Predicting the impact of mutations on protein-ligand binding affinity is crucial in drug discovery, particularly in addressing drug resistance and repurposing existing drugs. Conventional structure-based methods are often limited by their reliance on static cocrystal structures. To address this, we integrate AlphaFold 2 (AF2) subsampling with a Siamese neural network to predict mutation-induced changes in the relative binding affinity. By leveraging AF2 subsampling, we generated conformational ensembles for Abelson tyrosine kinase (ABL) mutants, shifting the paradigm from single-point predictions to an ensemble-based approach that accounts for intrinsic structural flexibility. Furthermore, we augmented the data set by pairing the generated conformations with reference states, followed by the identification of structurally relevant states via a most probable distribution analysis. To facilitate relative affinity prediction, we developed SIGMA-Net (Siamese structure and graph-aware multistructural affinity prediction network), which was employed to discern features between wild-type and mutant states, enabling free-energy predictions with chemically meaningful accuracy. Benchmarking on the tyrosine kinase inhibitors (TKI) data set and the refined set of PDBbind, our proposed approach achieves higher correlation coefficients for five of six TKI molecules across 31 ABL mutants, outperforming molecular docking and trichannel graph network (TriG-Net). By integrating conformational sampling with Siamese learning, our method enhances both the predictive accuracy and robustness. It achieves absolute binding free energy (ABFE) prediction performance comparable to that of state-of-the-art models such as Boltz-2, whereas Boltz-2 demonstrates better performance in relative binding free energy (RBFE) prediction in the evaluated systems. This framework effectively transcends the limitations of static structure dependence, providing a transferable solution for modeling protein-ligand interactions in highly flexible drug targets.
Xie et al. (Mon,) studied this question.