GSE120895 (platform: GPL570) and GSE17800 (platform: GPL570) datasets were collected from the Gene Expression Omnibus (GEO) database ¹ . The GSE120895 dataset contained 8 control samples and 47 cases of endomyocardial biopsies (EMBs) from DCM patients, which is used as the training set to identify hub genes. The GSE17800 included 8 control samples and 40 cases of EMBs from DCM patients, which is used as the validation set to verify hub genes. The “affy” (version 1.68.0) package in R was used for data processing. Gene expression data were normalized by Robust Multi-array Average (RMA) algorithm, which included background correction, quantile normalization, probe summarization and log2 transformation. All subsequent statistical analyses were performed based on the expression values after log2 transformation.

A total of 564 ferroptosis-related genes (FRGs) were acquired from the Ferroptosis database V2 (FerrDb, ² ). A total of 200 hypoxia-related genes (HRGs) were collected from the Molecular Signatures Database (MSigDB, ³ ). All data and statistical analyses were performed in R software (version 4.0.5).

The DCM-related differential expressed genes (DEGs) with false discovery rate (FDR) <0.05 were identified in the GSE120895 dataset using the “limma” (version 3.46.0) package in R. Then, 564 FRGs, 200 HRGs, and DCM-related DEGs were intersected to obtain ferroptosis- and hypoxia-related DEGs. The volcano plot was used to visualize these results. Subsequently, Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genome (KEGG) functional enrichment analysis of ferroptosis- and hypoxia-related DEGs were carried out through the DAVID database ⁴ . The screening criterion for terms was p-value <0.05.

The “WGCNA” package in R was utilized to perform WGCNA in the GSE120895. First, the “hclust” function was used to cluster all samples and detect the outliers. The high degree was set to 0.9, and the function “pickSoftThreshold” was utilized to calculate the soft threshold to construct a scale-free co-expression network. The weighted adjacency matrix was transferred into the topological overlap matrix (TOM) and 1-TOM. Genes with similar expression patterns were clustered, and modules were divided by default parameters according to the “cutreeDynamic” function. Since the modules identified by the dynamic tree cutting algorithm may be similar, these modules were merged at a height cutoff of 0.25. The “moduleEigengenes” function was used to calculate module eigengene (ME) and Pearson analysis was employed to identify DCM-related ME. Genes with |gene significance (GS)| >0.4 and |module membership (MM)| >0.4 in the hub module were selected as the DCM-related genes.

DCM-related genes, ferroptosis-related DEGs, and hypoxia-related DEGs were intersected to acquire intersection genes. Then, LASSO regression analysis was utilized to select the hub genes through the “glmnet” (version 4.1–1) package. Additionally, receiver operating characteristic (ROC) analysis was performed to evaluate the diagnostic value of hub genes in distinguishing DCM from healthy controls using the “pROC” (version 1.17.0.1) package. Area under the curve (AUC) was used to evaluate the diagnostic accuracy of hub genes.

The relative abundance of each immune cell was calculated using the single-sample GSEA (ssGSEA) algorithm. Signature gene sets marking various types of infiltrating immune cells were obtained from Charoentong’s study [12, 13], which covers a range of human immune cell subtypes, including activated CD8 T cell, activated dendritic cell, macrophages, monocyte, natural killer cell, and regulatory T cells. The difference in immune cell infiltration between DCM patients and healthy controls was compared using the Wilcox test. Furthermore, the correlation between differentially infiltrating immune cells and hub genes was explored.

The miRNA-expressed dataset GSE112556 comprised 3 healthy controls and 3 heart tissue samples from DCM patients. The differentially expressed miRNAs (DEmiRNAs) with p-value <0.05 between DCM patients and controls were obtained. The Encyclopedia of RNA Interactomes (ENCORI, ⁵ ) dataset was used to select hub gene-related miRNAs. Next, the predicted miRNAs were intersected with DEmiRNAs, and their intersection was taken to obtain miRNA-mRNA targeting pairs with negative regulatory relationships. Additionally, the miRNAs-related lncRNAs were attained through the ENCORI dataset. Next, the lncRNAs associated with the hub genes were defined as DCM-related lncRNAs by correlation analysis (p-value <0.05). Finally, the lncRNA-miRNA-mRNA network was constructed using the Cytoscape.

