Prediction of SARS-CoV-2 spike protein mutations using Sequence-to-Sequence and Transformer models

Abstract

In the study of viral epidemics, having information about the structural evolution of the virus can be very helpful in controlling the disease and making vaccines. Various deep learning and natural language processing techniques (NLP) can be used to analyze genetic structure of viruses, namely to predict their mutations. In this paper, by using Sequence-to-Sequence (Seq2Seq) model with Long Short-Term Memory (LSTM) cell and Transformer model with the attention mechanism, we investigate the spike protein mutations of SARS-CoV-2 virus. We make time-series datasets of the spike protein sequences of this virus and generate upcoming spike protein sequences. We also determine the mutations of the generated spike protein sequences, by comparing these sequences with the Wuhan spike protein sequence. We train the models to make predictions in December 2021, February 2022, and October 2022. Furthermore, we find that some of our generated spike protein sequences have been reported in December 2021 and February 2022, which belong to Delta and Omicron variants. The results obtained in the present study could be useful for prediction of future mutations of SARS-CoV-2 and other viruses.