Feature Learning of Virus Genome Evolution With the Nucleotide Skip-Gram Neural Network

• Published in: Evolutionary bioinformatics online Vol. 15; p. 1176934318821072
• Main Author: Shim, Hyunjin
• Format: Journal Article
• Language: English
• Published: London, England SAGE Publications 01.01.2019; Sage Publications Ltd; SAGE Publishing
• Subjects: Algorithm development for evolutionary biology; Algorithms; Alleles; Artificial intelligence; Artificial neural networks; Biology; Cluster analysis; Clustering; Datasets; Disinfectants; Evolution; Experiments; Gene frequency; Gene sequencing; Genomes; Machine learning; Mutation; Natural language processing; Neural networks; Nucleotides; Population; Principal components analysis; Recombination; Source code; Stochastic processes; Viruses; VP1 protein; feature learning; experimental evolution; neural network; allele embeddings; virus evolution; genetic interaction
• ISSN: 1176-9343; 1176-9343
• DOI: 10.1177/1176934318821072

Recent studies reveal that even the smallest genomes such as viruses evolve through complex and stochastic processes, and the assumption of independent alleles is not valid in most applications. Advances in sequencing technologies produce multiple time-point whole-genome data, which enable potential interactions between these alleles to be investigated empirically. To investigate these interactions, we represent alleles as distributed vectors that encode for relationships with other alleles in the course of evolution and apply artificial neural networks to time-sampled whole-genome datasets for feature learning. We build this platform using methods and algorithms derived from natural language processing (NLP), and we denote it as the nucleotide skip-gram neural network. We learn distributed vectors of alleles using the changes in allele frequency of echovirus 11 in the presence or absence of the disinfectant (ClO2) from the experimental evolution data. Results from the training using a new open-source software TensorFlow show that the learned distributed vectors can be clustered using principal component analysis and hierarchical clustering to reveal a list of non-synonymous mutations that arise on the structural protein VP1 in connection to the candidate mutation for ClO2 adaptation. Furthermore, this method can account for recombination rates by setting the extent of interactions as a biological hyper-parameter, and the results show that the most realistic scenario of mid-range interactions across the genome is most consistent with the previous studies.