Identification of Core Ferroptosis Regulatory Genes in Metabolic Dysfunction-Associated Steatotic Liver Disease Based on Single-Cell Transcriptomics and Multi-Cohort Machine Learning
Background: Metabolic Dysfunction-Associated Steatotic Liver Disease (MASLD) is one of the most common chronic liver diseases globally, yet its pathogenesis remains incompletely elucidated. Ferroptosis, an iron-dependent form of cell death driven by lipid peroxidation, has been demonstrated to play a crucial role in hepatocyte injury during MASLD and Metabolic Dysfunction-Associated Steatohepatitis (MASH). However, the molecular mechanisms regulating the susceptibility of MASLD hepatocytes to ferroptosis remain unclear. Single-cell transcriptomics provides a novel perspective for resolving the heterogeneity of the hepatic microenvironment in MASLD, but there is still a lack of standardized pipelines for systematically identifying core ferroptosis-associated driver genes from massive single-cell datasets. Methods: This study performed an integrated analysis of six MASLD-associated snRNA-seq samples from GSE189600 (n=48,081 cells). Following dimensionality reduction, clustering, cell annotation, and ferroptosis scoring for each cell type, the hepatocyte subpopulation was extracted for high-dimensional weighted gene co-expression network analysis (hdWGCNA) to identify the M3 module, which was most highly correlated with MASH. Using the 108 cross-dataset consensus genes from this module as features, 101 two-stage machine learning algorithm combinations were trained on the GSE48452 cohort (n=73) and externally validated in two independent cohorts (GSE135251 and GSE89632). Global interpretation was conducted using fastshap (nsim=100 Monte Carlo) combined with shapviz to dissect core driver genes. Slingshot pseudotime analysis was utilized to track the dynamic changes of genes along the MASLD progression trajectory. CellChat analysis was applied to explore the cross-talk between the ferroptosis gene regulatory network and intercellular communication within the hepatic microenvironment. Virtual knockouts were performed using scTenifoldKnk to analyze downstream target-associated genes. Finally, the DrugReflector platform predicted drugs from the L1000 database, and molecular docking via AutoDock Vina was used to verify the drug accessibility of the core targets. Results: Ferroptosis scoring revealed the ferroptosis susceptibility of MASLD hepatocytes. hdWGCNA identified the M3 module as the core MASH-associated module (r=0.60). The algorithm combination of glmBoost and random forest (randomForestSRC) demonstrated optimal performance (average AUC across 3 cohorts = 0.967). The fastshap SHAP analysis identified five core driver genes: ERBB2 (0.069), BACH1 (0.051), RBMS1 (0.050), NDRG2 (0.038), and CTSD (0.036). Pseudotime analysis showed that these five genes were temporally upregulated along the MASLD trajectory (Slingshot Lineage 2 correlation with MASH, r=0.702). BACH1 elevated first in early pseudotime, while ERBB2 and RBMS1 peaked in the intermediate phase. Conclusion: This study established a pipeline for identifying core MASLD regulatory genes, from single-cell transcriptomes to multi-cohort validation. A consistent SHAP importance ranking was validated through the glmBoost combined with random forest (randomForestSRC) algorithm and dual fastshap/kernelshap methods. This revealed that the five M3 module genes (ERBB2/BACH1/RBMS1/NDRG2/CTSD) constitute a multi-layered regulatory network for hepatocyte ferroptosis in MASLD, encompassing tyrosine kinase signaling (ERBB2), iron metabolism transcriptional regulation (BACH1), RNA-binding protein-mediated post-transcriptional regulation (RBMS1), metabolic stress adaptation (NDRG2), and lysosomal ferroptosis execution (CTSD). These five genes may serve as critical regulatory nodes and therapeutic targets for ferroptosis in MASLD.