The Drug Gene Interaction Database (DGIdb, ⁶ ) was utilized to obtain hub genes-related drugs. Molecular docking technology is to place small molecules (ligands) in the binding region of macromolecular targets (receptors) by computer simulation, and predict the binding energy (binding affinity) and binding mode (conformation) of the two by calculating physical and chemical parameters, and then identifies the lowest-energy conformations of ligand-receptor binding. Low binding energy is the basis of stable binding between molecules. The 3D structures of target proteins (receptors) and drugs (ligands) were downloaded from the RCSB PDB ⁷ and the PubChem database ⁸ , respectively. Subsequently, proteins and drug molecules were hydrogenated and other pretreated in AutoDockTools, followed by molecular docking. In the docking results, binding energy was used as a reference to screen the most active ligand molecules and target genes. Binding energy <0 means that the ligand and receptor can bind spontaneously. The smaller the binding energy, the more stable the binding between ligand and receptor. Binding energy less than −5.0 kJ/mol (Note: −5.0 kJ/mol = −1.19423 kcal/mol) is the basis for screening candidate targets of active ingredients/drugs. Finally, the molecular docking results were visualized by the PyMol software.

The human cardiomyocyte AC16 was purchased from Wuhan Pricella Biotechnology Co., Ltd. Doxorubicin (DOX; 0.3 µM) was used to stimulate AC16 cells at 37 °C in a humidified atmosphere of 5% CO2 for 24 h to induce cardiomyocyte injury model. ATrizol (TIANGEN, DP424), FastQuant cDNA first strand synthesis (TIANGEN, KR106) and SuperReal PreMix Plus (SYBR Green) (TIANGEN, FP205) kits were used for RNA extraction, reverse transcription and RT-qPCR, respectively. The hub gene TGM2 was selected for knockdown experiments, and the expression of TGM2 was detected by RT-qPCR and western blotting. The sequence of siRNA is shown in Supplementary Table S1. According to the manufacturer’s instructions, apoptotic cells were evaluated using the Annexin V-FITC/PI apoptosis detection kit (Solarbio, CA1020). In addition, immunofluorescence staining was performed on cardiomyocytes in each group. After fixing and permeabilizing cells, incubate them with mouse anti-α-actinin monoclonal antibody (Biodragon, BD-PM3612) and rabbit anti-troponin T-C polyclonal antibody (Biodragon, BD-PT5362), followed by fluorescently labeled goat anti-mouse IgG (red, Biodragon, BD9276) and goat anti-rabbit IgG (green, Bioss, bs-0295G-FITC). Finally, stain the nuclei with DAPI and observe under a fluorescence microscope. In addition, the dual-luciferase reporter assay was used to investigate the association between hsa-miR-296-5p and TGM2.

A total of 609 DEGs were identified in DCM patients compared to healthy controls in the GSE120895 dataset, including 102 up- and 507 down-regulated DEGs. Then, 564 ferroptosis-related genes, 200 hypoxia-related genes, and 609 DEGs were intersected (Figure 1A). Finally, 16 ferroptosis-related DEGs (SESN2, AGPAT3, MAP3K5, EIF2AK4, PRDX6, NR4A1, ARF6, KRAS, ABCC1, RGS4, LPIN1, YY1AP1, USP7, GPAT4, ACSL1, and ADAM23) and 2 hypoxia-related DEGs (PPP1R15A and TGM2) were acquired (Figure 1B).

In the GO:BP terms, these 18 genes were significantly enriched in triglyceride biosynthetic process, phosphatidic acid biosynthetic process and cellular response to amino acid starvation (Figure 1C). In the GO:CC terms, these 18 genes significantly enriched in cytosol, cytoplasm and membrane (Figure 1C). In the GO:MF terms, these 18 genes were significantly enriched in ATP binding, GTPase activity and 1-acylglycerol-3-phosphate O-acyltransferase activity (Figure 1C). The KEGG results showed that the DEGs were significantly enriched in mTOR signaling pathway, Glycerophospholipid metabolism and Glycerolipid metabolism (Figure 1D).

The GSE120895 dataset was used to carry out WGCNA. All samples were clustered, and the sample clustering tree is displayed in Figure 2A. Then, the soft threshold was set as 6 to construct the scale-free network (Figure 2B). The minimum number of genes in the module was set to 500, and 7 modules were attained (Figure 2C). Next, the brown and red modules were merged, as were the blue and turquoise modules (Figure 2D). Finally, 5 modules were identified. Subsequently, we found that the brown module was significantly correlated with DCM (Figure 2E). Then, a total of 315 DCM-related genes with |GS| >0.4 & |MM| >0.4 were acquired (Figure 2F).

The 315 DCM-related genes, 16 ferroptosis-related DEGs, and 2 hypoxia-related DEGs were intersected, and 8 intersection genes were acquired (Figure 3A). Subsequently, 6 hub genes (PPP1R15A and TGM2 for hypoxia and MAP3K5, USP7, SESN2, and ADAM23 for ferroptosis) were identified using LASSO regression analysis (Figure 3B). The expression levels of ADAM23, MAP3K5, and USP7 were significantly decreased in DCM patients, while the expression levels of PPP1R15A, TGM2, and SESN2 were significantly increased in DCM patients in both GSE120895 and GSE17800 datasets (Figures 3C,D). Additionally, ROC analysis showed that the AUC values of 6 hub genes were all more than 0.8 in both the GSE120895 and GSE17800 datasets (Figure 4; Supplementary Figure S1). Furthermore, a diagnostic model for DCM was constructed via LASSO regression based on the 6 hub genes. Results showed that the AUC of the model exceeded 0.9 in both the GSE120895 (Supplementary Figure S1A) and GSE17800 (Supplementary Figure S1B) datasets.

