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

Lim Heo; Michael Feig

Protein models were generated initially from inter-residue distance predictions rather than templatebased modeling because of a lack of experimentally determined close homolog structures. The resulting models were then further refined by a molecular dynamics simulation-based refinement protocol to improve the physical realism at the atomistic level of the structures. We also applied the same refinement protocol to the models predicted by DeepMind's AlphaFold method. The detailed procedures are described in METHOD section. We compared our models with the other available models, i.e. the original AlphaFold models and the predictions from the Zhang group. As shown in Table 1 , our models and the models from the Zhang group provide more complete sequence coverage than the AlphaFold models.