A Fine-tuned ProtGPT2 (transformer model) for Predicting more Virulent SARS-CoV-2 variants and understanding its virulence by biophysical methods.

Abstract

The emergence of Variants of Concern (VOCs) of SARS-COV2 with increased virulence and transmissibility has been linked to multiple mutations in the RBD region, altering their antigenic properties. In this study, we used a specialized ProtGPT2 model trained on the SARS-COV2 spike protein to forecast probable mutations on the spike protein that have not yet emerged. Upon prediction, we systematically studied the stability of single-site and multisite mutations using unbiased molecular dynamics simulations. Binding free energies were used to study the physicochemical significance of the mutations and their affinity to human ACE2 receptor. Our results show that there are specific hot-spots that mutate in the spike protein that enhance binding affinity through electrostatic and improved non-bonded interactions and highlight the role of specific energetic contributions in viral adaptation and infectivity. Our analysis revealed that the reduction of a disulphide bridge within sites 480-488 lowered the binding free energy and increased the flexibility of the loop region, enhancing its interface interaction with ACE, leading to a more virulent variant than Omicron.