital measurement apparatus, they explored a wide range of biological questions with their model.

This was true “model-driven biological discovery.” As Horowitz had envisioned, each discrepancy between prediction and experiment was a chance to move closer to a “complete” understanding of molecular biology.

And there were many discrepancies. The authors wholly acknowledged this, saying, “similar to the first reports of the human genome sequence, the model presented here is a ‘first draft,’ and extensive effort is required before the model can be considered complete.”

One major bottleneck in making further progress was the lack of experimental tractability of M. genitalium compared to other model organisms. The organism’s small genome helped reduce the scope of the challenge for the first whole-cell model. But M. genitalium’s small cellular size, penchant for antibiotic resistance, and lack of established tools made it impossible to validate certain predictions in the lab.

In the thirteen years since this work, Covert’s lab and others have marched forward on the problem of modeling E. coli. While E. coli is one of the simplest model organisms, its genome is nearly an order of magnitude larger than M. genitalium (4,641,652 versus 580,070 base pairs) and contains over 4,000 genes. In 2021, the E. coli whole-cell model represented 43% of these genes.

Covert thinks considerable progress has been made since then. He believes we are approaching a “Turing test moment” where no E. coli biologist could reliably detect the difference between the outputs of the simulation and a matched experiment.

For one of the most exhaustively studied microbes, simulation and reality may soon be indistinguishable.

For human cells, with billions of DNA base pairs and tens of thousands of genes, the timeline for mechanistic whole-cell models is much less certain. It may still be a leap of faith that it’s even possible.

Which brings us back to gradient descent. What if computer models of cells are the quintessential type of program that we should be learning from data rather than attempting to define and parameterize by hand?

The Bitter Lesson

Silicon Valley has changed since the 1980s. While the built environment has remained frustratingly consistent, tech has exploded. Of the top ten largest companies in the world, seven are now technology companies. A new cultural intensity exists—for technologists and for their kids—that Covert doesn’t remember being present growing up. Maybe the ratio of excellence and anarchy has shifted. Or maybe the expectations are heightened.

With great power comes great responsibility.

In the 20th century, legendary American philanthropists such as Andrew Carnegie, John D. Rockefeller, and Henry Ford contributed unprecedented sums of money to public works projects. Entirely new universities, research institutions, libraries, and hospitals were established. According to the historian Olivier Zunz, “The philanthropic projects were acts of generosity and hubris on a scale never before entertained.”

These are big shoes to fill for the new Tech Elite, who are now grappling with how to best exercise their newfound power. Much like their predecessors, an early focus of Silicon Valley philanthropists has been scientific research. History doesn’t repeat, but it often rhymes.

A central example is the Chan Zuckerburg Initiative. Founded in 2015, Mark Zuckerberg and Priscilla Chan announced they would be donating 99% of their Facebook (now Meta) shares via the CZI’s efforts.

The CZI’s first major project was the establishment of a $3B Biohub that would provide new resources and connective tissue between the Bay Area’s three major biomedical research institutions: UCSF, Berkeley, and Stanford.7

The Biohub’s founding mission is to “cure, prevent, and manage all diseases by the end of this century,” which was immediately polarizing.8 Zunz could have just as easily been writing about this project.

At first, scientists—including Steve Quake, the CZI’s first Head of Science—laughed at the idea. Quake joked, “I couldn’t say it with a straight face … I don’t know why you hired me, because I couldn’t say the mission.”

But gradually, the idea grew on him and other scientists. With most research relying on a bureaucratic and increasingly conservative NIH, a new institution was a good thing. And crucially, a longer time horizon for funding meant that scientists could take bolder bets. Stacking these types of projects together over a century, who knows what could happen?

Specifically, the CZI aims to think on ten year time horizons—over three times longer than the average NIH grant. But a decade isn’t long enough to completely omit deliverables and clear milestones.

According to Quake, the CZI’s “North Star for the next decade is understanding the mysteries of the cell.”

The rationale was twofold.

First, Quake and others at the CZI had seen the march of gradient descent that Karpathy so succinctly described. AI is happening. They felt that proteins were the first breakthrough application of these techniques in biology, but they wouldn’t be the last.

Second, AI models need a lot of data. AlphaFold and subsequent protein models were only possible because of the data bank of crystal structures that researchers had contributed to for decades. When it came to cellular data, the CZI had been investing heavily in the development of enormous atlases of single-cell genomics measurements.

Maybe these massive single-cell datasets could fuel a step change in the performance of AI models for predicting cell behavior.

In March of 2024, CZI’s AI team, led by Theofanis Karaletsos, got to work. The first order of business was to host a workshop at the San Francisco Biohub to start sketching out this thesis in more detail. They convened a group of leading AI researchers and single-cell biologists to consider what might be possible.

Ultimately, conversations from the workshop helped to unify ideas that were being developed in each lab. Over the course of several months, these ideas were distilled into a perspective in Cell entitled How to build the virtual cell with artificial intelligence: Priorities and opportunities.

In the introduction, the authors outline prior efforts in cellular simulation. Covert’s “pioneering work” modeling M. genitalium in 2012 is cited as a major milestone. But the roadblocks for bottoms-up mechanistic approaches are enumerated.

Cells are composed of a diverse set of exquisitely complex processes. Each process operates across scales ranging from atoms to entire tissue systems in the body. To make matters even worse, cellular behavior is often nonlinear, meaning subtle differences in signaling can drive massive changes downstream.

Can we ever hope to produce a complete bottoms-up mathematical description of mammalian cells? When?

Instead, the authors propose a different approach: “Two exciting revolutions in science and technology—in AI and in omics—now enable the construction of cell models learned directly from data.”

If they are right, it wouldn’t be the first time. In practically every domain of computer modeling with sufficient data, learned models have outperformed more detailed mechanistic models. It’s what Rich Sutton, a Canadian AI researcher, calls “the Bitter Lesson.”

In a 2019 blog post that’s become famous in the AI community, Sutton wrote:

The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin. The ultimate reason for this is Moore's law, or rather its generalization of continued exponentially falling cost per unit of computation. Most AI research has been conducted as if the computation available to the agent were constant (in which case leveraging human knowledge would be one of the only ways to improve performance) but, over a slightly longer time than a typical research project, massively more computation inevitably becomes available. Seeking an improvement that makes a difference in the shorter term, researchers seek to leverage their human knowledge of the domain, but the only thing that matters in the long run is the leveraging of computation. These two need not run counter to each other, but in practice they tend to. Time spent on one is time not spent on the other. There are psychological commitments to investment in one approach or the other. And the human-knowledge approach tends to complicate methods in ways that make them less suited to taking advantage of general methods leveraging computation.

Consider natural language processing. Decades of research building exquisitely detailed models of linguistics and semantics produced models that were superseded by the Transformer. The Transformer is a general method that can scale with data and compute to learn rich representations of language.

Internet-scale data and compute won.

The argument sketched out in the CZI paper is that we should test the same recipe for cell biology. But instead of Internet data, we have omics to build on top of. And the improvements in sequencing technology have outpaced even Moore’s Law:

With the exponential improvements in DNA sequencing since the Human Genome Project, we can now produce atlases of single-cell measurements at an extraordinary scale. Rather than painstakingly developing mathematical models based on human knowledge, what if we attempted to learn cellular dynamics directly from this data?

In the proposed AI Virtual Cell (AIVC) approach, we should focus on using general learning methods to capture universal representations (URs) of each major cellular building block—such as DNA, RNA, and proteins.

Again, think about this as a big Matryoshka doll of different BioML models. A model for DNA. A model for RNA. A model for proteins. And a model to integrate this information into a snapshot of cellular behavior alongside data types like microscopy, proteomics, or RNA sequencing. And ultimately, connecting these models together in graphs to represent multi-cellular interactions.

The major research questions become:

1. What types of data do we need?

2. What are the right models for each data type?

Probably the most counterintuitive part of this strategy is what success would look like. Imagine that this works. We learn a general model that can accurately recapitulate cellular behavior.

We would have a model of cell biology—we just wouldn’t understand it!

To remedy this, the authors propose developing a suite of tools to perform computational experiments.

First, we would need manipulators. These tools would learn a representation of a perturbation—such as a chemical or genetic manipulation—that could be applied to a UR. A manipulator would convert one UR into another, representing a change in cell state.

But we would also need tools that produce human interpretable outputs, not just unintelligible vector representations of cell state. This is where decoders come in. These tools would learn to map between URs and synthetic experimental measurements, such as generated microscope images or sequencing data.

Essentially, we’d be swapping out physical cells for their virtual simulacra. The major benefit would be that virtual experiments scale in a way that physical experiments don’t. In theory, we could test billions or trillions of hypotheses using these models. If a result looks promising, it could be validated in physical reality.9

Like most perspectives, the AIVC paper attempts to unite many ideas that had already been swirling around. Many of the authors had spent years testing related approaches. Models like Universal Cell Embeddings, published in 2023, showed how protein models could be integrated with RNA data to create cellular representations. Other architectures like GEARS and scGPT were already developed to simulate cellular perturbations.

The central message was that this research direction—testing the Bitter Lesson for cell biology—is the most promising path forward. Driving this point home, the paper concludes by saying, “We believe that we are stepping into a new era of scientific exploration and understanding. The confluence of AI and biology, as encapsulated by AIVCs, signals a paradigm shift in biology and shines as a beacon of optimism for unraveling multiple mysteries of the cell.”

Natural Predators

Extraordinary claims require extraordinary evidence. Whenever a new modeling paradigm emerges, other scientists immediately attempt to poke holes in it. It’s like a form of immune response. And typically, the grander the claims, the faster—and larger—the response is.

As soon as AI cell models emerged, third party benchmarks and evaluations started to trickle out. Some of the results were quite surprising. In September of 2024, a group of highly experienced biostatisticians from Germany published a preprint comparing many of the first AI models for perturbation prediction against “deliberately simplistic” linear models.

Predicting the response to perturbations has become a major focus area for AIVC researchers. With experimental tools like CRISPR, specific genes can be up- or down-regulated, like turning a nob on a big cellular control panel. Using RNA sequencing, it’s possible to see what other genes were turned up or down based on that manipulation. Methods like Perturb-seq were invented to do this on a massive scale.

Basically, how does a cell switch up its gene expression programs when it is poked?

So the German lab evaluated how accurately the models could predict these responses for perturbations where two genes were up-regulated at the same time.

Two baseline models were used:

1. No change: in response to a double perturbation, this model would predict there would be no change to gene expression.

2. Additive model: simply return the sum of the change in expression for each single perturbation.10

Counterintuitively, the simple additive model had lower error rates than the complex AI models.

Separately, Eric Kernfeld produced similar results across a broader range of datasets during his thesis work at Johns Hopkins. In an accompanying blog post, Kernfeld wrote, “The natural predator of the methods developer is the benchmark developer, and a population boom of methods is naturally followed by a boom of benchmarks.”

The natural predators had found vulnerabilities.

AI models for these types of prediction tasks are clearly promising. The developers of the linear baseline wrote, “We do not consider our negative results on the foundation models’ performance in the prediction task as arguments against this line of research ... the progress brought by the transformer architecture and the transfer learning paradigm for many machine learning tasks is real and substantial.”

But the first implementations still felt far from producing “non-trivial insights,” as the benchmark developers put it.

I asked Yanay Rosen and Yusuf Roohani, two of the primary authors of the AIVC perspective, about how they interpreted these results. Both researchers wholeheartedly acknowledged the limitations of these early prototype models.

Rosen also pointed out that for the task of creating universal cell embeddings, AI models are already producing results that can’t be captured with simpler methods. These approaches make it possible to represent cells across diverse datasets, tissue types—and even species—in a single shared coordinate space. Biologists are already using these representations to discover new cell types and answer fundamental questions in evolutionary biology.

Recently, the AI team at the CZI extended this research direction with the release of the Transcriptformer model. Trained on over 112 million cells collected from 12 different species—representing 1.5 billion years of evolutionary distance—the model was designed for a wide range of prediction tasks. In their release, the CZI team wrote, “Researchers can use TranscriptFormer to predict what different types of cells are, whether a cell is diseased, and how genes interact.”

For the task of positioning a cell in relation to all other cells, there seems to be enough data for AI models to shine. But for predicting the dynamics of how a cell changes when perturbed, existing data and models appeared to be insufficient.

This is starting to change.

Roohani, who now leads a machine learning research group at the Arc Institute, is very excited about more closely integrating with experimentalists. The Arc, which is another product of large-scale tech philanthropy for science, has launched its own Virtual Cell Atlas project. A major focus is large-scale data generation.

This week, the Arc team published new results from this research program. Their new model, called State, appears to be a major step forward for perturbation prediction. You can see the interplay between benchmark developers and methods developer in action: linear baselines are now front and center. And this time around, the models appear to be picking up “non-trivial insights” more effectively.

A key driver of the jump in performance comes from a more careful representation of the biological noise that is present in the data. Cells are fundamentally noisy, heterogenous systems—and the experiments interfacing with them bring their own limitations. The State model uses mathematical bookkeeping to account for the biological and experimental noise that can drown out true signal.

When they do that, something very important happens: the model’s performance appears to meaningfully increase with more data. In other words, Sutton’s Bitter Lesson starts to kick in.

Recently, Tahoe open-sourced a massive dataset measuring 60,000 drug perturbations in 100 million distinct cells—dwarfing all other publicly available single-cell datasets.11 After training on this scale of data and diversity of cellular contexts, the State model was able generalize to a wider range of perturbation predictions.

Signaling what we should expect over time, the authors wrote, “Overall, the performance and flexibility of State sets the stage for scaling the development of virtual cell models.”

So, what are virtual cells?

Let’s start with what they are not.

It’s hard not to hear this term and imagine a fancy graphical rendering of a cell in its gyrating three-dimensional glory. If you zoomed in, you’d see ribosomes meticulously examining RNA molecules as amino acid chains spooled outwards, or kinesin proteins marching along microtubule walkways.

Crucially, every piece of molecular machinery would behave the way that it would inside of an actual cell. It would be the Magic School Bus for grown-up cell biologists.

That is not what scientists are building—at least, not yet.

For Covert, the more appropriate analogy is a meteorological simulation. As he and his research team wrote in their own recent perspective, “numerical weather prediction is a comprehensive endeavor to integrate observations from around the world and in space, over multiple timescales, into a mathematical model which both holds an initial state of the global system and can produce forecasts of changes in the atmosphere several days in advance.”

He and others hope to produce an equivalent mathematical model capable of simulating microbial systems. And as Morowitz envisioned in the 1980s, every glitch in these models can be treated as negative space in our broader map of molecular biology.

Covert told me he “wants to find Neptune,” referring to the 19th century astronomer Urbain Le Verrier’s ability to discover an entirely new planet “with the point of his pen.”

AI researchers think about the problem differently. Guided by the Bitter Lesson, the goal is to specify as few priors as possible, letting the data speak. The first goal is prediction, not understanding.

The AIVC models are learning “the statistical shape of life,” as Karaletsos puts it. But this “shape” is imprinted in massive numerical vectors within these networks that are illegible to humans. In Geoff Hinton’s early deep learning courses, he would tell his students: “To deal with hyper-planes in a 14-dimensional space, visualize a 3-D space and say ‘fourteen’ to yourself very loudly. Everyone does it.”

What if these models achieve incredible predictive power? The real work would begin. Armed with virtual instruments, biologists would be able to manipulate cells at a scale that would be unimaginable in the lab.

If the Venter gang was like Zeppelin, these new research groups are more like Taylor Swift.12 Massive projects are produced back-to-back, accompanied by carefully coordinated press releases and social media announcements—which are essential for mimetic distribution in modern science.

Over time, it’s possible that these distinct sounds converge. The bottoms-up mechanistic modeling efforts may start to integrate AI methods into their systems. Covert’s lab was one of the first groups to develop deep learning methods for microscopy, after all.

Similarly, as AIVC research efforts expand beyond RNA, model architectures will become increasingly complex. Hierarchical systems with distinct modules for each biological process will start to resemble approaches from the world of whole-cell simulation.

For now, Roohani has a much simpler goal: make the existing models good enough that experimentalists adopt and use them. Like the “GPT Moment,” this may not require any semblance of perfection.

This could create a feedback loop of compounding progress as predictions are tested in labs around the world, creating more data to further refine the models.

Quake imagines a future where “cell biology goes from being 90% experimental and 10% computational to the other way around.”

The first step in that direction will be convincing researchers that they should at least run a query or two before picking up a pipette.

Notes and Further Reading

Much of this essay focused on whole-cell modeling efforts from the Covert Lab and early AIVC research from the authors of the Cell perspective:

• Whole-cell modeling:

  • A Whole-Cell Computational Model Predicts Phenotype from Genotype

  • A forecast for large-scale, predictive biology: Lessons from meteorology

• AI Virtual Cell:

  • How to build the virtual cell with artificial intelligence: Priorities and opportunities

  • Predicting transcriptional outcomes of novel multigene perturbations with GEARS

  • scGPT: toward building a foundation model for single-cell multi-omics using generative AI

  • Universal Cell Embeddings: A Foundation Model for Cell Biology

  • TranscriptFormer: A Generative Cross-Species Cell Atlas Across 1.5 Billion Years of Evolution

  • Predicting cellular responses to perturbation across diverse contexts with State

  • Tahoe-100M: A Giga-Scale Single-Cell Perturbation Atlas for Context-Dependent Gene Function and Cellular Modeling

One of the tough things when writing about such an active space is avoiding the urge to be exhaustive. There is a lot more research happening in this space—including results that came out as I was writing:

• Genentech has been pursuing this work for years under the direction of Aviv Regev, one of the AIVC authors:

  • Toward a foundation model of causal cell and tissue biology with a Perturbation Cell and Tissue Atlas

  • Gene-embedding-based prediction and functional evaluation of perturbation expression responses with PRESAGE

  • PerTurboAgent: A Self-Planning Agent for Boosting Sequential Perturb-seq Experiments

• Altos Labs has been working on perturbation prediction and producing benchmarks.

• Recursion has consistently prioritized virtual cell research.

• Last week, Xaira Therapeutics published a massive Perturb-seq dataset and announced their own efforts to build a virtual cell model.

• Noetik has been building virtual cell models for spatial biology.

• Theofanis Karaletsos also helped to lead research on perturbation prediction during his time at Insitro.

• While seemingly less interested in the “virtual cell” terminology, NewLimit has been developing AI models to predict the effects of combinations of transcription factors (TFs). The goal is to identify TF cocktails that drive safe cellular rejuvenation.

• This week, the Arc Institute announced a Virtual Cell Challenge.

• Demis Hassabis has also prioritized virtual cell research as the logical next step for molecular machine learning. In a recent interview, he outlined a research plan focused on continually moving upwards from proteins to protein complexes, molecular pathways, and ultimately, cells. In some ways, this is a hybrid of the different ideas outlined in this essay!

This list is still far from exhaustive.

Thanks for reading this essay on the rapidly evolving world of virtual cells.

I’d like to thank Markus Covert, Yanay Rosen, Yusuf Roohani, and Theofanis Karaletsos for their helpful conversations as I was doing research.

If you don’t want to miss upcoming essays, you should consider subscribing for free to have them delivered to your inbox:

Subscribe now

Until next time! 🧬

Subscribe now

Today, we’re going to briefly pause the hub-and-spoke series (first articles here and here) to consider the broader evolution of drug discovery technologies over time—and the impact this may have on biotech strategy.

Enjoy! 🧬

Ten years ago, the scientologist (no, not scientist) Bob Duggan sold his American biotech startup to AbbVie for $21 billion (no, not million). His company, Pharmacyclics, had developed a highly promising cancer drug and was handsomely rewarded for it. In fact, Duggan’s payout of over $3.5 billion was reported to be one of the largest returns from a public buyout in history.

Aside from the eccentricities of Duggan as a founder—and the other large characters involved that made all of this book worthy—this was a relatively normal deal. A Big Pharma company was under serious pressure to find a new source of revenue as one of its blockbuster products reached the end of its patent exclusivity. An American startup’s breakthrough drug became that new source.

Much of the biotech industry is predicated on this dynamic. Small startups have become the primary source of innovation, running the majority of early-stage clinical trials each year. Big Pharma companies buy these startups at a rich premium to continually replenish their pipelines.

But things are changing. A decade later, Duggan is back. This time, he’s got a new drug that beat Keytruda—Merck’s cancer immunotherapy drug that grosses $30 billion a year—in a clinical trial. Crucially, Duggan didn’t find this drug in an American lab. He licensed it from a Chinese company.

Unlike Pharmacyclics, this story didn’t end with a massive multi-billion dollar acquisition. Instead, Merck went to the same source, buying their own version of the same drug type for $500M from another Chinese company.

This is a really big deal. Much like we recently saw in AI, China has become a serious competitive threat in biotech, demonstrating the capacity to rapidly develop new drugs that can rival—or surpass—those being produced in American labs. In other words, “the drug industry is having its own DeepSeek Moment.”

Consider another example. As GLP-1 drugs became an explosive success, pharma companies raced to get their hands on their own next-generation versions of this product class to compete with Novo Nordisk and Eli Lilly for market share. Again, Merck looked to China, acquiring an oral GLP-1 drug for $112M upfront. The deal was backloaded with milestones up to $1.9B based on commercial success.

For context, Viking Therapeutics, an American biopharma company with an oral GLP-1/GIP agonist in its pipeline, currently has a $3.8B market cap. Rather than pursuing a wholesale buyout of Viking, why not just grab a molecule from China for a bargain and see if it works?

In a moment where the biotech market is depressed, the heightened external competition makes things even harder. While the M&A market has already been slow, now founders and investors alike are losing sleep over deals falling apart at the last minute because of stealthy Chinese competitors they had never even heard about.

This whole situation has spurred a lot of analysis. Some of my favorites so far are David Li’s reflection as a Chinese American biotech startup founder, his subsequent analysis in the Timmerman Report, Bloomberg’s coverage, and Alex Telford’s characteristically thoughtful take on the question: Will all our drugs come from China?

Here, I want to zoom out and consider how we got here.

In 1987, Merck was featured on the cover of Fortune Magazine as “America’s Most Admired Corporation” for “betting on magic molecules.”

In the decades prior, Merck scientists were responsible for producing breakthrough molecules for hypertension, some of the most successful vaccines in history (Maurice Hilleman, one of the most prolific vaccinologists of all time, was a scientist at Merck), the first statins, and entire new classes of antibiotics.

Now, for two of the hottest products in the market, Merck has looked outside of the American biotech startup ecosystem—which was still struggling to come into existence in 1987—and acquired molecules from the Chinese ecosystem—which wasn’t a major source of innovation until very recently.

Clearly, the global drug discovery industry has undergone considerable evolution. At this juncture, I think it’s worth considering the long arc of commoditization for drug discovery technologies, as well the implications of this historical pattern for the future of the industry.

The Long Arc of Modality Commoditization

Let’s think about biologics.

For most of history, nearly all drugs were plant-based chemicals with a useful impact on human physiology that humans had serendipitously discovered. Over time, tools were developed to more systematically search through chemical space for useful small molecules. Another small category was the set of proteins, such as insulin, that could be isolated from animals and used to treat disease.1

The recombinant DNA revolution that gave rise to Genentech’s founding in 1976 planted a seed of radical change. Using the tools of genetic engineering, it became possible to produce biologically derived molecules to mitigate disease in entirely new ways. Fifty years later, there are nearly as many biologics approved each year as there are small molecules.

But at first, few people believed this new type of drug would be possible. And a similarly small cohort of people in the entire world had any of the requisite skills necessary to even try. It’s easy to lose track of the fact that Genentech started out as a fringe operation full of scientists working day and night in a uniform of tee shirts, jeans, and running shoes.

As Peter Thiel would frame it, the company had an important shared secret that the rest of the world didn’t understand yet.

Success changed this. When Genentech achieved their first product breakthroughs, the market responded with fanatical optimism. Nothing—not even Princess Diana—seemed more exciting than engineering life.

When Genentech went public in 1980, the initial share price was $35. Within one hour of trading, the price nearly tripled, jumping to $88. It was an absolute frenzy of excitement.

