Univariate- and machine learning-based plasma metabolite signature differentiates PSC-IBD from IBD and is predicted to be driven by gut microbial changes

TL;DR

This study uses metabolomics and machine learning to identify a plasma signature that distinguishes PSC-IBD from IBD, suggesting gut microbial changes drive the condition.

Contribution

A novel plasma metabolite signature for PSC-IBD diagnosis and insights into gut microbial influences are presented.

Findings

A high-performing model differentiates PSC-IBD from PSC with changes in amino acid and vitamin metabolism.

Metagenomic analysis suggests shifts toward pro-inflammatory gut microbes in PSC-IBD.

The dataset is shared to support future IBD metabolomics research.

Abstract

Inflammatory bowel disease (IBD) is a group of chronic inflammatory conditions of the gastrointestinal tract comprising two major phenotypes, Crohn’s disease (CD) and ulcerative colitis (UC). Up to 8% of patients with IBD also develop primary sclerosing cholangitis (PSC), characterised by cholestasis and progressive destruction of the biliary tree, resulting in cirrhosis, end-stage liver disease and cholangiocarcinoma. Clinical outcome can currently not be improved through medication, denoting the importance of diagnosis prior to irreversible damage, which requires biomarkers of (early) disease.

We employed direct infusion mass spectrometry (DI-MS)-based metabolomics on plasma to build predictive, potentially diagnostic models for PSC-IBC and other phenotypes including IBD subtype, stricture and fistula presence and more. We used this dataset to simultaneously investigate aetiology of these phenotypes.

Samples of 348 IBD patients were included for analysis. The data was analysed using our previously reported tool, MetaboShiny. We built predictive models using Random Forest (RF), and subsequently combined with univariate statistics to rank m/z features connected to PSC-IBD. This ranking was used to perform mummichog enrichment analysis connected to metabolic and metagenomic changes.

The highest performing predictive model differentiated PSC-IBD from PSC. The metabolic signature was enriched in changes to amino acid and vitamin metabolism, alongside changes to the metagenome suggesting decreases in anti-inflammatory microbial species and increases in pro-inflammatory species.

These results demonstrate the potential of DI-MS-based metabolomics with machine learning to create diagnostic models and generate hypotheses on the metabolomic-metagenomic level. Sharing our dataset of patients will enrich future human IBD metabolomics research possibilities.

The online version contains supplementary material available at 10.1007/s11306-026-02420-w.

Figures