Protein-Ligand Binding Site Prediction and de Novo Ligand Generation from Cryo-EM Maps

Abstract

Identification of protein-ligand binding sites is one of the most challenging tasks in drug discovery and design. Recent advances in machine learning community, particularly deep learning, inspired considerable research into deep learning-based methods for protein ligand binding site prediction (PLBP) and have achieved promising results. However one limitation of these methods and models is that they are developed and evaluated using conventional databases consisting of relatively small protein structures determined from X-ray crystallography and NMR with high resolution. When current PLBP methods are directly applied to large protein complexes, they either fail or produce very poor results. Given the increasing popularity of protein structures determined from Cryo-EM maps, which usually capture large protein complexes at relatively low resolution, there is a strong need to apply PLBP methods to Cryo-EM maps. For this purpose, we created a novel database (EMD-Ligand Dataset) consisting of Cryo-EM maps and information about their ligand-binding partners. Our study of several state-of-the-art PLBP methods show that even though they perform reasonably well on conventional databases, they do poorly on the novel EMD-Ligand Dataset, which calls for more research and work to be done by the community. As current methods for determining protein structures experimentally from Cryo-EM maps are time-consuming and expensive, we integrated PLBP methods with DeepTracer, an automatic and fast, de novo Cryo-EM protein structure modeling method, and established an end-to-end system capable of predicting ligand binding sites directly from Cryo-EM maps. Additionally, we leveraged an existing ligand generator to generate drug-like ligands from our predicted ligand-binding pockets and demonstrated its effectiveness.