By 1983, American firms had invested half a billion dollars into new biotech firms. Two years later, the U.S. Department of Commerce estimated that there were roughly two hundred biotech firms that had raked in nearly $2B in investment in total.

In retrospect, this was clearly a hype bubble. Recombinant DNA technology was still nascent. More big wins like recombinant insulin didn’t quickly follow. There were still looming regulatory questions. Manufacturing biologics at scale was a challenge.

The initial boom was followed by a period of disillusionment and contraction. In 1985, a reporter in Massachusetts wrote,

If all this seems far too much for scientists with little or no business track record, you're right. For that reason alone, two trends have become dominant in the post-boom biotech industry.

First, at the few firms which are better capitalized, the scientists/founders/chief executives are hiring, or being forced to hire a traditional “numbers cruncher” as company president or chief operating officer. These are seasoned business executives from major companies, not the groves of academe. At BioTechnica International in Cambridge, they just hired a 20 year Dupont man. At Collaborative Research, in nearby Lexington, the new president formerly headed a Johnson and Johnson subsidiary. So did the new president of Damon Biotech in Needham Heights.

Second, momentum in the industry appears to have shifted in favor of large corporations. Especially those which supplied a big share of biotech's original funding through equity purchases and research and development contracts. In a sense, the good response to the first public stock sales by the new biotech firms only postponed the bad news for many of them. Now, when the young companies badly need another infusion of capital, corporations like Dupont, W.R. Grace and Co., Monsanto, and Eli Lilly & Co. are instead directing their investments toward inhouse biotech capabilities. As a result, industry experts are predicting a spate of mergers. Nelson Schneider, an E.F. Hutton analyst, believes that as many as two thirds of the biotech companies will either merge together or be acquired by major drug or chemical manufacturers.

In other words, the industry was forced to mature and actually realize revenues, and incumbent organizations wised up and started to build out their own biotech capabilities internally. By the end of the 1980s, most biotech stocks had lost three quarters of their value.

But there was an important kernel of truth at the center of this bubble: recombinant DNA really was a transformative tool for making new medicines. Even in a more sober environment, the companies with the resources, technology, talent, and grit to survive kept churning out products.

After synthetic insulin, Genentech produced seven more biologics throughout the 80s and 90s. Amgen, another early biologics pioneer founded in 1980, had their own string of breakthrough medicines that set them apart from the rest of the struggling competition. Regeneron, which was founded in 1988 after the initial boom, differentiated themselves over time with a keen focus on human genetics and a powerful technology platform for producing monoclonal antibodies—which proved to be one of the most important types of biologic.

While there were still many skeptics, the scope of biologics continued to expand across medicine, reaching the point in 2022 where their volume of approvals was first on par with small molecules. (The graph we looked at earlier.)

The commercial success for these first movers is also undeniable. Genentech was acquired by Roche in 2009 for $46.8B, where it still operates as a highly impactful independent subsidiary. Amgen currently has a market cap of $168B. Regeneron is now worth $78B, with a share price nearly 4000% higher than at their IPO.2

This evolution is almost a textbook example of the phenomenon known as the Gartner hype cycle. An initial “innovation trigger” causes a big spike in hype and excitement. When the hype isn’t immediately justified, the market cools. If the initial trigger has substance, there is a more gradual rebound over time.

And we do seem to have entered a “plateau of productivity.” The ability to produce biologics is no longer a secret guarded by a small set of companies. Scientists around the world have now spent decades refining the tools for developing these drugs. An entire wave of companies has cropped up to offer antibody development as a service.

Consider Adimab. Founded in 2007, the company used the next generation of antibody engineering tools—namely yeast surface display—to rapidly produce new molecules for a large number of partners. They’ve now worked with over a hundred different partners, with “more than 75 clinical programs originating from our platform.”

Given that Adimab is a private company, it’s hard to do an apples-to-apples comparison of their commercial success relative to first movers like Amgen or Regeneron. But extrapolating from their $1.1B valuation during a secondary transaction ten years ago, the rough ballpark is that they’ve grown into a ~$5-10B business.

Now, if companies want access to a different technology, or have distinct preferences for cost, speed, geography, or a multitude of other factors, they can choose to go with FairJourney Biologics (valued at ~$900M in 2024), OmniAb (~$400M market cap), Ablexis, Specifica, Creative Biolabs, Twist Biosciences, Alloy Therapeutics, or somebody else. This list is illustrative and far from exhaustive.

What’s interesting to note is that across each generation, the magnitude of the businesses decreases by roughly one order of magnitude. The first movers grew into ~$100B+ companies. The leading service providers that followed became ~$10B+ companies. Now, the new entrants in the discovery market are ~$1B+ companies.

To me, this looks like the textbook definition of commoditization, which is the gradual process of making a good or service into a commodity and competing on price. A commodity is a good or service that is interchangeable with other commodities of the same type.

Think about electronics. At the start, very few companies could produce the best TVs. These companies charged a large premium. Over time, this premium was competed away. Now, a large number of companies sell massive flat screens packed with smart features for hundreds of dollars at Costco. This is commoditization.

Similarly, it’s getting increasingly hard to distinguish between antibody discovery providers, with many companies using similar technologies to produce antibodies against the same drug targets.

So far, we’ve exclusively focused on the history of antibodies. But I want you take a leap with me that will potentially annoy some drug developers: no discovery technology is immune from the inescapable pull towards commoditization—like virtually every other technology.

For big and small molecules alike, once discovery technologies—whether it’s high-throughput screening, in silico screening, in vitro or in vivo models, or an analytical assay—become standardized, companies around the world will compete to offer them as a service.

This is The Long Arc of Modality Commoditization.

Over time, revolutionary ideas become universal building blocks for the next wave of innovation.

Biotech’s Strategic Evolution

In parallel to the standardization and commoditization of discovery technologies, biotech investing became professionalized. With many decades of refinement, the industry moved towards standardized models for valuing companies. New strategies emerged.

One strategy that has gained considerable popularity is the “fast follower” approach where a new drug is developed to be “best-in-class” for a target with other drugs on the market, rather than to be “first-in-class” for a wholly new drug target.

An analysis published in 2003 pointed out two key benefits to the “quest for the best.” First, these drugs clearly have a lower risk profile because the target has already been validated by a drug approval based on human evidence. Investors typically talk about the amount of “biology risk” they are underwriting for a new target hypothesis. Second, the difference in risk didn’t seem to be commensurately rewarded. In fact, looking at drug launches between 1991 and 2000, the majority of blockbusters were developed against known targets, and the fast-followers created more value than the riskier novel drugs.

A clear example of this strategy is Merck’s acquisition of the oral GLP-1 agonist we looked at earlier. Novo Nordisk and Eli Lilly took on a lot of risk proving out the efficacy of the first GLP-1 drugs. Now, other companies are competing for fast-followers with improved properties, like formulation in a pill rather than dosing via injection.

Many biotech investors have taken this type of analysis to its logical extreme. As the size of early-stage rounds—and the funds financing them—have swelled, it’s gotten harder to justify risking big pools of capital on totally unproven target hypotheses. In practice, this has led to substantial crowding around validated targets.

To make things even more efficient, “virtual biotechs” became prominent around the 2010s, where all research and development was outsourced to discovery partners like Adimab. The goal was often to rapidly produce a best-in-class molecule against a known target that could be sold to a Big Pharma for late-stage development and commercialization.

This industry history is essential for understanding the recent explosion of Chinese licensing deals, because many of the top outsourcing partners were Chinese Contract Research Organizations (CROs).

WuXi Biologics, a sprawling Chinese company offering a large suite of biologics discovery and manufacturing services, has become the second largest outsourcing partner in the world, capturing over 10% of the global market.

Now, the extremely logical strategic evolution from China, encoded in a new set of policies in 2015, is to develop their own drugs rather than just remaining a service provider. In a world where most people are using commoditized discovery technologies against the same drug targets, China has two key advantages:

1. Speed. The new set of reforms made it possible to launch clinical trials much more quickly.

2. Cost. Salaries for Chinese scientists are a fraction of those for American scientists. An army of highly-skilled—often American-trained—researchers can be thrown at far more problems.

With these advantages, Chinese startups and biopharma companies have seemingly saturated the space of known drug targets. Companies place cheap “call options” on a wide range of targets in the form of pre-clinical or early-stage assets. When a specific target or product idea gains traction with Big Pharma, these “options” can be exercised by pouring gas on the existing program to race it forward.

This puts intense pressure on the fast-follower strategy. When American scientists go to bed at night, the machines in the labs of their competitors on the other side of the world keep humming.

So far, we’ve traced the history of commoditization for drug discovery technologies and the subsequent professionalization of biotech investing. These changes help to contextualize the “DeepSeek moment” for the industry.

Now, let’s consider where value might accrue in the future.

AI Could Be The Last Wave of Commoditization

Over the last few years, a lot of money has been poured into companies with moonshot visions to transform drug discovery with AI. Some companies, like Xaira Therapeutics, which started with $1B in “Seed” funding, aim to develop their own medicines. But many others, such as EvolutionaryScale, Profluent, Chai Discovery, and Latent Labs, are considering strategies more akin to Adimab, where this new technology is offered as a broadly enabling piece of infrastructure.

When Latent Labs was launched, Tony Kulesa at Pillar wrote:

What emerged was a clear vision for democratizing access to advanced AI tools in drug discovery. While every biotech and pharma company searching for therapeutic molecules understands the role AI can play, most aren’t in a position to develop their own models and tooling at the cutting edge. Simon’s insight was that by giving partners instant access to the best tools, Latent Labs could accelerate drug design across the entire industry.

The combination of large funding rounds and new business models has drawn a mix of intrigue and skepticism. Andy Dunn at Endpoints wrote, “Latent’s launch shows how AI-focused startups can buck tradition in biotech. Most biotechs are formed around a molecule, research paper or key intellectual property. Instead, Latent’s investors are betting on AI talent in Kohl and Alex Bridgland, another ex-DeepMind developer of AlphaFold, to figure it out.”

Let’s consider the bear case and the bull case for this investment thesis.

In the bear case, none of these technical directions—whether the focus is on novel data generation, scaling models, architecture improvements, or some combination of all three—meaningfully move the needle when compared to commoditized technologies.

In his field notes from the Molecular Machine Learning conference at MIT, Simon Barnett from Dimension wrote, “my interpretation of [the co-founder of Adimab] Dr. Wittrup’s talk was that he views monoclonal antibody (mAb) discovery as a mostly solved problem and that the impact of machine learning (ML) to this domain is exaggerated.”

If AI techniques prove to only make a small quantitative difference on problems like antibody discovery, the companies offering these solutions could join the long line of companies competing to offer these services. We could expect to see <$1B companies rather than ~$50-100B+ generational behemoths.

What about the bull case? Squint with me for a second, and imagine a trajectory of AI progress that leads to a qualitative difference, genuinely moving us into a world of design rather than discovery. Imagine a model that spits out zero-shot predictions of the platonic antibody—perfect affinity and specificity, exquisitely optimized along every dimension—for any target. Put in a target product profile (TPP), get out a drug.

That would probably be a pretty big deal.

One commonly invoked comparison is Cadence Design Systems, which is a $66B company that generates most of their revenue from licensing their software and IP for electronic design automation (EDA) for the semiconductor industry. For high-value domains, the best design tools can be extremely valuable. Could the “Cadence for Pharma” be even bigger?3

Is there any evidence for this technological trajectory?

Last year in March, the Baker Lab at the University of Washington published a preprint entitled Atomically accurate de novo design of single-domain antibodies. Building on decades of leading work on computational protein design, they introduced an AI model that could effectively generate miniature antibodies (called VHHs or nanobodies) for a given target.

These results stirred up an enormous amount of excitement and interest—including the $1B bet to launch Xaira. But the work was a proof of concept, not a magical black box capable of spitting out perfect antibodies. Scientists pointed out that the affinity of the generated nanobodies for their targets were still too weak to become drugs. And nanobodies are strange proteins that don’t fully resemble human antibodies—again limiting their clinical utility in many applications.

Slightly less than a year later, the Baker Lab “significantly updated” their original preprint, renaming it to Atomically accurate de novo design of antibodies with RFdiffusion.4 As you might guess, the title was changed because the work was extended beyond just VHH design. The updated preprint also demonstrated the design of single-chain variable fragments (scFvs) which are another antibody format that have two variable domains rather than the single variable domain of a VHH.

Another important update was a response to the concerns about affinity. The authors wrote, “While initial computational designs exhibit modest affinity, affinity maturation using OrthoRep enables production of single-digit nanomolar binders that maintain the intended epitope selectivity.” In other words, AI doesn’t produce perfect binders yet, but they can be quickly tuned with existing experimental techniques.

So that’s about a year. At the risk of drawing a line between two data points, progress seems pretty rapid. Moving forward, what if somebody creates a giant set of training data for affinity using OrthoRep and that step moves out of the realm of atoms and becomes encoded in the world of bits in the form of improved model weights?

In the next five years, what will prevent the continued march from VHHs to scFvs to full-fledged monoclonal antibodies?

Again, just squinting, it appears that we could be on the cusp of digitizing the development of biologics. If the advantages in speed and cost—and potentially quality—are substantial, this could lead to consolidation in the discovery market, where new entrants quickly attract a sizable portion of outsourced work, displacing incumbents.

Now let’s consider what the world looks like if we take the concept of “foundation model” seriously. What if important underlying patterns of biological structure and function are learned across many tasks? As Simon Kohl from Latent Labs told Endpoints, “The vision is grander. I think we can expand from there, and we will find over time many other areas beyond the molecular interactions level will be steerable with generative models.”

So if this—or any part of what I just outlined—is true, some of these companies could become really big.

But one of the biggest threats is likely to be… commoditization! After all, the entire “DeepSeek Moment” framing comes from the abrupt leap in AI capabilities from Chinese research groups with less resources than their American counterparts.

And there are already signs of this.

So far, Demis Hassabis, the co-recipient of the Nobel Prize for protein structure prediction and CEO of both DeepMind and Isomorphic Labs, has been betting on algorithm innovation rather than the development of a proprietary data moat for model defensibility. In a recent interview, he said, “make your algorithms better, your models better. You do have enough data — if you were innovative enough on your algorithm side.”

It’s been amazing how quickly serious algorithmic competitors have emerged.

In May of 2024, Isomorphic and DeepMind published a paper describing AlphaFold3, their latest and greatest structure prediction model. In September of 2024, Chai Discovery published and open-sourced a state-of-the-art model. Roughly two months later, a research group at MIT produced another open-source version with comparable performance.

It will be interesting to see where value accrues in this new AI race.

No matter what, all of these advances will open up new opportunities for innovation in other parts of the drug discovery stack.

Biological Insights Could Become More Valuable

Not all biotech platforms focus on therapeutic modalities. Some companies focus on the other side of the coin: the identification of new biological targets to drug. In Steve Holtzman’s “typology of platform companies” these are “insight platforms.”

Focusing on disease insights comes with its own basket of strategic challenges. In his original post, Holtzman wrote,

HOWEVER, Genus 2, Species A Platform Companies face a set of challenges not faced by Genus 1 Platform Companies. These arise intrinsically from the fact that the output of the Genus 2, Species A Platform Companies is data/information/insights, NOT, as with the Genus 1 Platform Companies, New Chemical Entities (NCEs) and biologic therapeutics.

• The history of data in the biopharmaceutical industry is the history of its commoditization.

• Companies whose “life’s blood” is drugs/products have a vested interest in rendering data “pre-competitive” (or, at least doing so after they have had proprietary access for a time). They win based on their products; they don’t want to be held captive by the owner of the information.

• A more restrictive intellectual property (IP) environment: gone are the days when a transcriptional profile showing over-expression of a gene in a diseased tissue (or a genetic mutation in the diseased state) could get you an issued patent of the logical form, “a method of treating disease X comprising modulating target A by any means” (with dependent claims stating that the “means” could be an antibody, an antisense, an RNAi, a gene therapy, a small molecule, etc.).

• Moreover, the demands of the customer base became more expansive over time. Whereas, in the 1990s, most big pharma customers were willing to agree to terms that restricted their use of the data to the discovery and development of small molecule drugs (because that is all they did), in the present all pharmas/big biotechs would demand the right to exploit the data for all therapeutic modalities.

• Finally, the Genus 2, Species A Platform Company has and builds its expertise in the generation, curation, and interrogation of the data. It does not come into existence or can afford to cultivate/invest in significant capabilities in drug discovery/development in one or more therapeutic modalities, or in-depth biology/translational capabilities, in one or more diseases.

The Net Result: the Genus 2, Species A Platform Company ends up abandoning its data/information business to build and become a drug discovery and development business.

Let’s break this down.

First, it’s important to understand that data generation technologies have their own long arc of commoditization. (That’s technology, folks!) Next, the main issue in the past has been the asymmetric negotiation with larger partners—who used to have proprietary access to discovery technologies that gave them an unfair advantage in actually creating chemical matter against new targets.

This dynamic has already started to shift. The growth and maturation of the CRO industry has already made it possible for insight companies to show up to partner discussions with their own NCEs rather than just patents around target insights.

What if AI accelerates this dynamic? Over time, as modalities become increasingly commoditized, the time and cost from target insight to developable chemical matter could be compressed even further.

In this world order, the pendulum could swing. New disease insights could become more valuable than any incremental starting point in chemical space against a known target.

The $100B+ GLP-1 success story, after all, was built on a biological insight, not a modality advance.

There are a few economic and technical realities that could slow down progress on this front.

In terms of economics, as Yang nicely laid out, part of the reason GLP-1 was a pharma success rather than a biotech success is because of the centrality of M&A in the industry. Most early-stage biotech investors are banking on large acquisitions for liquidity, which means they keep a very close eye on the shopping list of pharma buyers. And pharma buyers really don’t like spending billions of dollars to test new biological hypotheses—especially when the size of the market opportunity is uncertain.

How can we change this and unleash a new wave of more innovative medicines? We’ll need to continue to decrease the time and cost for every part of the stack—from discovery, to development, to commercialization.

Doing this would make early-stage drug discovery a lot more valuable.

Improving discovery and commercialization both seem like technology problems. Accelerating development may require new technology and regulatory reforms. Taking a page out of China’s book (for a change!) and studying their recent reforms could be a good starting point for the latter.

On a technical level, it’s important to recognize that modeling human biology is a much more difficult AI problem than modeling a specific modality. Consider two questions. Does my antibody bind this target with higher affinity? What impact will activating the GLP-1 receptor have on the totality of human physiology? The tools to definitively answer the first question in the lab are readily available. The answer to the second question can’t fully be known until first-in-human trials, because our pre-clinical models are only rough approximations of human biology.

More effectively simulating human biology will likely require substantial data generation and continued progress in new AI paradigms.

Over time, tackling these economic and technical challenges could dramatically reshape the biopharma landscape, leading to a new wave of commercial biotechs advancing bold new therapies.

But all of these rapidly compounding technologies could also lead to an even more radical departure from traditional biotech business models in the long run.

Moats Could Look Very Different

In most industries, there are multiple viable strategies to create enduring competitive advantages over competitors. The canonical 7 Powers framework developed by Hamilton Helmer is an attempt to enumerate the most prevalent approaches.

The 7 Powers are:

1. Scale Economies — A business where per unit costs decline as volume increases.

2. Network Economies — A business where the value realized by a customer increases as the userbase increases.

3. Counter Positioning — A business adopts a new, superior business model that incumbents cannot mimic due to the anticipated cannibalization of their existing business.

4. Switching Costs — A business where customers expect a greater loss than the value they gain from switching to an alternate.

5. Branding — A business that enjoys a higher perceived value to an objectively identical offering due to historical information about them.

6. Cornered Resource — A business that has preferential access to a coveted resource that independently enhances value.

7. Process Power — A business whose organization and activity set enables lower costs and/or superior products that can only be matched by an extended commitment.

(For a deeper understanding, check out Helmer’s 7 Powers book, or read ’s great summary.)

At the risk of oversimplifying, only two of these Powers matter in biopharma. Big Pharma companies benefit from Scale Economies because they can amortize the costs of development and commercialization with revenue from their existing portfolio. For biotechs, basically the only real source of Power is control over a Cornered Resource in the form of new intellectual property (IP).

As Peter Drucker once wrote, “The pharmaceutical industry is an information industry.” The value of a small molecule drug has nothing to do with its physical instantiation, which is worth very little. The ability to charge the highest margins for any physical product in existence is purely a function of the IP. Once the IP expires, generic drug makers can swoop in and offer substantially lower prices.

This is “The Biotech Social Contract,” as Peter Kolchinsky at RA Capital would frame it. Scientists and entrepreneurs are rewarded patent exclusivity for innovation. But the exclusivity is finite. When it expires, the innovative drugs become cheap commodities for all future generations.

The looming concern with commoditization is that this could soon become China’s Social Contract with the biopharmaceutical industry—as long as they can produce equivalent IP cheaper and faster.

But what if there was a different way to build a defensive moat in biotech?

There are some early examples pointing to orthogonal forms of defensibility. It’s hard to know what genericization will look like for CAR-T therapies—a new form of medicine centered around a process for engineering a patient’s own cells to eliminate cancer. In Kolchinsky’s book on The Biotech Social Contract, he actually expressed concern about this. Companies marketing these drugs may derive moats from Process Power rather than from a Cornered Resource around IP.

Now let’s run this forward over time.

Personalized cancer vaccines are another form of therapy with this type of form factor. There is no single composition of chemical matter at the core of this type of medicine. Instead, each dose is produced by a complex combination of patient measurements, algorithms, and manufacturing steps.

This has really interesting consequences. For the first time, this could be a medicine with Network Economies. Because each dose is designed with an algorithm, the quality could be improved as more data is collected. Patients could benefit from taking the medicine produced by the company with the largest data moat. This approach also clearly benefits from Process Power. Over time, the winner for this new modality could even accrue a clear Branding advantage as the market leader.

If more forms of medicine start to look like this, we could see a wave of biotechs directly competing to establish themselves as an entirely new generation of pharmaceutical company.

In considering this, I can’t help but think of my good friend ’s writing around Vertical Integrators. In his words, there are several defining characteristics of these companies:

Vertical Integrators are companies that:

1. Integrate multiple cutting-edge-but-proven technologies.

2. Develop significant in-house capabilities across their stack.

3. Modularize commoditized components while controlling overall system integration.

4. Compete directly with incumbents.

5. Offer products that are better, faster, or cheaper (often all three).

For Vertical Integrators, the integration is the innovation.

Just as before, there are clear challenges with this strategy. There is no free lunch!

A massive hurdle will likely be financing and capital formation. This approach to company creation is totally distinct from how most biotech investors think about generating returns. It’s completely unclear if a Big Pharma company would be willing to buy a company with such a complex product without definitive proof of commercial viability.

Winners in this space may need to look elsewhere for capital. One possible option is to tap the growing pool of “Deep Tech” venture capital that is focused on underwriting breakthrough advances in hardware and innovation in the world of atoms. Later stage investment could come from generalist growth equity firms rather than traditional biotech crossover funds.5

Attempting to build a Vertical Integrator in biotech is not for the faint of heart.

Combining multiple technologies in new ways is hard. Financing will be hard. Scaling commercial efforts will be hard. It could take a lot longer to achieve success.

Because of all of these factors, biotech investing could start to mirror the rest of the private market’s evolution. Companies could stay private longer. Consider SpaceX, which has raised nearly $10B over its 23 years as a private company and is now valued at $350B. Liquidity for early investors and employees has come from secondary markets rather than M&A transactions or IPOs.6

Despite the hurdles, the prize is potentially massive.

Measurement tools that were once unimaginable are now commonplace in biology. The foundational insights that gave birth to the last wave of generational biotechs have been refined and commoditized. AI is accelerating biology’s transition into a predictive and quantitative discipline.

Tackling big problems that have evaded prior approaches—like cancer, infectious disease, and brain health—will likely require creative solutions that integrate multiple digital and physical building blocks.

