Machine learning prediction in genomic sequences of prokaryotic viruses from metagenomic datasets

• Ano de defesa: 2022
• Autor(a) principal: Amgarten, Deyvid Emanuel
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Aprendizado de máquina; Bacteriófagos; Fagos; Host prediction; Machine learning; Metagenômica; Metagenomics; Phage prediction; Phages; Predição de hospedeiro viral; Prokaryotic viruses; Virology; Virus; Vírus; Vírus ambientais; Vírus de procariotos
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/95/95131/tde-17022022-091454/

Environmental viruses are extremely diverse and abundant in the biosphere. Several studies have shown prokaryotic viruses (or simply phages) as major players in determining biogeochemical cycles in oceans as well as driving microbial diversification. Besides this ecological role, phages may also be used for clinical purposes since they can kill bacterial cells and terminate infections. A crucial step in this process is the isolation of new phages, which can target a specific bacterial pathogen. Thus, researchers employ screening techniques to find and isolate pathogen-specific phages from environmental samples, which are a rich source of new phages. However, this task remains mostly exploratory and laborious if the researcher has no detailed information about the sample and its potential viral diversity. Having this problem in mind, we propose the development of a bioinformatic workflow to identify genomic sequences belonging to phages in environmental datasets, as well as for host prediction of the identified phages based on their genomic sequences. To achieve this goal, we implemented a random forest classifier and created the tool named MARVEL (Metagenomic Analyses and Retrieval of Viral Elements), which is able to efficiently predict phage genomic sequences in bins generated from whole community metagenomic short reads. We also developed a toolkit, name vHULK (Viral Host Unveiling Kit), which can predict phages host given only their genome as input. vHULK presents higher accuracy than available tools and it can predict both host species and genus in a multiclass prediction setting. Data generated by the application of both tools in public and private composting metagenomic datasets is used for recovery, annotation, and characterization of phage diversity in composting environments. Both tools are publicly available through a GitHub repository: https://github.com/LaboratorioBioinformatica/.