Early progress on accelerating drug development with Claude
NewsResearch Policy Commitments
[](https://www.anthropic.com/)
Company
Try Claude
[](https://www.anthropic.com/search)
Chapter 01: Case studies
Early progress on accelerating drug development with Claude
- 02 Extending our work in the life sciences
!Image 1: Early progress on accelerating drug development with Claude
- Published
Oct 27, 2026
- Team
_Anthropic’s life sciences team is working to accelerate curing disease. Over the past several months, our scientists have engaged in deep research partnerships with six organizations developing new medicines: BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines, and The Column Group. We've also begun drug discovery efforts of our own. In doing so, we've learned a lot about how Claude can support drug development, and how we’ll need to improve Claude further to realize AI's potential in medicine. We plan to scale this work up substantially from here_.
There are many opportunities to speed up drug development, even beyond the most commonly discussed applications of AI, like proposing initial designs for new molecules. Developing a drug involves many varied tasks: doing early research to identify a target for a drug and the right type of drug to engage it, designing a molecule with the right properties, optimizing it to be manufactured reliably at scale, drafting the regulatory filings needed to run clinical trials, designing and administering those trials, and much more.
Our strategy to realize AI’s medical benefits is based on having AI increasingly involved in all of these tasks. There is no single problem to solve; the bottlenecks are distributed all throughout the drug development process. The best way to identify and work through these bottlenecks is to get deeply involved in real drug programs, across many different disease areas and stages of development.
So, over the past several months, we’ve partnered with six organizations to do so. We’ve also started drug discovery efforts of our own, including programs for rare childhood forms of muscular dystrophy, where Claude handles much of the scientific and operational work.
This work serves our goals in three ways. First, it lets us better understand the strengths and limitations of our models in relation to a much wider range of drug development tasks (we’ve already prioritized specific improvements to our models based on what we’ve learned so far). Second, these programs can serve as case studies for the wider life sciences community. If Claude is to significantly increase the rate of new cures, we need to work with many more biotechnology firms and integrate Claude more deeply into their development processes. Third, these projects should, if their positive trajectories continue, have direct impact themselves. We’re optimistic that some may result in effective new medicines for the diseases they target, including food allergies, rare childhood diseases, and several forms of cancer.
In this post, we provide an update on our progress with our partners, along with technical reports and preprints that describe our thought process, methods, and early results in more detail. We then briefly outline where our work is headed next.
Case studies
**Generate Biomedicines: modality engineering.** Generate Biomedicines has been working on increasing the usefulness of a cancer treatment known as CAR T-cell therapy. This therapy has historically helped treat blood cell cancers; Generate aims to reengineer it so that it can treat solid tumor cancers too. To do this, they’re aiming to shorten one part of the therapy’s genetic instructions, which would create room for additional instructions suited to other kinds of cancer. Six Generate scientists used Claude Code to propose new designs for this part. One design was about 70% shorter than the standard version, and in lab tests it was more effective: it allowed immune cells to make roughly 10 times as many receptors (the proteins they use to spot cancer cells), it killed tumor cells more efficiently, and the immune cells remained active for longer. Read more, and see Generate’s technical report.
**Divergence Labs: hit generation.** Divergence Labs is working on a novel way to treat food allergies. Food allergies are set off by a reaction from a molecule called IgE, some of which is attached to our immune cells and some of which floats freely in our blood. Divergence Labs is developing drugs that would detach the IgE from our immune cells entirely, rather than solely neutralizing the IgE circulating in our blood, as standard medications do. This could provide a much more effective treatment, but it requires a precise design, with two “arms”: one that would bind tightly to IgE, and another that would help 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 those early tests, Divergence found “hits” for both arms: 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. Read more, and see Divergence’s technical report.
**BigHat Biosciences: lead optimization.** BigHat Biosciences pairs AI with an automated lab to improve potential drugs (or “leads”), which must bind their target tightly, remain stable through manufacturing and storage, and avoid clumping together (among other criteria). BigHat’s scientists and an Anthropic researcher worked with Claude to improve 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. Claude, connected to BigHat’s lab data system, Reccy, 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 the 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. Read more, and see BigHat’s technical report.
**Enveda: manufacturing and filings.** 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. To ease the load, 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 to about three months by hand. For the other, Enveda’s standard human review process found one minor error. Read more, and see Enveda’s technical report.
