Abstract
Protein-peptide interactions mediate many biological processes. Accurate structural models of protein-peptide complexes, determined by experiment or computational prediction, are essential for understanding function and designing interaction inhibitors. AlphaFold2-Multimer (AF2-Multimer), AlphaFold3 (AF3), and related models such as Boltz-1 and Chai-1 can predict protein-peptide binding geometry, often with high accuracy. Using a dataset of experimentally resolved structures, we analyzed the performance of these four structure prediction models to understand how they work. We found evidence of bias for previously seen structures, indicating that models struggle to generalize to novel proteins or binding sites. We tested how models use the protein and peptide multiple sequence alignments (MSAs), which are often shallow or of poor quality for peptide sequences. We found weak evidence that models use coevolutionary information from paired MSAs, but both the protein and peptide unpaired MSAs contribute to prediction accuracy. Our work highlights the promise of deep learning for peptide docking and the importance of diverse representation of interface geometries in the training data for optimal prediction performance.