Frontier models move into pharma pipelines
-
Nº CI
- Date
- 17 Sep 2026
- Issue
- 101
- Stories
- Eight
- Editor
- ARC
Big pharma picks its model, Stanford makes papers executable, and a formulation lab runs itself.
Novo Nordisk adopts Anthropic's Claude for drug discovery
Novo Nordisk will run drug discovery work on Claude, the frontier model family from Anthropic, per a Euronews report picked up on Hacker News. Details are thin, but the shape of the deal matters: one of the world's largest pharmaceutical companies is betting on a general-purpose reasoning model rather than a bespoke biology-specific system. That choice puts general models in direct competition with the purpose-built discovery platforms pharma has been buying for a decade, and it sets a reference point for what a frontier-lab pharma partnership is worth. The open question is whether the work stays in literature synthesis and analysis or reaches target selection and candidate design.
Stanford tool turns research papers into working agents
Paper2Agent converts papers into working software agents, Stanford researchers report. Feed it a paper and its codebase and it produces a tested agent that reproduces the published results, then runs the same methods on new data. Reproduction has been the field's unglamorous bottleneck — most published pipelines take days to stand up, when they run at all. Turning the published artifact into something executable on demand changes what a methods section is expected to deliver, and moves reuse of published code from aspiration toward default.
Autonomous lab formulates drugs and cites its evidence
An autonomous lab formulates drug products using agents required to ground each decision in cited evidence, according to a new arXiv preprint. The setup pairs literature-grounded reasoning with physical experiments, so a proposed excipient or process parameter arrives attached to the source that justified it and the run that tested it. Formulation has stayed stubbornly manual while target discovery got automated. Building an auditable evidence trail into the loop is what makes closed-loop chemistry defensible to a regulator, not only to a reviewer.
Anew Labs links discovery steps into one loop
Anew Labs unveiled AnewDDE , an engine that wires structure prediction, binding affinity, molecular design, and wet-lab feedback into a single loop rather than shipping another standalone model. Integration between steps, not a better model at any one step, is increasingly what discovery platforms compete on.
Rare-disease agent asks for the symptoms it needs
A rare-disease agent asks for the phenotypes it is missing: HPOQuest chooses which Human Phenotype Ontology terms to acquire next instead of scoring a fixed feature list. Active questioning moves diagnostic agents closer to how a clinical workup actually proceeds.
Montana builds its own path for experimental treatments
Montana's new biotech law draws an inside account on Hacker News from a participant, describing how the state assembled its own access pathway for experimental treatments. Where a therapy can legally reach patients is becoming a planning variable in translation rather than a fixed constraint.
Small models split up single-cell annotation work
scACORN routes single-cell tasks to a set of specialized small language models under one orchestrator, rather than pushing every transcriptomic interpretation question through a single large model. Cheap specialist ensembles are a plausible answer to the cost of running frontier models over scRNA-seq at scale.
Agents design primers for qPCR and LAMP tests
Cooperating agents design primers for qPCR and LAMP diagnostics, splitting specificity checks, thermodynamics, and candidate selection across separate roles. Assay design joins the list of routine molecular work now getting the same agent treatment being applied upstream in discovery.
Reply with your discoveries. A human reads them. Forward freely.
|