**The Column Group (TCG): company creation.** Because drug discovery can be slow, expensive, and prone to failure, investors in biotechnology companies have favored safer, previously validated ideas over riskier new ones, even when the latter could help more patients. We’ve 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. This could be a model for how Claude can reduce both the scientific and the operational and administrative overhead required to run a biotechnology company. If this works, this 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; its very first Claude skill, which helps Claude assess whether drug targets are worth pursuing, has been evaluated on 45 target-disease pairs. Read more.
More recently, we’ve also begun working with Bristol Myers Squibb (BMS) 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, we 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._
In addition to our work with these partners, we’ve begun to run some research programs ourselves. We’ve 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 Mithril’s 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, Cody is continuing this research, and plans to start several new programs. If his drug candidates continue to hold up, we expect a drug to be ready to start clinical trials in humans as soon as late 2028. Read more and see our technical report.
Extending our work in the life sciences
All of this research is still at an early stage. In particular, so far it is mostly _preclinical—_ that is, for the most part, it isn’t yet at the stage where the drugs are being tested in humans. We plan to continue extending Claude’s role into clinical trials as these drug programs advance, and as we partner with more organizations whose drug development programs are already at the clinical stage.
This research is only one part of our efforts to accelerate drug discovery. Ultimately, we want to share our models with as many drug development researchers as possible. To that end, we’ve rolled out Claude Science, our AI workbench for researchers, and we’re providing Claude credits for life scientists. We’ve also created the Life Sciences Verification Program to allow trusted organizations to use our most powerful models with more permissive versions of our safeguards, which are necessary to prevent misuse but which can impede legitimate biological research. We’ve also recently launched an open call for guest investigators in the life sciences; we’re seeking proposals from research groups who want to use Claude to investigate key questions in drug discovery and other areas.
We’re looking forward to working with many more organizations as we expand our life sciences research. If you work at an organization focused on drug discovery and are interested in collaborating with us, you can get in touch below.
Get in touch with us
Please select
[](https://www.anthropic.com/)
Research
Policy
Commitments
Company
Products
Models
Solutions
Help and security
Terms and policies
Our mission
Anthropic is a Public Benefit Corporation. Our mission is to ensure the world safely makes the transition through transformative AI.
Log in
© 2026 Anthropic PBC
- [](https://www.youtube.com/@anthropic-ai)
- [](https://x.com/AnthropicAI)
- [](https://www.linkedin.com/company/anthropicresearch)
Generate Biomedicines
Generate:Biomedicines: Redesigning a component of CAR T-cell therapy
_The Generate:Biomedicines team shares how they worked with Claude to redesign a component of CAR T-cell therapy to make it more compact, freeing up space for new functions that could help treat cancer._
_Read Generate’s technical report and get in touch (info@generatebiomedicines.com)._
Generate:Biomedicines designs protein drug candidates using machine learning and frontier biohardware. Over the past eight months, a six-person team of Generate scientists redesigned a stubborn component of chimeric antigen receptor (CAR) T-cell therapy, a type of cancer therapy, with help from Claude.
CAR T-cell therapy uses genetically engineered versions of patients’ own immune cells to attack cancer cells. The part the Generate scientists redesigned is called the signaling domain—the section of the engineered receptor that tells the immune cell to attack. Making it shorter frees up space for other functions in the CAR’s genetic instructions, such as attacking several cancer markers at once or adding proteins that help the T-cell survive inside a tumor.
Working with Claude and other AI tools, Generate’s scientists built and directed a system of AI agents to design signaling domains that are much more compact than a standard version used in several approved therapies. The best design was about 70% shorter and no less effective. In laboratory tests, one of the top designs put roughly 10 times as much receptor on the cell surface—which may make a T-cell better able to recognize and attack cancer cells—and the lead designs killed tumor cells more efficiently than the standard version.
- _Read Generate:Biomedicine’s technical report, “Refactored intracellular domains enhance CAR expression and function.”_
Why improving CAR T-cell therapy is a challenge
CAR T-cell therapy has transformed the treatment of several blood cancers, but the drugs themselves are cumbersome to manufacture. Today, a patient’s T-cells are removed, engineered in a factory over several weeks, and infused back into the patient. A major goal for the field is to do that work _inside_ the body: delivering the genetic instructions for building the receptor directly to the patient’s T-cells. This could be a faster and cheaper way to deliver the therapy and could expand what CAR T-cell therapy can do. In particular, it could make the therapy better suited to treating autoimmune conditions, as well as early-stage solid tumor cancers, against which it has so far been less successful.
