Interformer: an interaction-aware model for protein-ligand docking and affnity

    This Paper of the Day describes Interformer a a deep learning model designed for protein-ligand docking and affinity prediction. It uses a Graph-Transformer architecture to capture non-covalent interactions between ligands and proteins, such as hydrogen bonds and hydrophobic interactions. Interformer employs an interaction-aware mixture density network to model these specific interactions, enabling it to generate physically plausible docking poses. Additionally, Interformer incorporates a negative sampling strategy to enhance its pose-sensitivity and improve affinity prediction, especially when dealing with less accurate binding poses. Experimental results on benchmark datasets and real-world drug design projects demonstrate Interformer's superior performance compared to state-of-the-art methods. The model's ability to interpret the internal mechanisms of protein-ligand interactions further enhances its practical applicability in accelerating the drug discovery process.