Machine learning-based diagnosis of inflammatory bowel disease using gut microbiome virulence factor profiles: multicohort integration and independent validation

Inflammatory bowel disease (IBD) is a chronic intestinal disease that includes Crohn’s disease (CD) and ulcerative colitis (UC). IBD severely affects the quality of life of millions of patients worldwide. Consequently, the timely diagnosis and treatment of IBD are critically important. However, current diagnostic methods are predominantly invasive and costly. In recent years, metagenomics-based studies, utilizing cross-cohort analysis, have successfully constructed biomarkers based on microbial composition or specific gene functions, enabling highly accurate, non-invasive diagnosis of IBD. The decreasing cost of metagenomic sequencing has increased its potential for the development of non-invasive microbiome-based diagnostic and risk-stratification tools. Recent studies have indicated that the structure and function of the gut microbial community are significantly altered in patients with IBD. Moreover, the abundance of certain virulence factors (VFs) is much greater in these patients than in healthy controls. High-resolution metagenomic data can aid in understanding the pathogenesis of IBD and have the potential to become a new generation of non-invasive diagnostic tools. This study established a Discovery cohort by integrating eight metagenomic cohorts. This cohort comprised 1,062 fecal samples, including 737 from patients with IBD and 325 from healthy controls. This study implemented a standardized virulence factor profiling workflow, termed VFfinder, to quantify gut microbial virulence factor profiles across multiple metagenomic cohorts. Using the species and virulence factor abundance matrices, we developed three machine learning models for IBD classification: the Species Model, the VFs Model, and the IBDseek, which integrates both feature types. The Validation cohort, comprising three metagenomic datasets with 170 IBD samples and 125 healthy controls, was used to evaluate the performance and generalizability of each model. Virulence factors in the gut microbiome were significantly enriched in patients with IBD. In the Validation cohort, the 79-species Species Model achieved an AUC of 0.94 and an accuracy of 88.81%. The VFs Model, a 45-VF-based machine learning model, achieved a sensitivity of 96.47% (95% CI, 92.48–98.69%) and an accuracy of 89.49% (95% CI, 85.42–92.75%) in the Validation cohort. Furthermore, the IBDseek Model, integrating 45 VF features and 79 species features, achieved a sensitivity of 96.47% (95% CI, 92.48–98.69%), specificity of 81.60% (95% CI, 73.68–87.96%), and accuracy of 90.17% (95% CI, 86.19–93.32%). In summary, our study demonstrates the cross-cohort enrichment of bacterial virulence factors in patients with IBD and supports their utility as stable diagnostic features. The IBDseek Model, which integrates VFs and species features, achieved the highest observed accuracy and shared the highest sensitivity with the VFs Model in the Validation cohort. These findings support the potential of virulence-associated microbial features as informative biomarkers for noninvasive IBD classification and provide a basis for the future development of microbiome-based diagnostic support and risk stratification tools. Prospective evaluation in clinically relevant populations is needed to establish their clinical utility and integration into existing diagnostic pathways. Not applicable.