But delivering genetic instructions into a living cell means working within tight constraints. Space is at a premium: the instructions are typically delivered either by a viral vector or a lipid nanoparticle carrying mRNA, both of which can only hold a limited amount of genetic material. It would be helpful to give the CAR more functions, or to bundle in protective proteins to help the T-cell survive inside a tumor. The more compact each part of the receptor is, the more room is left in the delivery vehicle for these extra functions.
One way of boosting a CAR’s functionality would be to improve the delivery system—for example, by designing a capsule that can carry more genetic instructions. To use a software analogy, it would be like making a larger flash drive that can carry more information. But another way would be to make the individual parts of the CAR shorter—like writing software more efficiently so more programs fit on a given drive. Just as a programmer refactors a codebase to make it more compact and efficient without changing what it does, Generate wanted to refactor the CAR to make it shorter but no less effective.
The team had already redesigned the part of the receptor that recognizes a marker on the surface of the cancer cell, but the signaling domain was a harder problem. Unlike the part that recognizes the marker, the signaling domain has no fixed three-dimensional shape to work from, and the way it does its job is not completely known. That made it a compelling challenge for Generate’s scientists: could they create and direct a system of reasoning agents to help redesign a protein whose structure and function were only partly understood? Would those agents be able to help the team explore novel designs that push the field’s understanding forward?
Building the agentic system
In February 2026, using Claude Code, the team built a system of AI agents running Claude and other models. Generate’s scientists directed the agents’ work through this system, giving them experimental lab results, prompting them based on specific hypotheses around CAR biology, providing feedback, and creatively exploring the design space together.
The agents then generated a set of proposals. Each presented a specific design for a shorter signaling domain, with the agents’ reasoning about what to cut, what to keep, and why the redesigned domain should still work effectively. The agents’ reasoning drew from published research, the scientists’ own insights, and Generate’s experimental data. Through rounds of review and steering, the team selected 186 designs for the first round of lab testing.
To test the designs, Generate’s scientists wrote the genetic instructions for each redesigned receptor, introduced those instructions into cells in the lab, and observed how the cells behaved. The first round of scientific hypothesis generation and experimentation took two weeks, from proposal selection to initial results.
Based on these experiments, the team identified several shortened signaling domains that worked better than the standard design and were less likely to fire when the target marker is absent, which can wear T-cells out over time. Generate’s scientists fed those results back into the system to direct the search toward a few strong leads.
What the team found
One of the strongest candidates, designed by Generate’s scientists working with Claude, had a signaling domain about 70% shorter than the standard version. When tested in human T-cells in the lab, it led to roughly 10 times as many copies of the receptor on the cells’ surface. Receptors are how a T-cell recognizes its target, so producing more receptors per molecule of drug delivered may improve the T-cell’s ability to find its target and work as intended.
The lead candidates killed tumor cells more efficiently than the standard version, and they released more of a key chemical signal that T-cells use to rally the rest of the immune system. Generate believes both improvements were driven in part by the higher number of receptors on each cell—which also means less mRNA is needed to make a working CAR-T cell—and in part by increased potency.
That surplus of receptors could also make new kinds of CAR T-cell therapy possible. For instance, instead of carrying many copies of a single receptor, a T-cell might carry several different types of receptor, each aimed at a different marker, with enough of each type to remain effective. That could be useful against solid tumors, in which different cells don’t always share the same markers. Separately, because receptors gradually wear away, having more of them could help each cell remain effective longer.
These are early laboratory results—Generate has not yet tested that the designs are safe or effective in patients. Read more about the methods and findings in Generate’s preprint.
Rethinking how the team works with Claude
Over the course of the project, the way the team worked with Claude changed too. At first, one of Generate’s wet lab biologists used Claude Code to build software tools for colleagues who didn’t code. Among these tools was a customized electronic lab notebook system—essentially, a digital record of experiments—that better served the team’s needs than existing off-the-shelf options. Within a few weeks, the rest of the biologists were building tools with Claude Code themselves.
A few months later, the whole team was working directly in Claude Code—it had become a “homepage” of sorts for the team, and was connected to Generate’s data, tools, and custom applications. They’d become accustomed to working with the agents to retrieve and compare experimental data, offer hypotheses about what might have gone wrong in a failed experiment, and suggest experiments to try next. The agents had become part of the team.
What’s next
The work is still at an early stage, and the next step is to test how the designs perform in more complex preclinical models, such as mice. Generate and Anthropic are now working together on the target and indication strategy for applying the redesigned CAR to solid tumors. They plan to work with Claude to reason across broad datasets—both public ones and Generate’s own—to select the optimal combination of tumor markers to target to effectively kill cancer cells. They hope this work will result in a lead drug candidate for solid-tumor cancer in the near future.
