Predicting Antibody Relative Binding Free Energies with MD-Derived Graph Neural Networks

Designing high-affinity antibodies is constrained by the vast mutational space and the time required for biophysical assays. Physics-based simulations, such as free-energy perturbation (FEP), offer a rigorous way to estimate relative binding free-energy changes (∆∆G) but are computationally intensive. We introduce a surrogate model to predict ∆∆G using graph neural networks and features extracted from equilibrium molecular dynamics (MD) trajectories. Training targets are generated with FEP for four antibody–antigen complexes from SKEMPI (PDB: 2B2X, 1JRH, 1MLC, 3HFM), spanning single- through sextuple-site substitutions. FEP shows Pearson r ~ 0.5 to experiment on these systems, motivating its use as a training signal. Our pipeline encodes time-dependent interface contacts into graphs to learn mutational effects on ΔΔG. We contextualize our methodology against structure-only approaches and highlight potential advantages of encoding conformational heterogeneity.