Beyond Molecule Design: How Claude Is Being Wired into Real Drug-Development Workflows with six organizations

By David Nakamura | AI & Data News — AIApplications

If your job is deciding where generative AI earns its keep right now, skip the model-architecture debate and hunt for insertion points. Anthropic’s latest life sciences update is useful in exactly that sense: Over the past several months, its scientists entered deep research partnerships with six organizations developing new medicines, and the company has also started drug discovery efforts of its own. The ground covered runs from target discovery and molecule design through optimization, manufacturing paperwork, filings, company creation and clinical trial planning — which makes the update a workflow map rather than a single demo.

Read as an applications brief, seven programs show where a model is placed inside an existing loop, what it is asked to produce, and what humans still verify. The partner set is Generate Biomedicines, Bristol Myers Squibb, Divergence Labs, Enveda, BigHat Biosciences and The Column Group. On the internal side are programs for rare childhood forms of muscular dystrophy, where Claude handles much of the scientific and operational work. Working inside real drug programs is how Anthropic learned how Claude can support drug development and what improvements are still needed to realize AI’s potential in medicine; it plans to scale this work up substantially from here.

Where the work stalls: five stages to instrument

Drug development offers many openings beyond the application people usually discuss — proposing initial designs for new molecules. Day to day the work is a chain of varied tasks:

• early research that identifies a target and the right type of drug to engage it
• designing a molecule with suitable properties
• optimizing reliable manufacture at scale
• drafting the filings required for clinical trials
• designing and running those trials

Anthropic’s strategy is to involve AI increasingly across these tasks, on the premise that there is no single problem to solve: bottlenecks are distributed throughout the process, and deep involvement in real programs across many disease areas and stages is how they get found and resolved. The approach serves Anthropic’s goals in three ways. First, it clarifies model strengths and limitations for a much wider range of tasks, and specific model improvements have already been prioritized. Second, the programs can act as case studies for the wider life sciences community — because if Claude is to significantly increase the rate of new cures, the work must involve many more biotechnology firms and integrate Claude more deeply into their development processes. Third, if their positive trajectories continue, the projects should have direct impact; Anthropic is optimistic that some may result in effective new medicines for food allergies, rare childhood diseases and several forms of cancer. Technical reports and preprints describe the thought process, methods and early results in more detail for each case below.

Seven deployment patterns, described as use cases

1. Generate Biomedicines — a shortened genetic instruction that beat the standard version

Use case: iterative design work inside a cell-therapy engineering loop. Generate Biomedicines is trying to make CAR T-cell therapy useful for solid tumor cancers, not only blood cell cancers. The first move was a targeted edit: the team shortened one part of the therapy’s genetic instructions, creating space for additional instructions suited to other kinds of cancer. Six Generate scientists then used Claude Code to propose new designs for that part. One design was about 70% shorter than the standard version, and in lab tests it was more effective: immune cells made roughly 10 times as many receptors, the proteins used to spot cancer cells, killed tumor cells more efficiently and remained active longer. Generate’s technical report gives more detail.

2. Divergence Labs — turning a stated goal into a candidate backlog at volume

Use case: high-volume candidate generation followed by human narrowing. Divergence Labs is exploring a new approach to food allergies caused by IgE. Standard medications only neutralize the IgE circulating in blood; Divergence wants drugs that detach IgE from immune cells entirely. That could be more effective, but requires a precise design with two arms: one that binds tightly to IgE, and another that helps prise IgE from the immune cell. Claude converted Divergence Labs’ goals into a detailed design plan and then, largely autonomously, generated roughly 35,000 antibody designs in about a month to be narrowed down and tested. In early tests the lab found “hits” for both roles: three designs that could bind to IgE, and 20 that could dislodge it. Three of those 20 appeared to dislodge IgE more than the standard medication. Its next step is to build a drug that combines the two arms. Divergence’s technical report gives more detail.

3. BigHat Biosciences — an AI loop plugged into the lab’s own data system

Use case: closed-loop property optimization. BigHat Biosciences combines AI with an automated lab to improve potential drugs called leads, which must bind their targets tightly, remain stable through manufacturing and storage, and avoid clumping together, among other criteria. Claude supported these three properties in two leads, both antibody fragments that could become parts of future drugs: one designed to target cancer cells, and one designed to keep a drug in the body longer. Connected to Reccy, BigHat’s lab data system, Claude proposed changes to each molecule, and BigHat’s lab made and tested about 1,010 variants in about 10 weeks. Partway through, the team suspected the standard heat-stability and clumping metrics were misleading. Claude and an Anthropic researcher reanalyzed the raw data, then proposed different metrics and a protocol change to test them. BigHat’s scientists refined and ran it, and the main predictions held. Compared with the originals, the best versions bound more than 10x more tightly, the cancer-targeting lead stayed in solution 18°C higher, and the lead for longer-lasting drugs clumped more than 100x more slowly. BigHat’s technical report gives more detail.