#### Acknowledgements
Generate would like to thank the scientists who contributed to this work, including Liam Bai, Hetvi Desai, Cloris Li, Aakriti Kumar, Wujie Wang, and Jason Yi.
#### Disclosure
As part of a research collaboration, Anthropic provides Generate:Biomedicines with Claude usage credits and engineering support.
Divergence Labs
Divergence Labs: Speeding up antibody design for food allergies
_The team at Divergence Labs describes how they worked with Claude to design early antibody candidates for a new food allergy drug._
_Read Divergence’s technical report and get in touch (careers@divergencelabs.com)._
Divergence Labs is a biotechnology company founded in 2025 by an interdisciplinary team of drug hunters, computer scientists, biologists, and hardware engineers. It recently identified an opportunity to engineer better medicines for food allergies, which affect around 33 million people in the US alone.
Food allergies are caused by an abnormal immune response in which the body mistakenly identifies proteins in foods such as peanuts or tree nuts as harmful. This has the potential to cause life-threatening immune reactions. In 2024, Xolair (omalizumab) became the first FDA-approved medication to reduce allergic reactions to multiple foods. Although this is a major advance, over half of patients on Xolair still cannot consume more than a 2,000 mg dose at a time, of two or more foods they’re allergic to (roughly eight peanuts). Xolair reduces the risk of a reaction after accidental exposure, but patients must still avoid their trigger foods and carry an EpiPen in case of a larger exposure. These limitations are related to the molecular design of Xolair, opening up an opportunity to design a better medicine from scratch.
Below, Divergence shares how its researchers worked with Claude to design early antibody candidates that could serve as the building blocks for a new, more effective type of food allergy drug. (Read the technical report for more details.) Claude did much of the work autonomously, and the team had lab results within weeks, compared with up to nine months in previous projects. The results suggest that AI could help researchers design more complex drugs and free up resources to work on more of them at once.
- _Read Divergence Labs’ technical report, “Towards a best-in-class IgE disruptor for food allergy with Claude-enabled antibody design.”_
Going beyond existing therapeutic mechanisms
Xolair is an antibody that neutralizes immunoglobulin E (IgE), the molecule that sets off the immune system’s reaction to food and environmental allergens. There are two predominant forms of IgE: free and bound. Free IgE circulates in the blood; bound IgE is already attached to immune cells. Xolair was designed to neutralize free IgE (i.e., it is an “inhibitor”). It can also dislodge bound IgE, but in patients this only occurs at concentrations far higher than one could practically receive by injection. Yet bound IgE is the form that sets off the allergic reaction: when a person comes into contact with an allergen, the IgE molecules on the surface of an immune cell cluster together, signaling the cell to respond. A more effective medicine would therefore need to both neutralize free IgE and disrupt bound IgE, making it both an inhibitor and a “disruptor.”
Disrupting bound IgE is unusually difficult, however. A drug has to attach at exactly the right spot to knock an IgE molecule off the immune cell it’s anchored to. What’s more, traditional antibodies have two arms, so a single antibody can grab two bound IgE molecules at once and pull them together. This causes the same clustering that sets off an allergic reaction—meaning an antibody-based drug could trigger the exact immune response it’s meant to prevent. Removing bound IgE safely and effectively therefore requires precise control over a drug’s shape and how it binds, which is hard to achieve with conventional methods of antibody discovery.
**Two ways an antibody can work.**Inhibitors (left) stop free IgE (orange) from attaching to its receptor on an immune cell, either by occupying the receptor or by latching on to the IgE itself. Disruptors (right) dislodge IgE that has already attached.
Finding a drug candidate requires improving many properties at once, including how well it works, how stable it is, and how easily it can be manufactured. Claude turned the research team’s goal of disrupting bound IgE into a detailed design plan consisting of discrete phases, then carried out most of that plan autonomously, pausing after each phase so Divergence’s researchers could review the results and decide on next steps.
Claude started by analyzing every known antibody that targets IgE, and pinpointed the structural features that allow some of them to dislodge bound IgE. Divergence’s research team chose a design with two different arms: an “anchor arm” to grip the IgE molecule, and a “disruptor arm” to break its hold on the immune cell. Because both arms attach to the same IgE molecule, this design is also intended to avoid the clustering problem described above. Each arm would be designed separately, then joined into a single antibody. Claude helped the team identify where on the IgE molecule each arm should attach, in order to accurately combine the two.
Claude then split the design work into two projects, one for each arm. For the disruptor arm, it set out to improve on a promising design from an earlier study. For the anchor arm, Claude had to create a design from scratch. The new anchor arm design also had to be compatible with the disruptor arm design, so that both could attach to the same IgE molecule at once.
