Abstract
Virtual screening of large compound libraries against specific RNAs offers a promising and cost-effective approach for identifying novel lead compounds. Developing a reliable scoring function for RNA-ligand interactions is crucial for effective computational drug screening. However, current RNA-ligand scoring functions are primarily designed to predict native binding poses for given RNA-ligand pairs and have not been thoroughly evaluated for their virtual screening capabilities. Here, by leveraging more accurate descriptions of stacking, solvation, conformational flexibility, and other key effects, we developed RLDOCKScore, an enhanced scoring function optimized for both pose prediction and virtual screening. RLDOCKScore demonstrates overall better performance compared with other tested scoring functions, effectively balancing pose prediction accuracy with virtual screening capabilities. In pose prediction evaluations using 122 RNA-ligand complexes, RLDOCKScore outperforms all of the tested scoring functions except SPRank. For virtual screening applications, RLDOCKScore showed overall better performance than other methods when tested against the HIV-1 TAR ensemble and four riboswitches from the ROBIN benchmark set, with the best top-2% enrichment factor and the AUC value reaching 25.0 and 0.86, respectively, for the HIV-1 TAR ensemble. These results establish RLDOCKScore as a valuable new method for computational drug discovery applications.