Three DCM datasets (GSE57338, GSE120895, and GSE116250) were obtained from NCBI’s Gene Expression Omnibus¹ (10). The details of each dataset are shown in Supplementary Table S1. The initial processing of GSE57338 and GSE120895 was carried out in R (version 4.1.1), which included dataset consolidation, background calibration, normalization, and log2 transformation (11). Multiple probes corresponding to the same gene were taken as the gene expression value. The “sva” R package was also used for minimizing batch impacts and other undesired variances (12). Moreover, DEGs for DCM were screened using the “LIMMA” package, with an adjusted value of p < 0.01 and |log2 Fold change (FC)| ≥ 0.5 (13). In addition, a separate dataset (GSE116250) was used for subsequent validation analysis.

A network of gene co-expression was created using WGCNA (9) to examine gene interactions. Genes with less than 25% variation between samples were first eliminated, and the remaining genes were then imported into WGCNA for subsequent analysis. Further, using the pick-Soft-Threshold function, adjacency was generated based on the soft thresholding power β, which was determined by co-expression similarity. The corresponding dissimilarity (1-TOM) was determined after converting the adjacency relationship into a topological overlap matrix (TOM). Hierarchical clustering and the dynamic tree-cut approach were employed to carry out module detection. We used average linkage hierarchical clustering with a minimum size (gene group) of 100 for the gene dendrogram to classify genes with comparable expression profiles into gene modules (14). Additionally, module membership (MM) and gene significance (GS) were calculated for modules associated with clinical attributes. A visualization of the eigenstate network was also performed. The module’s genetic data were used for subsequent analysis. By reviewing the literature, we obtained 172 genes that were of interest to us. These interesting genes (IGs) were subsequently used for the acquisition of common genes (CGs). Thereafter, CGs were identified by intersecting the DEGs screened in the combined dataset, the most important module’s genes, and genes of interest.

The web-based platform Metascape (version 3.5; http://metascape.org) provides a comprehensive tool for researchers to annotate and analyze gene functions (15). It has been used for enrichment analysis to carry out the Gene Ontology (GO) and the Kyoto Encyclopedia of Genes and Genomes (KEGG) for the CGs. The enriched GO terms included ontologies for biological processes, cellular components, and molecular structures. Here, a value of p < 0.05 and count ≥ 2 were set as the cutoff value.

The PPI interaction network of CGs was created using the STRING database (version 11.5; www.string-db.org), wherein a minimum interaction score of 0.15 was determined as the critical value (16). The PPI network was then visualized using Cytoscape. Three different algorithms of the Cytoscape plug-in, CytoHubba, were applied to obtain the hub genes of the PPI network. Further analysis was conducted using the Maximum Clique Centrality (MCC), Maximum Neighborhood Component (MNC), and Edge Percolated Component (EPC) algorithms. The top eight genes from each algorithm were plotted in a Venn diagram, and overlapping genes between the three groups were considered as the hub genes.

All animals were provided by the Animal Experimental Center of Chongqing Medical University (Chongqing, China). This study was approved by the Chongqing Medical University Animal Welfare Committee. As shown in previous studies, the DCM model was established by intraperitoneally injecting mice with DOX (17, 18). Male C57BL/6 mice (8 weeks; 22–24 g) were kept on a 12:12 h light–dark cycle in a room at 25°C and 50% relative humidity with free access to food and water. Mice were randomly assigned into two groups: control (n = 6) and DOX (n = 6). Mice in the DOX group received intraperitoneal administration of DOX hydrochloride (cat. no. 25316–40–9; Macklin; 4 mg/kg) once a week for 8 weeks. The control group received an identical dosage of saline. Successful establishment of the DCM model was confirmed by gradual cardiac functional deterioration and ventricular dilation in mice.