For both projects, Divergence provided Claude with in-house molecular design tools, along with existing protein modeling methods that the team had adapted for this problem with Claude’s support. One adapted method was particularly successful at increasing the functional activity of the disruptor arm. Divergence’s design model normally builds a protein from scratch, but the modified version could start from an existing one. This made it possible to make small, targeted changes to the earlier design and improve it quickly.
**How Claude designed the two antibody arms.**Claude planned the approach (left), then ran separate projects to improve the disruptor arm (center) and design the anchor arm (right), nominating the best designs from each project for experimental validation. Green outlines mark the stages where Claude worked with Divergence’s researchers. Blue outlines mark the steps it carried out mostly autonomously. BLI = biolayer interferometry.
In total, Claude generated about 35,000 designs over 22 separate stretches of autonomous work. This took around 2,750 hours of computing time spread across many Nvidia H100 GPUs. Claude then used biomolecular structure prediction models to predict how well each design would attach to IgE, and kept only the highest scoring ones—no more than 96 per arm—to make and test in the lab. Going from the initial idea to lab results took roughly three weeks for the disruptor arm and four and a half weeks for the anchor arm. Previously, comparable work had taken Divergence up to nine months.
**How Claude spent its effort.**The share of Claude’s reasoning that went into each type of task during the project. Running the design work and analyzing the results took up the largest share, about 42%.
Experimental results
The Divergence team then tested the disruptor arm designs in the lab. They used a technique called biolayer interferometry (BLI), which uses light to track molecules attaching to and detaching from one another in real time. This allowed them to see whether each design could knock IgE off the receptor that binds it to an immune cell.
In the first round of testing, 20 of the 91 designs worked as intended, measurably dislodging bound IgE within one to two minutes. In further testing, three of these appeared to outperform Xolair at concentrations patients could realistically receive. The 20 working designs were also quite different from one another, differing at as many as 8 positions in their protein sequences from the reference. This has given the team a varied set of starting points to refine into potential therapies.
**How the top three disruptor designs compare with Xolair.**Each line shows the share of IgE still attached to its receptor in the 10 minutes (600 seconds) after a drug was added, at three different concentrations. A lower line means more IgE was dislodged. By the end of the test, three designs (blue) had dislodged more IgE than Xolair (omalizumab, orange). The gray line is an inactive control. nM = nanomolar concentration; s = seconds.
The team also tested 57 of the anchor arm designs, and confirmed that three of them (around 5%) attached to IgE. The best of the three gripped it very tightly, with a binding affinity comparable to many approved antibody drugs.
**The three anchor designs that attached to IgE.**Top: computer models showing where each design (red) attaches to the IgE molecule. Bottom: BLI of each design attaching to IgE and then, after the dotted line, detaching from it. The drug-bound BLI probe measured attachment to IgE (560nM) during association. The KD value reflects how tightly each design grips IgE; a lower number means a tighter grip. Design 1 gripped the most tightly and barely detached during the test.
What’s next for the team
Divergence set out to design a drug that does what Xolair can’t: dislodge the bound IgE that sets off an allergic reaction, at doses patients could realistically receive. Working with Claude, the team now has early versions of the two parts such a drug would need. In lab tests, the best disruptor arms appeared to dislodge bound IgE more effectively than Xolair, and the best anchor arms gripped IgE tightly.
Some of the problems Claude solved were more conceptual. For example, computer models can’t directly measure whether an antibody will knock IgE off its receptor, so Claude chose a stand-in these models could measure: the shape of the IgE molecule, which is “open” when attached to a receptor and “closed” when a known disruptor has attached to it. After confirming that the models reproduced both shapes correctly, Claude discarded any design that didn’t hold IgE closed. It then analyzed the results to inform later phases of the project. These are the types of time-consuming tasks that limit the pace of protein engineering today. If researchers can confidently hand this work to Claude, they can begin to pursue several projects in parallel.
There’s still room for improvement. When given a design objective, Claude was good at setting up a round of design work, including deciding what to measure and what to filter for. In certain cases, however, the team found that Claude was less pragmatic and contextually grounded than a human operator in deciding between further investigating a result and moving on. In one phase, for example, the goal was to explore if a particular strategy could redesign one of the loops the disruptor arm uses to grip IgE while preserving the design’s function. The first set of results was definitive, but Claude proposed five follow-up design runs, each with slightly different settings that were unlikely to change the key conclusion. In contrast, the human operator suggested a focus on a previously unexplored, alternative strategy. Choosing the experiments that are most worth running remains a judgment call that researchers, at the moment, seem to be better positioned to make.
