Benchmarking Protein-Ligand Binding with Deep Learning-Based Models

Abstract

This paper investigates the predictive capabilities of three leading deep learning-based protein structure models — AlphaFold2, ESMFold, and RoseTTAFold — focusing on their accuracy in identifying protein-ligand binding sites and the subsequent accuracy of ligand docking. We achieve this by docking ligands of experimentally determined protein complexes to predicted structures. We also examine the ligand docking ability in the context of allosteric sites, which provides a further level of depth to our research. Through a systematic comparison across three different protein classes – GPCRs, kinases, and proteases – we were able to more concretely understand the relationship between structural accuracy measures and binding site and ligand pose accuracy. We find that generally, higher overall structural accuracy in predicted structures leads to higher binding site and docked ligand pose accuracy, although it has a lesser effect on the latter. Furthermore, our research confirms that deep-learning models struggle to be practical tools from a drug discovery perspective, as they are not yet reaching the accuracy rates to be reliable. Through this, we seek to understand the role of modern computational techniques within the realm of drug discovery and ligand docking.