AutoHMD: Scalable and Accurate Validation of Antibody-Antigen Complexes by Heated Molecular Dynamics

Antibody-antigen (Ab-Ag) interactions are central to immune recognition and widely exploited in therapeutic design, making accurate structural characterization of these complexes a central challenge in the field. While docking and artificial intelligence-based approaches can generate multiple candidate binding modes, reliably identifying near-native conformations represent a key bottleneck. Existing methods depend largely on static scoring functions, which fail to capture the dynamic nature of Ab-Ag interfaces and, as a result, offer limited power to discriminate between competing poses.

Here, we evaluate heated molecular dynamics (HMD) as a strategy to discriminate binding pose quality based on structural stability under thermal perturbation, implemented through the automated workflow AutoHMD. Across a diverse set of Ab-Ag complexes, HMD enabled clear separation between high-quality and incorrect poses. During simulations, near-native complexes maintained a stable interface, whereas incorrect ones showed progressive divergence, particularly at elevated temperatures. Interaction analysis revealed that high-quality poses maintain more stable and persistent interfacial contacts throughout the simulations. In contrast, incorrect poses display an unstable interaction network. Additional simulation replicates increased classification consistency, while a mean iRMSD cutoff of approximately 2.5 Å achieved 90% accuracy in cross-validation and retained robust performance on an independent external set. Its transfer to Boltz-2-generated complexes further supported the broader applicability of HMD to AI-predicted binding poses. Overall, AutoHMD provides a reproducible, scalable, and high-throughput framework for validation of predicted antibody-antigen complexes, with potential applications in structural bioinformatics and biopharmaceutical design.

Supplementary data are available at Bioinformatics online. The AutoHMD workflow scripts are available on GitHub at https://github.com/SFBBGroup/AutoHMD. The version used in this study is permanently archived in Zenodo at https://doi.org/10.5281/zenodo.22757672. All simulation and structural data supporting this study are publicly available in Zenodo at https://doi.org/10.5281/zenodo.20722229; https://doi.org/10.5281/zenodo.22750139; https://doi.org/10.5281/zenodo.20707612.