The results described here are a long way from a new medicine. Finding promising initial designs, known as _hits_, is only the first of many steps in drug development. Additional hit optimization work (e.g., avidity-based constraints, developability) is still required.
As a next step, the Divergence team plans to use computer modeling to determine which anchor and disruptor arms are most compatible, then build and test complete two-armed antibodies. The most promising of these two-armed combinations will be tested on the immune cells involved in allergic reactions, to show whether they can remove bound IgE without triggering the immune response they’re meant to prevent.
Divergence expects these methods to be useful well beyond food allergies. Indeed, they might be applicable to any condition that requires a drug to do something more complex than attach to its target. But for now, the team’s focus remains on the 33 million people in the US with food allergies, and on creating a treatment that could one day mean they no longer need to avoid trigger foods or carry an EpiPen.
#### Acknowledgments
Divergence would like to thank its entire team of determined enthusiasts, advisors, and consultants for advising and leading this work. Please reference our technical report for additional context on the campaign and contributions.
#### Disclosure
As part of a research collaboration, Anthropic provides Divergence with Claude usage credits and engineering support and reimburses certain third-party computing costs.
BigHat Biosciences
BigHat Biosciences: Optimizing antibody drug candidates with Claude
_BigHat Biosciences shares how an Anthropic researcher working with Claude improved two antibodies on binding, stability, and aggregation in about 10 weeks—and improved on how those properties were being measured._
_Read BigHat’s technical report and get in touch (partnering@bighatbio.com)._
BigHat Biosciences is a clinical-stage biotechnology company that designs and develops next-generation protein therapeutics: antibody-based drugs engineered to function in ways the body’s own antibodies can’t. BigHat pairs AI models with an automated lab: the models propose new molecules, and the lab makes and tests thousands of them a week. The results of every round of experiments are fed back to the models, so that each new set of designs is informed by the last.
Designing an antibody-based drug is a balancing act. An antibody works by binding to a specific target in the body, such as a protein on the surface of a cancer cell. But the drug also has to be sturdy enough to keep its shape through manufacturing, shipping, and storage, which scientists predict by measuring the drug’s “thermostability,” the temperature to which it can be heated before it unfolds. And the drug’s molecules have to avoid sticking to one another and forming clumps, a process known as aggregation. Changes that help with one of these properties often make another worse. For example, reshaping the part of the antibody that grips the target can improve binding, but it can also loosen the structure that keeps the molecule folded, reducing its thermostability.
Managing these trade-offs is the work of an antibody engineer. In a process called lead optimization, the engineer takes an antibody that already binds to its target and refines it to make a better drug. BigHat and Anthropic wanted to find out whether a scientist working closely with Claude could do this work. Could an experienced researcher, paired with Claude, improve an antibody’s binding, thermostability, and resistance to clumping, all at once?
To find out, the BigHat team chose two existing antibodies as starting points. For each one, a researcher at Anthropic worked with Claude to design a batch of 176 slightly altered versions, known as variants. BigHat’s lab then made and tested each variant, and the researcher and Claude used those results to design the next batch. After three rounds of this process for one antibody and four for the other, each had yielded a variant that outperformed the original on all three properties: it gripped its target more tightly, held together at higher temperatures, and resisted clumping better.
Claude’s contribution to this project went beyond designing antibodies. Partway through, the lab’s scores for stability and clumping proved erratic, and BigHat’s quality checks flagged many of them as unreliable. The researcher suspected the problem was in how the scores were being calculated. He and Claude went back to the instrument’s raw readings, where Claude worked out metrics better suited to these molecules and proposed a change to the lab’s protocol to check them. Once BigHat’s results bore them out, the teams used the revised scores to choose their final molecules. These results suggest that AI can speed up the search for answers, and also help scientists ask better questions.
- _Read BigHat Bioscience’s pre-publication research paper, “Reasoning in the loop: mechanism and measurement in antibody engineering.”_
Two molecules
For its pilot projects, BigHat chose two starting molecules, each a small antibody fragment that could become a building block in a future drug. The first, an anti-albumin VHH, is one of the smallest antibody fragments found in nature. It binds to albumin, the most abundant protein in human blood; attaching an albumin-binding piece to a drug is a common way of keeping that drug in the body for longer. The second, an anti-mesothelin scFv, is a stripped-down, lab-made antibody that targets mesothelin, a protein that is rare in healthy adults but common on the surface of many cancer cells. An antibody that sticks to mesothelin can deliver a cancer drug to a tumor or draw the body’s immune cells to the tumor to attack it.
