Auto-Research Agent on MSA Organization to Improve Protein-Protein Complex Structure Prediction
| Source: arXiv AI
Tags: AlphaFold 3, multiple sequence alignment, DockQ, FoldBench, protein structure prediction, AI agents
An autonomous agent that reorganizes MSAs per target improves AlphaFold 3 protein-complex predictions on 1,015 interfaces while using 60% fewer sequences on average.
Details
AlphaFold 3's protein-complex predictions depend on MSA content, depth and cross-chain pairing, yet pipelines apply uniform recipes. The authors show the value of depth and pairing is strongly target-dependent. Their agent treats a frozen AF3 as a black-box instrument: for each target it forms biological hypotheses, reorganizes the MSA and refines sequence selection and pairing using label-free confidence signals. On 1,015 interfaces from three benchmarks released after AF3's training cutoff, it beats the default pipeline on every benchmark with about 60% fewer MSA sequences. Gains are largest for inter-species heteromers, with mean DockQ rising from 0.465 to 0.612 on FoldBench and 0.482 to 0.562 on HD439.