When the AI antibodies were graded blind

Computational antibody design has no shortage of impressive numbers. Almost all of them are retrospective and self-reported meaning each group also chose the conditions under which its own method was evaluated. AIntibody, published in Nature Biotechnology in 2026, chose the opposite approach. Modeled on CASP, the blinded competition that made AlphaFold’s breakthrough credible, it collected 511 designed or predicted antibody sequences from 29 organizations, all against the SARS-CoV-2 receptor-binding domain. Then, it put every one of them through the same experimental pipeline: high-throughput surface plasmon resonance, KinExA, a five-assay developability panel.

Three tasks resulted in three different verdicts

Affinity maturation inside a defined library was where computation genuinely delivered. The winner reached 95 pM, statistically tied with the best experimentally derived antibody. Notably, this amounted to a roughly 2,000-fold improvement on the parent. A real result but the headline most coverage stopped at.

Ranking clones by affinity within HCDR3 clusters produced the opposite while randomly picking clones from a cluster gave a 39% hit rate. Between 9.8 and 13.8% of AI submissions beat the simple cluster control. All but one strategy landed below the random baseline. As a notable exception, a self-supervised antibody language model reached 50%. The authors read the task as the benchmark’s clearest failure mode: predicting affinity from clustered sequencing data seems to be a lot harder than what the available methods can handle today.

Designing CDRs outside the library where the sequence space is open and no survivorship bias present in the training data produced 30.4% non-binders. Of the 168 submissions to that task, only 53.6% were both binders and developable.

The most instructive finding seems to be a scoring bug

The nominal winner of the third task produced a 2.9 pM antibody which failed to elute from a hydrophobic interaction column - in normal therapeutic development, likely a disqualifying biophysical liability. It won anyway since the developability panel was scored as an offsettable composite: strong results on the other four assays kept the combined score under the passing threshold, so the candidate counted as developable and its affinity decided the rest. The authors flagged this as a design flaw in their own benchmark and committed to hard go/no-go criteria for the next round.

This is a lesson which is worth more than the leaderboard. A composite metric quietly assumes its components are exchangeable. For antibody developability (as for most multi-criteria decisions), they are not: some failures amount to hard vetoes. Any scoring scheme which cannot express a hard veto will eventually recommend something useless.

Two more findings deserve attention. A non-AI statistical consensus sequence placed third overall with perfect developability, and a comparatively simple Gaussian process regression tied for a cluster win. Baselines keep being competitive which is exactly why benchmarks need them. Intriguingly, no method class won across all three tasks, groups that led one challenge fell behind in another which suggests matching the method to the respective task rather than seeking one general-purpose model.

The authors are careful with regard to the scope: RBD is an unusually well-characterized target and two of the three tasks handed participants affinity data that a real program would not have obtained until a later development stage. Therefore, they read their results as an upper bound on current capability, not a general verdict. The benchmark mainly illustrated the general possibility of such an evaluation: 29 organizations submitted work to be measured under conditions they did not control. This is how a field builds a shared picture of where it actually stands.