Anthropic collaborations: Divergence Labs antibodies showed no improvement over the starting antibody, and a simple selection rule rivaled AI models in the Generate Biomedicines project, according to Abhishek Mahajan’s analysis

Anthropic collaborations: Divergence Labs antibodies showed no improvement over the starting antibody, and a simple selection rule rivaled AI models in the Generate Biomedicines project, according to Abhishek Mahajan’s analysis

On October 27, Anthropic described how its AI, Claude, worked with six partners on drug development. On October 28, OpenAI Foundation employee Abhishek Mahajan posted on X an analysis of three of these projects.

Divergence Labs is designing an antibody that removes IgE, the antibody involved in allergic reactions, from the cell receptor through which IgE triggers a response to an allergen. Omalizumab, the only antibody approved in the US for food allergy, binds IgE in the blood but removes it from the receptor only at concentrations far above those reached in patients. The new antibody has two “arms”: a “displacer” that detaches IgE from the receptor and an “anchor” that binds receptor-bound IgE and holds the displacer nearby.

In the Divergence report, all displacer variants were based on sdAb026, an antibody already known to displace IgE. Of 91 variants tested, 20 displaced IgE and differed from sdAb026 by no more than eight amino acid substitutions. Three variants displaced IgE more effectively than omalizumab at all five concentrations, but sdAb026 itself also outperformed omalizumab. The authors write that they “did not measure any variant with improved functional activity compared with sdAb026 at controlled doses.”

The anchor was designed from scratch: 3 of 57 candidates bound IgE, and the two-armed antibody has yet to be assembled. Mahajan concludes that, even with a known mechanism, AI appears to remain short of being able to fulfill a request to “design me a better drug.”

In the BigHat antibody optimization project, Claude helped establish that an instrument reading labeled “melting temperature” reflected protein aggregation. Mahajan sees this primarily as an interesting account, with little evidence of a new AI capability.

Generate Biomedicines is shortening the signaling portion of CAR, an engineered receptor that enables a patient’s T cells to recognize a tumor, although shortening it is expected to impair function. Over three rounds of experiments, AI agents proposed shortened variants using measurements from previous rounds. The selected variant consists of 45 amino acids, less than a third of the length of the clinical version. When delivered as RNA, it required a 4,6-fold lower dose than the clinical version to kill half the B cells in blood cultures from five donors.

Mahajan was struck by how poorly specialized protein models distinguished functional variants from nonfunctional ones, while general-purpose language models guided the project. The report notes, however, that without experimental results, the language models also predicted that most variants would be functional.

Mahajan asks whether another approach could have achieved the same result. He points out that in 2022, researchers assembled around 2300 combinations of signaling fragments and used neural networks to improve the receptors. He therefore sees Generate’s result primarily as a reduction in the time required, with limited evidence of a new capability, although “perhaps enough reductions in time will one day amount to a new capability…” According to the report, earlier work explored which blocks to combine, while Generate explored how much could be removed.

Mahajan’s second question, “What would the dumbest alternative look like?”, leads him to a rule in the Generate report: in the first round, variants failed when the first tyrosine in the activation region was closer to the cell membrane than the ninth amino acid position. This rule rivaled several models in selecting functional variants, and among variants that passed it, the models’ advantage largely disappeared. The report states that the rule was derived retrospectively, after all measurements were available.