If the companies solving these global problems establish moats in new ways, we could see the first $1T+ biotech firms come into existence.

The public biotech market is pretty gloomy right now. And for American biotechs, the steady uptick in Chinese acquisitions has further threatened their prospects of success. As Adam Feuerstein wrote, “Sentiment is lousy and the bad mood is relentless, to the point where people are seriously wondering if a sector turnaround is ever possible.”

At the same time, the early-stage market is brimming with potential. The pace of technological innovation is equivalently relentless. Entrepreneurs equipped with hard-earned lessons and powerful tools are pursuing wholly new ideas.

Perhaps the right question isn’t whether the market will rebound. Because it will. Markets are cyclical. Instead, the question is if biotech is on the cusp of a phase shift into something new altogether. If it is, it’s never been a better time to build.

In other words…

Biotech is Dead. Long Live Biotech!

Thanks for reading this essay on the commoditization of therapeutic modalities and its strategic implications.

If you don’t want to miss upcoming essays, you should consider subscribing for free to have them delivered to your inbox:

Subscribe now

Until next time! 🧬

Subscribe now

What I initially thought would be a summer hiatus stretched into a longer break from regular writing. But I’ve now defended my PhD (🍾🍾) and am excited to have more mental bandwidth for CoB again!

And I’ll have more to share about what I’m doing next… soon. 👀

For now, I’m back with a detailed profile of one of the companies I’m most intrigued by.

Enjoy! 🧬

In 1994, Jeff Bezos left a high-paying job at a Wall Street firm and moved across the country to start an Internet business. In his telling of the story, the logic was quite simple: Web usage was exploding—with 2300% annual growth—and he wanted a piece of the action.1

Clearly, he was directionally right. Amazon is now one of the biggest businesses in the world. The Internet did continue to exponentially grow. But he was also early—early enough that the dot-com bubble, caused by the “irrational exuberance” of other people chasing the same opportunity, was an existential threat to the company early on.

In the aftermath of the bubble, entrepreneurs got back to work discovering every viable Internet business. One interesting visualization of this search process is the saturation of every corner of the Craigslist home page with Internet companies.

Many of these businesses failed, some succeeded, and others succeeded massively. But the search was extensive. Name a sector—any sector—and there are probably dozens of attempts to build Internet businesses within it.

The genomic revolution has gone a bit differently.

We definitely have our own version of Bezos’s Internet growth statistic: the cost of DNA sequencing has dropped so rapidly that it makes Moore’s Law look linear.

We also had a bubble of “irrational exuberance” from investors followed by deflated hopes. But the cycle following the bubble played out differently. Some very large businesses did emerge from the chaos. At its highest share price to-date, Illumina had a $75B market capitalization. It’s not Amazon, but it’s a very big business—and its growth is evidence of the large-scale adoption of DNA sequencing by research labs and companies all over the world.

What we haven’t seen is the same type of frantic hunt for every nugget of value that could be commercialized as a result of this enormous secular trend. The set of diseases for which we have a molecular explanation—which resembles the equivalent of the Craigslist homepage for this analogy—has grown far faster than the corresponding set of new drugs to treat them.

I think this partially explains some of the discontent in the scientific community about the Human Genome Project. After billions of dollars that could have been spent in other places, the result is less transformative than initially advertised.2

For technologists who are more accustomed to seeing their discoveries or inventions transform the world more rapidly, this is confusing:

The responses are fascinating. Many scientists will point out the challenges in delivering gene therapy payloads to specific tissues or organs in the body, or the need for more efficient and safer editing technologies. This is all true—and the progress being made on this front is exciting.

But this doesn’t fully explain the disconnect. We now know far more about the genetic basis of thousands of diseases than we previously did. And most of these diseases still have no treatment of any kind—let alone sophisticated gene therapies.

Why is that?

In 2012, a group of economists from MIT proposed an explanation. Led by Andrew Lo, a brilliant financial economist who was one of the earliest people to sound the alarm on the compounding financial risk that ultimately grew into the 2008 financial crisis, the paper was an exegesis on risk in the biotech market.

In their view, the discrepancy between research progress and clinical progress could be explained by “the trend of increasing risk and complexity in the biopharma industry.” Both the science and the process of commercializing the science had gotten a lot more complicated, making returns much more uncertain.

Based on this diagnosis, Lo and his team proposed a financial engineering strategy to pool risk differently, allowing for more capital—and different kinds of capital—to be put to work accelerating biotech progress.

In 2015, a principal at a biotech venture firm called Third Rock Ventures named Neil Kumar quit his job. Kumar could viscerally feel the inefficiencies of the status quo approach to commercializing science. He was determined to test Lo’s ideas. The company Kumar founded is called BridgeBio.

In less than ten years, BridgeBio has launched over thirty drug discovery programs, receiving eighteen INDs and three FDA approvals—an impressive feat by any industry standards. Part of Kumar’s intensity is reflected in one of the company’s central mottos: every minute counts. Patients can’t wait.

Moving at this speed hasn’t always been smooth sailing. Commercially, the company has had a complex trajectory, oscillating between over $10B and less than $1B on the public market.

At the center of BridgeBio’s rapid trajectory is an important business innovation: the hub-and-spoke corporate structure. BridgeBio isn’t one company. It’s an archipelago of discrete subsidiaries—each focused on advancing new therapies for specific genetic diseases.

Here, we’re going to continue our study of hub-and-spoke biotechs by analyzing BridgeBio in detail. This business is based on a fascinating blend of financial theory and cutting-edge genetics. We’ll explore how BridgeBio aims to dramatically accelerate the development of treatments for genetic diseases—and how they could potentially build a generational biotech company by doing so.

Studying BridgeBio is different than studying other biotech companies. It’s an experiment testing a bold financial engineering theory in the real world—in real markets. For current and future founders, investors, or people simply trying to understand the economics of drug discovery, every alignment and gap between the company’s model and reality is a chance to learn something new.

The experiment is continuing as you read this. Kumar and the team are still relentlessly pushing forward. A drug called Attruby—which you’ll hear much more about soon—was approved during the time I was writing this.

Given that BridgeBio is a public company, I want to be very clear about what this post is and is not.

This is not financial advice, I don’t own any BridgeBio shares, and I didn’t receive any compensation from BridgeBio to write this post. My goal is to provide a retrospective analysis of how BridgeBio came to be, and to study their attempt to hyper-scale biotech via financial engineering.

What does it mean to pursue “Moneyball for biotech,” as Andrew Lo would put it?

Can Financial Engineering Cure Cancer?

At a high level, biopharma can seem like a pretty efficient market.

Labs in universities all across the country are constantly producing new research results, and the National Institutes of Health (NIH) funds this work—to the tune of ~$50B a year—with fairly bipartisan support.3 Hundreds of biotech startups are formed every year to commercialize all of this research. The FDA provides a clear roadmap for the development of new drugs. Big Pharma companies buy successful products for billions of dollars—providing a clear path to a return on investment. The result is a continuous stream of new medicines.

This is the Miracle Machine of American biomedical innovation.

As puts it, Health Technology is Weird. Compared to other areas of “deep tech” commercialization, the path from the lab to the market is straightforward.

For most of his life, Andrew Lo assumed pretty much the same thing. But on closer inspection, things were less efficient than they seemed. As he described it, “I started running into more and more examples of amazing medical opportunities that were laying on the shelves because the funding just wasn’t there.”

Beyond anecdotal examples, this phenomenon is numerically apparent. When Lo started asking these questions around 2010, about $48B was spent on biomedical research. Another $127B is spent on clinical development. But only $6-7B was spent on translation.

This translational chasm is called the “Valley of Death” for new biomedical advances.

This observation became the kernel of a new research project in his lab: Lo wanted to understand why this type of funding gap existed.

In his group’s 2012 paper—the same one that would serve as the inspiration for BridgeBio—they proposed an explanation: investing in biomedicine was getting riskier for both scientific and economic reasons.

Scientifically, new therapeutic ideas—like the excitement around precision medicine and the use of molecular biomarkers, or totally new therapeutic modalities like cell and gene therapies—were more complicated. This complexity was a sign of progress, but it also introduced more uncertainty and risk into translation.

From an economic perspective, things were also getting a lot more complicated. They wrote,

A host of economic and public-policy conditions has also contributed to this uncertainty, including declining real prescription-drug spending; rising drug-development costs and shrinking R&D budgets; the 'patent cliff' of 2012 during which several blockbuster patents will expire; increased public-policy and regulatory uncertainty after the Vioxx (rofecoxib) debacle; the potential economic consequences of healthcare reform; less funding, risk tolerance and patience among venture capitalists; narrow and unpredictable windows of opportunity for conducting successful initial public-equity offerings; unprecedented stock market volatility; and the heightened level of financial uncertainty from ongoing repercussions of the recent financial crisis. Consequently, the lengthy process of biomedical innovation is becoming increasingly complex, expensive, uncertain and fraught with conflicting profit-driven and nonpecuniary motivations and public-policy implications. Although other industries may share some of these characteristics, it is difficult to find another so heavily burdened by all of them.

In their view, all of this risk explained the aversion to financing translation. The prospects of success for any single project often seemed far too uncertain to justify investment.

To make investment in translation more compelling, Lo and his team proposed a financial engineering strategy for risk mitigation. Their idea was to pool together a sufficiently large basket of translational efforts such that success at the portfolio level was a statistical certainty.

Would you invest $200M with a 5% chance of hitting the jackpot and getting >$10B in returns for a blockbuster drug? Maybe. But also, maybe not. That’s a pretty low chance for such a big investment.

But if you could find the capital to take that same bet 150 times, your chances of success approach >99%. With a big enough portfolio, the risk profile fundamentally changes.

There was a second critical idea in their proposed solution: the portfolio could be funded with a combination of equity and securitized debt—which is when batches of debt are packaged into a single unit that can be sold to investors.

Why?

As Lo’s team put it,

These two components are inextricably intertwined: diversification within a single entity reduces risk to such an extent that the entity can raise assets by issuing both debt and equity, and the much larger capacity of debt markets makes this diversification possible for multi-billion-dollar portfolios of many expensive and highly risky projects. One indication of this larger capacity is the $1 trillion of straight corporate debt issued in 2011 versus the $41 billion of all initial public-equity offerings (excluding closed-end funds) that same year.

Again, the core idea is to pool risk to make biotech more investable. The decreased risk would open up additional pools of capital. More capital would make even greater diversification possible.

Now, if you’ve ever watched The Big Short, you may be thinking to yourself, “Isn’t this type of elaborate debt packaging exactly the sort of thing that caused the housing crisis?!?!” If you were, you wouldn’t be wrong.

This point wasn’t lost on Lo either—again, he was one of the people raising concerns about these practices years before the financial crisis happened. The authors conceded, “Our proposal is clearly motivated by financial innovations that played a role in the recent financial crisis; hence, it is natural to question the wisdom of this approach.”

But mathematical modeling and simulations seemed to suggest that this idea really did have legs. Based on historical likelihoods of success, they estimated that this basket approach would yield average investment returns of 8.9–11.4% for equity investors, and 5–8% for debt investors.

At a first pass, 5-8% doesn’t seem eye-popping. But for large pools of capital like pension funds that are responsible for investing trillions of dollars collectively, managers shoot for an 8% return rate.

On paper, it seemed like this might actually work!

There was one small problem. The simulations were run with pools of capital ranging from $5B to $15B. Getting to that enticing >99% probability of success at the portfolio level assumed $30B.

That’s a very expensive economic experiment. It’s interesting looking back at the early responses to this paper. The biopharma blogger Derek Lowe covered it with tempered optimism. Nature Reviews Drug Discovery published an article about it as well, quoting many pharma executives and investors who found the thesis compelling.

But expressed interest is a different thing than committed capital.

Leonard Lerer, the co-founder of Innovative Finance, was quoted saying, “it is going to be very difficult for investors to get their heads around it.” Proposing a complex securitization strategy in a complex industry—in the aftermath of the 2008 financial crisis—isn’t the easiest pitch.

But Lo was insistent that it should be attempted. After publishing another review paper entitled Can Financial Engineering Cure Cancer? he was approached to lead a series of small workshops at MIT called CanceRx. He closed one of these workshops saying, “We have to get some people to step up and actually try this model.”

This message didn’t fall on deaf ears. Neil Kumar was in the audience with his good friend Brian Stephenson. They remember looking at each other and thinking: “Why not us?”

Reduction to Practice

Out of everybody in the audience that day, Kumar had come with a uniquely prepared mind. While earning his PhD in chemical engineering at MIT, Kumar had received some help on mathematical modeling from Andrew Lo and taken a few of his classes.

After MIT, he’d worked in consulting and biopharma business development before taking a job as a principal at Third Rock Ventures. Although Third Rock was one of the most prestigious biotech firms in Boston, he was frustrated. His job was to translate advances in genetic medicine into new companies, but only a small fraction of the promising opportunities were actually being commercialized.

Kumar was a resident of the Valley of Death. Translation was happening, “It’s just that we were doing it so slowly,” he remembered.

So when Lo’s paper on biotech securitization came out, he got really excited. He’d accumulated firsthand experience with the problem that was being outlined. He reached out to Lo to talk about the ideas in more detail, which is how he ended up at the CanceRx conference.

Lo had also invited Stephenson because he thought he’d be uniquely interested in the topic. Kumar and Stephenson were no strangers. After first meeting in their graduate program at MIT, they had started their careers together at McKinsey. Together, they had played around with ideas for using portfolio theory in drug discovery, but none of their biopharma clients had the incentives—or frankly, interest levels—necessary to put the plans into action.

At this point, Stephenson was working at Capital IP, a firm that offers a suite of credit solutions to technology companies for avoiding the dilution of equity financing. His obsession with the intersection of technology and finance had persisted.

After the conference, Kumar and Stephenson got serious about brainstorming how to reduce these financial theories to practice. They’d meet up and work late into the night filling in different pieces of the strategy. As the pieces coalesced, they’d relay their drafts to Lo at MIT—who ultimately joined as a co-founder—and he would send back revisions.

Their goal was to identify a viable model that could start small and scrappy, but had a structure that could compound over time and support an increasingly diversified portfolio.

In other words, what was the minimum viable product for the biotech megafund?

Kumar, Stephenson, and Lo all agreed on one important choice early on: cancer wasn’t the right place to start. Cancer trials are expensive and have the lowest success rates in all of drug discovery. Importantly, from the perspective of portfolio theory, it’s hard to de-correlate the trials when they are all targeting variations of the same underlying disease process. A big component of portfolio theory is that the basket of bets should be as statistically independent as possible.

Lo and his research group had already started to model out a different therapeutic area: orphan diseases. Defined as diseases that affect less than 200,000 individuals in the United States, orphan diseases were consistently overlooked by pharma companies in the era where most blockbusters came from new medicines for common diseases with large patient populations.

This all changed in 1983 when the United States passed the Orphan Drug Act (ODA) into law. Blending a combination of research grants, tax credits, and an expedited regulatory process, the ODA sweetened the pot for drug developers looking to treat these rare diseases.

In addition to the increased commercial viability, both the molecular biology and genomic revolutions dramatically increased the scientific viability of treating these diseases. Around 80% of orphan diseases have a genetic basis. As measuring DNA became cheap and ubiquitous, the number of diseases with a clear molecular explanation exploded.

By 2024, the main database tracking genetic diseases had nearly 7,000 entries where the root cause was well understood.

This scientific revolution explains the chart we looked at earlier:

This opportunity was an ideal fit for a portfolio approach for several reasons. In theory, each disease has a distinct underlying mechanism, making each trial much more de-correlated compared to focusing on cancer. Importantly, drugs supported by human genetic evidence have much higher success rates, with Mendelian diseases caused by a single gene having the highest success rates in the entire industry. As Kumar frames it, genomics has turned treating these diseases into an engineering problem, not a science problem.

The major bottleneck was attracting enough capital to scale translation for thousands of diseases where each individual program still came with non-trivial risk. It seemed like the perfect proving ground for Lo’s megafund strategy.

This all sounded great—until it made first contact with reality. Building a fund was challenging for a few reasons. First, without a track record, raising enough capital to truly test the portfolio approach would be near impossible. (Even with a track record, ~$15-30B is a lot of money!) Second, it would be difficult to guarantee success in the ten year window typical of funds. Third, and perhaps most importantly, the most talented scientific minds he was recruiting wanted to make medicines—not just invest in them.

As every great entrepreneur does, he quickly iterated.

Instead, Kumar proposed building a company. More specifically, a holding company that owned and operated a portfolio of biotech firms making medicines for specific genetic diseases.

In other words, a hub-and-spoke biotech. This solved a bunch of problems. Now, BridgeBio could scale in a stepwise fashion, building up subsidiaries over time. Much less capital would be needed upfront. But over time, the central holding company could still issue debt because of its diversified portfolio. They could also recruit top talent into the subsidiaries, offering both equity and the freedom to drive progress on their programs.

It was a plan to build the megafund “from the bottom up,” as Kumar framed it in one of his early brainstorming emails. With the revised focus from cancer to genetic diseases and the transition from fund to company, all of the core ideas for BridgeBio were in place. It was time to build.

Still, doing new things is really hard.

At this point, Kumar had quit his job at Third Rock and was actively pitching the idea to prospective investors and employees. He was met with a giant wave of rejections—including from Third Rock. “There’s a lot of excitement after you quit your job to be an entrepreneur—and then you go out to pitch your company and you get punched in the face about 8,000 times,” he said.

Kumar wouldn’t stop. As Marc Andreessen once wrote, “The world is a very malleable place. If you know what you want, and you go for it with maximum energy and drive and passion, the world will often reconfigure itself around you much more quickly and easily than you would think.” By throwing himself at the problem with full force, Kumar found that doors started to open.

Reaching Scale

Once the strategy for BridgeBio was clear, they could focus on conversations with the most compatible investors. Biotech VCs still didn’t get it—they wanted a more traditional asset-centric approach. But Kumar was able to convince two successful businessmen, Peter Feinberg and Seth Merrin, to become the company’s first investors.

As they refined and simplified the value proposition, private equity (PE) firms and hedge funds became interested. While PE firms typically shied away from underwriting therapeutic risk, the diversified hub-and-spoke model was uniquely appealing. Kohlberg Kravis Roberts (KKR), an enormous New York–based investment company with over $500B under management, became BridgeBio’s first major investor. Soon after, Perceptive Advisors, a famous biotech hedge fund, also invested.

It was 2015, and BridgeBio finally had $40M to start assembling their portfolio. What should they prioritize first? Inspired by Strategy: A History, a 768 page opus by Lawrence Freedman, a famous Oxford war historian, Kumar felt the company needed a crystal clear objective. He jotted down a simple closed-hand formula and set of principles that would become BridgeBio’s objective function to this day.

They would prioritize the rapid development of genetic medicines with beautiful mechanistic science supporting their development. One of the first major opportunities that fit this objective was acoramidis, a small molecule medicine for transthyretin amyloid cardiomyopathy (ATTR-CM) being developed by a small company called Eidos Therapeutics.

ATTR-CM is a hereditary disease of the heart. After years of basic research, the underlying biology of ATTR-CM is well understood. The disease is caused by a mutation of the transthyretin (TTR) gene, which encodes a transport protein that is responsible for shuttling two molecules—thyroxine and retinol—to the liver.4 When mutated, TTR, which typically self-assembles into a complex consisting of four identical subunits (a homo-tetramer) becomes mis-folded and accumulates in the heart. This accumulation can ultimately lead to heart failure.

Acoramidis works by re-stabilizing the TTR complex, mimicking a known protective genetic variant for ATTR-CM. It’s a case of clear, beautiful science.

In 2016, BridgeBio decided to acquire Eidos and re-launch it as a subsidiary within their hub-and-spoke umbrella. Emphasizing speed—which was baked into the company’s objective function—they designed the Phase 1 trial for acoramidis to evaluate safety and early signals of efficacy. They quickly read out a positive Phase 1 result and financed the Phase 2 study, which again proved to be successful.

Outside of Eidos, the entire company started to gain steam. Several more rounds of private financing had brought in over $400M—with commitments from top tech investors like Sequoia as well as from blue-chip biotech investors and PE funds. The story was resonating. They had used this capital to pick up several more promising assets—both from academic labs and from large pharma companies like Novartis. As long as the genetic evidence was clear, the source—or even the drug modality—didn’t matter.5

At this stage, it was time to go public. There was a slight problem with this. If you’ve been reading along in this hub-and-spoke series, you may remember that Nimbus Therapeutics has remained private. One reason is that an LLC can’t be publicly traded.

BridgeBio wanted access to public equity markets, but wanted to retain the hub-and-spoke model. Looking for solutions, they considered going public via a special purpose acquisition company (SPAC), which is when a public shell corporation acquires a private company. This was in vogue in 2019, but came with drawbacks.

Instead, they performed some corporate jiu jitsu and reorganized the company for a traditional IPO. By tucking the LLC into a traditional corporation, they could pursue a public listing without canning their portfolio model.

With this maneuver complete, they were ready for the public market. And it turned out that the public market was very much ready for them as well. In June of 2019, BridgeBio offered just over 20 million shares at a price of $17 per share. With additional investment from their syndicate, they brought in just over $400M in new capital, making their IPO the largest of the year—and one of the largest biotech IPOs of all time. By the end of 2019, the share price had climbed to over $40, giving BridgeBio a market capitalization of over $4B.

At this point, BridgeBio was approaching $1B of capital raised. But to truly reach scale for their portfolio model, they knew they would need to realize Lo’s original model of tapping into debt financing in addition to equity financing. (Remember, in 2011 there was $1 trillion of corporate debt issued, compared to $41B raised for IPOs.)

Using their original simulations, they made the argument that despite being a pre-revenue biotech company—which is normally an extremely risky investment—their portfolio model gave the company a fundamentally different risk profile. Debt investors ran the models and bought the argument.

In March of 2020, BridgeBio levered up with $550M in debt in the form of a convertible note with a 2.5% interest rate. This is a form of debt where investors are paid interest over time, but the major incentive is the potential for conversion of debt into equity at a premium after a specified period of time. Normally, the term is five years, but BridgeBio was able to negotiate a seven year term to ensure time for more clinical readouts.

Next, BridgeBio raised another convertible note—this time with an eight-year term—for a whopping $750M at a 2.25% interest rate. This officially tipped the scales: by 2021 BridgeBio had raised a multi-billion dollar war chest, with the majority coming in the form of debt rather than equity financing.

Still, Kumar and Stephenson were wary of being undercapitalized to fully test their portfolio theory in the event of initial failure. The first readout for their Phase 3 trial of acoramidis was a month out. If the readout was negative, it could tank their ability to finance the rest of their pipeline. Effectively treating “debt as insurance” against this scenario, they raised $750M in additional debt, with $450M coming up front and $300M delayed until they had clinical readouts.

This was a total historical anomaly for two reasons. First, debt providers aren’t running around issuing massive chunks of capital to clinical-stage biotechs with no products. The ability to raise this debt was a testament to the belief in the fundamentally different risk profile afforded by diversification. Second, all of this lending took place at the absolute peak of the debt market driven by COVID-19.

What does that mean?

The spread of the pandemic distorted the market. Some companies saw massive growth—like video conferencing software and at-home exercise equipment—while other businesses saw customer demand disappear overnight. This type of distortion can lead to massive unemployment and cause big economic issues. So in response, the Federal Reserve slashed interest rates—even more steeply than during the 2008 financial crisis—to effectively make it cheaper to borrow money.

When interest rates are low, “safer” investments like government bonds are far less attractive because of their relatively low yields. So with the big drop, managers of big pools of money were suddenly incentivized to look elsewhere—which ends up moving a lot more money into debt markets.

I’m spelling all of this out to highlight the absolutely incredible timing for BridgeBio.

A company that was predicated on a strategy of raising debt found themselves—almost overnight—in a market with cash sloshing around in search of new opportunities to buy corporate debt. I obviously wasn’t in the room, but you can imagine Kumar and Stephenson hunched over their desks, furiously modeling how to best capitalize on this insane opportunity.

