Metabolomic Machine Learning Predictor for Adequate and Deep Response to UDCA in Non-Cirrhotic PBC
Fan Yang1, Jin Lin1, Yujie Zhang1
1Department of Gastroenterology and Hepatology and Laboratory of Gastrointestinal Cancer and Liver Disease, West China Hospital of Sichuan University, Chengdu, China.
Background:
Primary biliary cholangitis (PBC) patients exhibit heterogeneous responses to ursodeoxycholic acid (UDCA), with partial responders remaining at significant risk of disease progression. Current prognostic tools lack metabolic insights necessary for personalised management. This study aimed to characterise the metabolic landscape of UDCA response and develop prediction models integrating clinical and metabolomic features.
Methods:
A total of 347 non-cirrhotic PBC patients were classified as complete responders (n = 214) or non-responders (n = 133). Complete responders were further subdivided into deep responders (DR) and adequate responders (AR). Plasma metabolomics was performed to identify differential metabolites (VIP > 1, |log2FC| > 1, p < 0.05). Machine learning models integrating clinical and metabolomic features were constructed. Prognostic risk was assessed using the UK-PBC and GLOBE scores.
Results:
Metabolomic profiling suggested suppression of bile acid biosynthesis (e.g., taurochenodeoxycholate, log2FC = -1.29) and alterations in glutathione metabolism among responders. AR patients showed a higher risk of hepatic adverse events (p < 0.001), accompanied by differences in sphingolipid signalling and phospholipase D-related pathways. Key metabolic predictors were identified from metabolomic data. Prognostic scoring demonstrated significant differences in long-term outcomes (UK-PBC 15-year risk: DR 0.041 vs. AR 0.055, p = 0.01). The integrated prediction model achieved strong performance in UDCA response classification (AUC = 0.91) and DR stratification (AUC = 0.87), outperforming clinical-only models.
Conclusion:
This study establishes a metabolomics-informed framework for PBC management, identifying metabolic features associated with treatment response and developing predictive models integrating clinical and metabolic data. These findings may support improved risk stratification and personalised management in PBC, although further validation is required.
