An antibody design agent picked a winner. Should it go to the lab?

What the antibody benchmark tested

On October 31, 2026, Asli Alpman and colleagues submitted a preprint that asks whether AI agents can follow a molecular editing brief while improving a design objective. Their study uses ten expert-reviewed antibody affinity tasks with mandatory requirements and optional suggestions. In a separate setting, agents propose improvements from previously scored sequences without receiving the same task-specific editing instructions. Those two settings give researchers a way to distinguish following a design brief from searching for a higher score.

The authors report that access to antibody tools improved adherence to required edits in the task-guided setting, while scored design history improved optimization on their surrogate and could reduce adherence. Their combined agent loop performed comparably to a Bayesian optimization baseline on predicted affinity in the studied setup. Refused calls were excluded from the reported compliance analysis. The optimization benchmark used a model trained on campaign data as its scoring oracle. It reports no new wet-lab confirmation of the designs proposed in that loop.

BigHat, whose researchers helped write the preprint, separately announced an antibody optimization collaboration with Anthropic on October 27. The announcement helps explain the practical interest in the question. The preprint supplies the evidence discussed here, and the collaboration announcement comes from a participant.

A promising score can answer the wrong question

A founder reviewing an agent demo might see a shortlist ranked by predicted binding and assume the highest score marks the obvious next experiment. In practice, the team may be running a much narrower experiment. It might want to ask whether a particular loop can be changed without touching the parent framework. A proposal that alters both regions may be an interesting molecule, but building it would spend one of the available slots on a different question.

We would separate three decisions. Eligibility asks whether the sequence meets the agreed rules. Ranking asks whether it is promising enough relative to other eligible sequences. Experimental review asks whether the planned measurement would resolve the team's uncertainty. Each decision needs a different piece of evidence, and a high score has little bearing on whether the sequence matches the agreed brief.

This separation also makes failures easier to diagnose. If eligible variants bind poorly in the lab, the team can investigate the scoring model or the scientific hypothesis. If the actual edits departed from the brief, the result cannot answer the question the team thought it was testing. A single dashboard metric can hide that difference.

When a candidate breaks the brief, it may still be worth studying. Give it a different experiment.THE BIOMEDANA LENS

A hypothetical screen with one slot left

Imagine a team with capacity to express a small batch of VHH candidates and run a binding assay. It wants to test substitutions in a defined loop while preserving the parent framework and the length of the sequence. The highest-ranked proposal includes an extra substitution outside the authorized region. A second proposal passes the edit checks and sits near the seed in sequence space. A third also passes, with a different combination of substitutions and a weaker model score.

For the planned comparison, the first proposal stays out. It can remain visible for a separately approved experiment, with its extra substitution recorded. The second may be the best candidate if the team is trying to confirm a narrow hypothesis suggested by earlier data. The third may deserve the slot if it probes a plausible alternative that the other candidates barely cover. The decision depends on what else is already on the plate.

We would also reserve space for the parent control and agree on what would trigger a repeat before seeing the results. If a variant shows weaker binding and poor sample quality, the team needs a way to distinguish an uninformative measurement from evidence against the design. These are illustrative choices for one discovery screen. The authors did not report this experiment.

Write the checks against the sequence you will order

A check is only useful if it runs on the sequence that actually goes forward. We would keep the parent identifier and version alongside the candidate, make the numbering convention explicit, and bind the pass result to the exact sequence exported for synthesis. An agent's description of the edits can help a scientist navigate the comparison. The comparison itself should be computed from the parent and proposed sequence.

Numbering is a good failure case to test. A position in an aligned antibody numbering scheme can differ from the offset in a plain sequence string, particularly around insertions. The verifier should return the mapped position and observed change, and stop for review when the mapping is ambiguous. Include deliberately incorrect examples with a forbidden edit, an unapproved insertion and the wrong parent. A failure should be legible enough that a scientist can fix the input or revise the brief.

Some constraints are straightforward to evaluate with code. Others depend on interpretation, such as whether a predicted structure supports an intended binding mode. Give those questions a reviewer and an explicit unresolved state. A structure tool can support the review, but the team's record should preserve that it was a prediction and specify what assay could challenge it.

What we would pass on to another builder

Run the pilot beside a simple baseline under the same experimental budget. Start with one well-understood parent and one design brief. Give both approaches the same opportunity to use prior data. Before synthesis, count proposals that pass every required rule, refusals and invalid outputs separately, and compare the number of distinct designs that make it into the review. After testing, look at how many measurements were interpretable and what they changed about the next round.

Keep refusals and timeouts visible when evaluating cost and throughput. A configuration can appear reliable on the sequences it returns while producing too few usable proposals to fill a real batch. Likewise, an apparent improvement can disappear when the same families of nearly identical variants are counted as a single design option. Set the counting rules with the scientists before looking at the pilot results.

Treat the first pilot as a test of where the agent earns its place. If conventional search fills the shortlist with equally useful candidates, the agent may contribute more by proposing alternate briefs or assembling review evidence. If it consistently helps scientists find eligible designs worth testing, expand only after reviewing how that behavior transfers to a different parent or assay.

Deciding which design gets the next lab slot

In the hypothetical VHH screen, the team applies these checks in order and records why each candidate advances, waits or moves to another experiment.

The questions behind the headline

Follow the evidence

1. Alpman et al. — Can agents reliably design therapeutic antibodies? (October 31, 2026 arXiv submission) ↗
2. BigHat Biosciences newsroom — October 27 Anthropic collaboration (industry context) ↗

Original BioMedAna editorial analysis. The linked sources support the reported developments; examples, frameworks and product implications reflect our perspective.

Explore the underlying concepts

• Natural-language protein design and the experimental brief
• Antibody design, function and developability evidence