Optimizing variant-specific therapeutic SARS-CoV-2 decoys using deep-learning-guided molecular dynamics simulations

Treatment of COVID-19 with a soluble version of ACE2 that binds to SARS-CoV-2 virions before they enter host cells is a promising approach, but it needs to be optimized and adapted to emerging viral variants. The computational workflow presented here consists of molecular dynamics simulations for RBD-ACE2 binding affinity assessments of ACE2 or RBD variants and a novel convolutional neural network architecture working on pairs of voxelized force-fields for efficient search-space reduction. We identified hACE2-Fc K31W along with multi-mutation variants as high-affinity candidates, which we also validated in vitro with virus neutralization assays. We evaluated binding affinities of these ACE2 variants with the RBDs of Omicron BA.3, Omicron BA.4/BA.5, and Omicron BA.2.75 in silico. In addition, candidates produced in Nicotiana benthamiana, an expression organism for potential large-scale production, showed a 4.6-fold reduction in half-maximal inhibitory concentration (IC50) compared with the same variant produced in CHO cells and an almost six-fold IC50 reduction compared with wild-type hACE2-Fc.