AI Drug Discovery

The most ambitious bet in the space — $1B to build a vertically integrated AI drug company from scratch. The team is elite (ex-DeepMind, ex-Genentech leadership), but the thesis requires everything to work: data generation, model training, and clinical translation. No one has pulled that off yet.

Has the best foundation models in the world and the worst incentive to share them. Isomorphic's moat is AlphaFold, but their challenge is proving that structure prediction translates to clinical outcomes. Eli Lilly and Novartis partnerships are the first real test.

The only public AI drug discovery company with a genuine data moat — their phenomic dataset is unmatched. But being public means quarterly pressure on a 10-year thesis. The Roche-Genentech partnership validates the platform; pipeline clinical readouts will validate the company.

Daphne Koller's bet that the bottleneck isn't compute but data quality. Insitro generates its own cellular data at scale, which is the right instinct — but the gap between cellular models and human biology is where most drug discovery AI fails.

The leading generative protein design company, now backed by Novartis. Their Chroma model is impressive, but the real question is whether computationally designed proteins can survive the clinic. First IND expected 2026.

Smart positioning — instead of discovering new drugs, Cradle helps biotech teams engineer better versions of proteins they already have. Lower risk, faster feedback loops, clearer path to revenue. The un-sexy approach that might actually work.

Focused on the unsexy-but-critical middle of drug discovery: lead optimization and ADMET. Less flashy than de novo design, but this is where 80% of programs actually fail. Strong team, realistic scope.

Terray's bet is that better chemistry data beats better AI. Their microfluidic platform generates binding data at a scale no one else can match — and in drug discovery, data wins. The anti-thesis to pure computational approaches, and it might be right.

The sleeper pick. Cradle is doing for protein engineering what GitHub Copilot did for code — making the bench scientist 10x more productive. The SaaS model means they capture value across the entire protein engineering market without owning any single therapeutic. Smart capital-light strategy in a capital-heavy industry.

Generative biology is the most intellectually honest framing in this space — they're not pretending to have solved drug discovery, they're building a generative engine for protein therapeutics. The Novartis partnership de-risks the platform, but the real test is whether diffusion models for proteins can outperform directed evolution at scale.

Daphne Koller has quietly built one of the most defensible data assets in the industry. The iPSC-derived cellular models are the moat — every other AI biotech is training on the same public datasets. This is the one to watch if you believe data quality beats model architecture.

Alphabet is essentially funding its own subsidiary, but the pharma partnerships with Lilly and Novartis give this credibility that most DeepMind spinouts never achieve. The question is whether AlphaFold's structural insights translate to clinical candidates — structure is necessary but nowhere near sufficient.

The public markets are finally giving Recursion credit for the data flywheel, but the stock is still priced on vibes, not pipelines. They need a Phase II readout that proves phenomics-driven discovery actually works in patients, not just in pretty embeddings.

A billion dollars for a company with no clinical assets — this is either the most important bet in biotech or the most expensive science experiment in history. The team justifies the valuation; the timeline doesn't.