Claude Designed 35,000 Antibody Designs: Here's What Researchers Found

Claude is moving beyond the chat window and into drug discovery, helping researchers search for a potential food-allergy treatment.

In roughly a month, the AI generated about 35,000 antibody designs, and early laboratory tests identified several promising candidates.

Anthropic disclosed the results on Oct. 27 from its work with Divergence Labs, one of six drug-development partners testing Claude in real research programs. Other projects reported gains in cancer-cell engineering, antibody testing and pharmaceutical paperwork. 

For drugmakers, the question is whether faster research tasks can shorten the costly journey from promising design to useful medicine.

Claude's allergy project targets antibodies attached to immune cells

Immunoglobulin E (IgE) helps trigger allergic reactions. Some circulates in the blood, while some attaches to immune cells. Divergence wants to remove the attached IgE rather than only neutralizing what's circulating.

A treatment would need two parts: one to bind IgE and another to pull it from immune cells. Anthropic said early tests found three designs that bound IgE and 20 that detached it. Three of those 20 appeared to remove more IgE than a standard medication. 

Like earlier Claude-designed protein binders, these are early laboratory candidates, not finished medicines. Divergence still has to combine both functions and test whether they work together.

What eWEEK found: Claude shows promise beyond allergy research

Claude's reported contributions went beyond generating candidates. One partner used it to modify cancer-fighting immune cells, another to improve experimental antibody fragments, and a third to review manufacturing records. Anthropic outlined the results in its Oct. 27 report: 

The measures are different. According to Interesting Engineering, Generate and BigHat reported improvements in laboratory tests; Enveda measured time saved reviewing documents. None demonstrates that a new treatment works in patients.

The allergy experiment leaves a key question unanswered. Anthropic did not disclose how many of the 35,000 computer-generated designs were physically tested, so no success rate can be calculated. It also did not name the medication used for comparison.

Enveda reported the clearest measurable time saving. Most other projects remain preclinical, and their laboratory gains have not established better treatments or lower overall development costs.

The results show why AI's impact on drug development cannot be measured by a single benchmark. Generate Biomedicines and BigHat reported laboratory improvements that still require validation, while Enveda demonstrated measurable time savings in document review. Neither outcome establishes lower overall drug-development costs.

The projects also require different levels of validation. AI-generated antibodies must undergo laboratory testing to confirm their effectiveness, while document-review tools can be checked against existing records for errors and omissions. Both require expert oversight, but their paths to practical value differ. 

Drugmakers should measure progress beyond the design stage

Claude's results give biotechnology companies a reason to explore AI-assisted drug discovery, but generating thousands of designs is only one part of developing a medicine. Researchers still need to manufacture promising molecules, test their properties, and determine whether the results can be repeated.

Divergence Labs illustrates the challenge. Although Claude identified promising IgE-binding and IgE-removing designs, the company has not yet combined both functions into a single treatment. For drugmakers, the next experiments will help establish whether the approach offers a practical advantage over existing methods.

Before investing further in AI-assisted research, biotechnology teams should evaluate three areas:

• Experimental success rates: Track how many AI-generated designs undergo laboratory testing, how many demonstrate the intended effects, and whether those results can be reproduced. Comparing these figures with conventional research methods can help determine whether AI improves the chances of finding useful candidates.
• Total research costs: Faster design does not necessarily mean cheaper drug development. Companies should account for computing expenses, laboratory materials, unsuccessful experiments, specialist review, and additional testing. Cost per validated candidate offers a more useful measure of AI's financial value.
• Clinical readiness: Stronger antibody binding or increased cancer-targeting receptor production may support further research, but neither guarantees a successful treatment. Researchers must still examine safety, effectiveness, and whether promising laboratory findings translate into benefits for patients.

Progress in other areas of AI-assisted medicine provides useful context. Previous AI-designed vaccine research and an AI-assisted drug candidate have reached human testing, while Divergence's allergy project remains at an earlier stage.

For pharmaceutical executives deciding whether to expand AI research programs, the priority should be measurable improvements in research efficiency, including the time and money needed to produce candidates that pass laboratory testing. 

A successful pilot should demonstrate value beyond the number of designs generated.

For patients with food allergies, the next meaningful milestone will be a treatment that works safely and effectively in humans. Claude may help researchers explore more possibilities, but the real test is whether those possibilities eventually lead to better medicines.

Want to build your AI skills? Try our free six-minute How to Talk to AI course to learn a few simple ways to write better prompts and get more useful results from AI. eWeek readers also get seven days of free access to The Neuron Academy’s other AI courses. Check out all the lessons here →