Co-evolution integrated deep learning framework for variants generation and fitness prediction

Abstract

Pandemic caused by viral protein is characterized by waves of transmission triggered by new variants replacing old ones, resulting in immune escape and threatening public health. Therefore, there is an obvious need to accurately identify the vital mutation sites and understand the complex patterns of mutation effect of viral protein. However, existing work do not explicitly modelling vital positions functioning for virus fitness, leading to large search space with money- and time-consuming search cost. Here, we propose EVPMM (evolutionary integrated viral protein mutation machine), a co-evolution profiles integrated deep learning framework for dominant variants forecasting, vital mutation sites prediction and fitness landscape depicting. It consists of a position detector to directly detect the functional positions as well as a mutant predictor to depict fitness landscape. Moreover, pairwise dependencies between residues obtained by a Markov Random Field are also incorporated to promote reasonable variant generation. We show that EVPMM significantly outperforms existing machine learning algorithms on mutation position detection, residue prediction and fitness prediction accuracies. Remarkably, there is a highly agreement between positions identified by our method with current variants of concern and provides some new mutation pattern hypothesis. The method can prioritize mutations as they emerge for public health concern.