Machine Learning Uncovers Transcriptional Regulatory Networks of the Human Gut Symbiont Bacteroides thetaiotaomicrons

Prof. Byung-Kwan Cho and Prof. Suhyung Cho at KI for BioInnovation, in collaboration with the University of California San Diego and KAIST colleagues, have constructed a comprehensive transcriptional regulatory framework for Bacteroides thetaiotaomicron — a key human gut symbiont. By applying independent component analysis (ICA) to 461 RNA-seq datasets, the team identified 110 independently modulated gene sets (iModulons) that explain 72.9% of transcriptome variance, significantly expanding knowledge of the organism’s regulatory networks. The study was published in Nucleic Acids Research in November 2025.

Machine-learning Driven Elucidation of Transcriptional Regulatory Networks

Despite its prominence as a model gut symbiont and growing therapeutic relevance, the global transcriptional regulatory network (TRN) of Bacteroides thetaiotaomicron VPI-5482 had remained poorly resolved, particularly the functions of its unusually large repertoire of 54 sigma factors, including 50 extracytoplasmic function sigma factors (ECF-σs) predicted to coordinate responses to environmental signals. Deciphering such a complex regulatory landscape in a non-model organism demands approaches that can extract biological signals from large, heterogeneous transcriptomic data without prior mechanistic assumptions.

To meet this challenge, the team applied independent component analysis (ICA) — an unsupervised machine learning algorithm — to a compendium of 461 RNA-seq datasets spanning stress responses, in vivo gut colonization, diverse carbon sources, and targeted genetic perturbations, including CRISPRi-mediated repression of 39 ECF-σ factors. ICA decomposes transcriptome-wide expression data into statistically independent signals, each capturing a distinct regulatory program. This decomposition yielded the BtModulome: 110 independently modulated gene sets (iModulons) that collectively explain 72.9% of total transcriptome variance, a coverage comparable to that of ICA-based frameworks built for well-characterized model organisms, validating the approach for regulatory discovery in B. thetaiotaomicron.

Expanding the Known Regulatory Network and Characterizing ECF-σ Factors

Integration of iModulon analysis with the expanded TRN enabled the identification of 1,697 regulator–regulon associations, including 311 novel predictions — representing a 22.4% expansion of the known TRN of B. thetaiotaomicron. These associations encompass 34 transcription factors, including 11 previously uncharacterized ECF-σ factors. A key example is the PerR iModulon, which revealed additional regulatory interactions including co-regulation by SigR-1 and SigH-1. The expanded PerR network incorporates stress-response components such as thioredoxins, catalase, and alkyl hydroperoxide reductases.

The team performed detailed characterization of four ECF-σ factors. SigW-1 orchestrates transcriptional regulation of a PUL59 subset encoding sulfatases critical for cleaving sulfated caps from mucosal glycans, challenging the conventional view that PUL genes are regulated solely by their own embedded ECF-σ pairs. The SigW-1 iModulon showed progressive activation during in vivo adaptation in mouse intestines. SigH-1 mediates (p)ppGpp-dependent stringent responses and regulates amino acid biosynthesis genes — a function previously unseen in related sigma factor families.

Implications for Gut Microbiome Research and Microbiota Engineering

The B. thetaiotaomicron iModulon represents a foundational resource for understanding the adaptive mechanisms of a key gut commensal. The framework demonstrates that module-centric, machine-learning-driven analysis can uncover regulatory logic invisible to traditional gene-centric approaches, particularly in non-model organisms. The complete dataset, including iModulon gene memberships, activity metrics, and experimental conditions, is publicly available, providing a platform for hypothesis-driven research in gut microbiome biology. The modular structure of iModulons is expected to facilitate rational engineering of for therapeutic applications.

This work was supported by the Korea Bio Grand Challenge (RS-2018-NR029581), the Bio and Medical Technology Development Program (RS-2021-NR056597 and RS-2021-NR056566) through the National Research Foundation of Korea (NRF), and the Basic Science Research Program through NRF (RS-2023-00246928). This work was also funded by the KAIST Jang Young Sil Fellow Program.