Metagenomic Profiling in Preterm Birth, Periodontitis, and Colorectal Cancer

Abstract

The human microbiome plays a critical role in diseases, influencing immune response, metabolism, and disease progression. Recent advances in microbiome sequencing techniques have highlighted its potential as a diagnostic, prognostic, and therapeutic strategies in various diseases, including preterm birth (Section 2), periodontitis (Section 3), and colorectal cancer (Section 4). Dysbiosis, characterized by alterations in microbiome composition, has been linked to pathogenesis, disease progression, and treatment outcome, emphasizing the need for comprehensive metagenomic analyses. By investigating microbiome profiling, researchers can uncover microbial biomarkers and host-microbiome interactions that contribute to underlying mechanisms of disease. Thus, understanding these complex relationships not only enhances early detection and risk stratification but also paves the way for microbiome-based therapeutic interventions and personalized medicine strategies. Ultimately, as microbiome research continues to evolve, its integration with genomics, metabolomics, and immunology suggests promise for transforming disease management and improving treatment outcomes. Section 2 investigated the association between the prenatal salivary microbiome and preterm birth (PTB) using 16S ribosomal RNA (rRNA) gene sequencing and developed a random forest-based prediction model for risk of preterm birth. A total of 59 pregnant women were included as the study participants, with 30 in the PTB group and 29 in the full-term birth (FTB) group. Salivary microbiome samples were collected via mouthwash within 24 hours before delivery, and 16S rRNA gene sequencing was performed to analyze microbial taxonomic composition. Differentially abundant taxa (DAT) were identified by DESeq2, revealing the 25 significant taxa, including three PTB-enriched DAT and 22 FTB-enriched DAT, suggesting distinct microbial differences upon PTB. A random forest classifier was applied to predict PTB risk based on salivary microbiome composition, achieving the high balanced accuracy (0.765±0.071, mean±SD) using the nine most important taxa. These findings indicate that salivary microbiome profiling may serve as a novel predictive tool for PTB risk assessment, complementing existing clinical predictors. Section 3 characterized salivary microbiome compositions to classify peridontal health and different stages of periodontitis using 16S rRNA gene sequencing. A total of 250 study participants were included, comprising 100 periodontally healthy controls and 150 periodontitis patients equally classified into stage I, stage II, and stage III. Microbial diversity indices were calculated, and ANCOM was used to identify 20 differentially abundant taxa among the multiple periodontitis stages. A random forest machine learning model was developed to classify periodontitis stages based on the proportions of differentially abundant taxa, achieving an area-under-curve of 0.870±0.079 (mean±SD). Among the identified differentially abundant taxa, Porphyromonas gingivalis and Actinomyces spp. were the most important features in distinguishing periodontitis stages. Random forest classifier also effectively distinguished healthy indi- viduals from stage I periodontitis with an area-under-curve of 0.852±0.103 (mean±SD) and detected periodontitis patients from healthy controls with an area-under-curve of 0.953±0.049 (mean±SD). Exter- nal validation with Spanish and Portuguese datasets showed a slight performance decrease, likely due to ethnic variations in salivary microbiome composition, emphasizing the need for population-specific models. Finally, functional pathway enrichment analysis based on DAT was performed to derive potential microbial metabolic activities associated with periodontitis stages. These findings suggest that salivary microbiome composition profiling may serve as a non-invasive diagnostic technique for periodontitis, aiding in early detection and personalized dental care. Section 4 conducted a comprehensive metagenomic analysis of colorectal cancer using PathSeq, focusing on key clinical outcomes, including recurrence history and overall survival duration. Significant differences in alpha-diversity and beta-diversity indices were observed between tumor and its adjacent normal tissues, with further stratification revealing distinct microbial diversity patterns associated with recurrence status and survival outcomes. Differentially abundant taxa were identified, highlighting microbial signatures may influence CRC progression and prognosis. To evaluate the predictive potential of these selected differentially abundant taxa, we developed a random forest-based machine learning model for CRC recurrence risk and survival duration. While the classification model for recurrence prediction achieved moderate balanced accuracy (0.570±0.164, mean±SD), and the regression model of survival duration showed moderated mean-absolute errors (729.302±179.940, mean±SD), these results suggest that gut microbiome composition alone may not be sufficient for personalized clinical predictions. These findings emphasize the need for multi-omics integration, combining host genomic alterations, e.g. somatic and germline mutations, with gut microbiome compositions, to improve CRC risk stratification and personalized medicine applications. This study highlights the potential role of gut microbiome for biomarkers in CRC diagnosis and prognosis while underscoring the complexity of host-microbiome interaction in CRC progression. Together, these studies demonstrate the clinical relevance of microbiome profiling in three distinct yet interconnected diseases by analyzing microbial diversity, identifying differentially abundant taxa, and leveraging machine learning for predictive modeling. While each condition exhibited unique microbial signatures, the findings collectively underscore the broader impact of dysbiosis on pathogenesis and disease progression. These results suggest that microbial biomarkers could serve as valuable tools for early detection, risk assessment, and personalized medicine strategies across multiple disease contexts. However, the predictive performance of machine learning models highlights the requirement for multi- omics integration, incorporating host genomic data to improve the accuracy of disease prediction and personalized therapeutic interventions. Moving forward, further large-scale and multi-cohort validation studies will be essential to refine microbiome-based biomarkers and ensure their clinical applicability in therapeutic guidance. By deepening our understanding of host-microbiome interactions, this dissertation contributes to the growing field of microbiome-driven personalized medicine, paving a novel approaches in disease prevention and management.