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

For each simulation, we additionally report a convergence estimate in Table 1 , estimated as the mean pairwise RMSD of the top 10 cluster centers predicted by FARFAR2. Prior work aiming at accurate prediction of single native crystal structures has demonstrated that convergence is a predictor for modeling accuracy, 10, 15-16 with prior tests suggesting that models that have 7.5 Ã… convergence or lower have mean single-structure prediction accuracy of at least 10 Ã…, and models with 5 Ã… convergence or lower have single-structure prediction accuracy of at least 8 Ã…. 10 In this work, we are however not assuming that the RNA targets form a single 'native' structure. Instead, we take this convergence measure as a proxy for whether sampling may have been adequate to generate a useful model set. As a more direct measure of the thoroughness of sampling, we also present the occupancies of each of the top 10 clusters. Conformations sampled repeatedly in independent Rosetta-FARFAR2 runs (cluster membership greater than 1) indicate some level of convergence in sampling and those conformations are more likely to be realistic low-energy structures. In Figs. 1-4 below, we show cluster members as a cloud of translucent structures behind each cluster's lowest energy conformation to visually convey the level of convergence; lack of such a cloud indicates a 'singlet' in which the cluster involves only one member. We include up to 50 representative top-scoring models in each cluster as the model collection for each RNA element, with structures available in the Github repository: https://github.com/DasLab/FARFAR2-SARS-CoV-2.