Equivariant graph neural network-based accurate and ultra-fast virtual screening of small molecules targeting miRNA-protein complex

MicroRNAs (miRNAs) are small RNA molecules with significant therapeutic potential for treating various diseases, underscoring the need for effective methods to screen drugs targeting disease-associated miRNAs. In this study, we introduce miRPVS, a rapid virtual screening approach designed to identify small molecule drugs targeting miRNA-protein complex. miRPVS identifies binding pockets on the surface of these complexes, expanding the scope of potential small molecule targets. It employs an equivariant graph neural network model to extract 3D structure features of small molecules, enabling accurate prediction of docking scores. Using miRPVS, four complexes involved in pri-miRNA cleaving, pre-miRNA transport, and mRNA depress were identified as promising targets. For each target, hit compounds were screened from the ZINC20 database, which contains approximately 600 million druglike small molecules. MiRPVS predicted the docking score for these compounds, with Pearson correlation coefficients between predicted and experimentally docked scores comparable to those obtained through twice docking. Notably, the average deviation was only 0.67% across the four complexes. Remarkably, the entire screening process for all four complexes was completed in 14 h using just four V100 GPUs. Additionally, we integrated AlphaFold3-predicted structures into the miRPVS workflow, enabling virtual screening of small molecules against miRNA-protein complexes without experimentally determined structures. miRPVS demonstrated performance comparable to traditional docking methods while significantly reducing computational time and resource requirements. This innovative approach holds great promise for accelerating the discovery of small molecule drugs targeting miRNA-regulated pathways, addressing a critical gap in miRNA therapeutics.