Evaluation ofAlphaFoldantibody–antigen modeling with implications for improving predictive accuracy

Abstract

High resolution antibody–antigen structures provide critical insights into immune recognition and can inform therapeutic design. The challenges of experimental structural determination and the diversity of the immune repertoire underscore the necessity of accurate computational tools for modeling antibody–antigen complexes. Initial benchmarking showed that despite overall success in modeling protein–protein complexes, AlphaFold and AlphaFold‐Multimer have limited success in modeling antibody–antigen interactions. In this study, we performed a thorough analysis of AlphaFold's antibody–antigen modeling performance on 427 nonredundant antibody–antigen complex structures, identifying useful confidence metrics for predicting model quality, and features of complexes associated with improved modeling success. Notably, we found that the latest version of AlphaFold improves near‐native modeling success to over 30%, versus approximately 20% for a previous version, while increased AlphaFold sampling gives approximately 50% success. With this improved success, AlphaFold can generate accurate antibody–antigen models in many cases, while additional training or other optimization may further improve performance.

A Comparison of Antibody-Antigen Complex Sequence-to-Structure Prediction Methods and their Systematic Biases

“…Biases when comparing correct and incorrect models can indicate which factors are important for the success of a model, as well as features that may be used to discriminate good and bad quality models. AlphaFold-Multimer in particular relies on Multiple Sequence Alignments (MSAs), and richer MSAs have been reported to produce better antibody-antigen models (Yin and Pierce 2024). However, as MSAs provide co-evolution information to aid AlphaFold-Multimer in predicting binding, and antigens generally are co-evolving to evade antibodies as well as usually from different species, one would not necessarily expect accuracy in predicting their binding motif to correlate with MSA richness.…”

“…ML methods performing docking such as DiffDock-PP may be useful for predicting targets that AlphaFold-Multimer currently fails on, as suggested by our discovery of ClusPro’s orthogonal ability to succeed on some targets where AlphaFold-Multimer failed using even rigid-body docking. AlphaFold-Multimer has also very recently been shown to improve at its antibody-antigen predictions with multi-seed runs, on a small benchmark dataset (Yin and Pierce 2024).…”

“…AlphaFold v2.2.0 has been compared to ZDOCK in its ability to dock AlphaFold-generated antibodies and antigens (Polonsky et al 2023). AlphaFold-Multimer has also been compared to ClusPro’s ability to dock AlphaFold-Multimer generated antibodies and antigens, although that case only evaluated whether or not models met a highly stringent DOCKQ score cutoff of 0.49 or above (Yin and Pierce 2023). RoseTTAFold has not been evaluated in its capability to predict full antibody-antigen complexes, but has been evaluated in its ability to predict antibodies on their own—in which case it predicts the H3 loop crucial for most binding motifs (Regep et al 2017) better than ABodyBuilder and comparably to SWISS-MODEL, but the overall 3D structure worse (Liang et al 2022).…”