AI-driven predictions of binding trends of SARS-CoV-2 variants from atomistic simulations

Abstract

Binding processes are fundamental for cellular functions such as immune response activation, cell regulation, and signal transduction among others. However, experimental measures of binding free energies often depend on the system set up while in silico calculation at atomistic level of the binding process can be computationally demanding due to the long timescale of typical binding/unbinding events. The aim of our study is to leverage machine learning approaches to address this challenge and estimate binding affinity trends between two proteins using short atomistic simulations. Our technique uses a neural network algorithm applied to a series of images generated by the simulation data and representing the distance between two molecules in time. The algorithm is capable of distinguishing with high accuracy low vs high binding affinity of non-hydrophobic mutations, indicating that our method excels on the inference of the binding affinity trends for charged and/or polar amino acid mutations. Moreover, it shows high accuracy in prediction using a small subset of the simulated data. We apply our algorithm to the binding between several variants of the SARS-CoV-2 spike protein and the human receptor ACE2. *We acknowledge support from the IBM Research AI Hardware Center, and the Center for Computational Innovation at Rensselaer Polytechnic Institute for computational resources on the AiMOS Supercomputer. This material is based upon work supported by the National Science Foundation under Grant No. DBI-1548297