Metabolomics and Machine Learning Uncover 5′-Methylthioadenosine as a Metabolic Indicator of Mixed-Pattern Drug-Induced Liver Injury

Abstract

Background/Objectives: Drug-induced liver injury (DILI) remains a leading contributor to acute liver failure, yet phenotypic classification still relies on the R-value, leaving the mixed-pattern (MD) subtype a heterogeneous category between hepatocellular (HD) and cholestatic (CD) injury. We aimed to identify a metabolic signature that helps resolve this heterogeneity. Methods: Targeted LC-MS/MS metabolomics was performed on 79 patients with DILI (HD = 54, CD = 7, MD = 18) and 221 healthy controls. A 32-metabolite panel was used to train an XGBoost classifier with pre-specified, fold-invariant hyperparameters, evaluated via Leave-One-Out Cross-Validation and interpreted by TreeExplainer-based SHAP attribution. Results: SHAP identified 5′-methylthioadenosine (MTA) as the dominant contributor. A four-parameter logistic fit of its SHAP dependence curve gave an exploratory, model-derived transition point at a relative abundance of 93. Applying this value to the 18 held-out MD patients assigned 16 to a hepatocellular and 2 to a cholestatic trajectory. MTA correlated with total bilirubin (r = 0.53, p < 0.001) but not ALT (r = −0.19, p = 0.10), dissociating it from hepatocyte leakage. Conclusions: MTA emerges as an exploratory biomarker that uncovers the distinct metabolic trajectories hidden within mixed-pattern DILI. Furthermore, these findings highlight the MTA-methylthioadenosine phosphorylase (MTAP) axis as a potential mechanistic driver of cholestatic injury.