Right after this debt spree, the door swung shut. The Federal Reserve began to aggressively hike rates again and money flowed back into government bonds. But Kumar and Stephenson already had secured over $2B in debt capital to fortify against any hiccups in their clinical progress.

Little did they know that they were in for much more than a small hiccup.

The Acoramidis Saga

In the original simulations of Lo’s megafund strategy, the duration of each trial was observed to have a strong impact on performance—with faster trials delivering better returns. Optimizing for speed, the Phase 3 trial for acoramidis had an unusual structure. The readouts would come in two distinct phases: “Part A” and “Part B.” The first primary endpoint was improvement on a test of how far patients could walk in six minutes after a year on the drug. A year and a half later, the rate of mortality and cardiovascular hospitalizations would be assessed.

In 2021, two days after Christmas—and a month after issuing debt—the first endpoint for acoramidis read out. It was negative. There was no significant improvement on the walk test. BridgeBio’s shares immediately cratered to a third of their price.

Now, the portfolio model would be stress tested like never before.

On paper, the megafund model assumed nearly perfect conditions. First, get a mountain of cash up front. Next, check back in fifteen years after you’ve run your basket of de-correlated trials. Profit!

Of course, reality is much less rational.

BridgeBio is a public company whose valuation is contingent on market narrative. The huge swing after the first readout revealed a critical discrepancy between how the company was internally computing its own valuation and how the market was actually valuing it. In the eyes of public investors, most of the value could be attributed to future sales of the company’s lead asset—much like how any other biotech is valued. A missed endpoint threatened those future prospects.

To make matters worse, the entire biotech sector was getting crunched. Because biotech is so capital intensive, the sector trades inversely to interest rates. When cash becomes expensive, biotech stocks get punished.

Kumar and Stephenson battened down the hatches and did everything they could to weather the storm. To extend runway, they rapidly reduced the BridgeBio workforce by 30%. Internally, they worked to re-emphasize the importance of their mission for patients. The worked needed to continue.

They pulled every possible lever to stay lean. To avoid the additional costs of building out their commercial organization in this environment, they found external partners for Nulibry, a drug they had gained FDA approval for in 2021. The priority review voucher this approval brought in was sold off.

But all of these efforts paled in comparison to one simple fact: their war chest of very favorable debt from the prior year gave them sufficient runway to let their “beautiful science” speak for itself.

In March of 2023, a positive Phase 2 readout for infigratinib, another medicine in their pipeline, nearly doubled their share price to $18, roughly the value offered at their IPO. This gave the company a chance to scoop up $150M via a public offering to prepare for more upcoming readouts.

Finally, the moment of truth was around the corner.

BridgeBio braced for the Part B readout of acoramidis in July of 2023. The stakes couldn’t have been higher. Despite their portfolio approach, it had become clear a second failure could pose an existential threat to the company’s future. They’d taken a big bet on this program, having gone as far as using some of their debt to reacquire outstanding shares of Eidos. At the same time, the possible upside from approval had become even more clear.

A lot had happened since BridgeBio first scooped up Eidos in 2016. In 2019, Pfizer gained approval for tafamidis, which became the first drug on the market for ATTR-CM. In 2023, tafamidis delivered $3.3B in sales. As better diagnostics for the disease become more widely used, Global Market Insights estimated that the ATTR-CM could grow to $11.2B by 2032.

All of a sudden, BridgeBio’s pursuit of beautiful science had positioned them to compete for a slice of a fast-growing blockbuster market. Companies like Alnylam and Intellia were also hot on their heels with genetic medicines for ATTR-CM. The pressure was on!

On July 17th in 2023, the company hosted a call with investors to share updates on the results of the Part B readout for acoramidis. The drug appeared to be a decisive success, delivering positive readouts for reducing heart-related hospitalization events (50% relative risk reduction), survival (25% relative risk reduction), and even the 6-minute walk test that initially was negative. The drug also appeared to more effectively stabilize TTR when compared to Pfizer’s tafamidis.

The market responded.

BridgeBio’s share price jumped back to over $30. In November of 2024, after a tumultuous journey, the company gained approval for acoramidis (now marketed as Attruby), triggering evolution towards becoming a full-fledged commercial organization.

BridgeBio’s Current and Future Potential

Having gained three FDA approvals in only ten years, BridgeBio’s track record inspires confidence that they are capable of efficiently translating genetic insights into new medicines. But efficient R&D doesn’t always equate to efficient commercialization.

And making this transition is really hard. According to McKinsey, only 20-30% of first-time commercial organizations manage to exceed their launch expectations, whereas 40-50% of established organizations do so. Because of this, companies can find themselves in a “second valley of death,” where existing investors want to sell their shares to realize profits from clinical success before anybody else wants to step up to invest in a potentially risky new product launch.

This has been particularly acute for BridgeBio. Early on, analyst projections for Attruby were so low that Kumar pointed out they were effectively predicting “the worst second-mover launch of all time.” In typical Kumar fashion, he was undaunted by the challenge, saying, “There’s probably a degree of skepticism around ‘can a bunch of science nerds launch a drug’? I get that, and I’m going to show that we can do that.”

So far, that’s exactly what they’ve done. Kicking off 2025 at the JPM Healthcare Conference, BridgeBio announced that Attruby has already been prescribed over 400 times, providing evidence for adoption from the medical community—and for BridgeBio’s ability to deliver on it. Some analysts have now taken note and updated their forecasts with more optimistic projections.

But for BridgeBio to truly live up to Lo and Kumar’s initial vision, success can’t be measured by a single product launch. Successes will need to compound, snowballing into a giant diversified portfolio. To truly satisfy their objective function, they’ll need to translate genetic insights into approved drugs at a rate that requires adding an asterisk to charts of yearly approval rates to explain the leap in volume.

This year will be another important stepping stone towards that goal. The company is anticipating three distinct Phase 3 readouts in the second half of the year. If successful, the BBP-418 and Encaleret trials would deliver the first approved therapies for patients with the genetic diseases they target. The Infigratinib trial would deliver the first oral therapy for children with achondroplasia.

Kumar and the team continue to tweak and refine the hub-and-spoke model. While debt investors underwrote the portfolio approach in 2020 and 2021, it’s become clear equity investors are still wary. Most of BridgeBio’s current value is attributed to Attruby’s future revenue streams. To keep funding early R&D efforts while commercializing Attruby, BridgeBio spun out two new subsidiaries in 2024, BridgeBio Oncology and GondolaBio, attracting $500M in fresh capital from top investors.

It will be fascinating to see what comes of the relationship between the parent company and these subsidiaries over time.

Our industry is currently in flux. The American Miracle Machine has produced some objectively extraordinary results, but is clearly not perfect. The existing system is facing wholly new challenges, both internally and externally.

In the midst of all of this change, BridgeBio is running a fundamentally unique experiment. Grounded in first principles thinking, they are challenging existing assumptions around nearly every aspect of biotech financing and development.

Clearly, not every aspect of the new model has worked like they thought it would on paper. But new medicines are making their way to patients at an extraordinary rate. And each hurdle has provided valuable lessons for both the company and the rest of the sector.

In some ways, the BridgeBio experiment is like a health checkup for biotech. Every gap between the original “frictionless” megafund theory and the realities of the market can help diagnose why more drugs aren’t getting to patients faster.

And we need to keep testing creative ideas to solve these problems. Because as Kumar emphasizes, the patients can’t wait.

Sources and Further Reading

This was a particularly challenging company to write about. While I did my best to cover the most critical aspects of BridgeBio’s origins and journey so far, I had to omit a lot of interesting details.

I’m grateful to Neil Kumar and Anthony Bogachev at BridgeBio for talking with me and offering support in my research process.

If you want to go deeper, here are some particularly valuable resources:

• Commercializing biomedical research through securitization techniques — the original research paper that spurred the development of BridgeBio.

• Applications of Portfolio Theory to Accelerating Biomedical Innovation — a brilliant academic paper written by Neil Kumar, Andrew Lo, Chinmay Shukla, and Brian Stephenson detailing the company’s history from the perspective of the founders. This was one of the most valuable sources I came across.

• BridgeBio S-1 — a snapshot of BridgeBio’s state at the time of IPO. It also includes more details about how they took a hub-and-spoke company public.

• From Aha Moment To $10 Billion Biopharma — a nice profile of BridgeBio’s early history that I pulled several quotes from.

• Select coverage of the Acoramidis Saga:

  • BridgeBio's 'baffling' heart disease drug fail 'de-intensifies' competition for Alnylam

  • BridgeBio claims success in heart drug trial, rebounding from earlier failure

  • JPM25: After BridgeBio touts 'remarkable' early uptake for its Vyndaqel rival, analysts hail Attruby's debut as a 'launch to own'

• Earlier drug approvals that I didn’t cover:

  • 'Hub-and-spoke' biotech BridgeBio scores first FDA approval with Alexion castoff

  • BridgeBio keeps FDA approval spree rolling, this time with nod for targeted bile duct cancer med

• Challenges with reacquiring Eidos that I didn’t cover:

  • BridgeBio scraps plan to buy Eidos after bids rebuffed

  • Eidos Reportedly Received $4.6 Billion Takeover Bid From Glaxo

  • BridgeBio Wins Appeal Over Eidos Deal, Ending Pension Fund Suit

    • tldr: it doesn’t really benefit shareholders to sell off a valuable spoke and have all of that cash sitting on the balance sheet of a public company. This is a big difference when considering the strategy and corporate structure of Nimbus and BridgeBio.

Thanks for reading the second installment of the hub-and-spoke series. If you enjoyed this, considering reading the first chapter about Nimbus, another fascinating hub-and-spoke company.

If you don’t want to miss the next chapter, you can subscribe for free to have it delivered to your inbox:

Subscribe now

Until next time! 🧬

Subscribe now

Enjoy! 🧬

Biopharma is a profoundly acquisitive industry. Pharmaceutical giants need to maintain their massive cash flows by developing or acquiring new products—and this balance has swung overwhelmingly towards acquisition over time. This type of “Big Brother” relationship has been around since the biotech industry’s inception, when Genentech first partnered with Eli Lilly in 1978 to manufacture and market recombinant insulin. Genentech was ultimately acquired by Roche in 2009.

This is a blessing and a curse. The massive incumbents provide critical near-term resources but impose a cap on long-term growth—you get snapped up as soon as you have a great product. As Roivant CEO Matt Gline put it, “I think it is interesting the extent to which we lionize [M&A] as sort of the end all, be all. That what success looks like for a biotech company is getting sold.”

So maybe “Big Brother” isn’t quite the right analogy. Attempting to build a generational biotech is a bit like the journey of a baby crab avoiding being devoured by its own mother.

As David Yang traced in a great thread, the critical ingredient for transitioning from a biotech startup to an enduring business is the independent development of a blockbuster drug that brings in over $1B in sales. But this is incredibly expensive to do, so as soon as companies get data hinting at a blockbuster product, there is intense pressure—especially from investors—to sell to pharma, or another biotech looking to make the phase transition to become a “next-gen” biopharma acquirer.1

This is especially tricky for platform companies, which can get bought wholesale for their lead programs before ever getting a chance to compound as a business. Recent examples are the Carmot, Prometheus, and Karuna acquisitions. In each case, the founders decided to launch brand new companies to keep pursuing the original vision of their previous companies.2

What if there was a way to benefit from large M&A transactions while keeping the core platform and team of a company intact over time?

This is exactly the problem that hub-and-spoke biotech companies aim to solve. The innovation is in corporate structure, where a central hub company containing key technology or personnel is fire-walled from spoke companies centered around specific products. With this structure, spoke companies with valuable assets can be sold off to pharma—much like picking a ripe apple off of a branch instead of chopping down an entire tree for one fruit.

To better understand this model—and some of its different permutations—we’re going to cover three important case studies in a new series. We’ll be covering Nimbus Therapeutics, BridgeBio, and Roivant Sciences. Each company has created billions of dollars of value by putting their own spin on the hub-and-spoke structure.

As we compare and contrast these different businesses, we’ll consider many strategic questions. What problem is each business trying to solve with this model? Is the added complexity worth it? Which ideas were viable? Which ideas failed?

For now, let’s get a better sense of how this works with our first case study.

Nimbus Therapeutics

In March of 2011, Bruce Booth, a partner at Boston-based Atlas Venture and industry blogger, published a post with a compelling hook:

Bill Gates has just backed one our new startups – Nimbus Discovery LLC – as part of an extension to the seed tranche. Here’s the press release.

It might come as a surprise to some, but Bill Gates has been a long-time biotech supporter: he was a founding investor in ICOS and was on that Board for 15 years, and importantly, he’s also one of the largest investors in Schrödinger, the world’s leading computation drug discovery company, and our founding partner with Nimbus.

So, with this financing, we’re launching Nimbus out of ‘stealth mode’.  Here’s the story.

What? Bill Gates was investing in computational drug discovery back in 2011? It’s like reading a dispatch from the future—but it was written over a decade ago. This was one of the early waves of excitement centered around the idea that computers could revolutionize drug discovery.

This new company aimed to combine three central ideas into a unified whole.

The first big idea, as we can see above, was that advances in computational chemistry had made the technology ready for prime time. These new tools, like Schrödinger’s “WaterMap” software package for modeling the interaction of solvents (in the human body, this is water) with molecular structures, could provide a differentiated wedge for developing new drugs against “exciting but difficult-to-drug” targets.

Schrödinger, which is still a leading company in computational chemistry, was founded in 1990, nearly twenty years before striking this partnership to launch Nimbus. The early 90’s were a very exciting time for computer technology.

The same year Schrödinger was founded, Microsoft made over $1B in revenue—the first time a software company had ever achieved that feat. People were also really excited about the possibility that computers could make drug discovery much more efficient.

Vertex Pharmaceuticals, founded a year before Schrödinger, was based on Joshua Boger’s vision of “rational drug design” with a major focus on computer modeling. Here’s an interesting excerpt from The Billion Dollar Molecule by Barry Werth:

From a business standpoint, what differentiated Vertex’s story—and what Boger now intended to flog most assiduously—were three things: Vertex’s integrated approach, its first-among-equals attention to chemistry, and its lineage, which tracked impressively through the mainline of the pharmaceutical industry and Merck, then Wall Street’s favorite company. “Smart ex-Merck guys making drugs with computers,” Aldrich said, encapsulating.

Vertex wanted to become the next Merck. But Schrödinger wanted something different. Perhaps inspired by Microsoft’s success, the vision for the company was to sell their modeling software to drug discovery companies rather than making drugs themselves.

In 2010, just a year before the Nimbus deal, Bill Gates—one of the biggest believers, and beneficiaries, of the software revolution—invested $10M into Schrödinger. David Shaw, another billionaire very keen on the potential of scientific modeling, was also a key early backer.

I’m reading between the lines here, but the mentality at Schrödinger seemed to be, “We’ve got the resources to invest in hardcore R&D for this software. When it gets good enough, this is going to be an absolutely enormous software business.”

So the Nimbus deal probably seemed like a win-win. Schrödinger could keep focusing on what “made its beer taste better,” as Jeff Bezos would put it, by cranking on software R&D. Setting up a parallel entity that was run by drug discovery experts from Boston would get them closer to the downstream economics of pharma without too much distraction.

This new company would get access to Schrödinger’s scientists, compute, and unreleased algorithms. It would also get target exclusivity—meaning Schrödinger couldn’t turn around and work on the same targets with somebody else. In exchange for all of this, Schrödinger got somewhere around 10% ownership of the company as a founding partner.3

So that’s idea number one: build on top of Schrödinger’s cutting-edge computational chemistry platform.

The second central idea was about how to leverage the computational predictions that were produced. Nimbus was founded in 2009, a year after a global financial crisis. The company wanted to be as capital efficient as possible.4 To do this, their plan was to build a “virtually integrated, globally distributed” model.

First, get model predictions from Schrödinger without needing to build the computational platform. Next, send the first batch of predictions to offshore Contract Research Organizations (CROs) to make experimental progress as cheaply as possible.5 Then, feed the data into the models and make new predictions.

Wash, rinse, repeat.

As Booth put it, “We’re really pushing the envelope on virtual drug discovery.”

The third central idea was to use the hub-and-spoke model.

In theory, they had just established a repeatable process for rapidly producing new differentiated assets at a very low cost. Here, they thought about the problem we outlined earlier. Would they have to start and sell a brand new company for every asset? Yes, but they could do that from within a broader umbrella.

Here’s what they designed:

Let’s start at the center. The main hub of Nimbus is a passive limited liability corporation (LLC) holding company. Investors fund this entity. Shareholders like Schrödinger own part of this LLC in exchange for services. This LLC owns a series of distinct subsidiaries—one of which is the actual management company. The remaining subsidiaries house the actual IP related to any specific target.

Any target-specific subsidiary can be sold without impacting the others. These are the apples that can be plucked from the tree.

When apples are plucked (target-specific subsidiaries are sold) the money flows back through the LLC to shareholders and investors. Because the holding company is an LLC, it isn’t taxed like a C-corporation. So the sale is taxed once, not twice. In Booth’s words, this “solves the classic drug discovery illiquidity problem (where it takes 7-10 years to get liquid via M&A or IPO); this LLC structure enables us as shareholders to cycle capital back to our investors in a tax-efficient manner on a per project basis.”

Going a click deeper, here’s a look at the deal framework that this model makes possible:

Pharma partners (or investors) can make “equity or equity-like” investments in specific R&D programs. These deals provide capital from partners without diluting the parent company, and can even be structured around pre-negotiated terms based on milestones. Nimbus would know upfront how much they would make in a sale if R&D was a success.

It basically creates an “evergreen drug discovery stage biotech” that can continually recycle returns to investors—and to the hub company—and keep cranking on R&D for new projects.

This all sounds really cool in theory. But how does it actually work in practice? Studying this business in 2024, we have the benefit of historical evidence.

Historical Transactions

I won’t bury the lede: this model has worked really well. Nimbus has now sold multiple assets—including a sale worth $4B in cash and $2B in milestones—while keeping the team and platform intact. As the current CEO Jeb Keiper puts it, “The day after a transaction everybody at Nimbus walks back into work and keeps working on the rest of the pipeline.”

This is awesome!

But like any substantial journey, there have been real challenges along the way. The Nimbus team learned which of their initial ideas were viable—and which weren’t.

Early on, the idea was to use the Nimbus engine to crank out a large number of pre-clinical drug programs in parallel. “Invest $10 million to get 5 development candidates in 2 years.” Before even generating data for an Investigational New Drug (IND) application, they’d license these promising assets off to pharma. Simple as that!

But nobody bit. Just like today, the proof is in the clinic—not the computer or the lab. Ultimately, Nimbus needed to file an IND and generate Phase 1 data themselves. But with this data in hand, they were able to make their first major deal.

In 2016, Nimbus delivered it’s “Apollo mission” by selling their Acetyl-CoA carboxylase (ACC) subsidiary (literally named Nimbus Apollo) to Gilead for $400M in upfront cash and $800M more in potential milestone payments.6

This was a big inflection point. They’d now shown that they could rapidly move from one of Schrödinger’s virtual screening hits into the clinic. This sale also proved out the hub-and-spoke model. It truly “enables the sale of the eggs, not the goose.”

But could the goose lay another golden egg?

The Nimbus team got back to work to see. But as their CEO Jeb Keiper wrote, it wasn’t all smooth sailing:

The transition had its challenges though: we had begun working in clinical development, hired staff, and now were reset to an early-stage preclinical company.  All our resources in chapter 1 had begun funneling to the lead program, and with only $67 million raised over 7 years, Nimbus was not exactly “robust” at an enterprise-level.  We had just 22 people by the end of that year, 15% of the company having departed following the Gilead deal.

Despite the hub-and-spoke structure, the company still needed to concentrate on making sure their Apollo mission could actually reach escape velocity. Once it was gone, they needed to restructure and get back to R&D—which they used 5% of the Gilead proceeds to do.

What happened next highlights the amount of skill, perseverance, and ultimately, luck, that can be behind the big headline deals in the news.

In 2017, Nimbus arranged another spoke deal using the mechanics we discussed above. They gave Celgene options to buy two of their target franchises: TYK2 and STING. According to Keiper, “Nimbus retained full ownership and control of the programs in exchange for funding and pre-programmed exits of $400 million each for Phase 1b data in a few years.”

In theory, this was going to be a rerun of the Gilead deal. But that’s not how things went.

Instead, in 2019, Celgene was acquired by Bristol Myers Squibb (BMS) for $74B. Even in an extraordinarily acquisitive industry, this was big news: it was the largest pharmaceutical acquisition in history.

But for Nimbus, that wasn’t the fact they cared about. It was a much bigger deal that BMS already had a TYK2 program of their own.

This dynamic plunged the company into more troubled water. BMS would want to retain their option to buy Nimbus’s TYK2 program in the event that theirs failed. It was much less certain what would happen if BMS’s program succeeded.

Over the next few years—in the middle of a global pandemic—Nimbus and BMS entered into a legal battle over what would happen with these two programs. The major concern was that BMS would be incentivized to kill Nimbus’s TYK2 program to eliminate competition:

“There is a credible threat that, unless stopped, BMS will slow down or kill the development of Nimbus’s Tyk2 asset. First, BMS has the incentive to protect its leading Tyk2 inhibitor, deucravacitinib, from future competition. Second, BMS has an incentive to avoid additional payments to Nimbus,” the complaint says. “BMS does not have the same incentives to develop Nimbus’s Tyk2, as Celgene did before its acquisition by BMS, given that this would lead to the cannibalization of BMS’s own Tyk2 sales.”

Ultimately, the two companies settled out of court, leaving Nimbus with full ownership of its TYK2 program. At this point, the program had entered Phase 2b trials. With this major twist of fate, Nimbus prepared to finally scrap their hub-and-spoke LLC structure and go public as a clinical-stage biotech.

But that’s not how the story went either! Because of a massive black swan event—the largest pharma acquisition of all time—Nimbus had full ownership of this program. During COVID, Nimbus was also operating in the largest bull run for public biotech stocks of all time. But as their IPO plans crystallized, the market sharply reversed. The IPO window for biotechs swung firmly shut.

The Nimbus management decided to hang tight and continue privately financing this program. This proved to be a wise choice. That year, BMS gained full approval for their TYK2 drug for psoriasis. Most analysts expected that this drug (now marketed as Sotyktu) would come with a severe safety warning because of the risks associated with targeting this family of enzymes. But it didn’t. The allosteric inhibition of TYK2—which was the same mechanism of Nimbus’s drug—now had the potential to grow into a blockbuster autoimmune franchise.

Interest in Nimbus’s TYK2 franchise exploded. Ultimately, Takeda acquired the spoke for $4B in cash and $2B in milestones—ten times the return of the original pre-negotiated Phase 1 deal with BMS.

Today, Nimbus remains a private company. After two major sales, their hub-and-spoke LLC structure remains intact. The engine is still humming. Reflecting on the company’s North Star, Keiper says,

Our mission remains the same: We design breakthrough medicines.  Our objective in dollars and cents terms is to again shoot for the moon, to become again a multi-billion-dollar biotech.  But our ambition is far greater than that.  Nimbus has an opportunity to build a legacy R&D institution.  A paradigm of excellence in small molecule drug discovery and development.

With this history in mind, let’s consider the lessons that can be drawn from the journeys of Schrödinger and Nimbus.

Lessons From History

One important nugget from this history is that selling software to pharma is really hard. In February of 2020, Schrödinger went public. Thirty years after their founding, we could see their growth. In their S-1, they wrote, “In 2018, all of the top 20 pharmaceutical companies, measured by revenue, licensed our solutions, accounting for $22.0 million, or 33%, of our 2018 revenue.”

This is a solid business. Schrödinger is currently trading at ~$17 a share, with a $1.29B market cap. Building a billion dollar business is incredible, but it’s a far cry from Microsoft’s growth. The company is also a fraction of the size of Vertex, who pursued a vertically integrated model based on similar ideas. Vertex is now worth over $100B.

