De novo 3D models of SARS-CoV-2 RNA elements and small-molecule-binding RNAs to guide drug discovery

Ramya Rangan; Andrew M. Watkins; Wipapat Kladwang; Rhiju Das

While there are not yet extended benchmarks for using FARFAR2 model ensembles for small molecule virtual screening, recent RNA-puzzles blind prediction trials have shown that FARFAR2 models in combination with conservation information enables manual identification of ligand binding sites in 3D models of bacterial 'riboswitch' aptamers at nucleotide resolution. [11] [12] [13] To guide use of FARFAR2 models for virtual screening, we have therefore collected a data set termed FARFAR2-Apo-Riboswitch, containing models of RNA elements that are known to bind small molecules, depicted in Fig. 5 . These targets include binders for diverse ligands, from S-adenosyl methionine (SAM) to glycine, cobalamin, 5-hydroxytryptophan, the cyclic dinucleotides c-di-AMP (the ydaO riboswitch), the alarmone nucleotide ZMP (AICAR monophosphate), glutamine, and guanidinium. Each RNA was modeled without ligands present during fragment assembly or refinement, to mimic the protocols that would be used in virtual drug screening in which modeling of apo RNA structures are used for computational docking of ligands. Three of these model sets for SAM-I, SAM-I/IV, and SAM-IV made use of homology to previous riboswitch structures' ligand binding sites (Homology, Figs. 5A-C), for historical reasons: the actual RNA-Puzzles challenges (or in the case of SAM-IV, an 'unknown RFAM' challenge for the RNA-Puzzles, involving tests based on cryoEM 27 ) were posed at times in which crystal structures of riboswitch aptamers with homologous SAM binding sites were available. These model sets should therefore serve as 'positive controls' for virtual drug screening protocols, which should unambiguously identify homology-guided SAM binding sites as good aptamers for SAM.