Machine Learning Predicts Treatment Response and Prognostic Pathways From Whole-Blood Transcriptome in Primary Biliary Cholangitis.

BACKGROUND AND AIMS: Prognostic biomarkers that link disease progression and/or responses to therapeutic interventions in patients with primary biliary cholangitis (PBC) remain undefined. In this study, we used a machine learning (ML) approach with whole-blood transcriptomic data to predict clinical outcomes in response to obeticholic acid (OCA) and disease progression. METHODS: We developed an ML model that incorporated whole-blood RNA-seq analyses using longitudinal samples from the POISE study (discovery cohort) and a prospective group of PBC patients who remained unchanged or progressed to liver transplantation/hepatic decompensation (independent validation cohort). The model identified 1200 candidate genes predictive of outcomes, which were investigated using pathway analysis. A genetic algorithm refined the panel to a 105-gene list that was trained on the POISE end-of-treatment cohort. RESULTS: The algorithm's generalisability was tested by assessing the baseline POISE patients as responders or non-responders to OCA (AUROC 0.93) and independently validated by differentiating liver disease-related survival between progressors and non-progressors (AUROC 0.94). Pathway analysis of the ML candidate genes identified enrichment of FXR regulated genes, autoimmune disease-related inflammation, immune regulation, fibrosis and integrated stress response pathways as prognostically related to PBC. When we compared these ML-derived prognostic pathways with RNA-seq analysis of PBC versus healthy controls, the most relevant transcripts were those involved in the integrated stress response and metabolic remodelling in PBC. CONCLUSION: The ML-derived score links both prognostic and disease pathogenesis pathways and can be used in future studies to better understand the pathophysiology and management of PBC.