Background: Primary Biliary Cholangitis (PBC) is a cholestatic liver disease characterized by chronic inflammation and granulomatous destruction of interlobular bile ducts. Currently ursodeoxycholic acid (UDCA) is first-line therapy for PBC, which up to 40% of patients with PBC respond only partially to (termed “UDCA non-responders” or “progressive”). Based on current PBC management guidelines, after one year of first-line treatment initiation, the need for second-line therapies should be assessed. Several risk scoring systems including GLOBE have been developed to estimate the risk of adverse outcomes (progressive or non-progressive PBC). The purpose of this study is to calculate a differentiation model based on the concentration of essential compounds in urine samples to evaluate UDCA response in PBC patients. This helps to identify next steps in disease management.

Methods: Urine metabolomics analysis by ¹H NMR spectroscopy was performed in a case-control study on 53 healthy controls and 45 patients with PBC. GLOBE scores were calculated to differentiate PBC patients into progressive (18 patients) or non-progressive (27 patients). Univariate, multivariate, and machine learning analysis were performed on the metabolites dataset (71 metabolites per sample). Subsequently two multilayer perceptron (MLP) neural network classification models were developed to differentiate PBC patients from healthy controls, progressive disease from non-progressive, and to identify associated significantly important metabolic biomarkers.

Results: Univariate analysis on metabolites dataset identified 20 compounds with significantly different levels of concentration in normal versus PBC. 3 identified metabolites with significantly different concentration levels in progressive versus non-progressive PBC groups include ethanolamine, methanol, and urocanate. PLS-DA analysis was performed to generate PCA score plot and classification. Results showed a clear separation in both normal versus PBC and progressive versus non-progressive groups. MLP neural network classification method and ROC analysis were used to identify biomarkers and establish classification models. The optimized MLP classification model calculated using significantly important metabolites for normal versus PBC differentiation showed 90% and 0.93 classification accuracy and the area under the curve (AUC) values, respectively. The differentiation accuracy and AUC value of the optimized MLP classification model for progressive versus non-progressive PBC groups was 86% and 0.88, respectively. MLP model was calculated using 6 most important metabolites selected using correlation-based feature selection algorithm.

Conclusion: Our MLP model uses quantification of essential biomarkers for differentiation of progressive from non-progressive in patients with PBC with high accuracy.

Sandra O'Keefe

Appearances