The Lock of Food Allergy: Claude Generated 35,000 Designs in a Month, and 3 That Could Pry It Open Were Left

Food allergy affects more than 200 million people. The only approved antibody drug, Xolair, sequesters free IgE in the blood, but IgE already bound to FcεRI on mast cells does not leave immediately. IgE and FcεRI are one of the tightest-binding protein pairs in immunology, with dissociation so slow that it is measured in weeks or even months. After patients stop treatment, mast cells remain sensitized.

Divergence Labs and Anthropic collaborated to try a different action: instead of waiting for IgE to let go on its own, use an antibody to directly pull it off FcεRI. In his tweet, Bowen Jing compared the process to picking a lock, requiring the right geometry and force. Their approach is a two-arm antibody, with the two arms having different functions and binding different epitopes. One arm grabs IgE, and the other pries at its binding to the receptor.

Here Claude took on expanding the design space and narrowing down the options. Anthropic's case study describes it concretely: Claude translated the team's goals into a detailed design plan and, over about one month, generated about 35,000 antibody designs for subsequent narrowing and experimental testing. Early tests found candidates for both arms: 3 designs could bind IgE, and 20 could pry IgE; among them, 3 had stronger prying effects than the standard drug. The next step is to assemble the two arms into a complete therapeutic format.

The video shows how Xolair works for comparison: it mainly neutralizes free IgE, and mast cells that have already been sensitized still take a long time to recover.

This is still preclinical data and far from reaching the market. Some users in the replies also questioned it, saying that candidate drugs that downregulate the IgE signaling pathway are already in clinical trials, and arguing that this work is not a big enough change. The original tweet did not respond. This criticism is a reminder of one thing: good early metrics do not equal clinical endpoints. Another user asked whether, after combining the two arms, the harder part is maintaining prying activity or avoiding mast cell activation. This question is closer to reality than “AI invented a new drug.”

From a development perspective, the signal in this case is not the number 35,000 itself. Generating designs is cheap; experiments to express, purify, and verify that the two arms do not interfere with each other are still expensive. What is truly scarce is the closed loop that connects design, synthesis, and testing. Claude's value may lie here: rapidly turning hypotheses into testable lists of molecules, then letting experimental data come back to correct them. It does not make clinical judgments for people, nor does it skip wet-lab experiments.

Anthropic's blog also lists other collaboration cases, including Generate Biomedicines, BigHat Biosciences, Enveda, and others, showing that they are placing Claude into different stages of drug development. The Divergence project is just an early slice.

Reference links:

Anthropic blog: https://www.anthropic.com/research/early-progress-drug-development

Divergence technical report: https://divergencelabs.com/publications/ige-preprint.pdf