Early progress on accelerating drug development with Claude

An unreviewed summary of a Science blog post reporting six drug-development partnerships and Anthropic's own early programmes, in which Claude was applied across the pipeline from molecule design to manufacturing paperwork and trial planning, with most of the work still preclinical.

Anthropic · 2026-10-27 · original

Not reviewed. This note reports what the post itself says. Unlike the other notes in this collection, its claims have not been checked against the underlying papers, reports or data. The work concerns biological design that could be misused, and in-depth analysis of it triggers the safety safeguards of the AI system used to prepare these notes, so it was left at the level of the post's own summary.

What the post reports

The argument is that there is no single bottleneck in drug development, so there is no single thing for AI to fix. A drug program spans target identification, molecule design, optimizing for manufacture, regulatory filings, trial design and administration, and the post says the obstacles are spread across all of it. Anthropic's response, it reports, has been to join real drug programs at several stages and disease areas: deep partnerships with six organizations (BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines and The Column Group), plus programs of its own.

The case studies the post gives, each with a linked technical report, cover very different stages. Generate Biomedicines worked on making a cancer cell therapy usable against solid tumors; six of its scientists used Claude Code to propose a shortened version of one component, and the post says one design, about 70% shorter than the standard, performed better in lab tests on three counts. Divergence Labs, pursuing a food-allergy treatment, had Claude turn its goals into a design plan and then generate roughly 35,000 candidate designs in about a month; early testing found three hits for one requirement and 20 for the other. BigHat Biosciences used Claude against its own lab data system to improve two candidate molecules, making and testing about 1,010 variants in about 10 weeks; partway through, the post says, the team suspected two standard measurements were misleading, and Claude and an Anthropic researcher reanalyzed the raw data and proposed different ones.

The Enveda case is about paperwork rather than molecules. Its manufacturing chemists spent up to 75% of their time on documentation; connecting those documents to Claude, over tests on 1,061 pages, caught all 11 issues human reviewers had found and flagged seven they had missed, cutting human review time by 89%, from over 25 hours to under three. Claude also drafted two regulatory documents, one in four hours against roughly three months by hand, with Enveda's review finding one minor error. The Column Group is establishing a company, Pleiades, built around running more programs than a small team normally could. Bristol Myers Squibb is using Claude in clinical trial planning, with a quoted executive saying a single prompt produced an initial analysis that would otherwise have taken days. Anthropic has also acquired Mithril Medicine, a one-person company working on rare childhood muscular dystrophies, whose founder took a candidate from idea to promising results in mice in under 18 months; the post says a trial in humans could begin as soon as late 2028.

Who did the work and who checked it

Anthropic publishes the post; the work is joint with the six named partners, each of which the post credits with its own technical report, and the BMS quotation is attributed to Michael Pourdehnad. The commercial relationships are stated in the post: these are partnerships, Mithril Medicine is an acquisition, and the post closes by soliciting further collaborators. No peer-review status is claimed for any of it. The post also mentions a Life Sciences Verification Program that gives trusted organizations more permissive safeguards, which it frames as necessary to prevent misuse while not impeding legitimate research.

Caveats the post itself gives

The post calls the research early and says it is mostly preclinical, for the most part not yet tested in humans, and frames the partnerships as a way to learn where its models fall short, noting that it has already prioritized specific model improvements from what it has learned. The hopeful outcomes are hedged: the programs may result in effective medicines if their positive trajectories continue, and a human trial date is given only as a possibility conditional on candidates holding up.

How it connects

References