BPS2026 – Leveraging neural networks to correct FoldX free-energy estimates

Fast and accurate prediction of how mutations reshape protein energy landscapes, affecting folding stability and binding affinity, remains a central challenge in biophysics. While experimental methods are accurate, they are often time-consuming and costly. Computational approaches like FoldX offer fast and inexpensive alternatives by estimating free-energy changes (ΔΔG) from structure, but their accuracy remains limited. In this work, we present dFX, a novel neural-network-based approach that leverages the rich biophysical insights embedded in FoldX energy terms to improve the accuracy of the free-energy predictions. We first assembled the largest available collections of experimental ΔΔG values and their corresponding 3D structures. Using those structures, we performed ΔΔG baseline predictions via FoldX that also outputs its contributing energy terms. We then trained dFX using those FoldX energy terms as input features. Our dFX models significantly improve the prediction accuracy beyond current capabilities. They increase the Pearson correlation coefficient between experimental and estimated ΔΔG for both folding and binding across different mutation types, outperforming FoldX alone, and overall generalizing well from lower to higher order mutations. Consistently, our dFX models also improve epistasis prediction by more accurately capturing non-additive effects. On an external SARS-CoV-2 data set, the dFX model trained on single mutations for binding outperformed FoldX and other machine-learning approaches. Once trained, our neural network models add minimal computational time but provide a substantial increase in accuracy, especially for higher order mutations and epistasis, making them a valuable addition to any FoldX free-energy prediction pipeline. Our dFX approach could be further optimized to predict antibody escape, aiding in the efficient development of watch lists.