17 Comments
User's avatar
Will G.'s avatar

Love this!

Expand full comment
Akshay's avatar

Hugely optimistic about the future!

Expand full comment
V Allen's avatar

Great essay with great examples. Lots to be excited about in biology!

Expand full comment
John Michael Brummer's avatar

As an investor in Illumina and a few other small biotech companies, and as a non-biologist, I welcome this very interesting article. It explains a lot to me and I keep fingers crossed that I will recieve at last some return on my investment….. thanks and keep going!

Expand full comment
Görlitz's avatar

Great read - thanks! In relation to “searching for new platform companies with technologies that can make otherwise impossible drugs“ - are you familiar with the cell-free AI-enhanced enzymes platform, eXoZymes? Here’s a third party report: https://www.slack-capital.com/p/exozymes-snapshot-investment-case

Disclaimer: I work for this company and as we’re already public on Nasdaq (EXOZ), I’m not trying to pitch here - just making sure you’re aware that what I’m hearing you talk about, across you excellent writing, is that this is a platform that can make new-to-nature molecules, fast and cheap, compared to traditional biotech. Regardless: Thanks for sharing your thoughts - very inspirational!

Expand full comment
Elliot Hershberg's avatar

Thanks! I haven't studied eXoZymes closely, but I'm a fan of cell-free so I'm curious to check out this write-up.

Expand full comment
Eric Kernfeld's avatar

Really nice essay! Thank you for sending it out!

I'm thinking about Brandon White's three truths. Does #3 not contradict #1 and #2? Recapping:

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

2. The further you 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.

Isn't choosing a target, or a phenotype, or a population of likely responders, very far from the commercial and clinical stages? Those sound like academic research questions to me. I've never worked in drug development so I'm curious to hear your take.

Expand full comment
Elliot Hershberg's avatar

Thanks for reading, Eric. I liked David's response, but I'll add my two cents as well.

It would be worth checking out Fig. 2 in Scannell's Eroom's Law paper: https://doi.org/10.1038/nrd3681

There are a ton of assays and steps in discovery before the clinic. (We mainly add more steps over time, and so far none of the steps have radically changed success rates.)

Toxicity studies come in the last portion of pre-clinical development once you've already done a lot of work to discover and develop a molecule. So it's closer to helping with a gating decision to advance a development candidate (DC) into Phase 1 studies, versus going from hit to lead.

Your point about right target/phenotype is a good one. That needs to be made at the start of a program—or else you don't have a target to develop a drug against! But maybe one way Brandon is thinking about it is that it is a decision around what program to even start or prioritize in the first place.

Efficacy prediction would be the Holy Grail of having better information to advance from Phase 1 to Phase 2, and identifying the right target population is a clinical development question.

Expand full comment
David Borhani's avatar

Hi Eric,

Good question!

3 is actually the consequence (or corollary) of 1 and 2.

The most costly failure mode in drug development is a Phase 3 clinical trial failure, usually due to insufficient efficacy relative to safety (i.e. adverse events or side effects, especially longer-term or rare side effects that only become clear in Phase 3).

Thus, choosing the right drug target/phenotype that can provide both adequate efficacy and safety margins when used in the right target population becomes critical.

But, the industry is generally very poor at this: only ~10% of drugs entering Phase 1 trials ever reach commercialization. And most of those failures are due to insufficient efficacy or safety or both, even though the drug looked good enough before the clinical trials (else no one would have taken it to the clinic...).

So, choosing wisely early on, and doing so continuously as preclinical data accumulate (#3), keeps you at the program decision points (#1) and keeps you closer to the end goal (#2).

Expand full comment
Saul's avatar

Agree completely. Changing the probability of clinical success in a meaningful way would be quite revolutionary (and with significant pricing implications too).

Expand full comment
Ben's avatar

Excellent article, do you think model architecture, the training data, or a combination is the moat for these Bio x AI companies? The switching costs for NGS services are high, but I’m not so certain that will be the case for AI models (see how people switch so easily between Google/OpenAI/Anthropic based on cost/performance). And as long as big pharma remains the largest data generators they will likely have the best training data. As you highlight Amgen has 20-30 internal models and many of the big pharmas have their own large AI groups (GSK, and Roche/Genentech make a lot of press on this front) and others have large deals with major data generators( AZ + Tempus). I guess I wonder if there will be space for a sustainable business model for many Bio x AI companies.

Expand full comment
Interesting Engineering ++'s avatar

Anyone feel like samping or "playing with a bit of code" I have a beginners journey/guide to appreciating why and how Alphafold Matters. Preamble to this amazing read! : https://open.substack.com/pub/interestingengineering/p/the-alphafold-journey-from-grand?utm_source=share&utm_medium=android&r=223m94

Expand full comment
ASHWIN RAJAN's avatar

Beautifully written

Expand full comment
Jamin Roh PhD's avatar

I suspect that at the very least new cancer drugs using AI methods continue to have high toxicity in patients. The major issue I see with this is that they large data models all treat cells like uniform spheres because they are only interested in abundance or expression level and less about the heterogenity of the cells. A prime example is the apical side of the GI tract, mucus secreting cells have a distinct distribution of membrane bound proteins. I hope the silico cell model will be able to capture that diversity in the future. I think the older imaging techniques while more costly in time and human effort may yield better targets for patients in the future. But I totally agree with your take on reagents!

Expand full comment
Rainbow Roxy's avatar

It's interesting how you contextualize the current biotech landscape, making me wonder what specific AI infrastructure gaps you believe are most primed for disruption given this scientific rennaisance, which truly is a brilliant take.

Expand full comment
Dhruv Ghulati's avatar

This is amazing

Expand full comment
Shekhar's avatar

Nice.

Expand full comment