Machine learning-assisted analysis of serum metabolomics for identifying biomarkers in intrinsic and idiosyncratic drug-induced liver injury

TL;DR

This study uses machine learning and metabolomics to identify blood-based biomarkers that can distinguish between two types of drug-induced liver injury.

Contribution

The study introduces a machine learning-assisted metabolomic approach to differentiate intrinsic and idiosyncratic drug-induced liver injury mechanisms.

Findings

Four differential metabolites were identified that distinguish intrinsic from idiosyncratic DILI.

Machine learning models achieved high diagnostic performance with AUC values above 0.8.

Key metabolic pathways linked to DILI include amino acid metabolism and related biosynthesis processes.

Abstract

This project aims to employ high-performance chemical isotope labeling (HP-CIL) liquid chromatography–mass spectrometry (LC-MS) to conduct a metabolomic study on the mechanisms underlying intrinsic and idiosyncratic drug-induced liver injury (DILI). By comparing the metabolic characteristics between these two types of DILI, we seek to identify biomarkers for predicting intrinsic and idiosyncratic DILI using machine learning strategies.

Based on the diagnostic criteria outlined in the EASL clinical practice guidelines on drug-induced liver injury, a review published in NEJM, enrolled DILI cases were classified according to the pathogenic mechanism into an intrinsic type (n = 17) and an idiosyncratic type (n = 27). Serum samples were collected from both groups. Metabolomic profiling was performed using high-performance chemical isotope labeling liquid chromatography–mass spectrometry (HP-CIL LC-MS) to identify differentially expressed metabolites between the two groups. Metabolites that showed significance in both univariate and multivariate statistical analyses were selected for further receiver operating characteristic (ROC) analysis. Machine learning approaches were employed to develop diagnostic models for distinguishing intrinsic and idiosyncratic DILI. These models were compared to identify potential biomarkers capable of discriminating between the two types of DILI, and the diagnostic performance of these candidate biomarkers was evaluated.

Serum metabolomic profiling identified four differential metabolites that distinguished intrinsic from idiosyncratic DILI through multivariate and univariate statistical analyses, followed by ROC curve analysis and machine learning-based selection. These potential biomarkers included Alanyl-Glycine (level 1),N2-Acetyl-L-Cystathionine (level 2a), Isomer 1 of 5-Hydroxyindoleacetic acid (level 2a), and Isomer 1 of 5-Hydroxyindoleacetic acid (level 2a). ROC analysis using multiple machine learning models yielded area under the curve (AUC) values greater than 0.8 for all models, indicating high diagnostic performance. Under a multivariate regression model, internal cross-validation (CV) within the training set demonstrated robust model tuning and stability, with an AUC of 0.983. Holdout validation further confirmed model reliability with an AUC of 0.935. Metabolic pathway analysis of these metabolites revealed that the most significantly associated pathways affecting intrinsic and idiosyncratic DILI were primarily related to amino acid metabolism, including tryptophan metabolism, tyrosine metabolism, cysteine and methionine metabolism, and the biosynthesis of phenylalanine, tyrosine, and tryptophan.

This study demonstrates that machine learning-assisted serum metabolomics can effectively characterize currently well-established intrinsic and idiosyncratic drug-induced liver injury, reveal metabolic disparities between the two types, and identify differential metabolites associated with their respective pathogenesis. These findings provide a valuable reference for predicting the mechanistic type of liver injury induced by various hepatotoxic drugs in the future.

Figures