Schrödinger appears to have arrived at a similar conclusion. Over time, they’ve evolved to partner on drug development programs—which is where the other two thirds of their revenue comes from. In 2018, they also launched their own pipeline of internal programs. It’s hard to imagine that this wasn’t influenced by their experiences with Nimbus. One of their largest revenue windfalls has been the distributions they received from the TYK2 sale.

As “generative AI” comes of age in biology, many new companies are exploring pure software business models to enable broad access to this technology. History would suggest this strategy captures less value—but purely indexing on history is a foolish exercise in technology. Things change! Studying history helps clarify what would need to be different. If we are ever going to see trillion dollar biotech software companies, the predictions they generate will need to be orders of magnitude better than any existing tools.

And what can we learn from Nimbus’s journey?

The company has proven that the combination of computation and experimental out-sourcing can produce real value. Consider their first ACC transaction with Gilead. Reflecting on the sale, Booth shared some impressive numbers:

Using the ACC program as an example, we achieved a bona fide Development Candidate in roughly 2 years and ~$8M from a novel virtual screen hit at an allosteric pocket. We achieved definitive clinical proof of mechanism in our Phase 1b for an aggregate, fully loaded project spend of ~$20M. These metrics for a first-in-class, chemically novel program are impressive by any measure.

Nimbus also proved that the hub-and-spoke model works. They’ve made two major asset sales while remaining intact as a business, a unicorn feat in the world of biotech platform companies.

I shared some of the gory details of their history because it’s important to understand that their success wasn’t pre-ordained at incorporation because of a clever business model. There wasn’t a smooth pre-meditated path from $400M to $4B between ACC and TYK2. Their trajectory required a lot of hard work, guts, and luck.

And while the model has worked, it didn’t work quite the way they’d originally hoped. When their ideas made contact with reality, they realized that they’d have to prioritize getting their Apollo Mission off the ground. This required ruthless prioritization of their best programs, much like any other biotech.

It’s also interesting to reflect on why the Nimbus team is doing this. Reading through their thinking, it feels like a blend of pragmatism and an intense commitment to an aesthetic ideal.

Pragmatically, it’s extremely challenging—and prohibitively expensive—to build a new commercial organization that supports a blockbuster product.

As Keiper put it, “We knew if we were successful in psoriasis, the implications would require a large multinational company to create the value of global registrations in multiple indications.  Given the value of established infrastructure in pharma, it was clear that an M&A acquisition of our TYK2 subsidiary was likely.”

This is part of the deep truth tucked behind the centrality of M&A in the biotech industry. As we saw earlier, most of the generational biotechs were built in the 1980s. At this time, it cost much less to independently develop and commercialize new drugs.

Here’s a crucial excerpt from Matt Herper’s excellent piece on the trial bottleneck:

“Certainly the cost of clinical trials has become so outrageous we have to do something to change it,” said George Yancopoulos, Regeneron’s co-founder and chief scientific officer. He remembers that when he started it cost $10,000 per patient to conduct a clinical trial. Now it can cost $500,000, he said. “Just think how expensive that can be. It really limits what we can do.”

This really needs to change. And it can! As far as I know, the laws of physics haven’t changed since trials cost $10,000 per patient. It’s possible.

But faced with the world as it is, the Nimbus team knew they’d ultimately need to sell the assets they developed. The strong aesthetic ideal expressed in the company is the desire to build an enduring “legacy R&D institution” despite this. They didn’t want to be a one hit wonder—they wanted to build something with staying power.

And the shape of that institution is very specifically designed. In Keiper’s words, “Nimbus is committed to the notion that “small is beautiful” in drug R&D: breakthrough small molecules designed by a small expert team.”

The hub-and-spoke model let them express that vision.

As we’ll see over the course of this series, that’s not the only type of business that can be built with this structure. Over the course of Nimbus’s history, they flirted with going public several times. This would have required changing up their corporate structure and winding down the LLC—which legally can’t conduct a public offering.

This isn’t the only possible outcome. BridgeBio and Roivant Sciences, the next companies in this series, are both public hub-and-spoke biotechs. Although all of these companies have structural similarities, they’ve used this business model to solve very distinct challenges.

Next time, we’ll continue our analysis with another case study.

What other types of enduring businesses are made possible with this structure?

Further Reading

This post wouldn’t have been possible without The Book of Nimbus that Bruce Booth, Jeb Keiper, and many members of the Nimbus team contributed to over the years on LifeSciVC. Most of the source links and quotes derive from these materials. If you want to go deeper on this history, I highly recommend studying this team’s writing.

This type of knowledge sharing is incredibly important for our industry. As I support more biotech companies writing the next chapter of biotech history, I hope to pay it forward by sharing more of what I learn.

Speaking of paying it forward, I hope that we can reduce the cost associated with building businesses like Nimbus over time. This business required a lot of legal and accounting work that many founders don’t want to deal with.

How could we make launching a hub-and-spoke platform as easy as using Stripe Atlas?

Thanks for reading the first installment of the hub-and-spoke series. If you don’t want to miss the next chapter, you can subscribe for free to have it delivered to your inbox:

Subscribe now

Until next time! 🧬

Subscribe now

During my summer hiatus, I made a lot of progress on my research, and started to lay the foundation for what will follow. I’ll be excited to share more on both fronts in the coming months.

For now, let’s celebrate the autumnal equinox with a story about the power of plants.

Enjoy! 🧬

As the Chief of Staff at OpenAI, Maddie Hall learned an important lesson firsthand: a small team of mission-driven researchers can turn science fiction into science fact on an extremely short timescale.

For decades, AI was a backwater academic discipline thought to be chasing a pipe dream—until it wasn’t. Now it’s a major driver of the economy.

Maddie remembers a colleague recommending her Profiles of the Future by Arthur C. Clarke. Written by a prolific science fiction author with an uncanny track record of future predictions—including telecommunication satellites and the Internet, to name a few—the book was an exercise in thinking big. It was also a reminder that even the best futurists didn’t see everything coming.

Observing the company’s rapid internal progress, Maddie constructed a new prediction problem for herself. Assume ubiquitous AI. What extremely important problems still wouldn’t be solved?

Probably messy problems rooted in the physical world. And which of these problems might have existential consequences if it remains unsolved?

Climate change.

After exhaustively researching options for making a planetary scale impact on atmospheric carbon levels, Maddie settled on her approach: engineering biology.

So in 2019, Maddie left OpenAI and co-founded Living Carbon, Public Benefit Corporation (PBC). In just five years, Living Carbon has developed photosynthesis-enhanced trees using biotechnology, planted hundreds of thousands of trees in four different states in the U.S., built out an experienced carbon project development team, and raised over $36M from leading investors including Temasek, Lowercarbon Capital, Toyota Ventures, Day One Ventures, and Felicis Ventures.1

Let’s explore why they are doing this, how they are doing it, and how the business model of a climate biotech works in practice.

A Journey Into Deep Time

In 2024, artificial neural networks (ANNs) need little introduction. They are the core algorithmic unit underpinning self-driving cars and AI chatbots. But twenty years ago, the idea that ANNs would be a promising path towards developing learning machines was obvious to few very people. Yann LeCun was one of them.

One of the core tenets of LeCun’s belief in ANNs was a simple existence proof: biological organisms—including humans—use neural networks to learn new things every day. If we could scale computational versions of these types of “deep” networks, they should be able to learn increasingly rich representations of the world.

He was right. In 2018, he shared the Turing Award with Yoshua Bengio and Geoffrey Hinton for his contributions to deep learning. But even at this stage, few people truly internalized the importance of continuing to massively expand the scale of these networks.

OpenAI emerged as one of the teams most zealously committed to this idea. For years, they had been pouring an outsized portion of their resources into compute for training bigger and bigger neural networks. As the size grew, performance continued to improve. Just think of the difference between the first GPT in 2018—which was interesting but not particularly useful—and GPT-4, which can solve challenging engineering problems with very minimal prompting.

OpenAI was able to make this non-obvious investment because they are a research-first company. Their first breakout product came years after their inception. Before they rocketed past $3B in revenue, they had a simpler selling point to prospective researchers: a mission.

Researchers could choose between staying in academia and testing new ideas at a small scale, joining a company to work on an applied machine learning team, or joining an industrial research lab like OpenAI that was entirely dedicated to achieving the full vision of Artificial General Intelligence (AGI).

Many of the most talented researchers in the field—who were also the biggest believers in the vision of AGI—chose the latter. The centrality of this massive vision played an important role in achieving early talent density and justifying continued R&D investment.

Maddie took this to heart. She observed a parallel in a different field: plant biotechnology. Academic scientists were achieving amazing results using synthetic biology to rewire photosynthetic pathways in plants, but their work would never be translated out of field trials. Companies that were pursuing translational work were mainly large corporations plugging away at making better crops.

What if there was a “third place” for the most talented plant biotechnologists to work on their actual mission? The type of place where they could do their life’s work. Where the most cutting-edge techniques could be used to mitigate the looming climate crisis.

Living Carbon was founded to be this third place.

Much like LeCun’s early research, the company is based on an existence proof. But it isn’t based on something that can be directly observed. It requires a journey backwards into Deep Time—a central obsession of Patrick Mellor, Maddie’s co-founder.

The story starts 4.5 billion years ago (4.5G) at the Earth’s formation:

A billion years after the Earth was hit by a protoplanet the size of Mars—ultimately forming the Moon and establishing tides in the ocean—the first major terraforming event occurred: Evolution gave rise to the first photosynthetic microbes.

The result was extreme. These microbes pumped out enormous volumes of oxygen as a byproduct of photosynthesis, which eliminated nearly all anaerobic life forms. The oxygen also converted methane into CO2, which is worse at blocking heat from escaping the atmosphere.

This triggered our planet’s first ice age. All because of photosynthesis.

Across several billion more years, the Earth’s climate continued to be shaped by evolutionary advances in photosynthesis. The emergence of seaweed froze the entire planet. Photosynthesis. Organisms emerged to eat the seaweed. Terrestrial plants emerge and trigger another ice age. Photosynthesis. Again and again.

Living Carbon draws the following conclusion:

“Every major ice age over geological history has been triggered by an evolution in photosynthesis.”

So, what is the plan of action?

Starting with fire and leading to our industrial revolution, humans are the first species to release carbon stored outside our bodies over deep time, as only volcanic eruptions have before. However, unlike a volcano, we have the capacity to do something about it. We could be the first organism to intentionally rebalance the effects of our own success.

We can learn from deep time how every major ice age has been triggered by plants. The current crisis can lead humans to work with plants to accelerate photosynthesis, sequester carbon, and preserve the biosphere on which we both depend.

If we do this successfully, we will be the first species on Earth to intentionally instantiate metabolism on a planetary scale.

AGI is a really big vision.

But so is rebalancing our climate by engineering another major evolutionary event for photosynthesis.

And so far, the vision seems to be resonating. I asked their Chief Science Officer, biotech veteran Yumin Tao, why he joined. He told me, “If I could finish my career making this type of impact on the climate, it would be incredible.”

Living Carbon has recruited a scientific team full of PhD plant scientists, ecologists, and synthetic biologists with decades of experience in the domains needed to tackle this problem.

Let’s explore their early results.

Tree Engineering

Photosynthesis is absolutely miraculous. It is a process carried out by genetically encoded nano-scale machines that organisms—microbes and plants—use to convert light and atmospheric carbon into organic matter that can fuel their growth.

Collectively, terrestrial and oceanic photosynthetic organisms convert over 100 billion tons of carbon into biomass every year.2 Every year!

This is worth restating. Biomass is the direct consequence of atmospheric carbon capture.

In theory, if we could double the speed of this process, we’d be doubling the amount of carbon that photosynthetic organisms are sucking out of the atmosphere.

A lot of research has gone into accelerating photosynthesis—primarily motivated by accelerating crop growth. One of the major strategies has been to increase the efficiency of the RuBisCO enzyme, which is responsible for catalyzing the reaction that fixes CO2. RuBisCO is the most abundant enzyme on the planet, but it’s also one of the slowest.

Tinkering with RuBisCO has proved to be a daunting challenge. Evolution has been trying this for billions of years and hasn’t had much luck either.

In 2019, researchers from the University of Illinois published promising results based on an alternative strategy. Instead of worrying about RuBisCO’s catalytic efficiency, the goal was to circumvent photorespiration—where the enzyme binds to oxygen rather than CO2—which emits CO2 (and uses energy) rather than accumulating it.

Testing a variety of approaches, the group from Illinois identified an alternative pathway (AP3) that consisted of two genes and one RNAi molecule in total. The two genes encode enzymes that convert the byproduct (glycolate) back into C02 inside of the chloroplast. The RNAi, on the other hand, blocks the transporter (PLGG1) that shuttles glycolate out of the chloroplast.

The idea is to convert unwanted byproducts back into CO2—all inside of the chloroplast—giving RuBisCO another chance to fix carbon.

In field trials, this pathway increased the biomass of tobacco crops by 28% and increased photosynthetic yields (measured by starch accumulation) by more than 40%.

Instead of crops, what if we could achieve the same type of photosynthesis enhancement in trees? Sticking to our earlier logic, any new reforestation effort would be removing 40% more CO2 from the atmosphere!

This is what Living Carbon set out to do—and it seems to be working.

Only three years after being founded, the company posted a preprint on bioRxiv describing the creation of poplar trees engineered to bypass photorespiration. Some of the engineered trees accumulated up to 53% more biomass in growth room experiments—an even bigger increase than previous efforts. A year later, the paper was accepted in Forestry.

The results are fairly easy to interpret: the engineered trees get bigger faster!

One major question about these results was how well they would stand up in field experiments—the plant science equivalent of translating in vitro pre-clinical results into actual demonstrable benefits. So far, that is going well too. They’ve launched multiple field trials in different locations, and their collaborators have already seen positive results in the field. Living Carbon will have more exciting news to share on this front soon.

Another major question is whether faster growing trees directly correlate with durable carbon sequestration. Nature operates with cycles. Plants draw down carbon, but their respiration—and ultimately, death—causes a lot of that same carbon to end up back in the atmosphere. This is called the fast carbon cycle.

Reducing respiration should make Living Carbon’s trees tip this cycle towards more carbon capture than emission, but they aren’t stopping there. In parallel, Living Carbon is engineering trees capable of accumulating an order of magnitude more copper through their roots. Copper helps to block fungal infections and slows down wood decay—ideally lengthening the time carbon remains captured.

But ultimately, these two traits are just the start. Living Carbon is a platform company. The scientific team has built out the capabilities to identify traits, develop genetic constructs (now focused on gene-editing technologies), transform a wide variety of plant species, and to quickly move from the greenhouse into the field.

Let’s recap. Inspired by her time at OpenAI, Maddie co-founded a research-centric company focused on a massive mission. The goal was to attract the talent and resources necessary to compress the time needed to solve a hard problem. In three years, Living Carbon genetically engineered a tree that accumulates over 50% more biomass.

That’s an exciting start. But like OpenAI, the company is growing a viable commercial organization around its research-centric core.

In other words, getting this right will require a marriage of tree engineering and financial engineering.

Financial Engineering

Sometimes capitalism is portrayed as a mindless sprawling machine that destroys Nature while siphoning wealth into the hands of a select few. But this is a caricature. In reality, it’s just a collective story that we tell each other to collaborate. When the story causes problems, we can change it. It’s not easy, but it’s possible.

This is what carbon credits aim to do.

One problem with capitalism is that it’s hard to measure and price in negative externalities. Try drawing the exact causal link between an oil refinery and the specific economic consequences of its emissions in the future.

Proponents of carbon credits argue we should use financial engineering—specifically, the development of new financial instruments and markets—to solve this problem. Carbon credits are a mechanism for people and corporations to invest in projects that offset emissions through the reduction, avoidance, or removal of carbon.

Starting new markets and measuring offsets are both hard problems. Last year, a team of journalists reported that more than 90% of carbon offsets large companies are purchasing to prevent deforestation are effectively worthless. In response to the study, a scientist at Oxford made the following point: “The challenge isn’t around measuring carbon stocks; it’s about reliably forecasting the future.”

How much of an impact does not cutting down a tree actually make?

As carbon markets continue to grow and develop better standards, we need new categories of interventions with more directly measurable benefits. This is what Living Carbon hopes to deliver.

Rather than preventing deforestation, Living Carbon focuses on reforestation efforts, specifically on land that has been degraded and is sitting idle. With their enhanced trees, they aim to convert decimated land—like old mining sites in Ohio—into rapidly growing forests demonstrably sucking carbon out of the atmosphere.

Achieving this is a two-sided problem. In the middle, there is Living Carbon’s growing suite of enhanced trees and their field team capable of delivering “Projects as a Service” (PaaS).

One one side, there are landowners. The pitch to the owners is that instead of donating their land to a trust or investing in reforestation themselves, Living Carbon will convert their idle acres into a valuable resource, generating meaningful revenue per acre annually.

On the other side, there are corporations looking for more effective carbon offsets to purchase. The pitch to the companies is that Living Carbon’s reforestation projects will produce high quality removal offsets that can be directly measured and verified.

To repeatably establish these types of projects, the company has built out a commercial team with backgrounds in project development, carbon markets, and real estate. Each project produces three types of revenue:

1. Carbon credits

2. Seedling sales

3. Contract value for PaaS

Part of what’s unique about Living Carbon is how quickly they’ve put this strategy into action. They’ve raced past other GMO tree projects in terms of real-world data collection, becoming the first group to plant multi-site, multi-year field trials of genetically modified trees last year.

If some of the large-scale projects they are currently negotiating come into fruition, Living Carbon could become one of the early leaders in the carbon market.

But there’s another commercial benefit for Living Carbon: trees have value beyond carbon markets. Unlike most of synthetic biology—where engineered organisms are used to produce products like chemicals or proteins—their engineered organisms are the product.

As Maddie puts it, “At the end of the day, we are selling biomass.”

This biomass can be monetized via carbon markets, where the sole purpose of the trees is to sequester carbon. Fast-growing trees capable of growing on previously idle land is also of obvious interest to timber companies. Woody biomass can also be sold as feedstock to producers of biomass-based electricity and sustainable aviation fuels (SAF).

So the business model generalizes to:

1. Biomass (sold for carbon credits, timber, feedstock for SAF, etc.)

2. Seedling sales

3. Contract value for PaaS

Given how nascent carbon markets are, this offers them flexibility to expand their business in multiple directions.

To make all of this work, the team sweats the details of their techno-economic analyses. They are exploring new corporate structures—much like hub and spoke biotechs—to finance the upfront costs of their project with loans rather than venture dollars.

As Maddie told me, “We don’t want to just be a feel good climate startup.” The goal is to build a real business that makes a planetary scale impact on the climate.

Achieving an impact on this scale will require fast-growing trees, and a fast-growing business to sustain the research and mission.

Over two years ago now, I wrote an essay entitled Viriditas, outlining why I care so much about biotechnology. Part of my argument was that we need to remember that AGI won’t solve all of our problems:

Perhaps it is myopic, but there is something special about biological life. About the prospect of more green, wet planets in the universe. The chaos and complexity of evolved cellular machinery. I’m arguing that if AGI is a rocket ship with any chance of launching sentient life into the Cosmos, something deeply important and beautiful would be missing if it weren’t launching biological organisms.

The digital realm is increasingly created unconsciously. Most code may soon be written by unconscious AI agents. As a programmer, I’m still working to adjust to this new reality. Imagine a future where AI realtors and lawyers negotiate with each other, and AI stock pickers reach out to their AI brokers.

Many wonder and worry: where will I be in this AI future? Where will my meaning and value come from?

People that are feeling this way should consider looking where Maddie looked: the physical world. Achieving the type of verdant and abundant future most people really want will require a modern renaissance of artisanship and engineering in the world of atoms.

This future will require real human blood, sweat, and tears—supplemented by AI advances—and continued progress in bioengineering.

Part of what I find compelling about Living Carbon is that the innovation is equal parts scientific and cultural. The company is founded around a mission that puts biologists—and the trees they engineer—at center stage in the story we tell ourselves about the future.

They are forcing the issue. If we want our planet to remain hospitable to humans, we need to start planting seeds of change now.

Thanks for reading this essay about Living Carbon’s efforts to rebalance our planet’s carbon cycle with plants.

If you don’t want to miss upcoming essays, you should consider subscribing for free to have them delivered to your inbox:

Subscribe now

Until next time! 🧬

When I was a little kid, I wanted to be a writer. After having my mind thoroughly blown by Star Wars, I filled countless marbled composition books with my own stories of imaginary alien worlds. In my teens, reading Dostoevsky novels made me want to capture deeper truths about the human condition, and my interests drifted towards poetry.

Initially intending to study English, a series of introductory biology courses at my university changed my life. I was exposed to a world that felt as wonderful and alien as Star Wars, but was real, and was moving incomprehensibly quickly inside every cell in my body. I fell in love with the world of the cell.

I dove into the world of research, spending all of my hours in the lab.

I’ve now spent over a decade as a student, scientist, and engineer in biotech. I’ve gotten the chance to work in a number of great labs and to publish some interesting technologies and results along the way.

For most of this time, it wasn’t obvious to me that I should focus any of my effort on writing about biotech. While doing so would represent the intersection of my two primary obsessions, the poems about immune cells that I submitted to JAMA Oncology were never accepted and didn’t seem to get me anywhere. So I stopped reading as many novels and focused on scientific papers.

During COVID, things changed. I had decided to defer from starting my PhD for a year, and was working remotely as a scientific software engineer. Suddenly, there was no journal club or departmental seminar with cookies and coffee on Wednesdays. My consumption of science was done in isolation.

Because I couldn’t find any outlets covering the science that I was most interested in, I started this newsletter as a virtual journal club, breaking down a new bioRxiv preprint every Sunday.1 The wave of initial interest caught me off guard. My goal was to reach 100 weekly readers in the first year. I reached that number in the first 24 hours.

After over a year of this format, this newsletter has mushroomed into something else altogether. Aside from two brief breaks in August of 2021 (starting grad school) and June of 2022 (getting married!), I’ve written continuously about a wide range of topics in biotech—including science writing, business strategy, biotech philosophy, techno-economics, and in-depth company storytelling.

Now, I’ve written over 200,000 words and have nearly 200 times the number of readers that I’d originally hoped for.

Before I write the next 800,000 words, I’m taking what I sincerely hope will be my last hiatus from writing this summer. My goal is to complete the last chapter of my research career: my PhD thesis at Stanford.

Once I complete my research, I have a vision to scale the impact of The Century of Biology and to place it on a sustainable foundation that will give me greater bandwidth to write more essays of higher quality.

Why am I doing this?

Over the course of this century, our ability to engineer living systems will continue to improve dramatically. This has extraordinary consequences for improving human and planetary health. We’re developing the tools to cure genetic diseases, create sustainable alternatives to petrochemicals, and engineer our immune cells to detect and destroy tumors.

My mission is to distill the most important data, companies, and ideas from the frontier of this revolution into coherent essays. The beauty of writing is that it is telepathy. I can beam these essays across the Internet into your inbox, and ultimately into your mind. This can lead to learning, agreement, disagreement, new ideas, collaboration, behavior changes, investment, and policy decisions.

In other words, my mission is to accelerate this revolution by telling its story.

I want to sincerely thank you for joining me on this journey so far. If you feel that you’ve derived value from this newsletter, the best way to support my work is to share your favorite essay with a friend who you think would enjoy it. Let’s accelerate the biotech revolution together.

Until next time, in the Fall! 🧬

Thanks for reading. If you don’t want to miss the return of this newsletter in the Fall, you can subscribe to have the next essay delivered to your inbox for free:

Subscribe now

Viriditas!

Subscribe now

Enjoy! 🧬

With modern tools, a small molecule drug discovery campaign often starts by screening millions—or even billions—of distinct chemicals. The same is true for antibodies. Blockbuster drugs like Humira that are sifted out of these enormous “libraries” of candidates are very lucrative needles in a molecular haystack.

For CRISPR medicines, things are a bit different.

Nearly identical CRISPR medicines can be reprogrammed to treat entirely new diseases. Let’s consider a pair of examples.

