Identification and multi-layered validation of seven diagnostic biomarkers for dilated cardiomyopathy via integrative machine learning, single-cell transcriptomics, and Mendelian randomization

<h4>Background</h4>Dilated cardiomyopathy (DCM) is the most common non-ischemic cardiomyopathy and a major cause of heart failure, but disease-specific molecular biomarkers remain limited. This study aimed to identify and prioritize tissue-level, disease-responsive candidate biomarkers for DCM using an integrative multi-omics bioinformatics framework.<h4>Methods</h4>Bulk myocardial transcriptomic data from GSE57338 were used as the discovery cohort, and GSE26887, GSE42955, and GSE79962 served as external microarray validation cohorts. GSE116250 was used for independent RNA-seq validation. Differentially expressed genes were intersected with WGCNA hub genes to define candidate genes. Four machine-learning algorithms, including LASSO, random forest, SVM-RFE, and XGBoost, were applied to identify core diagnostic candidates. Tissue-level model performance was evaluated by ROC analysis, calibration assessment, nomogram visualization, and decision curve analysis. Orthogonal validation was performed using GTEx, HPA, and snRNA-seq data. Immune infiltration, bidirectional Mendelian randomization, and CellOracle-based GRN analysis with a co-expression-based functional importance score were used as hypothesis-generating analyses. The workflow explicitly separated diagnostic performance, localization evidence, and exploratory mechanistic context in myocardial tissue.<h4>Results</h4>Integration of 309 DEGs and 2,093 WGCNA hub genes yielded 270 candidates. Seven candidates-HMGN2, AQP3, SERPINA3, FREM1, HMOX2, CSDC2, and TUBA3E-were selected by at least three algorithms. In the discovery cohort, the RF model achieved an AUC of 0.985 and the logistic model achieved a C-statistic of 0.993; however, these estimates were interpreted as potentially optimistic upper bounds because feature selection was not nested within cross-validation. External validation showed uneven robustness: SERPINA3, HMOX2, FREM1, and HMGN2 were consistently supported across microarray and RNA-seq cohorts, whereas AQP3, CSDC2, and TUBA3E were exploratory. GTEx, HPA, and snRNA-seq supported cardiac expression and cell-type localization, including cardiomyocyte enrichment of CSDC2/HMOX2 and fibroblast enrichment of FREM1. MR and GRN analyses suggested disease-responsive rather than disease-driving biology, including possible heart failure-associated AQP3 downregulation and a putative PPARGC1A-CSDC2/HMOX2 regulatory context.<h4>Conclusion</h4>This study identifies seven prioritized, predominantly disease-responsive tissue-level molecular candidates for DCM. These findings provide candidates and testable hypotheses for future translational research, rather than disease-driving therapeutic targets or a directly applicable clinical test.