Immune cell infiltration analysis results showed that CD56dim natural killer (NK) cells, macrophages, monocytes, NK cells, and NK T cells were significantly highly infiltrated in DCM patients in the GSE120895 dataset (Figure 5A). Additionally, the correlation between differentially infiltrating immune cells and hub genes was explored. CD56dim NK cells were significantly correlated with 6 hub genes (Figure 5B). Macrophages were significantly associated with SESN2 and ADAM23 (Figure 5C). Monocytes showed significant correlations with PPP1R15A, TGM2, and SESN2 (Figure 5D). NK cells and NK T cells were significantly correlated with MAP3K5 and ADAM23 (Figures 5E,F). Additionally, NK cells exhibited a significant correlation with SESN2.

A total of 21 DEmiRNAs were acquired in the GSE112556 dataset, including 7 up- and 14 down-regulated DEmiRNAs (Figure 6A). Then, 11 miRNA-mRNA negative regulatory relationship pairs were obtained, including 8 DEmiRNAs. Subsequently, 11 lncRNAs were identified that interacted with 7 DEmiRNAs (hsa-miR-770-5p, hsa-miR-16-5p, hsa-miR-139-5p, hsa-miR-296-5p, hsa-miR-338-3p, hsa-miR-148a-3p, and hsa-miR-363-3p) and 3 hub genes (ADAM23, SESN2, and TGM2). Finally, the ceRNA network, including 11 lncRNAs, 7 miRNAs, and 3 mRNAs, was constructed (Figure 6B).

A total of 16 drugs were obtained (Figure 7A). Among these, 1 drug may target the MAP3K5 gene, 15 drugs may target the TGM2 gene, and the remaining genes were not matched to drugs. Literature retrieval revealed that atorvastatin, a drug associated with TGM2, has significant benefits in treating DCM, and low-dose use significantly reduces levels of inflammatory cytokines and uric acid, improves hemodynamic parameters, and increases the 5-year survival rate of DCM patients [14, 15]. Subsequently, molecular docking between atorvastatin and TGM2 was performed. Molecular docking between atorvastatin and TGM2 was performed. The binding energy was −2.79 kcal/mol by binding LYS-173 with 1 hydrogen bond (Figure 7B).

The expression of the highly expressed hub genes (TGM2, PPP1R15A, and SESN2) in the DCM group was verified through RT-qPCR in AC16 cells treated with DOX. Compared with the control group, the expression levels of TGM2, PPP1R15A, and SESN2 in the DOX group were significantly increased (Figures 8A–C), which was consistent with the expression trend in the public database. Hub gene TGM2, a potential diagnostic biomarker, is not only associated with immune cell infiltration but also may play a crucial regulatory role in DCM via the lncRNA-miRNA-TGM2 axis. Moreover, drug prediction revealed that atorvastatin, a drug associated with TGM2, has significant benefits in treating DCM. Therefore, TGM2 was selected for subsequent experiments.

TGM2 was knockdown via siRNA transfection, and TGM2 expression was quantified by RT-qPCR to identify effective siRNA targets. The results showed that the knockdown effect of target 2 was the best (Figure 8D). Therefore, siRNA-2 was selected for subsequent knockdown experiments. Subsequently, the expression of TGM2 in each group was detected by RT-qPCR and western blotting. The results showed that the expression of TGM2 was significantly increased after DOX treatment, while the expression of TGM2 was significantly decreased in the knockdown group (Figures 8E–G). Subsequently, immunofluorescence staining was performed to investigate the effects of TGM2 on cTnT and α-actinin (Figure 9). Compared with the control group, the expressions of α-actinin and cTnT in the DOX group were decreased. After knocking down TGM2, the expressions of α-actinin and cTnT were increased. Furthermore, after knockdown of TGM2, the expression level of HIF-1α was inhibited (Figures 10A,B). The complementary binding sites between hsa-miR-296-5p and the 3′ UTR of TGM2 were predicted through the ENCORI database (Figure 10C). Based on the psiCHECK 2 vector, we constructed TGM2-WT (wild-type) and TGM2-Mut (mutant) dual-luciferase reporter plasmids. Dual luciferase assay showed that hsa-miR-291-5p exerted its regulatory effect by directly binding to TGM2 (Figure 10D). In addition, flow cytometry results showed that TGM2 had no significant effect on the apoptosis of AC16 cells (Figures 11A–E).