Precise Antigen-Antibody Structure Predictions Enhance Antibody Development with HelixFold-Multimer

TL;DR

Antigen-antibody structure prediction is challenged by the distinct evolutionary patterns of antigens and antibodies. The authors introduce HelixFold-Multimer, a targeted extension of AlphaFold-Multimer fine-tuned on SabDab antigen-antibody data, achieving a median DockQ of $0.469$ and a $58.2%$ success rate on 141 interfaces, outperforming AlphaFold-Multimer and AlphaFold3. They demonstrate that epitope information and high-precision structures enable improved docking, interaction prediction, and antibody design, including integration with FoldX and a Masked MSA design module. The results suggest substantial potential to accelerate antibody development and therapeutic design by providing reliable structure-based guidance and confidence metrics.

Abstract

The accurate prediction of antigen-antibody structures is essential for advancing immunology and therapeutic development, as it helps elucidate molecular interactions that underlie immune responses. Despite recent progress with deep learning models like AlphaFold and RoseTTAFold, accurately modeling antigen-antibody complexes remains a challenge due to their unique evolutionary characteristics. HelixFold-Multimer, a specialized model developed for this purpose, builds on the framework of AlphaFold-Multimer and demonstrates improved precision for antigen-antibody structures. HelixFold-Multimer not only surpasses other models in accuracy but also provides essential insights into antibody development, enabling more precise identification of binding sites, improved interaction prediction, and enhanced design of therapeutic antibodies. These advances underscore HelixFold-Multimer's potential in supporting antibody research and therapeutic innovation.

Figures (8)

• Figure 1: Performance comparison of HelixFold-Multimer and the leading structural prediction models for antibody-antigen interfaces. (a) Median DockQ scores. (b) Percentage of predictions exceeding a DockQ threshold of 0.23. (c) Performance of HelixFold-Multimer and AlphaFold 3 across different species. (d) Relations between the DockQ scores of HelixFold-Multimer prediction and confidence levels outputted by HelixFold-Multimer.
• Figure 1: Prediction accuracy for VH-VL interfaces. (a) DockQ scores for VH-VL antibody interfaces. (b) Percentage of predictions with DockQ scores exceeding 0.8, indicating high-quality interface predictions.
• Figure 2: Impact of Epitope Information on Antibody-Antigen Docking Accuracy. (a) Scatter plots comparing DockQ scores for HelixFold-Multimer predictions with and without epitope inclusion. (b) The DockQ bar plots of antigens with different homology. (c) The influence of epitope size on the prediction accuracy of antigen-antibody interactions.
• Figure 2: Performance comparison of the general version and antigen-antibody version of HelixFold-Multimer for antibody-antigen interfaces. (a) DockQ score for antibody-antigen interfaces. (b) Percentage of predictions achieving a DockQ score exceeding 0.23, indicating correct predictions.
• Figure 3: Comprehensive evaluation of HelixFold-Multimer's performance in antigen-antibody interaction prediction. (a) Binder recognition capacity for HelixFold-Multimer confidence metric across four antigens. (b) Affinity prediction capacity for HelixFold-Multimer confidence metric across various influenza strains. (c) Affinity prediction capacity for energy-based method FoldX using the HelixFold-Multimer or AlphaFold-Multimer predicted conformations as input. (d) Correlation between ESM-IF scores using input structures derived from AlphaFold3, HelixFold-Multimer, and crystallographic structures, and the experimentally measured affinities. (e) Integrated enhancement of HelixFold-Multimer and energy-based method for binder recognition task. (f) Integrated enhancement of HelixFold-Multimer and energy-based method for binding affinity prediction task. The color bar indicates the experimental binding affinity of various antibody-target pairs.