Claude enters pharmaceutical laboratories, with measured results on antibodies and experimental treatments

Anthropic presents initial results for Claude in pharmaceutical research, from antibody design to preclinical trials.

An antibody component shortened by 70%, approximately 35,000 candidate designs generated in a month, and an 89% reduction in document review time. Anthropic has published early results from collaborations with six drug development organizations, alongside research programs it has begun operating internally. The company is seeking to expand Claude's role across the pharmaceutical development process, from initial biological hypotheses to regulatory documentation and clinical trial planning.

The projects cover cancer immunotherapy, food allergies, therapeutic antibodies, pharmaceutical manufacturing, and rare genetic disorders. Several findings have been tested experimentally, but most of the candidate treatments remain in preclinical development. None of the reported experiments demonstrates that a Claude-assisted treatment has already been proven safe and effective in humans. Shortening a genetic component used in cancer immunotherapy

At Generate Biomedicines, six scientists used Claude Code to redesign a component of CAR T-cell therapies.

These treatments involve modifying immune cells so they can recognize and destroy cancer cells. They have demonstrated value against certain blood cancers, but extending their effectiveness to solid tumors remains challenging.

One limitation involves how much genetic information can be introduced into the engineered cells. Shortening certain components could create room for additional instructions designed to improve their effectiveness against different cancers.

The researchers asked Claude to develop alternative designs for one part of the therapy's genetic instructions.

One resulting design was approximately 70% shorter than the standard component. In laboratory experiments, it allowed immune cells to produce roughly ten times as many receptors, the proteins used to recognize cancer cells.

The researchers also reported more effective tumor-cell killing and longer-lasting immune-cell activity.

These are laboratory findings. They do not yet establish that the modification will improve the clinical safety or effectiveness of CAR T-cell therapy in patients. Approximately 35,000 antibody designs for food allergy research

The collaboration with Divergence Labs explores a different therapeutic strategy, focused on food allergies.

Allergic reactions involve immunoglobulin E, or IgE. Some existing treatments neutralize circulating IgE in the blood. Divergence is investigating a more ambitious approach: physically detaching IgE already bound to immune cells involved in allergic reactions.

The intended drug would have two functional components. One would bind to IgE, while the other would help remove it from the cell.

Claude converted that biological objective into a detailed design plan and generated approximately 35,000 antibody candidates over one month, operating with substantial autonomy.

Following selection and experimental screening, the team identified three designs capable of binding to IgE and twenty capable of dislodging it.

Three of those twenty showed greater IgE-displacing activity than the reference medication in preliminary tests.

The researchers have not yet produced a complete treatment combining both functions. Their next step is to build and evaluate molecules that both bind and dislodge IgE without triggering the allergic reaction they are intended to prevent.

The work also revealed a limitation in Claude's scientific judgment. In one experiment, the model suggested multiple follow-up studies even though the existing evidence was already sufficient to guide the next decision. The human researchers instead chose to explore a different approach. More than 1,000 antibody variants tested at BigHat

At BigHat Biosciences, Claude was integrated into a research process combining computational design with automated laboratory testing.

The work involved two antibody fragments intended for potential future medicines. One targeted cancer cells, while the other was designed to keep a therapeutic molecule in the body for longer.

A viable therapeutic molecule must do more than bind to its intended target. It must also remain stable during manufacturing, storage, and administration.

Connected to BigHat's Reccy laboratory data system, Claude proposed molecular modifications that the scientists then produced and tested.

Approximately 1,010 variants were evaluated over ten weeks.

The best-performing versions exhibited more than tenfold improvements in binding affinity. The cancer-targeting fragment remained in solution at a temperature 18°C higher than the original design.

For the second fragment, the rate of molecular aggregation was reduced by more than a factor of one hundred.

The experiment also involved more than generating candidate sequences. When the team questioned the usefulness of conventional stability measurements, Claude and an Anthropic researcher reexamined the raw data, proposed alternative metrics, and recommended a modified experimental protocol. BigHat's scientists refined and implemented those tests.

These improvements concern properties relevant to drug development, not demonstrated clinical effectiveness. Enveda reduces 25 hours of document review to under three

Some of Anthropic's most immediately measurable results involve operational work rather than new molecules.

At Enveda, pharmaceutical manufacturing chemists must maintain extensive records to document quality controls, production conditions, and compliance with regulatory requirements.

The company reported that some chemists spent up to 75% of their working time on paperwork.

Enveda developed a system connecting its manufacturing documents to Claude, allowing the model to identify potential inconsistencies for human review.

Across 1,061 pages of documentation, Claude identified all eleven issues previously found by human reviewers, along with seven additional issues they had missed.

Human review time fell from more than 25 hours to less than three, representing an 89% reduction.

Claude also prepared first