4. Enveda — start where the calendar is, not where the science is

Use case: document review and regulatory drafting. To conduct clinical trials, companies need records showing that each batch of a new drug has been made to strict quality standards, and to submit an application to regulators, among other requirements. At Enveda, manufacturing chemists spent up to 75% of their time on paperwork. The fix was an integration rather than a prompt: Enveda built a system that connects all of its manufacturing documents to Claude, then lets Claude flag any possible issues. Over tests on 1,061 pages of documentation, Claude caught all 11 issues that human reviewers had spotted — and flagged seven they’d missed. In the process, this method reduced human review time by 89%, from over 25 hours to less than three. Separately, Claude wrote the first drafts of two important regulatory documents. One took four hours, compared with about three months by hand. For the other, Enveda’s standard human review process found one minor error. Enveda’s technical report gives more detail.

5. The Column Group (TCG) — redesigning the operating model, not one task

Use case: company creation and overhead reduction. Drug discovery can be slow, expensive and prone to failure, so biotechnology investors have favored safer, previously validated ideas over riskier new ones, even when new ideas could help more patients. Anthropic has been working with TCG to see whether Claude can change this by cutting the time and cost of drug discovery so that investors can confidently support more ambitious ideas. Specifically, TCG is establishing Pleiades, a new company that will use Claude to run many more — and more complex — drug programs than is usual for a small team. That could be a model for how Claude reduces both the scientific and the operational and administrative overhead required to run a biotechnology company; if it works, the model could be replicated across the industry and direct investment toward harder problems in medicine. Pleiades is currently developing new tools and workflows for Claude, and its very first Claude skill, which helps Claude assess whether drug targets are worth pursuing, has been evaluated on 45 target-disease pairs.

6. Bristol Myers Squibb — analysis that took days, generated from one prompt

Use case: decision support inside clinical trial planning. More recently, Anthropic began working with Bristol Myers Squibb to find and address bottlenecks in clinical trial planning. BMS is using Claude to assess plans of an early-stage drug in clinical trials, and to estimate the likelihood of those trials succeeding. Together, the teams plan to expand this work in 2027 to understand its potential for increasing the probability of success in early development.

We use Claude at the clinical trial stage to analyze proprietary data, create a clinical study plan, and weigh the study’s chance of success. From a single prompt, Claude generated a robust initial analysis that helped inform our assessment of different development options, which would otherwise take days to complete. We did all of this while ensuring the security of our highly sensitive data. We look forward to expanding this work across our pipeline to support portfolio-level decisions.

Michael Pourdehnad, MD, Senior Vice President, Head of Early Clinical Development, Hematology, Oncology and Cell Therapy.

7. Mithril Medicine — one operator delegating contracts, logistics and lab coordination

Use case: a single person running science and operations together. In addition to the partner projects, Anthropic has begun to run some research programs itself. It acquired Mithril Medicine, a one-person company that has been developing treatments for two rare forms of muscular dystrophy. Its founder, Cody Tranbarger, used Claude to organize contracts, logistics and experiments with 11 external labs, and delegated many other aspects of operations to Claude. As a result, Mithril took a possible treatment for one form of muscular dystrophy from an idea to promising results in mice in less than 18 months. Now at Anthropic, Tranbarger is continuing this research and plans to start several new programs; if his drug candidates continue to hold up, Anthropic expects a drug to be ready to start clinical trials in humans as soon as late 2028.

