Divergence Labs Reports Claude-Assisted Antibody Candidate Aimed at Food Allergy

Divergence Labs has published a preprint describing an agentic computational design campaign, run with Claude, toward an antibody that the company says physically removes bound IgE from the high-affinity receptor FcεRI. The preprint, dated 29 October 2026 and hosted on the company's site, is aimed at food allergy.

"Agentic" here means an AI system that carries out a sequence of design steps toward a goal, rather than answering a single prompt. IgE is the antibody that drives allergic reactions. FcεRI is the high-affinity receptor it binds on mast cells and basophils, the immune cells that release allergy mediators.

The company describes itself as a collective of drug hunters, computer scientists, biologists, and hardware engineers working to systematize the creation of new medicines. It says it designs and develops product candidates across all areas of medicine by addressing molecular engineering inefficiencies. It lists offices in New York, Minneapolis, San Francisco, and Cambridge, UK, and gives careers@divergencelabs.com as its careers contact.

Anthropic, in a report dated 27 October 2026, said Divergence Labs worked with Claude to design early antibody candidates for a new food allergy drug. The preprint is the company's own technical account of that work.

The preprint's starting point is the existing therapy. It identifies Omalizumab (Xolair) as the sole FDA-approved biologic for food allergy, meaning a drug made from living cells. According to the document, Omalizumab neutralizes circulating IgE but does not effectively remove IgE from FcεRI. The candidate the company describes is designed to do the second thing: act on IgE that is already engaged with the receptor, not only on the free antibody in circulation.

The distinction matters mechanistically. A molecule that sequesters free IgE interrupts the supply of antibody reaching mast cells and basophils. A disruptor has to engage a receptor-bound complex, which presents a different binding surface and a different set of structural constraints. The preprint's framing rests on that difference, and the design problem it poses is considerably harder than neutralization.

In my view, the relevant question for computational chemists and protein engineers is less whether a language model can propose sequences and more how those proposals were filtered, scored, and tested. Agentic workflows compress the early search, but they do not remove the need for wet-lab confirmation, and the claim that a candidate removes receptor-bound IgE is a claim the data must carry.

The broader context here is a sector in which AI-assisted design is moving from target identification into molecular optimization. Anthropic's report and the Divergence Labs preprint put a named model at the center of a food allergy program, which is a more specific case than most public statements about AI in drug discovery. Food allergy is also a therapeutic area where the existing biologic has a defined limitation, so an alternative mechanism has a clear benchmark against which to be judged.

If the disruptor approach holds up under experimental and eventually clinical scrutiny, it would give developers a second mechanism for a condition with one approved biologic. If it does not, the preprint still documents a concrete attempt to apply an agentic design process to a hard receptor-level problem, with the methods and claims open for other groups to test.