Drug Discovery
AI models can explore chemical and biological space at a scale that physical screening will never reach. We build the systems that narrow that space to the handful of molecules worth making.
Capabilities
Rank and characterize targets from multi-omic, literature, and structural evidence.
Screen ultra-large libraries with learned scoring and physics-based rescoring.
Generate candidates under multi-objective constraints rather than filtering a fixed catalogue.
Improve potency, selectivity, and developability together instead of one at a time.
Estimate absorption, distribution, metabolism, excretion, and toxicity early.
Match approved molecules to new mechanisms and patient groups.
Search interaction space where single-agent screening cannot go.
From billions of possible molecules to a small number worth testing.
How we choose
Screening capacity is no longer scarce. We look for the questions that stalled because the calculation was too expensive to run at scale.
Where a property can only be settled by a calculation nobody could afford to repeat, scale changes the answer rather than the speed.
A model is only as good as the measurement under it, so we invest in producing the measurement rather than reusing a thin public one.
Frontier, open-weight, and domain-specific scientific models are compared on the same task before any of them is trusted.
Why AI
Make-on-demand libraries grew from millions to tens of billions of compounds. One 2019 screen docked 138 million molecules containing 10.7 million scaffolds that could not be bought anywhere. No physical screening deck reaches that, so most drug-like space had simply never been looked at.
Medicinal chemistry works outward from families that already worked. A model scores a failed diabetes compound and a novel scaffold on exactly the same footing, which is how an antibiotic was found inside a discontinued metabolic drug.
One screen ranked 12 million compounds for antibacterial activity and human-cell toxicity together. A human pipeline resolves that trade-off late and one molecule at a time.
Substructure rationales turn a lucky hit into a structural class. Chemists get something to develop, not a single compound to defend.
Evidence
Results from peer-reviewed literature, each one a thing a conventional pipeline was not set up to find.
A model trained on 2,335 molecules flagged a compound no antibiotic chemist would have tested. It is structurally unlike any existing class and cleared pan-resistant A. baumannii in mice. Of 23 predictions tested from a 107-million-molecule screen, 8 were antibacterial.
Stokes et al., Cell, 2020 ↗Graph neural networks scored twelve million compounds for activity and for human-cell toxicity at once, then reported the substructures behind each prediction. 283 compounds were tested; the resulting class is selective against MRSA and vancomycin-resistant enterococci.
Wong et al., Nature, 2024 ↗Two unrelated targets were screened against a virtual library richer in scaffolds than any physical collection. The campaign produced a precedent-free 77 nM β-lactamase inhibitor and 81 new chemotypes for the D4 receptor, including a 180-picomolar selective agonist.
Lyu et al., Nature, 2019 ↗Narrow spectrum is commercially unattractive and normally screened out. A model trained on a small in-house set was optimised for selectivity instead of breadth, and returned a compound that acts almost only on A. baumannii.
Liu et al., Nature Chemical Biology, 2023 ↗TNIK was nominated as an anti-fibrotic target from omics data, and generative chemistry designed the inhibitor; target to candidate took about 18 months. In a 71-patient Phase 2a the treated arm gained lung capacity where placebo lost it. The trial's primary endpoint was safety, so read this as encouraging, not as proof.
Xu et al., Nature Medicine, 2025 ↗The honest other half: no AI-discovered drug has been approved by a regulator. Halicin, six years after it made headlines, has still never been given to a human because it is poorly absorbed and quickly cleared. An analysis of AI-native pipelines found Phase I success of 80 to 90 percent but Phase II success near 40 percent, which is the industry norm. AI is demonstrably good at designing drug-like molecules; it has not yet shown that it picks targets that work in patients.