Matching a bottleneck to an insertion point

Workflow bottleneck	Where the AI sits	Evidence from these programs	What to verify before scaling
Design iterations are limited by scientist hours	Design assistant inside the engineering loop, via Claude Code	Six Generate scientists proposed new variants; One design was about 70% shorter, with roughly 10 times as many receptors in lab tests	Wet-lab confirmation, not the proposal alone
Candidate generation is too slow to cover a hard design space	Goal-to-plan conversion, then autonomous bulk generation	roughly 35,000 antibody designs in about a month	Whether your narrowing and testing capacity can absorb the backlog
Optimizing several properties at once	Assistant connected to the lab data system (Reccy)	about 1,010 variants in about 10 weeks; best versions bound more than 10x more tightly, 18°C higher in solution, clumped more than 100x more slowly	Whether the standard metrics themselves are misleading
Manufacturing documentation consumes expert time	All documents connected; AI flags possible issues for human confirmation	1,061 pages of documentation: all 11 issues caught plus seven they’d missed; review time down 89%, from over 25 hours to less than three	Keep the standard human review gate in place
Regulatory drafting is a months-long manual job	First-draft generation for filings	two important regulatory documents; one in four hours vs about three months by hand; the other had one minor error in human review	Review turnaround, and who signs the draft
New programs cannot be launched cheaply enough to attract risk capital	Company-level skills and workflows (Pleiades)	first Claude skill evaluated on 45 target-disease pairs	Replication economics across the industry
Trial planning analysis takes days per option	Single-prompt analysis of proprietary data plus study-plan creation	BMS: assessment of an early-stage drug’s plans and chance of success; expansion planned in 2027	Security of highly sensitive data

A six-step playbook other teams can reuse

1. Inventory the chain, not the headline task. Write out the stages — target identification, design, optimization at scale, filings, trial design and execution — because bottlenecks are distributed rather than concentrated in the step everyone talks about.
2. Pick the stage where expert hours are already disappearing. At Enveda the answer was manufacturing chemists spending up to 75% of their time on paperwork; at BMS it was analysis that would otherwise take days.
3. Connect the model to the system that already holds the data. BigHat connected Claude to Reccy, its lab data system; Enveda connected all of its manufacturing documents. Integration is what turns a chat answer into a workflow step.
4. Ask for a plan first, then for volume. Claude converted Divergence Labs’ goals into a detailed design plan before generating roughly 35,000 antibody designs; at Generate, the plan preceded new designs for one part of the therapy’s genetic instructions.
5. Keep the human decision gate explicit. BigHat’s scientists refined and ran the changed protocol; Enveda’s standard human review process still found one minor error; hits still have to be narrowed down and tested.
6. Re-audit your metrics, then measure time and quality together. Partway through, the BigHat team suspected the standard stability and clumping metrics were misleading, and reanalysis produced different metrics. Track both the 89% review-time reduction and the 1,061 pages where all 11 issues were caught plus seven they’d missed.

Common misconceptions to design around

• “AI in drug development means molecule design.” The update is explicit that the work spans many tasks, and that the strategy is broader involvement across distributed bottlenecks.
• “There is one big problem to solve.” The stated premise is the opposite: bottlenecks are distributed throughout the process.
• “Faster drafts mean less review.” Enveda kept its human review layer, which caught all 11 issues on its own and where the model added seven they’d missed; one regulatory document still carried one minor error.
• “The metrics you trust must be correct.” In the BigHat program the standard heat-stability and clumping metrics were suspected mid-run of being misleading, and different metrics plus a protocol change had to be proposed.
• “These are already clinical wins.” All of this research is still at an early stage, mostly preclinical; for the most part the drugs are not yet at the stage where they are being tested in humans.
• “Safeguards only block work.” They are described as necessary to prevent misuse but able to impede legitimate biological research — which is why the Life Sciences Verification Program exists for trusted organizations to use its most powerful models with more permissive versions of its safeguards.

Action recommendations for application owners

1. Start with the paperwork and data-review loop before the novel-science loop: it is where the measured gain (from over 25 hours to less than three per review pass) is easiest to replicate in any regulated workflow.
2. Define what counts as a “hit” and who narrows the list, so bulk generation does not simply move the bottleneck downstream.
3. Instrument the loop with a real connection to your system of record, then plan for a mid-run metrics audit as part of the protocol, not as an exception.
4. Use the access routes now available to researchers: Claude Science, Anthropic’s AI workbench for researchers; Claude credits for life scientists; the Life Sciences Verification Program; and the recently launched open call for guest investigators in the life sciences, which seeks proposals from research groups that want to use Claude to investigate key questions in drug discovery and other areas.
5. Treat each program as a published case study — read the technical reports and preprints before copying the setup, and watch the 2027 expansion for how far the model moves into clinical-stage work.

Read the status line before you copy the numbers

The honest framing matters for anyone planning a rollout: this is early, mostly preclinical evidence gathered by doing the work, and Anthropic plans to keep extending Claude’s role into clinical trials as these drug programs advance and as it partners with more organizations whose programs are already at the clinical stage. For application teams, the transferable part is not any single result but the pattern — map the distributed bottlenecks, connect the model to the system that holds the data, let it plan and generate at volume, and keep humans on the review and metrics gates. That is the part you can adopt this quarter, whether or not you ever design a molecule. AIdataNewsGlobal will keep tracking how these workflows scale.