Big pharma picks its AI model vendor
-
Nº LXXXII
- Date
- 21 Aug 2026
- Issue
- 82
- Stories
- Seven
- Editor
- ARC
BMS names a structure-model supplier, Anthropic starts writing to chemists, and an agent learns when to quit.
BMS bets on Chai's models
BMS is deploying Chai's models across therapeutic discovery, cofounder Joshua Meier said, one of the clearest signals yet that AI structure and interaction models have moved from evaluation to production inside big pharma. Chai Discovery builds models that predict biomolecular structure and binding, the layer sitting between target selection and a real candidate molecule. The announcement's framing is blunt: frontier drug programs require frontier capabilities. What shifts is the buying question — pharma is no longer asking whether these models clear a benchmark, but whose models get written into the discovery pipeline. Model access, not model existence, becomes the competitive variable in early discovery.
Anthropic talks binder design
Anthropic walked through binder design in a public post explaining that many drugs work by binding a target tightly, and that designing that first tight-binding molecule is the hard opening step of a program. The content is introductory; the sender is the news. A general-purpose AI lab writing directly to a drug-discovery audience rather than a developer one signals molecular design being treated as an application area in its own right, which changes who biology gets to count as a discovery-tools supplier.
Agent evolves its own algorithms
Autonomous algorithm evolution arrives in drug development in a new bioRxiv preprint, describing a system that generates and improves its own algorithms rather than executing fixed ones. Detail beyond the design claim is thin, but the direction is the notable part: most discovery tooling treats the method as fixed and the data as the variable. Handing method development to the system itself shifts the capability question from how good today's algorithm is to how fast the algorithm can be improved.
Cryo-EM agent learns to stop
CryoForge knows when to stop building. The self-correcting agent handles cryo-EM model building and decides when further refinement helps versus when to halt, putting an explicit stopping criterion into a workflow where iteration still runs largely on human judgment.
Feasibility built into peptide design
FAR-DPO steers peptide generators toward molecules that can actually be made, folding feasibility into direct preference optimization, the training method that pushes a model toward preferred outputs. Synthesizability moves inside the objective for cyclic-peptide design rather than acting as a downstream filter.
Test-time tuning for virtual screening
PETA adapts screening models at test time by updating only a small share of parameters, aimed at the distribution shift that degrades virtual screening on unfamiliar target classes. Cheap adaptation lowers the retraining cost that keeps screening models pinned to the chemistry they were trained on.
OpenAI's GPT-5.6 builder guide
OpenAI published a builder's guide to GPT-5.6, covering model selection and new Responses API capabilities aimed at cheaper agents. Cost per agent run is the binding constraint on long-horizon discovery loops, so routing guidance lands on the budget line.
Reply with your discoveries. A human reads them. Forward freely.
|