[Nature Machine Intelligence, 2024] This paper introduces PSICHIC, a graph neural network framework that leverages physicochemical constraints to predict protein-ligand interactions directly from sequence data. PSICHIC achieves state-of-the-art accuracy in binding affinity prediction, even surpassing existing structure-based methods in certain cases. Furthermore, its interpretable fingerprints illuminate the specific protein residues and ligand atoms involved in these interactions, offering a promising tool for virtual screening and enhancing our understanding of protein-ligand mechanisms.