BigHat set the same goals for each pilot. The new molecule had to reach a target level of binding about 10 times tighter than the original molecule’s, survive being heated to between 65 and 70°C without unfolding or clumping, and still be producible in BigHat’s lab. A molecule that balances binding, stability, and clumping well enough to be manufactured as a drug is said to be “developable.”
The teams measured stability and clumping with a single heating test, in which an instrument heats each sample and records two signals: the color of the light given off by the protein itself, which shifts as the molecule unfolds, and the light scattered by the sample, which increases as molecules clump together. The instrument’s software then reduces each test to two numbers: an estimated melting temperature (the temperature at which half of the molecules have unfolded), taken from the first signal, and the temperature at which it starts to clump, taken from the second. The teams relied on these two numbers to decide which variants to move into the next round.
A different way of working
In many labs, improvements to a molecule are found largely by trial and error. One method is to try each of the 20 amino acids—the building blocks of proteins—at a single position in the antibody and see which version scores best across the three properties. Another method, known as directed evolution, is to make random changes to the molecule, keep the versions that perform best, combine the helpful changes from different versions into new ones, and repeat.
More recently, teams have trained specialized AI models to help with this. Some propose new amino acid sequences, and others predict how a sequence will score. Some labs now connect these models directly to their equipment. But in these setups, researchers decide at the outset which tests to run and what scores qualify as success. The model’s job is to find molecules that score well, not to question whether the tests are measuring the right thing or to suggest running them in a different way.
These pilots worked differently. BigHat runs its lab through a custom-built system called Reccy, which keeps a centralized record of every molecule, physical sample, instrument run, and result. Reccy has a built-in connector for AI agents, which Anthropic used to link it to Claude Code. This gave Claude instant access to any of the data it might need during the experiments.
A researcher at Anthropic worked with Claude to design each round of variants. Claude proposed the changes itself, without relying on specialized tools. It chose them by reasoning about how each change would affect the molecule’s structure and physical behavior, and by studying the lab’s results from earlier rounds. When it needed a deeper analysis, it built simple statistical models from the results. The researcher asked Claude to explain its reasoning for each proposed change, and the two of them decided which variants to send to the lab. BigHat’s lab then made and tested the molecules, and the results were recorded in Reccy, where Claude could read them.
Improving an antibody is a highly iterative, time-consuming process. Antibodies are usually made by putting DNA instructions into living cells and growing them until they have produced enough protein, which adds roughly a week to every round. To eliminate this bottleneck, BigHat’s lab uses an approach called cell-free protein synthesis. It takes the protein-making machinery out of cells and supplies it directly with the DNA instructions for an antibody, so there is no waiting for cells to grow and no need to break them open to recover the protein. This lets the lab make thousands of antibodies a week and measure several properties for each. Because the lab runs dozens of programs in parallel, these pilots used only a small part of that capacity.
In each round, the lab made 176 variants and measured how tightly each one bound its target, how it behaved when heated, and other properties, including its purity and how much was produced. Results typically came back in nine to 10 days: seven to produce the DNA for each variant, and the remainder to make and test the antibodies. Over about 10 weeks and seven rounds, the two pilots tested about 1,010 unique variants across 1,232 samples.
Claude quickly recognized that both starting molecules closely resembled antibodies already described in published patents. Comparing them with the published antibodies allowed Claude to suggest changes that were likely to improve binding. For the scFv, the cancer-targeting molecule, Claude identified a single amino acid change that it expected would tighten binding, a hypothesis which was confirmed in the lab. But binding was only one of the goals: Claude still had to improve stability and reduce clumping by changing other parts of each molecule without losing the gain in binding.
Working with Claude also allowed the teams to explore bolder designs. Antibody engineers typically change only one or two amino acids at a time, because the combined effect of many changes is hard to predict. Here, the teams were able to go much further, because Claude proposed precise changes and reasoned about which to combine and which to test separately. About 150 of the variants they tested differed from the starting molecule by 10 to 13 amino acids. In one round of the pilot for the VHH, the albumin-binding molecule, 65 variants had 10 or more changes, and every one of them still bound its target.
The teams also tested more than 200 variants with just one change each, to learn which positions in the molecule were most important for binding and which could be changed freely. And they ran small controlled experiments to test specific ideas about the best candidates. The two final molecules differ from their starting points by eight (for the VHH) and six (for the scFv) amino acids.
