Model evolution in SARS-CoV-2 spike protein sequences using a generative neural network

Modelling evolutionary elements inherent in protein sequences, emerging from one clade into another of the SARS-CoV-2 virus, would provide insights to augment our understanding of its impact on public health and may help in formulating better strategies to contain its spread. Deep learning methods have been used to model protein sequences for SARS-CoV-2 viruses. A few significant drawbacks in these studies include being deficient in modelling end-to-end protein sequences, modelling only those genomic positions that show high activity and upsampling the number of sequences at each genomic position for balancing the frequency of mutations. To mitigate such drawbacks, the current approach uses a generative model, an encoder-decoder neural network, to learn the natural progression of spike protein sequences through adjacent clades of the phylogenetic tree of Nextstrain clades. Encoder transforms a set of spike protein sequences from the source clade (20A) into its latent representation. Decoder uses the latent representation, along with Gaussian distributed noise, to generate a different set of protein sequences that are closer to the target clade (20B). The source and target clades are adjacent nodes in the phylogenetic tree of different evolving clades of the SARS-CoV-2 virus. Sequences of amino acids are generated, for the entire length, at each genomic position using the latent representation of the amino acid generated at a previous step. Using trained models, protein sequences from the source clade are used to generate sequences that form a collection of evolved sequences belonging to all children clades of the source clade. A comparison of this predicted evolution (between source and generated sequences) of proteins with the true evolution (between source and target sequences) shows a high pearson correlation (> 0.7). Moreover, the distribution of the frequencies of substitutions per genomic position, including high- and low-frequency positions, in source-target sequences and source-generated sequences exhibit a high resemblance (pearson correlation > 0.7). In addition, the model partially predicts a few substitutions at specific genomic positions for the sequences of unseen clades (20J (Gamma)) where they show little activity during training. These outcomes show the potential of this approach in learning the latent mechanism of evolution of SARS-CoV-2 viral sequences.