DeepDOX1: A Dual-Drive Framework Integrating Deep Learning and First-Principles Quantum Chemistry for Drug-Protein Affinity Prediction.

In recent years, there has been a surge in artificial intelligence (AI)-based drug-protein (or pesticide-protein) affinity (DPA) prediction tools. The field has been primarily driven by evolving deep-learning architectures and increasingly complex representations, leading to a growing demand for training data sourced from limited experimental data. In this work, we present DeepDOX1, a dual-drive DPA prediction tool featuring the tight integration of a concise AI architecture and an interpretable, quantum chemistry-based representation of protein-ligand interactions. To be more specific, the first-principles quantum chemistry-generated features incorporating the interactions between the drug and the protein binding pocket allow a relatively simple convolutional neural network (CNN) model trained on a relatively small training set (9,938 binders) to exhibit exceptional generalization capabilities across extensive testing involving 1,281 binders. Notably, DeepDOX1 outperforms popular AI-based DPA prediction methods in the tests simulating real-world hit-to-lead optimization (HLO) scenarios and a highly challenging test set featuring covalent ligands, halogenated ligands, and metalloproteins, even though its training set does not contain any covalent ligands. To further validate its practical utility, we designed a series of novel covalent inhibitors targeting the diabetes target molecule Hu-FBPase using DeepDOX1. Subsequent experimental validation, including enzyme-level bioactivity assays and crystal structure determination, revealed strengthened activity of the newly designed compound and confirmed DeepDOX1's effectiveness in real-world drug design applications. It is conceivable that the combination of deep-learning architecture and first-principles quantum chemistry might be one of the next breakthroughs in DPA prediction.