Comparative Assessment of the Utility of Co-Folding and Docking for Small-Molecule Drug Design

Abstract

Deep-learning based co-folding methods predict the structures of proteins interacting with metal ions, small molecules, nucleic acids, peptides, and other proteins. One of their main objectives is their application for drug design, predicting the structure of small-molecule ligand/protein complexes. It has been shown that at present these models memorize ligand poses from the training data but do not generalize effectively to novel complexes and lack in the adherence to physical and chemical principles. Here, we use the recently introduced Runs N Poses benchmark set of 2,600 protein-ligand systems annotated by their similarity to the training data, to show that the physics-based docking algorithms Attracting Cavities and AutoDock Vina outperform co-folding methods for novel ligands and novel binding pockets. In addition to predicting ligand poses and at variance with co-folding methods, they provide a physical rationale on why a ligand binds (or does not bind) and insight into experimental structural model deficiencies.