Transfer learning across molecular graphs for predicting protein-ligand affinities and their changes upon mutations

Predicting protein-ligand binding affinity and mutation-induced affinity changes ({Delta}{Delta}G) remains challenging due to limited data and complex interaction mechanisms. Here we present DDMuffin, a deep learning framework that integrates structural, sequence, and interaction graph features, employing transfer learning and stringent dataset partitioning to achieve reliable generalization. DDMuffin demonstrates strong predictive accuracy on the rigorous LP-PDBBind benchmark (Pearson r up to 0.70 after excluding top 10% outliers; RMSE = 1.48 kcal mol-{superscript 1}). In evaluating mutation-induced affinity changes for clinically relevant kinase inhibitors, DDMuffin achieves competitive average performance (mean Spearman{rho} = 0.39), outperforming or matching several established methods. The approach provides valuable insights into mutation-driven drug resistance and interpretable ligand binding mechanisms, particularly advantageous for guiding inhibitor design and personalized therapeutic strategies. To facilitate broad application, we deployed DDMuffin as an accessible web server at https://biosig.lab.uq.edu.au/ddmuffin/, promoting systematic exploration of protein-ligand interactions in drug discovery research.