Integrated Machine Learning and Metagenomics Approaches for Predicting Microbial Environmental Niches From High-Dimensional Functional Genomic Data

Predicting the environmental niches of microbial species from their genomic functional content is a fundamental open problem in microbial ecology. This thesis presents the first large-scale, globally scoped machine learning framework designed to address this problem using metagenome-assembled genomes (MAGs) and their associated functional annotations. We frame niche prediction as a positive-unlabeled, multilabel classification problem over a high-dimensional functional feature space (p = 26,150 features; n (train) = 9144, n (test) = 2335 samples). Two interpretable classifiers, logistic regression and random forest, are trained and evaluated against empirical chance-level baselines under a rigorous grouped, stratified cross-validation scheme designed to prevent data leakage from non-independent samples. The study explicitly accounts for known sources of bias in metagenomic research, including taxonomic imbalance between bacteria and archaea, gene functional annotation incompleteness, and environmental sampling bias. Despite these challenges, our results demonstrate that functional genomic data carries a learnable signal for environmental niche prediction across 26 habitat classes. Each of these describes the environments from which each species was collected, defined using ontology terms. Feature importance analysis reveals that environmental discrimination is supported by a broad functional signal spanning both annotated KEGG pathways and a substantial proportion of genes lacking assignment to curated pathways. A focused analysis of the human gut model identifies biologically plausible functions associated with host adaptation, providing interpretable insights into the functional basis of niche prediction. The framework is validated on an independent dataset, and its limitations and generalizability are discussed in depth. We provide a documented pipeline that can serve as a benchmark and foundation for future work in computational microbial ecology.