Machine Learning Identifies Gut Biomarkers for IBD

BERLIN — Gut microbial biomarkers identified using machine learning can differentiate patients with inflammatory bowel disease (IBD) from healthy control individuals, according to a study presented at the European Crohn’s and Colitis Organisation (ECCO) 2025 Congress.

Of the four techniques the researchers tested, a “machine learning approach achieves the highest diagnostic accuracy, effectively distinguishing IBD and particularly differentiating Crohn’s disease from healthy controls in independent cohorts,” said study presenter Jee-Won Choi, Department of Biology, Kyung Hee University, Seoul, Republic of Korea.

“Integrating microbial markers with conventional diagnostics could enhance [their] clinical utility,” Choi said. However, further research is needed to determine the long-term validity of the biomarkers, she added.

Some experts questioned the reliability of the markers for IBD diagnosis due to the makeup of study populations, which included patients with known IBD who likely have undergone treatment that may have altered their gut microbiomes.

Biomarkers Found and Tested

The gut microbiota exists in two states: Eubiosis, which supports health, and inflammatory dysbiosis, an imbalanced state associated with disease, most notably IBD, Choi noted.

Although many studies have explored the differences between these two states, there have been three major challenges in identifying IBD biomarkers: The studies have had small sample sizes, they’ve concentrated on a single analytical approach, and they’ve had low reproducibility.

To overcome those challenges, researchers used a large-scale dataset and used multiple methods to determine which analytical approach yielded the most reliable results, Choi said. They validated their results in three independent cohorts with diverse populations, she added.

The study included 414 patients with Crohn’s disease, 880 with ulcerative colitis, and 2467 healthy control individuals from 21 centers in the Republic of Korea. Their gut microbiota profiles were analyzed from stool samples using 16S rRNA gene sequencing.

Researchers used four techniques to identify potential IBD biomarkers in the samples: Differential abundance analysis, supervised random forest machine learning, unsupervised network analysis, and literature-based curation.

Biomarker candidates generated by these methods were then compared for their diagnostic ability using a machine learning model. The findings were tested in three independent cohorts — one domestic and one international population, both of which included patients with IBD and healthy control individuals, and one dataset of patients without IBD.

The results showed that there were distinct differences in the microbial composition between healthy control individuals and patients with Crohn’s disease and with ulcerative colitis. Patients with IBD, particularly those with Crohn’s disease, consistently had a significantly higher prevalence of dysbiosis, Choi said.

Each of the four analytical techniques revealed distinct microbial biomarkers associated with IBD in general, as well as with Crohn’s disease and ulcerative colitis individually.

When comparing IBD patients overall with healthy control individuals, supervised machine learning resulted in the most effective biomarker sets for distinguishing between groups, with the area under the receiver operating characteristics curve (AUC) reaching 0.971. By comparison, the AUC results were 0.94 for literature-based curation, 0.924 for differential abundance analyses, and 0.914 for unsupervised network analysis.

Supervised machine learning also outperformed the other techniques when distinguishing between healthy control individuals and patients with ulcerative colitis (AUC, 0.958), and between patients with ulcerative colitis and those with Crohn’s disease (AUC, 0.902).

All the techniques performed strongly when distinguishing between healthy control individuals and patients with Crohn’s disease, with AUCs ranging from 0.911 to 0.95.

When the researchers turned to the independent datasets, they found that the biomarkers were able to distinguish between healthy control individuals and patients with IBD in general and particularly between healthy control individuals and those with Crohn’s disease, with AUCs of 0.969 in the domestic cohort and 0.848 in the international cohort.

The non-IBD cohort also demonstrated that the biomarkers were able to differentiate patients with metabolic dysfunction–associated steatotic liver disease, colorectal cancer, rheumatoid arthritis, and irritable bowel syndrome from those with ulcerative colitis and Crohn’s disease with a high degree of accuracy (AUCs ranging from 0.97 to 0.999).

Diagnostic Utility Questioned

Speaking from the audience, James Lindsay, PhD, professor of inflammatory bowel disease, Barts and The London School of Medicine and Dentistry, London, England, questioned the utility of the findings.

“Obviously, all these patients had IBD, and so they will have had treatment with antibiotics, etc.,” he said. “Surely the right validation cohort would be a group of people who have not yet been diagnosed with IBD to see whether your biomarker is able to separate those because the reason that people with IBD will have a difference is all the reasons that you have explained, ie, these patients were on treatment at the time that you took the samples.”

As a result, the biomarker panel isn’t for diagnosis but to confirm known disease, he added.

It’s important to look for microbiome signals of IBD, session Co-Chair, Lissy de Ridder, MD, PhD, associate professor of pediatric gastroenterology, Erasmus MC Sophia Children’s Hospital, Rotterdam, the Netherlands, told Medscape Medical News.

De Ridder agreed that the biomarkers need to be validated in patients who aren’t on treatments that could affect their gut microbiomes. Not only do medications for IBD make a big difference but also do other drugs such as proton pump inhibitors and antibiotics, as well as dietary interventions, she said.

“Having said that, because it’s a large population, that’s always a good start to take lessons from and then go more into the details” in further analyses, de Ridder added.

This research was funded by a grant from the Korea Health Technology R&D Project through the Korea Health Industry Development Institute, funded by the Ministry of Health & Welfare, Republic of Korea. No relevant financial relationships were declared.