Three rounds of design for the cancer-targeting scFv. Each dot is one variant, plotted by how tightly it bound its target (further right is tighter) and the temperature at which it started to clump (higher is better). The star marks the starting molecule, and each stepped line traces the best combinations of the two properties reached in one round. By the third round (DR2), five weeks in, several variants had reached the shaded region, where they met BigHat’s goals for both properties.
From antibody design to general scientific reasoning
Stability and clumping proved much harder to improve than binding. By the second round of results, Claude could predict binding reasonably well from one round to the next. It struggled to do the same for stability and clumping, especially on the VHH, where the scores didn’t seem to follow any pattern.
BigHat's scientists had already cautioned that these scores needed care, and the Anthropic researcher running the pilots suspected the same thing: that the problem lay not in the molecules but in how the instrument’s readings were being turned into scores. Because Reccy stores each instrument’s raw readings alongside the summary numbers, he and Claude could go back to the raw data and investigate.
The raw data for each sample consists of two curves, one for each signal, which show how the signal changes as the sample is heated. When Claude examined them, it found that the two could not be read separately: unfolding and clumping affect each other, so the curves have to be analyzed together. Read one at a time, the two summary numbers were misleading.
The scFv’s variants began to clump as soon as they began to unfold, and the clumps fell out of solution before unfolding was complete. The number being used as the “melting temperature” was really recording that collapse of the solution.
The numbers for the VHH variants were misleading in a different way. Most VHH variants began clumping at around 40°C, long before they unfolded. This could distort the unfolding signal, and some reported melting temperatures were more than 10°C too low. BigHat’s software had flagged more than half of these samples as unreliable, which was correct, but that left most of the molecules without a usable number. And because those variants all started clumping at about the same temperature, the clumping score could not tell good variants from bad ones. The relevant measure was how quickly they clumped. Telling the two processes apart matters, because a molecule that unfolds too easily is usually repaired by strengthening its interior, whereas one that clumps is usually repaired by changing its surface. A score that confuses the two will direct the designer to the wrong part of the molecule.
Claude proposed different scores that better matched how each molecule actually behaved. For the VHH, it proposed measuring clumping by how fast it happened, not by the temperature at which it started. For the scFv, it proposed measuring the temperature at which the sample collapsed (unfolding and clumping together), adjusted for how concentrated the sample was. To arrive at these recommendations, Claude ran an autonomous research session. It tested many possible explanations of how the molecules behaved as they heated up, generated the curve each explanation would produce, and kept those that best matched the real data.
To check that these scores reflected real properties of the molecules, the researcher and Claude proposed a change to the lab’s testing protocol. The standard protocol tests each sample under one heating rate and one concentration. Instead, the teams tested one set of samples at two concentrations and another at two heating rates. A true melting temperature would stay the same under all of these conditions, but a number that was in fact tracking clumping would change, because clumping depends on how concentrated the sample is and how long it spends at each temperature. Because BigHat’s lab is automated, the new conditions were incorporated into the very next round of tests.
Before the results came in, Claude predicted which numbers would hold steady and which would change, and its main predictions proved correct. The VHH’s melting temperature, read from the full curve, barely changed, which was indicative of a true melting temperature. The temperature at which the scFv collapsed shifted by two to five degrees, as would be expected if the instrument were really detecting clumping.
The revised scores made a practical difference. On the VHH, the instrument's clumping score could not separate molecules. No design in four rounds received a passing value that also cleared quality checks, including the final molecule, whose scattering stayed near baseline. With the revised scores, the teams could better distinguish design variants, and the final VHH clumped far more slowly than the one they started with.
Results and next steps
Each pilot produced a candidate that improved on its starting molecule in all three properties. Both reached the level of binding that BigHat had set as its goal. Compared with their starting molecules, they bound six to nine times more tightly, held together at temperatures 8 to 18°C higher, and resisted clumping better. Both could still be made in pure form and in sufficient quantity. These two were chosen for their balance, from a wider set of improved variants. The tightest binders in both pilots improved more than tenfold, about half of the final cancer-targeting variants bound too tightly for the assay to resolve, and the most stable cancer-targeting variants stayed in solution more than 20°C higher than the starting molecule. The work took about 10 weeks: one researcher and Claude designed the molecules, and BigHat’s team advised on the process and made and tested the designs.
The stability results in particular stood out. Small antibody fragments like the two in this pilot often begin to unfold somewhere between 50 and 65°C, and the industry generally wants a molecule to remain stable up to about 65°C before deeming it developable. The starting VHH melted at about 62°C, and the starting scFv collapsed at about