Abstract:
The rapid mutation of SARS-CoV-2 leads to the emergence of new strains. Previous studies have shown that the binding affinity between the S protein receptor binding domain (Receptor Binding Domain, RBD) of SARS-CoV-2 and host ACE2 is related to the infection ability of the virus. With the continuous outbreak of SARS-CoV-2 in the world, a large number of variants with multiple mutations in the RBD region have emerged. It is time-consuming and laborious to obtain the RBD-ACE2 binding affinity of variants by bioassay, which lags far behind the accumulation of variants, and cannot meet the demand for real-time surveillance of this virus. In order to rapidly predict the binding affinity of strains with multiple mutations, a deep neural network model is designed. This model learns the key features from RBD sequences to predict the RBD-ACE2 binding affinity by combining convolutional neural network, recurrent neural network and attention mechanism. The model is trained and evaluated on real data sets. The results demonstrate that the model can effectively predict the RBD-ACE2 binding affinity of variant of concern, and monitor the transmission ability of SARS-CoV-2 mutant strains