Auto-Research Agent on MSA Organization to Improve Protein-Protein Complex Structure Prediction

Auto-Research Agent on MSA Organization to Improve Protein-Protein Complex Structure Prediction. By Zilu Zhang and six others, from the Beijing University of Posts and Telecommunications, Shanghai Jiao Tong University, and two other institutions. Published on the second of November.

When AlphaFold 3 predicts the structure of a protein complex, the result depends on the multiple sequence alignment it is given: its content, its depth, and how sequences are paired across the chains. Current pipelines apply one uniform recipe to every target, typically maximizing depth and pairing sequences by species. Yet it is still contested whether cross-chain pairing helps AF3 at all. The authors show that the value of both depth and pairing is strongly target-dependent. If that is true, no single recipe can be optimal across all interfaces.

Their response is to stop looking for the one right recipe and let an autonomous agent choose per target. The agent treats a frozen AF3 as a black-box experimental instrument. It does not retrain or modify the predictor. It runs experiments with it, the way a researcher would, and decides what alignment to feed it.

For each target, the agent formulates biological hypotheses, then analyzes and reorganizes the MSA. It iteratively refines which sequences are selected and how they are paired. Throughout, it relies only on signals from the predictor that need no labels, so it never sees the true structure while deciding. The abstract describes this as self-directed, confidence-guided experimentation.

The test set pools protein-protein interfaces from three benchmarks, all of them released after AF3's training cutoff, so none of the test complexes could have been seen in training. On every one of those benchmarks, the agent improves overall interface DockQ over both AF3's default pipeline and fixed MSA recipes. It does so while using on average sixty percent fewer MSA sequences than the default.

The largest gains come on inter-species heteromers, where pairing by species is unreliable. On FoldBench, mean DockQ rises from zero point four six five to zero point six one two. On HD439, it rises from zero point four eight two to zero point five six two. The improvements come from both how sequences are constructed and the pairing policy, and that policy includes choosing, deliberately, not to pair.

The result recasts alignment design as a target-specific problem rather than a preprocessing step with one correct setting. And it shows an agent can adapt that design at test time, through biological reasoning and its own experiments, with less input than the standard pipeline and without touching the model's weights.