In June of 2021, Intellia Therapeutics, in partnership with Regeneron Pharmaceuticals, dosed six patients suffering from a rare genetic disease called transthyretin amyloidosis (ATTR) with an experimental medicine called NTLA-2001. This medicine consists of three distinct parts: an mRNA molecule encoding the Cas9 enzyme, a guide RNA molecule that targets Cas9 to the transthyretin (TTR) gene, and the sphere of lipids that encapsulates them called a lipid nanoparticle.

When delivered through a vein, the nanoparticles and their cargo circulate through the blood to the liver—where they are swallowed up by liver cells. Once inside of cells, the Cas9 mRNA is translated into the Cas9 enzyme and forms a sequence-specific editing complex with the guide RNA.

Transthyretin amyloidosis is caused by the accumulation of mis-folded TTR proteins which can lead to nerve and heart failure. In the case of NTLA-2001, the guide RNA programs Cas9 to insert mutations into the TTR gene—causing it to be effectively silenced. When patients were dosed, they experienced up to a 96% reduction in the levels of TTR proteins in their serum when compared to their diseased baseline levels.

Three years later, Intellia dosed ten patients suffering from an entirely different genetic disease called hereditary angioedema with another experimental medicine called NTLA-2002. Patients with hereditary angioedema experience painful swelling attacks that can be fatal if they occur in the airway and block breathing. NTLA-2002 was designed to target the kallikrein B1 (KLKB1) gene, which encodes a critical step in the molecular cascade that leads to the swelling attacks. The results were remarkable: “Among all the patients, the mean percentage change in the number of angioedema attacks per month from baseline through the latest assessment was −95%.”

NTLA-2001 and NTLA-2002 target different genes in different pathways that cause different diseases. But they are nearly identical. The lipid nanoparticles are the same. The Cas9 mRNA molecule is the same.

The only difference between the medicines is 20 nucleotides in the guide RNA.

Unlike small molecule drug discovery or antibody development, designing these critical 20 nucleotides is not a massive and intractable search problem. It can be done on a laptop computer, and the designs can be verified with DNA sequencing.1

For blood, eye, and liver diseases, there are already established delivery technologies and entire clinical protocols in place for testing new medicines. As we’ve seen, new gene-editing technologies are being invented—and even AI-generated—on a continual basis. Labs and companies around the world are hard at work expanding the set of tissues that these systems can be safely delivered to.

This is why Fyodor Urnov, the widely cited Berkeley scientist responsible for coining the term “genome editing” back in 2005, calls CRISPR “a platform for cures.” In theory, we could scalably test and produce these cures—even for patients with mutations found nowhere else in the human population—for every blood, eye, and liver disease with a genetic basis. Over time, with second-generation tools and delivery technologies, CRISPR could be used to treat the 400 million people worldwide suffering from genetic diseases—and could even extend into the realm of complex diseases.

So why aren’t we doing that?

This is the question that Urnov has been asking for many years, and the answer is complicated. Currently, there is a multifaceted thicket of regulatory, economic, and technological challenges preventing this vision from becoming a reality.

Instead of focusing on scaling the infrastructure necessary for personalized medicines, CRISPR’s translation into the clinic is being pursued in a business-as-usual approach by a finite set of for-profit companies. Each of these companies is focusing on diseases with economically justifiable markets, which has caused substantial crowding around a small set of blood disorders such as sickle-cell disease and beta thalassemia.

And it hasn’t been smooth sailing. As the biotech sector has struggled to reset after the “sugar rush” of funding that flowed into the market during the COVID pandemic, nearly every public CRISPR company has undergone rounds of layoffs.2

When medicines do make it into the clinic, they are expensive. Casgevy, the first FDA-approved CRISPR-based therapy, costs 2.2 million dollars. While there is a real pharmacoeconomic argument to be made that these medicines will save the medical system money in the long run by replacing a lifetime of complex medical care for very sick patients, it makes access difficult in the near term and causes healthcare spending to become much spikier.

Every time one of these million-dollar gene therapies gets approved, Peter Marks, the director of the Center for Biologics Evaluation and Research (CBER) within the FDA, gets an email from the agency’s commissioner, Robert Califf. As Marks puts it, they usually have “a lot of exclamation points.”

Urnov—now the Director of Technology and Translation at the Innovative Genomics Institute (IGI) founded by CRISPR Nobel laureate Jennifer Doudna—isn’t content with the status quo either. Recently, the IGI launched a new collaboration with Danaher aiming to “scale up the development of CRISPR cures.”

The ambitious vision is to lay the foundation to go from “one to a thousand” new therapies. Doing this will be incredibly challenging. We’re going to explore why the system currently works the way it does—and the problems they will need to solve in order to change it.

But first, let’s imagine how it could work if they succeed. As Urnov’s Lennon-inspired article asks of us, we’re going to imagine CRISPR cures.

CRISPR On-Demand

In a 2021 opinion article in Molecular Therapy and a subsequent New York Times article in 2022, Urnov laid out a vision for what a future with ubiquitous CRISPR medicine could look like if we work for it.

For longtime readers of this newsletter, it’s a clear 4-S story. The advances made in DNA Sequencing and Synthesis, in combination with new Software tools, could be used to deliver personalized medicines on an entirely new Scale.

To start, we could sequence the genome of every patient. Even this first statement would have been science fiction at the turn of the century. We had only just completed the Human Genome Project, which took over a decade and cost billions of dollars.

Consider how quickly technology has improved even since Urnov wrote this article three years ago. In 2022, researchers at Stanford broke the world record for the fastest genetic diagnosis, sequencing a patient’s complete genome in five hours and two minutes. With renewed competition in the sequencing market, the race is on to deliver complete genome sequences for $100 or less.3

For patients suffering from Mendelian disorders where a single gene is mutated, this can produce an immediate diagnosis. Next, the gene would be visualized in a genome browser where its full sequence could be collected.

With the gene sequence in hand, a scientist or physician could run the same standard software tools that Intellia used to design their guide RNAs and screen them for potential off-target edits. The guide RNAs could then be ordered online from a commercial provider. Turnaround times are typically days, not weeks. In the not-too-distant future, machines for synthesizing the guides could be distributed in hospitals around the world if DNA printers continue to improve and are adapted to produce RNA.

For patients with severe need and no other option, this could be all that is required to produce a cure.

The guide RNA would be packaged into a delivery vehicle, and the patient would be dosed with their personalized therapy. Sequence. Design. Print. Dose.

As the software tools for CRISPR design continue to improve and we accumulate more data, this could prove to guarantee sufficient safety. For additional certainty, we could develop protocols for using simple blood samples and DNA sequencing to verify on- and off-target editing with a patient’s own genome.

What if CRISPR becomes so safe and standardized that we use it to treat additional diseases?

Consider statins. Hundreds of millions of people take these small molecule pills to manage their cholesterol levels and risk of heart disease. They are one of the most commonly prescribed drugs in the world. But if Verve Therapeutics succeeds in achieving their goals, that may no longer be the case.

Verve is currently developing a CRISPR therapy for a rare form of heart disease, but the company’s long-term vision is to expand their efforts to also treat the more common type of heart disease. As Urnov argues, in the future, “one gene edit could replace daily medications such as statins.”4

Clearly, this is a very different future from the present. It’s a world where every patient’s genome is routinely sequenced, and personalized genetic medicines are as commonplace as small molecule drugs.

So, what’s preventing this from happening? Let’s examine the first major roadblock in the current system: the time and cost associated with clinical development.

New Drug, New Trial

Modern clinical trials are a fairly recent invention. Up until the turn of the 20th century, there was effectively no regulation for the development and sale of new drug products. The FDA only came into existence in 1906 and gradually evolved into its current form after two major belt tightenings in 1938 and 1962.

Now, new drugs are brought to market through an intensely standardized and regulated gauntlet of preclinical and clinical phases.

These regulations were largely successful in their goal of increasing the burden of evidence required to market new drugs and make medical claims about their benefits. Understanding the full consequences of this system—both good and bad—and why modern trials have gotten so expensive is a massive topic.

But when it comes to scaling CRISPR cures, there’s one particularly important detail to know:

No matter how similar two drugs are, the entire set of preclinical and clinical phases of development need to be repeated. New drug, new trial.

By Urnov’s estimates, this entire process would be “a four-year journey likely to cost at least $8 million to $10 million.”5 This makes the development of personalized CRISPR treatments effectively impossible. No company can justify these types of costs for rare—potentially even n-of-1—mutations, and patients often don’t have four years to wait to be cured.

The current regulatory framework was established in an era when nearly all new drugs were small molecules—before even antibodies were a class of medicine. From this perspective, it made a lot of sense to treat every new “test article” as its own unique entity. Even minor changes to the structures of small molecules can dramatically alter their impact on human physiology. In medicinal chemistry, there is a saying, “Change a methyl, change the drug.”

But is this true for CRISPR?

Are NTLA-2001 and NTLA-2002 really completely different medicines because of the twenty differing nucleotides in their guide RNAs? The answer is likely “no,” and Peter Marks and the FDA appear to agree.

On May 29th, the FDA issued a new draft guidance for a Platform Technology Designation Program to ease the regulatory burden for new drugs that use the same underlying technology. As Peter Marks put it, “We have to be thinking now because if every base pair changes we're going to have people come back to the agency, there are a lot of nucleotides out there that can go wrong in the genome. And I don't think we can have 3 billion visits to the FDA a year, so we'll have to figure this out.”

This is a really big deal. In this new potential framework, a company could recycle data between drug programs. We already know how the lipid nanoparticles and mRNA molecules behave in humans. Instead of reproving this, the focus can shift to proving the safety of the new guide RNA. As Jorge Conde of a16z put it, “We may be entering the 'reusable rocket' era for biotech.”

To fully understand the ramifications of this proposed regulatory update, let’s consider the current economic incentives of drug developers without any ability to rapidly reprogram existing medicines.

Producing Blockbuster Hits

In the world of tech investment, returns follow a steep Power Law distribution. Nearly every new startup fails, but a small handful of massive outliers manage to make up for this fact. Strangely, venture capital is one of the best-performing asset classes despite having the highest failure rate for any given investment.

When it comes to drug development, the Power Law may be even steeper. Not only do most drugs fail to gain approval, but most approved drugs still fail to recoup their development costs. Alex Telford describes this phenomenon nicely, saying, “This extreme unevenness of potential returns means that pharmaceuticals is fundamentally a hits-driven business. Because drug development is so costly and most launches are duds, the industry as a whole relies on blockbusters—drugs with annual global revenues of $1 billion or more—to cover the flops and keep the entire enterprise afloat.”

Historically, producing these types of blockbuster hits required developing treatments for common diseases with large patient populations. Lipitor—one of the most widely sold types of statins that we looked at earlier—broke records in the early 2000s for its multi-billion dollar sales.

Over time, the pharmaceutical industry’s focus has shifted towards rare diseases for producing new blockbusters. This shift can largely be attributed to three primary drivers:

1. Regulation. The U.S. passed a new law called the Orphan Drug Act of 1983, aiming to incentivize drug developers to make drugs for underserved patient populations. By and large, the incentives worked.

2. Increasing R&D costs. Clinical trial costs have steadily swelled over the years, disincentivizing development for drugs that would require larger cohort sizes to gain approval. Costs per patient have climbed from $10,000 up to as high as $500,000.

3. Pricing evolution. Biopharma companies tested the waters with new strategies, charging massive prices for small patient populations. The most extreme example so far is Hemgenix, which is a $3.5M gene therapy for Hemophilia B—which has less than 300,000 occurrences worldwide.6

As Matthew Herper incisively described it,

Instead of dealing with the difficulty of collecting data on new medicines, both society and government sidestepped it by focusing on treatments for much rarer illnesses: relatively rare cancers, rheumatoid arthritis, and multiple sclerosis to name a few, not to mention very rare illnesses such as cystic fibrosis or paroxysmal nocturnal hemoglobinuria.

The current “goldilocks zone” for new blockbusters is the set of rare diseases that are not too rare—the equation is still X patients times Y price >= $1B; even with a large Y, X still needs to be big enough—have sufficient unmet need, and have well-described underlying mechanisms that can be addressed by existing technologies.

These are the constraints that the current CRISPR companies are working with, and their clinical pipelines reflect this.

In the future that Urnov dreams of—and that the FDA is increasingly sympathetic towards—this all could change. Imagine if the draft guidance for platform approvals is passed into law. A company like Intellia could gain approval for platforms rather than products. The fundamental technology underlying NTLA-2001 and NTLA-2002 could be approved as a platform for curing genetic diseases of the liver, dramatically reducing the time and development costs associated with curing any one disease.

What if organ-level platform approvals for programmable medicines became the new hot blockbuster category?

But for any of this to come to fruition, there’s a looming technical bottleneck: we’ll need to solve the manufacturing challenges associated with producing personalized genetic medicines on demand.

The Last Mile of Manufacturing

Not all drug modalities are equal. Different types of medicine—such as small molecules, biologics, and cell and gene therapies—have different tradeoffs in terms of clinical benefit, trial risk, and manufacturability. Small molecules like Lipitor aren’t curative but can be cheaply and abundantly produced. Cell-based therapies like CAR-T can cure some cancer patients but are a nightmare to formulate. Biologics like Novo and Lilly’s GLP-1 agonists have caused enormous headaches when attempting to scale to large patient populations.

If personalized gene-editing medicines like CRISPR are ever to become as commonplace as small molecules, they’ll need to become equivalently straightforward to make.

There are good reasons to think that this is possible. If these medicines—like the ones produced by Intellia—are simply formulations of lipids and RNA, the COVID mRNA vaccines provide an existence proof that drugs with identical nucleic acids can be scaled to dose billions of people. Seemingly, the major manufacturing hurdle will be producing drugs with different nucleic acids for different patients. This is why it’s so interesting that Danaher is the company collaborating with the IGI.

Danaher is the 800-pound gorilla of biomedical tools and manufacturing. As a ~$200B conglomerate company, Danaher has steadily acquired many of the leading infrastructure providers for the life sciences ecosystem. In 2018, they acquired Integrated DNA Technologies, Inc. (IDT), which is one of the leading providers of synthetic DNA and RNA—including CRISPR guide RNAs. In 2021, they acquired Aldevron, a major manufacturing provider for nucleic acids, proteins, and biologics—including “High-quality, scalable mRNA manufacturing meeting a wide range of requirements from linear template through LNP formulation & encapsulation.”

With all of these different pieces, it’s hard to think of a company better equipped to tackle this challenge. I’ll be curious to see what this Danaher and IGI initiative is able to achieve in potentially vertically integrating these different technologies into cost efficient systems that can be more widely adopted by hospital systems.

Thinking back to our future scenario with ubiquitous CRISPR cures, there will also need to be continued adoption of genome sequencing and rapid genetic diagnoses. Again, with turnaround times already being measured in hours rather than days, this seems like more of a question of achieving widespread deployment rather than needing further technical leaps.

We’ve now looked at the regulatory, economic, and manufacturing challenges associated with scaling CRISPR cures. In each case, there is a reason to be optimistic that the future will be better than the present. Let’s zoom back out to see what all of this could lead to in the future.

Putting the Pieces Together

CRISPR therapies—along with other types of genetic medicines—are a step towards making programmable cures for disease. By changing twenty nucleotides in a drug for transthyretin amyloidosis, Intellia designed a potentially curative treatment for hereditary angioedema. This is not how drug development typically works.

But despite the promise, bottlenecks in the existing drug discovery ecosystem have rate-limited the scalability of these approaches. The same companies that developed these potential “platforms for cures” are laying off employees from their research teams. Regulatory hurdles established for a prior generation of technologies have come to dictate the business strategies of drug developers. We still can’t produce genetic medicines with the same speed and cost as small molecules.

Even in the face of these setbacks, there are many reasons to be optimistic that the future of these technologies is bright. Danaher and the IGI—world leaders in the manufacturing and development of CRISPR medicines, respectively—have joined forces in a new initiative to go from “one to a thousand” cures for genetic disease. Just this week, the FDA released a draft guidance indicating that we could soon operate in a world where platform technologies have accelerated pathways to develop medicines for drugs outside of the current “goldilocks zone” where drug developers are hunting for the next blockbuster drugs.

Where could things go in the future?

In the near term, these technologies and new regulations are likely to have the biggest impact on ultra-rare diseases. But as if this essay wasn’t already speculative enough, let’s dream a little bit bigger. A lot of new biomedical infrastructure will need to be built to scale the development of personalized medicines for genetic diseases. What if these new ultra-rare cures are like the pricey Tesla roadster that paved the way for the more broadly affordable Model 3?

Once the infrastructure is in place for rapid and inexpensive testing and dosing of n-of-1 medicines around the world, shouldn’t that drive down the costs of clinical development for all subsequent genetic medicines? Consider one of Alex Telford’s predictions for a new source of blockbusters in the future:

The biotech toolkit (incl. gene therapy) expands into more common acquired diseases as our understanding of complex biology improves and manufacturing costs decline. This leads to a revisiting of highly-genericized out of favour indications with remaining unmet needs (e.g. cardiovascular)

Already, we’re seeing companies like Verve pursue this strategy. In a world where companies like Danaher—or startups with new manufacturing technologies and solutions—offer a cheap and globally distributed infrastructure layer for any company with a compelling new product to build on top of, we could see a new golden era of investment in cures for more common diseases.

If even part of this vision becomes reality, the future of medicine could look very different than it does today.

Thanks for reading this essay about scaling the development of programmable genetic medicines.

If you don’t want to miss upcoming essays, you should consider subscribing for free to have them delivered to your inbox:

Subscribe now

Until next time! 🧬

Subscribe now

Enjoy! 🧬

After roughly four billion years of stumbling search, Evolution has churned out an absurd diversity of organisms. Microorganisms that can convert light into energy, trees that grow to hundreds of feet in height, insects capable of farming other insects, and mammals capable of forming technological civilizations. All of this diversity is encoded in genetic information, which primarily produces a universe of molecular machines—proteins.

If the scale of organismal diversity is absurd, the universe of proteins is practically incomprehensible. The millions of unique species on Earth collectively produce somewhere around a trillion different distinct proteins. In this enormous set, there are some very pragmatic proteins responsible for building and maintaining cells and replicating DNA, but there are also more exotic machines, like the bi-directional nano-scale motors that bacteria use to navigate towards chemicals.

Still, evolved proteins represent an infinitesimal fraction of the total theoretical space of possible proteins. Consider the following thought experiment: how many possible proteins with 200 amino acid residues are there? Since there are 20 possible amino acids for each position, the answer is 20 raised to the power of 200. This number is so big that it makes the number of stars in the observable universe look small.

One of the most promising fields of biotech is de novo protein design, which aims to explore the full space of possible proteins for new types of molecular machines beyond what Evolution has already produced.

This is a really big idea. Directed evolution—the set of laboratory techniques for rapidly sampling new beneficial mutations to known proteins—was a Nobel-worthy idea that has already produced a lot of value. Sampling beyond known proteins opens up an entirely new world of possibilities.

Recently, de novo protein design has gotten a big boost from AI. Taking inspiration from the wave of exciting results in using AI to generate new text, images, audio, and video, scientists have retrofitted and repurposed the underlying algorithms to generate new proteins.

This isn’t just a big idea—it’s apparently a billion dollar idea. Last week, a new startup called Xaira Therapeutics announced its $1B Seed round of financing. David Baker, the senior author of the paper with the figure above, is a co-founder. While the vision for the company spans everything from basic biological research to clinical trial design, a major focus is to design new antibody therapies by improving recent generative AI algorithms from the Baker Lab.

So, when people say “generative AI for biology,” they are talking about the development of deep learning models capable of generating new types of biological outputs—like new DNA sequences or proteins. In the future there may be models trained on a variety of inputs capable of producing a variety of outputs, but right now, most efforts are focused on using specific algorithms for specific biological inputs/outputs.1

These models and their outputs have been getting a lot of attention. Last week, a bioRxiv preprint from Profluent Bio describing new AI-generated CRISPR proteins took the Internet by storm. And it wasn’t just my corner of the Internet with other biotech nerds—The New York Times covered the work the same day it was posted.

According to the NYT article, Profluent’s model “analyzes the behavior of CRISPR gene editors pulled from nature and learns how to generate entirely new gene editors.” I think this is close to true, but not quite right.

In a clever piece of detective work, Brian Naughton, a biological data science guru, asked a simple question: how distant are these new gene editors from multiple natural sequences? When lining up the new editors against naturally occurring proteins, he realized that he could “recreate 98% of the sequence (all but 24 amino acids) with 3 Streptococcus sequences.”

I don’t think Brian intended for this to be a dunk on Profluent’s work, and I don’t either. It would be effectively impossible for even the best CRISPR biologist to sit down at their keyboard and type out this new sequence, let alone confidently predict that it would actually work in cells. Clearly, in its exhaustive training on natural sequences, Profluent’s model is learning something really interesting about how CRISPR proteins work.

But the generated results aren’t entirely new. They are like one of Dali’s surrealist paintings, synthesizing disparate existing objects into a new unified whole. The novelty is in the composition, not the ingredients—which come from Nature.

One way to look at this is that it’s like Digital Evolution, where the AI algorithm is sampling from the space of all possible CRISPR permutations.

This makes a lot of sense when considering how these models work. To understand all of this, we’re going to look under the hood of this specific approach. This will help bootstrap intuition for what we should expect—and not expect—from this first wave of generative AI models for biology.

A Quick Primer on Attention

When analyzing the growth of scientific knowledge, there are often fundamental results that seed important branching points in our understanding or capabilities. For the current wave of AI progress, that critical node is a 2017 paper from Google entitled Attention is All You Need.

This paper described a new neural network architecture called a Transformer. At its core, a neural network is a big pile of linear algebra that learns the relationship (function) between inputs and outputs. These learned functions can be used to solve hard problems like language translation that are difficult—or impossible—to produce solutions for by explicitly writing code.

In theory, given enough time and training examples, even the simplest neural networks can learn to approximate nearly any function. But we don’t live in the world of theory, so AI researchers need to invent new architectures, which are specific combinations of the underlying linear algebra operations that converge on interesting solutions in a reasonable amount of time.

So in reality, one of the layers in a Transformer model look something like this:

And much of the magic happens in the attention matrices (highlighted), hence the name of the paper. The query (Q), key (K), and value (V) matrices keep tabs on the relationships between each chunk of the input that is being processed. Visually, the attention patterns of a trained Transformer look like this:

Here, one of the attention heads—a specific copy of one of the matrices above—is detecting a relationship between “animal” and “it” in the sentence. Another is picking up on the relationship between “it” and “tired,” which is used to describe the animal later on in the sentence.

There’s an obvious question at this point: how do all of these matrices work together to actually learn that “it” is referring to “animal” in a sentence like this?

These matrices are parameterized by weights, which are just numbers that are multiplied with the inputs that can be tuned to improve the performance of the model. The training—or tuning of these weights—happens by computing the changes that would reduce the error rate based on known examples of inputs and outputs to the model.2

One of the ways to generate a lot of training examples for these networks is to mask parts of the input, and ask the model to guess the correct value that is hidden. This masking can be done a few ways. The mask can be applied in the middle of a sentence, like this:

Or the last token can be masked, which can be used to train the model to learn the relationship between all of the preceding chunks of the sentence and what the next one should be. This is called causal masking.

I realize that we’ve covered a lot of ground and that this explanation will be unsatisfying if you want to deeply understand Transformers. If you’d like to learn more, I’d recommend starting with the Illustrated Transformer article, or the 3Blue1Brown series on neural networks. Both are great starting points for building intuition about how these models work. They also both provide a nice rabbit hole of resources for further learning.

For now, here are the three takeaways we need to understand:

1. Neural networks learn to approximate complex functions, like language generation, by repeatedly tuning weights to perform better on training examples.

2. An important neural network architecture is the Transformer. These models have specific sets of matrices that track relationships between the elements of their input. This is called attention.

3. One of the best ways to train Transformers is to mask parts of the input and make the model correctly guess what is being hidden.

With these basics in mind, we’re in a better place to understand how these protein language models are generating new CRISPR sequences.

