A comparison of antibody–antigen complex sequence‐to‐structure prediction methods and their systematic biases

Abstract The ability to accurately predict antibody–antigen complex structures from their sequences could greatly advance our understanding of the immune system and would aid in the development of novel antibody therapeutics. There have been considerable recent advancements in predicting protein–protein interactions (PPIs) fueled by progress in machine learning (ML). To understand the current state of the field, we compare six representative methods for predicting antibody–antigen complexes from sequence, including two deep learning approaches trained to predict PPIs in general (AlphaFold‐Multimer and RoseTTAFold), two composite methods that initially predict antibody and antigen structures separately and dock them (using antibody‐mode ClusPro), local refinement in Rosetta (SnugDock) of globally docked poses from ClusPro, and a pipeline combining homology modeling with rigid‐body docking informed by ML‐based epitope and paratope prediction (AbAdapt). We find that AlphaFold‐Multimer outperformed other methods, although the absolute performance leaves considerable room for improvement. AlphaFold‐Multimer models of lower quality display significant structural biases at the level of tertiary motifs (TERMs) toward having fewer structural matches in non‐antibody‐containing structures from the Protein Data Bank (PDB). Specifically, better models exhibit more common PDB‐like TERMs at the antibody–antigen interface than worse ones. Importantly, the clear relationship between performance and the commonness of interfacial TERMs suggests that the scarcity of interfacial geometry data in the structural database may currently limit the application of ML to the prediction of antibody–antigen interactions.