Integration of Metabolomics, Lipidomics, and Machine Learning for Developing a Biomarker Panel to Distinguish the Severity of Metabolic-Associated Fatty Liver Disease

Abstract

Metabolic-associated fatty liver disease (MAFLD), a global health challenge linked to metabolic syndrome, requires accurate severity stratification for clinical management. Current invasive diagnostic methods limit practical implementation. This study integrated multi-omics and machine learning to establish a non-invasive biomarker panel for the early stage of MAFLD assessment. Transcriptomic analysis of GEO datasets via weighted gene co-expression network (WGCNA) and differential expression revealed critical pathways associated with MAFLD progression. Subsequently, LC–MS/MS-based widely targeted metabolomics and lipidomics were conducted on plasma samples from 40 healthy controls and 120 patients with MAFLD at varying stages of severity (40 mild, 40 moderate, and 40 severe). Machine learning algorithms (LASSO regression, logistic regression, decision trees, and XGBoost) were then applied to these datasets to identify critical biomarkers linked to disease severity. Ultimately, a biomarker panel comprising 5-aminolevulinic acid, mesaconic acid, shikimic acid, PC O-35:3, and PI 36:2 exhibited outstanding diagnostic performance in detecting the prevalence and severity of MAFLD, achieving an accuracy of 88.3% in the training cohort and over 91.7% prediction accuracy in the independent test cohort. The identified biomarker panel offers a promising non-invasive approach for assessing MAFLD severity, paving the way for precision medicine and treatment in MAFLD.