Generating (Somewhat) New Proteins

Biology and computer science have a unique bidirectional relationship. Studying biological computation gives us new ideas for how to program computers. The connections between biological neurons provided inspiration for artificial neural networks. Algorithms from computer science are also useful tools for learning new things about biology.

As the mountain of new machine learning techniques grows exponentially, cleverly repurposing these tools for biological purposes has become its own kind of art form. Around the time that language models started to really work, computational biologists started to have more success modeling proteins using these types of approaches.

Interestingly, many of the earliest results can be traced directly to new biotech startups. Surge Biswas, one of the lead authors of a 2019 protein modeling paper from the Church Lab, is now the co-founder and CEO of Nabla Bio. Alex Rives, one of the leader authors of a 2019 protein language model paper, is now the co-founder of EvolutionaryScale. And Ali Madani, lead author of a 2020 paper on protein language models, is the CEO of Profluent, the company that announced the AI-designed CRISPR proteins last week.3

In the 2020 paper and a pair of follow-up papers in Cell Systems and Nature Biotech in 2023, Madani and his colleagues developed and scaled causal language models for the task of generating new protein sequences.

Let’s think back to how this strategy works in the context of natural language. The causal training setup involves masking the last word in a string of words, and making the model guess what is based on the attention patterns in the rest of the string.4 Once the model is trained, this prediction step can be repeated to complete an entire sentence.

For example, the following prompt:

"Somatic hypermutation allows the immune system to"

Could generate the sentence:

"Somatic hypermutation allows the immune system to be able to effectively reverse the damage caused by an infection."

Instead of predicting the next word in a sentence, the model—called ProGen2—is predicting the next amino acid building block in a protein.

And just like a language model can be used to generate new sentences, ProGen2 can be used to generate new proteins.

This is how Profluent generated new CRISPR proteins last week. There’s only one more core detail we need to understand about their process: fine-tuning. The masking trick we’ve talked about is typically used as a pre-training step in developing Transformer-based Large Language Models (LLMs) for language applications. This step learns a general representation of language, before training a model on a more specific downstream task.

For proteins, ProGen2 was first pre-trained on a giant database of 500 million different proteins. To specifically design new CRISPR proteins, Profluent created two additional databases of examples of CRISPR proteins found in Nature for fine-tuning.5

The intuition here is that the pre-training process is learning deep principles of protein evolution, while the fine-tuning step is teaching the model to become the world’s expert in CRISPR biochemistry. It is tabulating the relationship between every pair of amino acids in every single CRISPR sequence that we know of.

With the final model, Profluent generated 4 million different versions of the CRISPR-Cas9 protein. Half of these were directly generated with no prompt. The other half were given prompts of 50 amino acids from different members of the Cas9 protein family.

When put to the test in cells in the lab, some of the generated editors actually worked. Not only did they work, but they may be safer and more efficient gene-editing systems than spCas9.

But let’s revisit our earlier question: are these actually entirely new CRISPR proteins?

This is where things get fuzzy. One of the final generated designs, called OpenCRISPR-1 by Profluent, differs by 403 of its 1,380 amino acid residues compared to spCas9. That’s substantial. But when searching for related sequences, the closest result was 86.3% similar. And when comparing three or more sequences like Brian did, the number quickly approaches 100%.

When considering how the model is designed and trained, isn’t that exactly what we would expect?

Each time it adds a new amino acid to a sequence, the model is predicting the highest probability choice based on its knowledge of every known CRISPR. Again, this is more akin to Digital Evolution than it is to de novo design.

And that’s a good thing, because we currently don’t know how to actually design something as complex as CRISPR from scratch. In their introduction, the authors said that “for the design of complex functions as embodied by the gene editors in this work, structure-based approaches do not offer a straightforward solution. By contrast, language models provide an implicit means of modeling protein function (and thus structure) through sequence alone.”

So while these aren’t entirely new CRISPR systems—which is a tricky thing to really measure—this is an impressive demonstration of the generative capacity of protein language models.

They are powerful new tools in the growing computational protein design toolkit.

Final Thoughts

You may be wondering why Open-CRISPR-1 is named the way it is. It’s because you can download the sequence from GitHub and do whatever you want with it. According to Profluent, “OpenCRISPR is free and public for your research and commercial usage.” After filling out a simple Google form, the only requirement for use is to sign a simple non-exclusive license agreement.

I’m not a CRISPR patent lawyer—which is something I’m glad about—so I have no idea how this will hold up over time. If the CRISPR patent battles weren’t already messy enough, this adds an entirely new dimension: how do you prove that you computationally generated the new sequence and defend its novelty? This was the basis of an academic feud at MIT that dragged on for nearly four years.

Ultimately, I hope this CRISPR protein and it’s successors do open up access to gene-editing.

But as I mentioned when discussing other new tools for programming biology, there is still a suite of regulatory and financial challenges in the way of a future with ubiquitous programmable CRISPR medicines.

Even for companies that already have their own CRISPR intellectual property, it’s not all smooth sailing:

We could transition from modalities that require intensive development for every new application, to programmable modules where only the instructions need to be updated to make entirely new medicines.

This type of technology sounds like a cash printing machine. Surprisingly, nobody seems to have solved the business model problem for curing thousands of genetic diseases with CRISPR. Intellia, the company with some of the deepest expertise in this technology, had to lay off staff earlier this year.

Squaring these facts requires a consideration of the current economics of drug development. In an upcoming post, we’ll explore this in detail. Is there a platform strategy that could realize the full potential of CRISPR?

As the entomologist E.O Wilson once framed it, “The real problem of humanity is the following: we have Paleolithic emotions, medieval institutions, and god-like technology.” As we come to learn the language of Evolution and refine our capacity to steer it, we need to solve the complex human challenges that are rate-limiting broader access to better health.

Soon, we’ll dive into the details of how we might get to a “world where we have CRISPR on demand within weeks.”

Thanks for reading this overview of how protein language models were used to generate (somewhat) new CRISPR proteins.

If you don’t want to miss upcoming essays, you should consider subscribing for free to have them delivered to your inbox:

Subscribe now

Until next time! 🧬

Subscribe now

Enjoy! 🧬

It’s hard choosing between present and future value. Would you pick one marshmallow now, or two if you wait? This type of question is what sits at the core of the perennial product vs. platform debate in biotech—it’s just drugs instead of marshmallows.

Opinions swing dramatically based on market conditions.

When conditions are favorable, investors extend their time horizons. With a decade of 0% interest rates building to a climax of accelerated biotech investment during COVID, money was poured into moonshot platforms. When the tide went back out, some of these companies were the farthest from the shore. All of the value was in the future.

When this happens, biotech investors become pessimistic about the possibility of building generational companies. Public markets value positive clinical results—everything else is a rounding error. Every result is one step closer towards cash flows from a drug, or more realistically a pharma acquisition. The goal is to drive towards present value.

I say when this happens, because it’s cyclical. In an interview in 2012, the biotech veteran Steve Holtzman expression his frustration with the “continued pessimism about the possibility of building great companies that are sustainable as opposed to merely short-term plays which can make money for investors.” It’s the same phenomenon. He was just talking about the previous market downturn.

What does it take to build a sustainable company?

All value comes from clinical products. But it’s only possible to build a generational biotech company with a foundational technology platform capable of developing multiple products. This is why The Column Group—one of the best performing biotech venture firms of all time—invests in “unique scientific platforms” based on their “potential to deliver multiple breakthrough therapeutics.”

Great platforms require big ideas and long-term thinking. It isn’t possible to build a company like Genentech, Regeneron, Vertex, or Moderna without both ingredients.

Each of these companies used strategic pharma partnerships to scale. Partnerships are a way to trade some future value in exchange for near-term sustainability. As Steve puts it, “You need to use equity capital, but you also need to use pharmaceutical partnerships where they are your near-term market. You’re selling things that are on the way to product, or technology that they wish to access.”

But it’s important to avoid selling away too much future value.

Let’s look at an example. In a previous post about platform strategy, and I compared the growth of Ionis and Alnylam over time. Both companies focus on developing medicines that target RNA rather than proteins, using different types of short RNA molecules instead of small molecules. This is a really big idea—it’s a radically different approach to drug development.1

While both companies had clinical success, Alnylam is now much bigger. One major reason is that Ionis partnered away more of their products, while Alnylam aggressively forward integrated to commercialize their own drugs more quickly.

In their 2012 10-K filing, Ionis described their model, saying “Our strategy is to do what we do best—to discover unique antisense drugs and develop these drugs to key clinical value inflection points. We discover and conduct early development of new drugs and, at the key clinical value inflection points, out-license our drugs to partners.”

Alnlyam was thinking about things differently. In their filing for the same year, they said “Our core product strategy, which we refer to as “Alnylam 5x15,” is focused on the development and commercialization of novel RNAi therapeutics for the treatment of genetically defined diseases with high unmet medical need. Under our core product strategy, we expect to have five RNAi therapeutic programs in clinical development, including programs in advanced stages, on our own or with one or more collaborators, by the end of 2015.”

Empirically, Alnylam’s concentrated strategy has captured more value so far.

But history isn’t a perfect predictor of the future. There’s a chance that this lesson may not generalize to biotech platforms focused on compounding data and machine learning advances. These companies have a different trade-off between present and future value to consider.

What are the defining characteristics of this type of company?

As defined by Jacob Kimmel, a co-founder of NewLimit, this new “species” of biotech has three primary traits:

1. Builds an In silico Model of the biological process sufficient to predict the effect of changes to key engineering parameters

2. Collects & curates a Data Corpus describing a biological system more completely than ever before

3. Generates value from the model by Predicting and Validating useful modifications to a biological process to make it faster, cheaper, or more effective

In this framing, partnerships have a different impact. Historically, partnerships have been viewed as selling away future value to gain present value. In theory, if a platform’s value compounds over time with data, a partnership creates near-term value and makes future programs more valuable by increasing the likelihood of future success.

If this is the case, a business model that more closely resembles Ionis may be the right starting point. By focusing on partnerships, a data-centric company is effectively being paid to compound.

One example of this strategy in practice is AbCellera. Founded in 2012, AbCellera focused on building an antibody discovery engine and making it broadly available to partners. All the way up to their IPO, there was no internal pipeline. Instead, they had 94 discovery programs with 26 different partners.

Unlike nearly every other biotech, AbCellera was cash flow positive early on. In their S-1 filing, they said “Our business has historically been both high growth and capital efficient. Revenues have grown at a 109% CAGR since 2014. We have generated positive operating cash flow cumulatively since our inception in 2012 and in every year since 2018.”

While the core engine was compounding, they avoided forward integrating towards developing an internal pipeline. After nearly a decade of steady growth, the COVID pandemic sparked a major inflection point. AbCellera quickly launched efforts to develop antibodies to neutralize SARS-CoV-2. And just as quickly, they had their first results.

With a single sample from a COVID patient, AbCellera became the first company in the world to identify a monoclonal antibody targeting the virus.

Only ninety days later, Eli & Lilly dosed the first patients with this antibody in a Phase 1 trial. With an acceptable safety profile, it was granted an emergency use authorization (EUA) by the FDA and marketed as Bamlanivimab. Over 700,000 COVID patients were dosed with Bamlanivimab over the course of the pandemic.

Without a pause, AbCellera used their platform to discover a next-generation SARS-CoV-2 antibody. Again, it was granted an EUA and marketed by Eli & Lilly. In a statement, Carl Hansen, CEO of AbCellera, said “The discovery of two authorized therapeutic antibodies within a year of each other demonstrates the power of our platform, and its potential to quickly generate best-in-class therapeutics for our partners.”

After COVID vaccines hit the market, the emergency authorizations for the antibodies were revoked just as quickly as they were granted. The massive revenue spike came and went. But AbCellera was left in an interesting position.

After the pandemic, AbCellera had their discovery engine and over $1B in cash. With this foundation, they moved into the last phase of their platform buildout by investing in “manufacturing, regulatory, and clinical capabilities.”

As their platform build nears completion, AbCellera is starting to advance an internal pipeline of antibody drugs. They aim to file their first two IND submissions by 2025. As they put it, “We are evolving into a vertically integrated biotech.”

To be clear, none of this is investment advice. This is a thought experiment about the compounding value of platforms, and AbCellera is a really interesting example. By broadly partnering instead of immediately building an internal pipeline, AbCellera was able to invest in a decade of technology development with impressive capital efficiency. Now, they’re evolving to keep growing.2

What if AbCellera—or another company with this strategy—starts to produce differentiated products at a rate that no other company can keep up with? This type of technological inflection point hasn’t happened for biotech yet. But reasoning by (very) loose analogy, we know that sustained platform investment can lead to massive growth. Consider NVIDIA’s trajectory.

At NVIDIA GTC this year, Jensen talked about the early years of the CUDA platform for GPU programming, which has been instrumental in their rapid growth. He said they thought the project would be “an overnight success” back in 2006, and that “almost twenty years later it happened.”3

In their long journey through the desert, NVIDIA was busy making machines that make their machines. They were refining their simulation software and design tools. They were laying the foundation for GPU programming.

Now, they have a massive technical lead that will only be overcome by another technological paradigm shift.

Maybe this is why Jensen is so excited about “digital biology,” as he calls it. He’s survived a twenty year platform buildout that ultimately transformed his industry. Maybe he thinks this new wave of tech-centric bio companies could do the same.

This is effectively making the TechBio worldview as explicit as possible. It’s a belief that technology can transform drug discovery. From that perspective, the logical choice is to invest in sustained technology development long enough that this bet pays off in a big way. Everything else would be a rounding error.

Achieving that may require a creative business model and partnering strategy. A technology company needs to be able to keep compounding despite market fluctuations and binary clinical readouts. AbCellera was founded in 2012—the same year that Holtzman expressed his frustration with the industry’s short-term thinking.

But it didn’t matter, and that’s the point. They were able to stay cash flow positive while building towards vertical integration with their engine over time.

The goal is to continue compounding until inflection.

Thanks for reading this essay about long-term platform strategy. I’d like to thank Patrick Malone, Pablo Lubroth, Michael Fischbach, and Packy McCormick for their conversations and feedback on these ideas.

One quick note: the annual SynBioBeta synthetic biology is coming up fast. This year, it will be May 6-9 at the San Jose Convention Center in California. SynBioBeta has been a great supporter of this newsletter (to be clear, not a sponsor!), and is offering a $200 discount for readers who use the code centuryofbio when registering. If you join, feel free to reach out and say hi!

If you don’t want to miss upcoming essays, you should consider subscribing for free to have them delivered to your inbox:

Subscribe now

Until next time! 🧬

Subscribe now

Enjoy! 🧬

In June of 2012, a research paper describing the discovery of CRISPR landed in Science Magazine.

In November of 2023, a CRISPR-based cell therapy for blood disorders called Casgevy gained regulatory approval.

Taken in context, this is extraordinary. Most new drugs take between ten and fifteen years to make their way onto the market. This is the time estimate for drug modalities that we have many decades of experience developing—biologics and small molecules. Eleven years ago, CRISPR was a purely academic result in bacterial cells. It was a promising result, for sure, but nowhere near a drug.

Having evolved as a bacterial immune system against viruses, CRISPR can make programmed breaks in DNA—breaks that happen at a specific sequence in the genome. From the outset, it was clear that this type of system had huge potential. If we could supply the repair materials at the site of the break, we could controllably introduce new genetic changes into living cells. A world of programmable gene editing seemed to be on the horizon, but you still had to squint to see it.

The magnitude of the opportunity—again, programmable gene editing—was matched by the magnitude of academic and commercial interest.

The 2012 Science paper has now been cited over 10,000 times, indicating the escalation in follow-up research studies. Jennifer Doudna and Emmanuelle Charpentier were awarded the 2020 Nobel Prize for their initial discovery. All of this progress has sparked a long and complicated patent battle over commercial rights to CRISPR gene editing. DLA Piper estimated that there are now over 11,000 patent families covering CRISPR-related applications.

Commercially, venture capitalists invested over a billion dollars into CRISPR startups in 2021 alone, with many leading researchers founding competing companies. All of this research and investment explains the dramatically compressed timeline from proof-of-concept to life-changing clinical product.1

But what’s unusual about CRISPR isn’t just the speed to the clinic. It’s the type of discovery that is clearly the root of an entire technology tree, rather than a standalone tool for gene editing. A programmable system for targeting nucleic acids is a general purpose tool for engineering biology.

These types of general purpose tools build on each other. The biotechnology industry owes its origin to recombinant DNA technology. It’s impossible to imagine modern genomics without PCR. Now, with these tools in our toolkit, the pace of exploration of the CRISPR tech tree is hard to even keep up with.

At its core, CRISPR is an enzyme system that is guided by RNA. Change the guide RNA, and you’ve programmed the enzyme to target another sequence of nucleic acids. The first application was for programmed DNA breaks via the Cas9 enzyme—opening the door for a modification to be made.

Very quickly, people figured out additional ways to use this system. Refinements were made to avoid the need for making DNA cuts to make edits.2 The discovery of the Cas13 enzyme made it possible to programmatically target and edit RNA in addition to DNA. The same enzymatic machinery has been adapted to interfere with genes (CRISPRi) or activate them (CRISPRa) across the entire genome.

Our toolkit for programming biology grows by the day.

Each week, new inventions are announced. The past few weeks have been no different. While I had a different essay in the works, I realized it might be worth coalescing my notes on some of the recent results into a shareable post.

Today, we’re going to quickly examine:

• HACE — a tool for making lots of DNA edits in a targeted region of the genome

• A trio of new “RNA writing” preprints (RESPLICE, PRECISE, CRAFT)

• Bridge RNAs — “a third generation of programmable biology”

• Bonus: more retrons!

• Looking forward: from therapeutic modalities to therapeutic modularity

I don’t normally do research roundups like this, so if you enjoy it and want to see more of these, let me know in the comments.

Let’s jump in! 🧬

Helicase-assisted continuous editing (HACE)

In 2021, Tuuli Lappalainen and Daniel G. MacArthur wrote,

The human population, through explosive growth, has performed a comprehensive saturation mutagenesis experiment on itself. It is now the case that any single base substitution that is compatible with life is expected to be present somewhere among the nearly 8 billion living humans. Humanity has thus, in effect, done many of the natural experiments required to understand our own genotype-phenotype map; this leaves geneticists to catalog the outcomes of those experiments, and to leverage both observational and experimental approaches to understand the mechanisms by which variants alter biology.

I can’t think of a more beautiful framing for the discipline of human genetics. Natural genetic variation—every mutation present in the human population—is a massive resource for understanding the function of the genome.

Each mutation is like an experiment testing the role that a letter in the genome plays in human physiology. It’s why human genetics is a seemingly inexhaustible resource for discovering new drug targets—and why targets with genetic evidence are more likely to succeed in the clinic.

While we can catalog the outcomes of these experiments through sequencing the human population, it also helps to be able to generate and test genetic variants in the lab. A new preprint led by Dawn Chen and Zeyu Chen from the Bernstein and Chen labs introduced a CRISPR-based tool for creating targeted genetic variation in cells.

The technology is called HACE, and here’s how it works:

The programmable component is a variant of Cas9 that make DNA nicks—single-stranded breaks—instead of full-fledged double-stranded breaks. The guide RNA (sgRNA) recruits this enzyme to a specific sequence in the genome where the nick is made.

The second component is the HACE editor, which is a fusion of a helicase enzyme that traverses DNA, and a base editing enzyme. The HACE enzyme loads onto the site of the nick, and drags the base editor along the genome making new mutations as it goes.

The result is something like this:

A sprinkling of new genetic variation at a programmed target position in the genome. Obviously, you don’t want to do this inside of a patient’s cells. But given the power of genetic variation for deciphering function, it seems like a powerful new research tool for functional genomics.

In this initial study, the authors used this tool to drill into mechanisms of drug resistance, RNA splicing, and the role of genetic variation in the immune system. I’m excited to see what else this will be used for in the future.

It’s a great example of how CRISPR is a general purpose tool with many possible uses.

Advances in RNA writing

The first wave of CRISPR technologies focused on making permanent changes to the genome—editing DNA. As we saw with the discovery of the Cas13 enzyme, some focus has shifted towards targeting and editing RNA.

This is interesting for several reasons.

From a therapeutic perspective, genetic diseases may be treated the most effectively by a permanent DNA edit. Consider Casgevy—the first approved CRISPR-based product. Permanent genome edits are made to a patient’s blood cells, fixing the mutation that causes their sickle shape.

But not every disease has this clear of a genetic basis. In many cases, it makes sense to introduce an intervention that is temporary rather than permanent. Technologies for “transcriptome engineering”—editing many RNA transcripts in cells—could deliver this type of effect, without the concerns around making on-target or off-target edits that persist indefinitely in the genome.

Over the span of one week, three new tools were introduced describing new approaches to RNA writing.

On January 30th, the Hsu and Konermann labs—the two founding groups of the recently formed Arc Institute—published a preprint describing a tool for “rewriting human [RNA] transcripts” via trans-splicing.3

Briefly revisiting the central dogma, a DNA-encoded gene is transcribed as an RNA, before being translated into a protein product. In between transcription and translation is the wild world of RNA splicing, where RNA molecules are chopped up and pre-processed. As the authors put it, “Splicing is a fundamental biological process that plays an important regulatory role at the midpoint of the central dogma of molecular biology, enabling the same genetic sequence to give rise to a dynamic set of protein isoforms that can have distinct biological functions, subcellular localizations, or stabilities.”

Normally, this happens in cis—meaning all the splicing is done to one RNA molecule. In very rare cases, two distinct RNA molecules can be spliced together. This is called trans-splicing.

In this study, the authors invented a system for programmatically controlling trans-splicing using CRISPR machinery:

This technology, called RESPLICE, has two modules.

The first module is somewhat complex. It’s a variant of Cas13 (CasRx) that is catalytically inactive—meaning it binds but doesn’t cut RNA. The guide RNA brings along a new exon to be spliced in. This is called the trans-splicing module (TSM).

The second module is a catalytically active CRISPR enzyme (Cas7-11) that is used to cleave the RNA transcript downstream of where the new exon is supposed to be added. This two module design reflects the fact that to get this to work, you need to:

1. Create proximity between the exons you want to be spliced together

2. Outcompete the normal splicing product the cell is trying to produce.

So, what could this tool be used for? It could be used to “replace disease-causing exons, repair truncated proteins, or engineer protein fusions.” Again, engineering the transcriptome of a cell rather than the genome should have uses in both medicine and engineering.

This broad utility explains why the labs at the Arc weren’t alone in their efforts to get this to work.

On February 1st, the Abudayyeh-Gootenberg Lab at MIT published their own preprint describing the independent development of a technology for programming trans-splicing called PRECISE

Here, all the logic is similar, but the system relies on RNA hybridization instead of Cas13 to bring the new exon into proximity of the molecule being edited. The same enzyme (Cas7-11) is used to cleave the unwanted exons.

After getting this to work, the authors explored ways to program new splicing events without CRISPR. While CRISPR is powerful, introducing foreign bacterial proteins into the body comes with the risk of triggering an immune response, and their size can make therapeutic delivery tricky. To do this, they used ribozymes, which are RNA molecules that can catalyze enzymatic reactions.

But the excitement around trans-splicing doesn’t stop there.

On February 8th, the Asokan Lab at Duke University published their own preprint describing a CRISPR-based system for trans-splicing.

The logic is very similar: an RNA-targeting Cas13 enzyme is used to stitch exons coming from distinct RNA molecules together.

This trio of new studies highlights the growing importance of preprints. It’s encouraging to see scientists gain more control over their own publishing timelines. Instead of feeling “scooped,” or having editors potentially reject some of this work based on primacy, we’re seeing a full-on RNA writing bake off.

Science gets to move at the speed of the Internet. These methods will be tested in labs around the world and refined. The most compelling ideas will be mixed and matched, and we’ll have improved tools as a result.

It was the kind of week that made one RNA biologist say, “this is nearly as important as the first reports on CRISPR/ Cas9.”

