Drug development is a multi-step process that extends from identifying disease-related biological targets and designing suitable molecules to optimizing manufacturing processes and planning clinical trials. In addition to scientific analyses, evaluating experimental data and preparing technical documentation represent a significant workload in this process. Anthropic has shared early findings from its collaborations with six organizations to explore how its artificial intelligence model, Claude, can be utilized across different stages of the drug discovery workflow.
The aim of these collaborations is to explore how artificial intelligence can help address scientific and operational bottlenecks throughout drug development, rather than being used solely to propose new molecules. These collaborations cover work on cancer, food allergies, and rare childhood diseases, as well as the optimization of drug candidates, the review of manufacturing documentation, and the planning of clinical trials. Anthropic is also conducting its own drug discovery programs targeting certain rare types of muscular dystrophy.
In these studies, Claude has been integrated into various research and development processes. Generate Biomedicines worked to shorten part of the genetic instructions used in CAR T-cell therapy, with the aim of expanding its potential application to solid tumors. One of the designs developed by six Generate scientists using Claude Code was approximately 70% shorter than the standard version; in laboratory tests, it was associated with immune cells producing nearly 10 times more receptors, killing tumor cells more effectively, and remaining active for longer periods. Divergence Labs is developing a new approach to treating food allergies by removing IgE from immune cells, rather than simply neutralizing circulating IgE in the blood. Claude generated approximately 35,000 antibody designs in about one month. Initial tests identified three designs that could bind to IgE and 20 that could dislodge IgE from immune cells. Three of the 20 designs appeared to dislodge IgE more effectively than the standard medication. The next step is to combine the two antibody arms into a single drug candidate.
BigHat Biosciences aimed to optimize the properties of two antibody fragments by combining Claude with automated laboratory data. In approximately 10 weeks, 1,010 variants were generated and tested; as a result of re-analyzing raw data and making changes to the measurement protocol, the most successful variants were reported to bind to their targets more than 10 times more strongly. The lead compound targeting cancer remained in solution at temperatures 18 °C higher, while the aggregation rate of the other lead compound slowed by more than 100-fold. Enveda used Claude for reviewing manufacturing documentation and preparing regulatory submissions. In a 1,061-page review, Claude identified all 11 issues detected by human reviewers, as well as seven additional issues that were missed. The time required for human review dropped from over 25 hours to under three hours, a reduction of 89%. In addition, first drafts of two regulatory documents were prepared; one was completed in four hours, whereas manual preparation was stated to take approximately three months. The Column Group (TCG) is founding a company named Pleiades, aiming to execute more and more complex drug development programs with Claude. The company’s first AI capability, which evaluates whether drug targets are worth pursuing, was tested on 45 target-disease pairs.
Bristol Myers Squibb (BMS) uses Claude to evaluate clinical trial plans for an early-stage drug and to estimate the probability of success. It is planned to expand the collaboration in 2027.
The shared early results demonstrated measurable changes across different application areas. One of the designs tested in the laboratory by Generate Biomedicines contained a genetic instruction sequence approximately 70 percent shorter compared to the standard version; this design was associated with immune cells producing nearly 10 times more receptors, killing tumor cells more effectively, and remaining active longer. Divergence Labs identified three designs capable of binding to IgE and 20 designs capable of removing IgE from immune cells among the 35,000 antibody designs created in about one month. In Enveda’s evaluation performed on a 1,061-page document, Claude identified all 11 issues detected in human reviews plus seven additional issues, reducing the time required for human review by 89 percent. In BigHat Biosciences’ study, the most successful molecule variants were also reported to bind to their targets more than 10 times more strongly or show marked improvements in other metrics.
These examples demonstrate that artificial intelligence can support different stages of drug development, from molecule design to experimental analysis, technical documentation to clinical trial planning. However, most of the studies are at the preclinical stage; the reported results do not prove that the developed approaches are safe and effective in humans. Anthropic plans to continue its efforts to refine Claude’s role in these processes.
Translated by: Elinsu Ak
Editor: Hazel Birgül
Reference: Anthropic. Early progress on accelerating drug development with Claude. 27 October 2026. https://www.anthropic.com/research/early-progress-drug-development