Mitochondrial dysfunction contributes to maladaptive macrophage polarization in atherosclerosis (AS), but the specific mitochondrial genes that shape disease-relevant macrophage states in plaques remain incompletely defined. The authors used an integrative strategy combining co-expression network analysis, machine learning, genetic epidemiology, single-cell mapping, and functional perturbation to prioritize mitochondria-related regulators acting within macrophage polarization states. The objective was to identify mitochondrial genes linked to macrophage mitochondrial integrity and to test genetic and functional evidence for roles in AS.
The analysis used multiple publicly available cohorts. Four GEO datasets (GSE28829, GSE41571, GSE43292, GSE163154; 147 samples total) were integrated as a discovery bulk expression matrix representing related AS phenotypes. Additional bulk datasets (GSE97320, GSE226790, GSE104140) were combined for meta-analysis of signatures. Single-cell RNA-seq data came from GSE159677 (3 healthy and 3 AS carotid samples). An independent bulk cohort (GSE111782) was used for validation. Bulk microarray and RNA-seq data were processed with standard pipelines (limma, normalization), and batch effects across the discovery bulk datasets were corrected with ComBat. Blood cis-eQTL instruments were obtained from eQTLGen and GTEx V8 for Mendelian randomization analyses, and CAD/CHD GWAS summary statistics were accessed via MRC IEU OpenGWAS.
Weighted gene co-expression network analysis (WGCNA) was applied to the 25% most variable genes in the integrated discovery matrix to identify AS-associated modules. Genes in a trait-associated module were intersected with 1,136 MitoCarta3.0 mitochondrial genes, producing 51 candidates. A feature-selection workflow using LASSO logistic regression and SVM-RFE prioritized candidates; 27 and 10 genes were selected by each method respectively, with eight genes retained by both for downstream Mendelian randomization and validation. The machine-learning steps were used for prioritization rather than as a deployed classifier, and later evidence guided final focus.
Two-sample Mendelian randomization used blood cis-eQTL instruments and GWAS summary statistics for CAD (test set) and CHD (validation set). Six of eight candidates had eligible cis-eQTLs (PPIF, PMAIP1, MRPL33, GATM, BCL2A1, and MGST3). MGST3 demonstrated the strongest protective genetic signal: training set OR = 0.989 (95% CI 0.978–0.999, P = 0.042) and test set OR = 0.977 (95% CI 0.960–0.995, P = 0.013). Instruments were blood-derived; thus MR results were interpreted as supportive genetic evidence rather than definitive proof of plaque-cell-specific causality.
Immune-cell deconvolution of the integrated discovery bulk matrix used CIBERSORTx with 1,000 permutations to estimate relative immune abundances across AS and control tissues. MGST3 co-expression was assessed in the 147-sample matrix by Spearman correlation with FDR correction; the top co-expressed genes were evaluated for enrichment within MitoCarta-derived mitochondrial modules and stability assessed by bootstrap resampling.
Single-cell RNA-seq processing followed a Seurat-based workflow with PCA, Harmony integration, and clustering. Major cell types were annotated using marker sets and CellMarker. Macrophage states were annotated as M1-like or M2-like along a continuum rather than strict binaries. M1-like cells were supported by inflammatory markers (e.g., IL1B, TNF, CXCL10), and M2-like cells by reparative and lipid-handling markers (e.g., MRC1, CD163, APOE). MGST3 expression was visualized on UMAP and compared across macrophage states and sample groups.
Analyses focused on AS M2-like macrophages (n = 6,419). Cells expressing MGST3 were identified; where >85% of AS M2-like cells expressed MGST3, top and bottom tertiles defined MGST3-high and MGST3-low groups. Functional module scores (OXPHOS, TCA, mitochondrial translation, antioxidant/ROS, mitophagy, inflammation) were computed with AddModuleScore and compared by paired tests across samples. A composite MitoHealth score was defined as mean z-score of OXPHOS, TCA, mitochondrial translation, and antioxidant scores minus the z-scored inflammatory score. MGST3-high M2-like cells displayed higher OXPHOS and antioxidant programs and higher composite MitoHealth scores, and they exhibited a less inflammatory transcriptional profile compared with MGST3-low cells. The MGST3-high M2 signature (top differentially expressed genes) was scored in bulk samples for disease-level assessment.
Single-cell mediation analysis indicated that the MitoHealth composite score statistically mediates the association between higher MGST3 expression and lower inflammatory activity, reduced M1-like polarization, and decreased atherogenic macrophage programming. Virtual knockdown using scTenifoldKnk generated predicted network perturbations; genes with strong |Z| scores were tested for GO enrichment as computational predictions rather than direct loss-of-function evidence. Cell-cell communication was inferred with CellChat to summarize signaling strengths and dominant ligand–receptor interactions distinguishing MGST3-high from MGST3-low M2-like cells.
Human carotid plaque validation confirmed reduced MGST3 at both mRNA and protein levels in whole atherosclerotic lesions. Notably, single-cell and disease-level analyses revealed an expanded MGST3-high M2-like macrophage population within AS plaques despite this whole-plaque downregulation. In vitro, MGST3 knockdown in M2-polarized macrophages reduced transcripts associated with M2 polarization and mitochondrial/antioxidant programs while increasing pro-inflammatory transcripts, providing transcript-level functional support for a role of MGST3 in maintaining a reparative, mitochondrial-health–associated M2-like state.
The MGST3-high M2 signature was evaluated across multiple bulk cohorts (meta-analysis of seven cohorts and independent validation), and co-expression stability was assessed with bootstrap resampling. A phenome-wide association study (PheWAS) detected no significant phenotype associations at the predefined threshold. Robustness checks included cohort-stratified analyses, random-effects meta-analysis, leave-one-out sensitivity, and use of independent validation cohort GSE111782 where appropriate.
The integrated multi-omics framework prioritized MGST3 as a mitochondrial gene associated with a protective genetic signal for coronary disease and as a marker of a mitochondrial-health–associated M2-like macrophage state expanded within atherosclerotic plaques. MGST3-high M2-like cells were OXPHOS-oriented, antioxidant-enriched, less inflammatory, and had higher composite MitoHealth scores. Single-cell mediation and in vitro knockdown provided further evidence linking MGST3 to mitochondrial preservation and lower pro-atherogenic activation. The authors propose the MGST3–MitoHealth axis as a compensatory macrophage program meriting further diagnostic and therapeutic investigation in AS.
Key limitations reported in the source include reliance on blood-derived cis-eQTLs rather than plaque- or macrophage-specific instruments, which limits cell-type-specific causal inference from Mendelian randomization. Virtual knockdown results were treated as computational predictions and not as substitute for in vivo loss-of-function validation. The authors note the need for activity-based and in vivo validation to determine mechanistic causality and therapeutic potential.