RNA writing is a space worth watching closely. Two companies—Wave Life Sciences and Ascidian Therapeutics—are already preparing to test RNA editing medicines in the clinic. Another company, Amber Bio, emerged from stealth to commercialize their own recently published Splice Editing technology.

It’s exciting to see the progress in genome engineering spilling over into our ability to engineer the transcriptome. Hopefully patients benefit from these tools in the coming years.

Bridge RNAs

Bioengineering is a strange discipline. CRISPR wasn’t invented. It was discovered.

A lot of brilliant engineering was required to shape it into a reliable gene editing tool, and ultimately a medicine. But this key insight is important: powerful tools for programming biology are buried in genetic databases just waiting to be found.

A week before publishing the RESPLICE preprint the Hsu Lab at the Arc Institute shared a “detective story” that led to the discovery of a new system called bridge RNAs.

The group hunted through a massive bacterial database looking for a family (IS110) of mobile genetic elements (MGEs). MGEs, sometimes referring to as “jumping genes,” are stretches of DNA that paste and cut themselves in and out of the genome.

By carefully analyzing the flanking non-coding regions of this MGE family, they found something crazy. These stretches encode a set of RNA molecules—what they now call bridge RNAs—that contain hairpin loops that recognize target and donor DNA sequences.

Not only are they likely a critical part of how these elements pop in and out of the genome, they showed that they are capable of programming the removal, addition, and flipping of DNA sequences.

These are all fundamental operations for engineering DNA.

It’s still early, but these bridge RNAs may prove to expand the genome engineering toolkit—making it possible to mix, match, and rearrange large segments of DNA programmatically.

It’s remarkable to think about what other tools might be lurking in these databases.

Bonus: more retrons!

Speaking of genome editing tools hiding in plain sight, let’s briefly turn our attention to retrons. Like CRISPR, retrons are a type of bacterial immune system that have evolved to detect viral infection.

Retrons highlight how wild molecular warfare between bacteria and viruses can be.

Retrons are genetically encoded in bacteria. The system consists of a non-coding RNA molecule that gets expressed. Sitting in the same operon, there is a reverse transcriptase enzyme that produces a copious volume of single-stranded DNA molecules—the retrons—by transcribing the RNA molecule back into DNA. It’s currently understood that viruses chop up these molecules, which produces a toxic byproduct that serves as a suicide switch for the host cell. When the cell dies, the phage is unable to reproduce and neighboring cells are spared from infection.

From a technological perspective, retrons are a useful tool for producing abundant single-stranded DNA inside of cells. This DNA can serve as a template for genome engineering. Instead of needing to add a repair template for CRISPR (or something like bridge RNA), it can be made directly inside the cell using retrons.4

The Shipman Lab at UCSF is one of the groups most actively researching retrons, and I’ve been excited about their work for quite a while. In a preprint posted last month, they conducted the most extensive experimental characterization of retrons to date.

There are thousands of computational predictions for retron sequences, but few have been tested in the lab before. This study developed an experimental approach to test over a hundred new retrons experimentally.

If you’ve been reading CoB for some time, you may know what I’m thinking: this is a beautiful example of the 4-S model at work. Modern DNA Sequencing made it possible to sequence hundreds of thousands of genomes. Software tools were used to make functional predictions. Now, DNA Synthesis tested these predictions at an unprecedented Scale.

This census will likely be a gold mine for future retron engineering. Already, it appears to have uncovered several substantially more efficient new retrons.

While they haven’t yet become a standard part of the genome engineering toolbox like CRISPR, more detailed work like this could establish retrons as an important tool for programmable DNA production in cells.

Looking forward: from therapeutic modalities to therapeutic modularity

We’ve just covered a lot of ground. A lot of new acronyms and molecular mechanisms.

Let’s look at one more set of new results that helps frame where all of this might be going.

On February 1st, Intellia Therapeutics published new clinical results for CRISPR editing of the liver to treat hereditary angioedema—a rare disease where patients suffer from “angioedema attacks” of severe swelling and fluid accumulation, most often in the limbs, face, intestinal tract, and airway. So far, patients receiving the editing treatment have experienced a 95% reduction in these attacks.

Taken on their own, these trial results are impressive. But there’s another layer to this story. As Fyodor Urnov pointed out, there is only one difference between this medicine and another Intellia therapy for a completely different disease: the guide RNA.

This is what a future with programmable medicine looks like.

In biotech, we often talk about drug modalities, which are specific modes of treating disease like small molecules and antibodies. Every new tool we looked at today was published in the span of a few weeks. What does the world look like in a decade? Two decades? A century?

We could transition from modalities that require intensive development for every new application, to programmable modules where only the instructions need to be updated to make entirely new medicines.

This type of technology sounds like a cash printing machine. Surprisingly, nobody seems to have solved the business model problem for curing thousands of genetic diseases with CRISPR. Intellia, the company with some of the deepest expertise in this technology, had to lay off staff earlier this year.

Squaring these facts requires a consideration of the current economics of drug development. In an upcoming post, we’ll explore this in detail. Is there a platform strategy that could realize the full potential of CRISPR?

Until then, thanks for joining me on this tour of the powerful new tools for programming biology.

Thanks for reading this roundup of some of the most exciting advances in genome and transcriptome engineering.

If you don’t want to miss upcoming essays, you should consider subscribing for free to have them delivered to your inbox:

Subscribe now

Until next time! 🧬

Subscribe now

Enjoy! 🧬

Our economy runs on the decay of this planet’s first life forms. Fossil fuels. We use one third of the planet’s available land to raise more life forms, and then slaughter them for their meat and biomaterials. Animal farming.

None of this is great.

Imagine explaining our fuel and food sources to intelligent alien visitors. It would be pretty embarrassing.

Industrial biotechnology has a simple thesis. We should transition our economy onto a more sustainable foundation: biomanufacturing. Instead of ripping ancestral biological matter out of the ground or raising sprawling farms of sickly livestock, we should grow the fuels, chemicals, materials, medicines, and food that we need using the tools of biotechnology. We already have a Bioeconomy—this would make it more rational and sustainable.

From first principles, the argument is sound. It’s just expanding Genentech’s founding vision to include the rest of the sectors of the economy that biotech could impact—which turns out to be a lot.

While the logic checks out, the initial economic performance has been underwhelming.

Many companies have lost considerable value over time in the public markets. Others have spiked before crashing and being sold for a fraction of their peak valuation.

Some of the earliest companies have now gone bankrupt.

Yikes!

Why is this the case? One could argue that it’s a matter of technological readiness levels. Synthetic biology—the modern descendant of genetic engineering—was simply too nascent a technology to reach industrial scale when these companies launched.

There is a kernel of truth to this. Each year, progress in organismal engineering compounds. Over time, the R&D costs for making new bioproducts will continue to decline.

But there’s another layer to this problem. Culturally, synthetic biology has focused primarily on engineering, with a secondary emphasis on product obsession. Economic considerations have come after that.

This needs to change. If we want to build a biotech-powered Bioeconomy, it will require as much economic ingenuity as technical ingenuity. We will need to design businesses capable of being financially competitive—even against incumbent giants who possess the advantages of regulatory capture and Scale Economies.

There is good news: there are tools for achieving this.

One of these tools is called Techno-Economic Analysis (TEA). TEA is a modeling framework that aims to help forecast a new technology’s ability to compete in an industrial market. This grab bag of financial tools will be as essential as any genetic circuit in jumpstarting the Bioeconomy.

Tracking inputs and outputs

As a species, accurate financial record keeping is a brand new technology. At the beginning of the 20th century, not a single city in the United States kept a budget. This dramatically hampered the effective allocation of resources.

Since then, things have gotten much better. More standardized accounting practices spread across the business sector and into the government. Now, in the Information Age, we can barely keep up with the amount of data being stored each day.

But there is a lesson from history. The creation of new systems—including cities, government, and industries—often precedes the measurement systems that emerge to help them run more smoothly.

This was also true for synthetic biology. As the story goes, Pat Brown sold Vinod Khosla on the vision for Impossible Foods with an incredibly simple market analysis: “It’s a trillion-and-a-half-dollar global market being served by prehistoric technology.”

But would Impossible burgers truly outcompete meat products in a trillion dollar market, or would they primarily be competing with veggie burgers? What is that market size? And what would the cost of goods sold (COGS) be compared to other meat and veggie burgers? What about the margins—especially accounting for the use of biomanufactured ingredients?1

These are the types of questions that TEA aims to answer.

Most TEA begins with a process flow diagram that describes each step of the production process in as much detail as possible.

Next, the process needs to be overlayed with a careful accounting of what goes in and out of each step. This is the mass and energy balance of the technology.

From here, we transition into the world of economics. The goal is to quantify the actual costs associated with the system as it is described. These costs come in two flavors.

First, there are capital costs. Working backwards from the process diagram, what are all of the pieces of required equipment, and what are their prices? What will the cost of the land be where the facility is constructed? How much will the construction cost? These are all of the upfront costs that can be paid back across the full lifecycle of production once things are up and running.

Second, there are operating costs. How much will the inputs into each step cost in total? This includes human inputs—the labor required to run the process. Even seemingly small details like waste removal matter. Everything needs to be accounted for.

With a more detailed description of a technology—and the costs associated with it—an important question needs to be answered. Will this new process compete in the market?

A new bioproduct realistically needs to be able to compete on performance and price. Consumers and investors are both wary of “green premiums.” Achieving viability on both fronts is a nontrivial task. A good TEA model can serve as a debugger for doing so.

First, TEA helps you realize your starting point. It’s important to know the costs of your process to the highest possible degree of accuracy. If you aren’t cost competitive at a first pass, the model should help point you to the knobs worth tuning. Where can the greatest amount of economic leverage be gained through improvement?

TEA also provides a basis for economic creativity. With an entire process described in as much detail as possible, it’s possible to start to reimagine it. One company is trying to do this for fermentation at large.

Changing the economics of fermentation

Tool building is great. People building tools for you is even better. I’ve been obsessed with building Web tools for biologists for quite a while. One of my ideas for my PhD project was to build a simple and effective tool for synthetic biologists to quickly gauge the economic viability of their inventions. I was excited to see that a company called Synonym had already done this—and that they had done a better job than I could have on my own.

The tool is called Scaler. With a small set of form inputs, Scaler generates a reasonable base TEA to gauge feasibility. Information can be progressively layered on to the base model for greater accuracy. Many of the complex details of getting a TEA right are abstracted away and handled for the user.

How is this abstraction possible?

The central assumption is that many bioprocesses are quite similar. Many synthetic biology products use fermentation—the conversion of feedstock into products via microbial metabolism—as the means of production. Synonym views fermentation as the “platform technology of the Bioeconomy.”

If everybody is using the same platform, why should each company painstakingly model the unit economics of a fermentation facility in gory detail? Just refine the parameters that are specific to your project.

Synonym takes things a step farther. Scaler is one part of a broader effort to make fermented products economically viable. The team is also working to modularize facility design and make infrastructure more reusable across projects. They also want to make it possible for multiple companies to share the same facilities—further spreading costs and increasing margins.

All of these improvements play a part in a larger goal: changing the financing structure of synthetic biology.

So far, many companies have exclusively relied on VC funding. This is dilutive, because it involves selling company equity. In financial terms, this cost is borne on the balance sheet of the company. Investors are wary of this right now too—I’ve heard many say they “won’t finance high tech infrastructure that produces commodities.”2

In other words, it’s an expensive financing method for companies, and it is difficult for venture capitalists to get a sufficient return on their investment.

But what if several companies share reusable infrastructure? What if each product has a viable TEA thanks to Scaler? This could start to look like an appealing investment for other sources of capital—like project finance—that have different return expectations. Facilities could start to be built off of companies balance sheets.3

This is a big idea.

It’s an attempt to fund biomanufacturing infrastructure in a similar way to how we fund solar panel installation.

The goal of TEA is to help new bioproducts win in big markets.

Guided by TEA, Synonym is attempting to reimagine the economics of fermentation. Other companies like Solugen—which is built upon a strange and beautiful marriage of chemical engineering and synthetic biology—developed their own TEA models to validate the feasibility of their ideas. Additional approaches beyond fermentation may also radically shift the economics of bioproduction.

It’s hard to overstate the scale of the opportunity here. We’re talking about moving the world away from fossil fuels and animal farming, while truly disrupting the multi-trillion dollar markets that Pat Brown described.

But doing so requires technology and economics.

It requires spilling the TEA.

Thanks for reading this essay on the importance of techno-economic analysis.

Thanks to the Synonym crew—including Edward, Joshua, Brentan, and Crystal—for their conversations during my research. As always, this wasn’t a sponsored post, and I’m not an investor in Synonym. If I was, I would tell you!

If you don’t want to miss upcoming essays, you should consider subscribing for free to have them delivered to your inbox:

Subscribe now

Until next time! 🧬

Subscribe now

Enjoy! 🧬

It’s been seventy years since the image of DNA’s double helical structure entered our collective consciousness. In the subsequent years, we’ve cracked the genetic code, invented powerful DNA sequencers, and discovered molecular machines capable of manipulating and programming nucleic acids.

All of this has wide-ranging consequences for life as we know it, but one application is immediately obvious. If some human diseases are caused by genes, can we use the tools of modern genetics to find out which genes are broken and ultimately perform molecular surgery to correct them?

The answer is yes, but it’s hard. In total, the FDA has approved thirty-three cell and gene therapy products for clinical use. Of these medicines, only six work by delivering a functional copy of the defective gene directly to cells in vivo—inside the body.1

While this number isn’t zero, there are thousands of known human genetic disorders, including hundreds of disorders with well-described underlying mutations. What causes the disconnect? Why aren’t there hundreds of approved gene therapies?

There are several reasons. Resources for clinical development are finite, and some of these diseases are so rare that they are less prioritized by drug developers. Still, there is a looming technical challenge that explains a non-trivial amount of the discrepancy:

Delivering gene therapies safely and specifically inside the body is a really hard problem.

Imagine this problem from a gene’s-eye view. You’re a gene that’s just been injected intravenously into a patient with a neurodegenerative disease. You need to make your way from the bloodstream to the brain. The problem of crossing the blood-brain barrier is stressing you out. Assuming you overcome this obstacle, you need to enter the diseased cell type where the defective gene is causing problems. You need to do all of this while evading the immune system, which evolved to seek out and mitigate foreign nucleic acids. Not only that, you need to avoid ending up in other tissues or organs, causing more problems for the patient. It’s a hard job, and you’re starting to wonder why you signed up for it in the first place.

One of the main strategies for shuttling genes into the nucleus of cells in the body is to use viruses. After all, this is what viruses evolved to do. Still, using viruses presents several challenges. They didn’t evolve to be highly specific therapeutic vectors, and there is a constant arms race to avoid immune detection.

To circumvent these problems—and to mitigate the potential risks associated with using a virus—gene therapy developers turn to the tools of molecular engineering. For the most widely used viral vectors, scientists strip out the original viral genome, replacing it with the therapeutic payload. This avoids risks associated with unwanted viral replication. Next, the viral protein shell—called a capsid in virology—needs to be engineered to target specific cells in the body. This is easier said than done.

Dyno Therapeutics was founded with a singular purpose: to solve the problem of in vivo gene delivery by cracking the capsid design problem.

Today, we’re going to dive deep into Dyno’s strategy for doing this. Dyno is a fascinating company. It is riding the tailwinds of exponential progress in genomics, synthetic biology, and AI in the pursuit of solving a Holy Grail problem for gene therapy. To do this, Dyno has created a purpose-built and somewhat contrarian platform strategy—they focus entirely on the problem of solving in vivo gene delivery instead of making their own gene therapies.

Dyno has moved quickly on this vision. Since their seminal publication and company launch in 2019, they’ve announced four partnerships with leading drug makers and have raised a $100M Series A led by Andreessen Horowitz to scale their design efforts. Recently, they’ve also announced two of their first off-the-shelf gene delivery products. This progress hasn’t gone without notice—Dyno has been ranked as one of the top emerging biotech companies to watch.

In order to understand all of this, I’ve spoken with many of the company’s co-founders and investors, visited their labs, read every paper they’ve published, and reviewed countless private and public materials on the company.

Studying Dyno’s thesis and initial efforts in detail reveals many lessons. We’ll examine new measurement technologies, the challenges of biological sequence design, how to pick problems worth solving, and what the horizontal biotech platforms of the future might look like.

The ultimate puzzle

If there’s one thing Eric Kelsic loves, it’s a good puzzle. At first, this fascination was as literal as it could be. As a precocious high schooler in Colorado, he became obsessed with Karakuri boxes—which can only be opened after solving a three-dimensional puzzle, often involving hidden mechanisms or a clever twist. He quickly tired of solving existing boxes and started a small business marketing his own inventions.

Kelsic’s love for tinkering and engineering led him to the sunny campus of Caltech for college. Nestled in Pasadena, Caltech looks and feels like a Temple of Science. Unlike the entrepreneurial hubs of Boston and the Bay Area, companies are few and far between. Scientific knowledge is front and center. And out of all of the sciences, Caltech’s impact on physics has been the most formidable.

Under the leadership of Robert Millikan—one of the greatest experimental physicists of the 20th century—Caltech’s warm campus served as a central meeting hub for many of the field’s greatest minds. Paul Dirac, Erwin Schrödinger, Werner Heisenberg, Hendrik Lorentz, Robert Oppenheimer, Niels Bohr, and Albert Einstein all paid visits to the “Millikan Institute” for varying lengths of time. After World War II, Caltech’s prowess in physics only continued, with Murray Gell-Man and Richard Feynman laying the foundation for the Standard Model of particle physics as faculty members.

Kelsic heard Feynman’s echoes. Excited by the pursuit of “getting to the bottom of things,” he decided to major in physics. Caltech also offered Kelsic his first taste of basic research. He was completely hooked. Enthralled by the need for open-ended problem-solving, Kelsic proceeded to wander through a variety of research fields, including materials science, cryptography, and complex systems theory.

Over the course of the 20th century, the quantitative and systematic thinking from physics began to spill over into the field of biology. Again, Caltech served as a central hub for a brewing scientific revolution. Many of the foundational results in the early days of genetics and molecular biology—including Thomas Hunt Morgan’s pioneering work in genetic mapping using fruit flies, Theodosius Dobzhansky’s foundational contributions to evolutionary theory, Linus Pauling’s seminal advances in the nascent field of structural biology, and Max Delbrück’s careful examination of the physical structure of genes—all happened at Caltech.

Like in physics, all of this activity caused many of the era’s great biologists to make a pilgrimage to this highly productive Temple. Jacques Monod, Jim Watson, Seymour Benzer, and Sydney Brenner all either visited Caltech or conducted their primary research there. Even the great resident physicists—including Feynman—spent time dabbling in molecular biology.2

Over time, Kelsic honed his taste in the types of research problems that he was drawn to. He began to hear the echoes of Delbrück. A course in biophysics fascinated him. Unlike high school biology—which consisted of unappealing rote memorization—this field seemed ready to get at the fundamentals of living systems. An additional bioengineering course with Christina Smolke sealed the deal.

Whereas the frontiers of physics had become concerned with extremes—the smallest possible things in the universe, the conditions of black holes, the behavior of matter at the coldest and hottest temperatures—biology was full of fresh problems. The notion that living systems are molecular machines filled Kelsic with excitement.

Attempting to understand—and perhaps one day design—these machines felt worth devoting his life to. In other words, he realized that it is The Century of Biology.

As he prepared to graduate, Kelsic had a choice. He had a job offer as a programmer at a startup. He also had the opportunity to join the Smolke Lab as a technician. Hearing Kelsic wax poetically about the beauty of molecular machines, the startup founder rescinded Kelsic’s offer. At this point, it was clear to the rest of the world that Kelsic was destined to make his impact in biology.

After a year of research, Kelsic headed across the country to start graduate school at Harvard. He was drawn to the Department of Systems Biology, which seemed to have the right balance between experiment and theory. After his first-year research rotations, Kelsic found his home in Roy Kishony’s lab.

Like Kelsic—and the founders of molecular biology before him—Kishony had originally trained as a physicist. He became obsessed with evolution and the problem of antibiotic resistance and decided to hop fields to focus his research efforts on finding a solution. Kelsic was drawn to Kishony’s style of doing “extremely simple experiments but on a really large scale.”

A beautiful example is one of the projects Kelsic contributed to in the lab. The setup was simple. A giant agar dish was prepared with regions of varying concentrations of antibiotics. The concentration was the strongest at the center of the dish. Microbes were plated on the far edges. Over time, they crept closer and closer to the center. With a single elegant experiment, the landscape of evolved resistance to antibiotics was directly visible.3

As his PhD progressed, Kelsic had to design his own thesis project. A few emerging technologies caught his eye. First, the Church Lab at Harvard had developed a technology called Multiplex Automated Genome Engineering (MAGE). While the antibiotic resistance experiment relied on random mutations to occur, MAGE repurposed a bacterial system called Lambda recombination to introduce billions of engineered variants into microbial cells. Critically, the variants were encoded in short stretches of synthetic DNA—called oligonucleotides, or oligos—that can be made at a massive scale in parallel. If the agar plate experiment was Evolution, this was Super Evolution.

Second, the high-throughput DNA sequencing cost decline was well under way. Sequencing was no longer a government mega-project—individual labs were using the new technology in increasingly creative ways. MAGE was powerful, but until then, it was prohibitively expensive to measure which variants were actually being introduced—and which were disappearing or being selected for—at each round of perturbation. Kelsic wondered: would it now be possible to comprehensively sequence a MAGE experiment, providing a glimpse of the underlying evolutionary fitness landscape at each step of the experiment?

The idea for MAGE-seq was in place. Using clever molecular biology, Kelsic devised a way to use MAGE to introduce every possible codon substitution into an essential E. coli gene, and then used high-throughput DNA sequencing to measure the fitness of each variant across the different “pools” of mutants.

In the initial study, Kelsic used MAGE-seq to probe the relationship between RNA structure and codon usage. While it may provide insights for gene design in the future, it’s a fairly academic problem. But buried in the supplement was one of the core insights that would lead to Dyno’s founding.

While this may look like a somewhat confusing heat map to the casual observer, Kelsic could see the underlying principles of biochemistry popping right out at him. This map shows the amino acid substitution at each position in the E. coli gene infA. Below the map, a PCA analysis showed that only a few principal components were necessary to explain much of the variance in the data. On closer examination, each principal component seemed to correlate directly with an underlying biochemical rule. For example, PC1 was most strongly associated with residues that were buried inside the three-dimensional structure of the encoded protein.

Put more simply, a small number of biochemical rules were highly predictive of which amino acids could be tolerated at each position in the gene.

Kelsic realized he was on to something. This experiment was an example of a multiplexed library. Each variant in the pool could be identified and tracked using DNA sequencing.

Despite the technical complexity under the hood, the experiment itself was simple. Swap every codon with every other possible codon. See what happens. Like with the agar dish, deep principles fell out of a simple experiment. Both studies used a simple idea at a massive scale to offer a glimpse of the evolutionary landscape Darwin could only theorize about.

It was clear to Kelsic that MAGE-Seq—and other technologies like it—could have a massive impact on the field of protein engineering. One day, we might be able to understand what a protein does just by looking at its 3D structure on a computer. But that day is far away. In the meantime, we have tools to test potentially billions of designs empirically using the awesome power of genomics.

Multiplexed libraries offer a data-driven way to explore and engineer complex protein systems.

As 2015 began, Kelsic tucked this idea in the back of his mind and prepared to defend his PhD. When thinking about his next steps, the allure of startups tugged at him. He and a few friends had earned runner-up at the Harvard President’s Challenge for Social Entrepreneurship. While their genome browser startup fizzled out, the experience had been a highlight of grad school. As the former captain of his college soccer team, Kelsic relished being a part of a tight-knit group working towards a common goal.

With this interest in mind, Kelsic reached out to George Church. Between his Darwinesque beard, his lanky 6-foot-5 frame, and his lab’s seemingly endless stream of inventions, Church is a towering figure in the world of biotech.4 One of the many quirks of the Church Lab compared to its peers is that it also tends to churn out more young biotech CEOs than academic professors.