Can oral samples tell us about Crohn's disease in the gut?
MSc thesis · UMCG, Groningen · Jul–Dec 2025
- The question
- Fecal samples are unpleasant to give and slow to process. Oral samples are easy to collect. Could a model built on oral microbiome data help spot Crohn's disease, and does a model trained on one body site still work on the other?
- What I did
- I built a Snakemake pipeline in Python and R for 256 oral and fecal metagenomic profiles from three cohorts (UMCG, Dutch Microbiome Project, Human Microbiome Project). It covers filtering, CLR transformation, diversity analysis, differential abundance adjusted for medication and other covariates, and four classifiers with nested cross-validation.
- What I found
- Within one body site the models scored almost perfectly (AUC close to 1.0). I don't trust that number much, because patients and controls came from different cohorts. The more useful result came from training on fecal samples and testing on oral samples. Only regularised logistic regression carried over (AUC 0.88), and the tree models dropped below chance. After adjustment, 204 fecal taxa still differed between patients and controls, and no oral taxa did.
- So what
- Oral samples alone are unlikely to replace fecal samples for Crohn's screening. For me the bigger lesson was that a high AUC is not enough: checking for cohort effects and testing on other data is what tells you whether a model can be used.
Patient data is covered by data-use agreements, so the repository only has the pipeline and aggregate results.
