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

In each of these RNA elements, stereotyped configurations of apical loops are modeled by Rosetta-FARFAR2 (Figs. 1B-G). However, it is important to note that our modeling of SL2 does not recover structures with the atomic level features of the previously reported canonical CUYG tetraloop, whose structure was determined by NMR for the SARS-CoV-1 sequence 4 (Fig. 1C ). We present a homology model of SL2 based on the SARS-CoV-1 sequence as an additional cluster in our data set ( Fig. 1C , see Methods). Our de novo FARFAR2 models approach 3.1 Ã… RMSD to this homologydirected model, consistent with our convergence estimate above. Nevertheless, there is still an atomic level discrepancy between the NMR structure and FARFAR2 clusters. This difference likely reflects limitations in Rosetta's high-resolution energy function, with the E-gap value for the homology-directed SL2 structure over 20 R.E.U. worse than the top-scoring de novo model. The primary issue is that the Rosetta energy function assigns un-stacked base pairs, as in the C and G nucleotides of the NMR structure of the CUYG tetraloop, a substantial desolvation penalty without stacking bonuses to compensate. The fine structural discrepancy may additionally reflect limitations in Rosetta FARFAR2 sampling, which might be resolved by bringing to bear more computational expensive sampling methods like stepwise Monte Carlo (SWM). 23 The SL5 element is a long stem loop in all betacoronaviruses whose tip has been elaborated into a 4-way junction in SARS-CoV-2 and related subgroups. 2 Due to the larger size and complexity of SL5, our 3D models for this domain did not converge sufficiently, i.e., each of the top 10 lowest energy models were 'singlets' with no other conformations discovered within 5 Ã… RMSD among the top 400 lowest energy models (Table 1) . Nevertheless, the modeling did suggest the potential for drug-binding pockets between helices that are brought into proximity by the four-way junction ( Fig. 2A) . It was of interest to test whether the secondary structure of SL5 might be well-defined even if . CC-BY 4.0 International license author/funder. It is made available under a The copyright holder for this preprint (which was not peer-reviewed) is the . https://doi.org/10.1101/2020.04.14.041962 doi: bioRxiv preprint our 3D modeling did not converge. To test the secondary structure of this stem-loop, we collected DMS and SHAPE data for this construct (see Methods). Secondary structure predictions from RNAstructure 24 guided by these DMS and SHAPE reactivity data (Fig. 2B ) recovered the same secondary structure developed from homology modeling to SL5 elements in other betacoronaviruses, as depicted in Fig. 1A , supporting the accuracy of this model as the dominant secondary structure. Fig. 3 presents Rosetta-FARFAR2 models for the SARS-CoV-2 frameshifting element (FSE). In SARS-CoV-1, the FSE pseudoknot structure and its dimerization domain have been shown to be critical for a (-1) ribosomal frameshifting event that leads to the production of ORF1a and ORF1b proteins from the same genomic region (Fig. 3A) . For this reason, we present models for both the monomer frameshifting element along with a dimerized system. We generated over 200,000 FARFAR2 models for the monomeric frameshifting element, with 100,000 models separately modeled with each of two similar secondary structures reported in the literature, which differ by a single base pair (Fig. 3) . 1, 25 The frameshifting element sim