Modeling of Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) Proteins by Machine Learning and Physics-Based Refinement

Lim Heo; Michael Feig

When comparing the models from our protocol with the (refined) AlphaFold models and the models from the Zhang group, we find that the resulting models do not reach a high degree of consensus for most of the modeled proteins (Figures 1-7) . This may be expected as the modeled proteins were very difficult to predict. However, there is consensus on some of the proteins. For the domain of Papainlike proteinase (PL-PRO), residues 1763-1927, ours and AlphaFold's model resemble each other with a Cɑ-RMSD of 2.96 Å. Moreover, for one of the domains of nsp4, residues 274-399, ours and AlphaFold's model have the same topology, an α-helix bundle with the only significant difference being the orientation of the N-terminal helix (residues 274-309) that led to an overall difference between the structures of 8.27 Å in Cɑ-RMSD. Moreover, the membrane glycoprotein structure was also predicted with a similar topology between our protocol and AlphaFold with presumed trans-membrane Nterminal helices in a similar orientation and a globular C-terminal domain for residues 117-203 with a similar β-strand topology. The difference between our model and the AlphaFold prediction was 8.69 Å in Cɑ-RMSD mainly because of the orientation of two beta-strands (residues 117-133). Except for these cases, none of the structures shared a same topology. The copyright holder for this preprint (which was not peer-reviewed) is the . https://doi.org/10.1101/2020.03.25.008904 doi: bioRxiv preprint