Host–pathogen interactions and pathogen evolution are underpinned by protein–protein interactions between viral and host proteins. An understanding of how viral variants affect protein–protein binding is important for predicting viral–host interactions, such as the emergence of new pathogenic SARS-CoV-2 variants. Here we propose an artificial intelligence-based framework called UniBind, in which proteins are represented as a graph at the residue and atom levels. UniBind integrates protein three-dimensional structure and binding affinity and is capable of multi-task learning for heterogeneous biological data integration. In systematic tests on benchmark datasets and further experimental validation, UniBind effectively and scalably predicted the effects of SARS-CoV-2 spike protein variants on their binding affinities to the human ACE2 receptor, as well as to SARS-CoV-2 neutralizing monoclonal antibodies. Furthermore, in a cross-species analysis, UniBind could be applied to predict host susceptibility to SARS-CoV-2 variants and to predict future viral variant evolutionary trends. This in silico approach has the potential to serve as an early warning system for problematic emerging SARS-CoV-2 variants, as well as to facilitate research on protein–protein interactions in general.
Based on a newly developed deep-learning model for predicting the binding affinities of protein–protein interactions, the effects of SARS-CoV-2 spike protein variants on binding to the ACE2 receptor and neutralizing antibodies were analyzed and used to predict immune escape and viral evolution.