AI Was Supposed To Level Drug Discovery

This week, Anthropic's latest protein-design result started making the rounds all over social media. Claude was used to design protein binders against 15 different targets, of which, fourteen worked.

That sounds like another AI drug discovery story, except that’s not really what Claude did.

Claude was given access to the same kinds of specialist protein design tools that researchers can already use. It chose where to bind, decided which models to run, generated sequences, screened the results, and eventually decided which proteins were worth sending to the lab for testing. Across 1,320 designs, 354 actually bound.

So the interesting part is not that Claude built the world's best protein model, because it didn't. The interesting part is that Claude increasingly knows which tool to use and which experiment to run.

And if that workflow keeps getting better, I believe, could create problem for AI drug discovery startups.

For years, the pitch was basically this. Pharma has money and laboratories. We have the intelligence.

That made sense when building a sophisticated computational drug discovery platform required a large team of machine learning scientists, custom infrastructure and proprietary models. A small biotech that had all of that could genuinely do something Pfizer could not easily replicate.

But what happens when everyone gets Claude? Everyone can access increasingly good structure prediction. Everyone can use protein design models, rent compute, search PubMed, ChEMBL and the PDB. The intelligence part starts getting cheaper and more standardized.

At first, you would think that sounds great for the little guy, then you look at what everyone does NOT have.

Eli Lilly has more than 20 years of proprietary preclinical data behind its TuneLab models, including over 500,000 measurements across pharmacokinetics and toxicology. Lilly estimates the experiments behind the initial models cost more than $1 billion to generate.

Across ten pharmaceutical companies, the MELLODDY project brought together 2.6 billion confidential experimental measurements covering more than 21 million molecules and 40,000 assays. The companies did not have to reveal their data to one another. Their models could still learn from the larger pool, and predictive performance improved.

Think about what is buried inside those databases.

The compound that looked perfect until the liver signal appeared. The chemical series that failed after 200 analogues. The target that produced beautiful mouse data and went nowhere in humans. The thousands of molecules around a scaffold that nobody bothered publishing because they did not work.

PubMed contains a lot of what science learned.

Big Pharma owns a lot of what science learned not to do.

That difference is so much more valuable as AI improves.

Give a startup and Pfizer the same “AI scientist” and they may technically have the same intelligence. But Pfizer's version can potentially read decades of internal experiments, failed programs, toxicology, and clinical observations that the startup's version has never seen.

They are running the same model, but they’re not running the same “scientist”.

The only thing to consider here is that old pharma data can be messy, inconsistent and trapped across systems that were never designed for machine learning. A startup can intentionally generate clean, high-throughput experimental data that might have a better dataset for a particular problem.

So the moat could be moving away from who has the smartest model and toward who has the best experimental memory.

AI was supposed to erase Big Pharma's advantages, instead, it may have just made 40 years of failed experiments incredibly valuable.

Insilico Medicine reported $106.3 million in first-half revenue and $35.5 million in net profit, marking its first profitable half-year. Most of that revenue came from drug discovery and pipeline deals, not software, giving AI drug discovery one of its clearest commercial validation points yet. [Link]

HHS is reportedly considering creating a new FDA deputy commissioner role focused on technology, healthcare and AI. The appointments are not final, but a dedicated AI role at that level would signal how quickly AI is becoming a core regulatory issue across drugs, devices and healthcare. [Link]

Chinese AI biotech Matwings says it used four consecutive AI-to-wet-lab cycles to optimize a protein drug that has now been cleared for clinical trials in China. The team tested 222 protein variants and repeatedly fed experimental results back into the models to improve potency, stability, expression and formulation. [Link]