Computational discovery and characterization of microbial functions in health and disease
Open Access DepositedMicrobiome profiling offers a rich source of biomarkers for precision medicine, yet reported signatures frequently fail to replicate across studies. This fragility reflects three recurring obstacles
immune checkpoint inhibitor (ICI) response in advanced melanoma. A harmonized meta-analysis of stool shotgun metagenomes from 15 cohorts, including trials combining fecal microbiota transplantation with ICI, integrates species-level composition, functional pathways, and biosynthetic gene cluster (BGC) features. The results reveal context-dependent yet reproducible associations, while also showing that cross-study prediction remains modest, underscoring heterogeneity in microbiome contributions to ICI efficacy and prioritizing candidates for mechanistic follow-up. Finally, to strengthen the link from association to function, the dissertation introduces \textit{bgcLens}, a nucleotide-level genomic language-model framework for detecting biosynthetic gene cluster loci directly from raw genome sequences. By complementing rule-based genome-mining pipelines and reducing dependence on upstream annotation, \textit{bgcLens} supports scalable functional microbiome genomics and helps translate microbiome signals into testable genomic hypotheses. These contributions provide an end-to-end framework for AI-enabled microbiome biomarker discovery
establish robust baselines, quantify reproducible signals across cohorts, and connect statistical patterns to mechanistic genomic candidates.
substantial baseline variability across populations and body sites, technical and analytical heterogeneity that introduces batch effects, and an incomplete bridge from statistical associations to mechanistic, function-level hypotheses. The work in this dissertation addresses these challenges by developing computational strategies that improve reproducibility while moving biomarker discovery toward actionable biology, progressing from principled modeling and baseline definition to multi-cohort synthesis, and ultimately to genome-level functional mining. The dissertation begins by consolidating practical guidance for machine-learning-enabled biomarker discovery in high-dimensional omics, emphasizing study design, harmonization, external validation, and interpretation to support trustworthy and reproducible inference. Building on these principles, it then quantifies baseline structure in the upper-airway microbiome by profiling paired oral and nasal communities from healthy adults across four winter (January) sampling campaigns (2019, 2020, 2022, and 2023). These data demonstrate strong, consistent niche differentiation between oral and nasal ecosystems, alongside measurable year-to-year variation within each site, providing a needed reference for interpreting microbiome shifts in disease and intervention settings. With baseline variability explicitly characterized, the dissertation next tackles cross-study reproducibility in a clinically consequential application
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.