Predicting Mutation-Induced Relative Protein-Ligand Binding Affinity Changes via Conformational Sampling and Diversity Integration with Subsampled Alphafold2 in Few-Shot Learning

Predicting the impact of mutations on protein-ligand binding affinity is crucial in drug discovery, particularly in addressing drug resistance and repurposing existing drugs. Current structure-based methods, including deep learning and physics-based computational techniques, are constrained by their dependence on known protein-ligand complex structures. The lack of structural data for mutated proteins, coupled with the potential for mutations to alter protein conformational states, poses challenges for reliable predictions. To address the challenge, we propose a few-shot learning approach that integrates AlphaFold2 subsampling, ensemble docking, and Siamese learning to predict the relative binding affinity changes induced by mutations. We benchmark our approach for predicting binding affinity changes in 31 variants of the mutated Abelson tyrosine kinase (ABL). The conformation ensembles for each ABL variant are generated using AlphaFold2 subsampling. Subsequently, the most likely conformational states are selected through Gaussian fitting to construct the pairwise structural inputs and compute relative binding energies. To leverage the structural information, we utilize the multimodal inputs including contact maps and molecular graphs as the molecular representation in the Siamese learning network. The relative binding affinities are derived by averaging the values predicted by the Siamese learning model across the constructed conformational pairs. Benchmarking on the Tyrosine kinase inhibitors (TKI) dataset and the direct binding affinity using the refined set of PDBbind database, our proposed approach achieves increased prediction accuracy in estimating relative binding affinities by using the conformations generated from AlphaFold2 subsampling, which includes structural diversity in the prediction and avoids reliance on crystal structures. We anticipate the approach will be valuable for predicting binding energy changes induced by mutations in proteins or ligands, facilitating drug discovery efforts.