PWRules Framework Applies Protein Words to Predict Small Molecule Binding with Interpretability

The PWRules framework improves the interpretability of protein-small molecule binding predictions by pinpointing favored small molecule fragments and establishing pairing rules with protein words—semantic sequence units. By utilizing binding affinity data, the framework ranks word-fragment rules via the PWScore function to highlight active compounds. Assessments on benchmark datasets reveal that PWScore performs competitively, on par with the physics-based model Glide and the deep learning model PSICHIC. This framework demonstrates wide applicability for protein targets beyond the training dataset, such as the SARS-CoV-2 main protease. Importantly, PWScore captures complementary interaction data, mitigating the dependence on opaque deep learning models in drug discovery while integrating principles and heuristics of protein-ligand interactions. This research was shared on arXiv under identifier 2604.16550v1, emphasizing the enhancement of binding prediction interpretability through this innovative method.