Enhancing Challenging Target Screening via Multimodal Protein-Ligand Contrastive Learning

Recent advancements in genomics and proteomics have identified numerous clinically significant protein targets, with notably 85% categorized as undruggable. These targets present widespread challenges due to their complex structures and dynamics, rendering conventional drug design strategies not always effective. In this study, we introduce Uni-Clip, a contrastive learning framework that incorporates multi-modal features of proteins (structure and residue) and ligands (conformation and graph). Optimized with a specifically designed CF-InfoNCE loss, Uni-Clip enhances the modeling of protein-ligand interactions for both undruggable and druggable proteins. Uni-Clip demonstrates superior performance in benchmark evaluations on widely acknowledged datasets, LIT-PCBA and DUD-E, achieving a 147% and 218% improvements in enrichment factors at 1% compared to baselines. Furthermore, Uni-Clip proves to be a practical tool for various drug discovery applications. In virtual screening for the challenging protein target GPX4 with flat surface, it identified non-covalent inhibitors with an IC50 of 4.17 M, in contrast to the predominantly covalent inhibitors currently known. Through target fishing for benzbromarone, Uni-Clip identified the intrinsically disordered protein c-Myc as a potential target, highlighting benzbromarones potential for repurposing in cancer therapy. Explainable analyses effectively identified binding sites consistent with molecular dynamics and experimental results, even for challenging undruggable targets.