Analysis and classification of human microbiomes: detection of bioindicators and optimisation through machine learning

Our understanding of human-associated microorganisms has evolved substantially since Theodor Escherich first documented Escherichia coli in the gut flora of healthy children. Today, we recognise these microorganisms as crucial contributors to human homeostasis. Among the techniques that emerged to study these microorganisms, metagenomics enabled the sequencing of all DNA within environmental samples without the need for culturing. In particular, the recovery of metagenome-assembled genomes (MAGs) allows direct genome investigation of previously uncultured organisms in their environmental context. Genome-resolved metagenomics allows the connection of functional potential to specific microorganisms, capturing subtle genomic adaptations to specific ecological contexts and providing a powerful approach to identify precise bioindicators with implications for epidemiology, drug discovery, and personalised medicine, for instance. However, despite the years of progress, the transformation of metagenomic data into biological insights is complex, causing the samples to be underused and relevant patterns to be overlooked. Metadata inconsistencies hinder systematic sample selection, genome recovery requires complex computational pipelines, and the high-dimensional nature of metagenomic data constraints comparative analysis. This dissertation addresses these challenges through integrated computational frameworks that bridge the critical gap between exponential data generation and actionable scientific insights while advancing our understanding of the composition and function of the human microbiome. First, we developed the HumanMetagenomeDB, a curated database of standardised metadata from 69,822 human metagenome samples collected from public repositories. This resource enabled systematic sample selection based on host characteristics, medical conditions, and technical parameters, focusing on facilitating large-scale meta-studies. The database also revealed significant geographical biases in microbiome sampling and gaps in disease coverage, highlighting areas that require additional research effort. Building on this foundation, we created MuDoGeR (Multi-Domain Genome Recovery), a streamlined, reproducible pipeline for genome recovery at scale, reducing the genome recovery from metagenomes technical barriers for non-bioinformaticians. Afterwards, to ensure reliable results for genome-centric studies, we conducted a systematic evaluation using simulated communities that revealed factors influencing genome recovery success. The evaluation established an optimal sequencing depth of 60 million reads and demonstrated that current methods can achieve up to 92\% recovery precision while highlighting specific limitations in recovering closely related species. After recovering the genomes, we next focused on streamlining comparative genomic analysis by developing gSpreadComp. In particular, gSpreadComp focuses on analysing antimicrobial resistance genes (ARGs) and virulence factors (VFs). We used human gut microbiome samples from subjects with different diets as a use case. gSpreadComp revealed that while overall resistance patterns were similar across diets, specific differences emerged - such as increased tetracycline resistance in omnivores and elevated bacitracin resistance in vegans. Notably, vegan and vegetarian diets showed significantly higher potential for plasmid-mediated horizontal gene transfer. gSpreadComp focuses on guiding researchers in generating testable hypotheses. The culmination of this work is an approach for biologically interpretable feature selection that combines autoencoder-based dimensionality reduction with density-based clustering to encode the functional potential of MAGs. Throughout the process, and as a consequence of it, we also assembled a comprehensive collection of 514,932 prokaryotic MAGs of the human microbiome, representing 6,763 bacterial and 31 archaeal species. The resulting interactive AutoML-driven platform allows researchers to explore microbiome patterns, generate hypotheses, and select relevant MAGs for further investigation. Naturally, the frameworks and databases developed here contribute to transforming metagenomic data into biological insights. However, the evolution toward more integrative and translational applications requires continued interdisciplinary collaboration between microbiologists, computer scientists, and clinical researchers.