Mice were initially anesthetized with 5% isoflurane (cat.no. 792632; Sigma-Aldrich; Merck KGaA), and the anesthetic effect was sustained at a concentration of 1.5% isoflurane. Thereafter, transthoracic echocardiography (cat. no. VINNO 6LAB) was performed in M-mode with a 23-MHz transducer. Both long-axis and short-axis images were used to evaluate the left ventricle (LV) echocardiography. The end-systolic and end-diastolic dimensions were assessed by the phases corresponding to the ECG T and R waves, respectively (19). M-mode LV internal diameter end systole/diastole (LVIDs/d), LV posterior wall end diastole (LVPWd), and interventricular septal end diastole (IVSd) were measured thrice. Fractional shortening and ejection fractions were calculated, as mentioned previously (19).

Mice were euthanized by general anesthesia isoflurane followed by cervical dislocation. The hearts were quickly extracted by opening the thorax and placed into the cold (4°C) saline. The total RNA was isolated from the heart tissues using the Animal Total RNA Isolation Kit (cat. no.RE-03014; Foregene). For reverse transcription, 5X All-In-One RT MasterMix (cat. G492; Abm.) was used. The cDNA was kept at room temperature for 5 min. The reaction was inactivated at 85°C for 5 min followed by incubation at 25°C for 10 min and at 42°C for 15 min. Also, qPCR was performed using the EvaGreen Express 2 × qPCR MasterMix-No Dye (cat. MasterMix-ES; Abm) and SLAN-96S Real-Rime PCR system (Shanghai Hongshi Medical Technology Co., Ltd.). The primers for GAPDH (internal control), latent transforming growth factor beta binding protein 2 (LTBP2), asporin (ASPN), immunoglobulin superfamily contain leucine-rich repeat (ISLR), microfibril associated protein 4 (MFAP4), fibronectin type III domain containing 1 (FNDC1), podocan (PODN), HtrA serine peptidase 1 (HTRA1), and fibroblast activation protein (FAP) were synthesized by Sangon Biotech (Shanghai) Co., Ltd. (Table 1). The 2^–ΔΔCT method was used to standardize all samples to GAPDH levels, and the process was repeated three times for each sample (20).

We performed a multivariate logistic regression analysis based on the experimentally validated key genes and constructed a risk prediction model using the “rms” package (21). A nomogram was also developed to predict the risk score of each individual. Additionally, the sensitivity and specificity of the model were assessed using ROC and calibration curve analysis. Similarly, an artificial neural network model of five validated key genes was constructed using the R software package “neuralnet” (22). A classification model for DCM was established utilizing data on gene weights and four implicit layers as model parameters. The disease classification score in this model was calculated by adding the weight scores multiplied by the relevant genes’ expression levels. The ROC curve was then plotted to assess the sensitivity and accuracy of the model. A different dataset (GSE116250) taken as a test group was used for the validation of the two models mentioned above, and their respective ROC curves were plotted separately.

To better understand the diagnostic efficiency of nomogram constitutive genes for DCM, the expression of each gene was compared between the control and experimental groups using the Wilcoxon sum-rank test, and their corresponding ROC curves were plotted. Additionally, these genes were verified in the validation dataset (GSE116250).

We evaluated the relative abundance of immune cells infiltrating between samples from DCM patients and healthy individuals using the ssGSEA approach to determine the composition of immune cells in the cardiac immune microenvironment. In addition, a correlation test between immune cells was conducted using the “corrplot” R package and a heatmap depicting immune cell correlations was plotted (23). The correlation between different immune cell types was analyzed by Spearman’s method and illustrated as a heatmap. The Wilcoxon test was used to examine the immune cell differences between samples from DCM patients and those from healthy individuals. The results were then presented using a violin diagram. Additionally, the correlation between five validated hub genes and immune cells was calculated using Spearman’s method. Based on the results obtained from gene and immune cell correlation, the lollipop diagram was plotted to highlight the gene with the highest correlation to immune cells.

GraphPad Prism (version 9.3.0) was used for statistical analysis. Data are presented as the mean ± standard deviation (SD). The two groups were compared using an independent sample t-test. In addition, the Wilcoxon sum-rank test was performed for the boxplot of the average difference in the expression of hub genes. The relationship between immune cells and between immune cells and hub genes was examined using Spearman’s method. The Wilcoxon test was performed to identify immune cell differences between samples from healthy individuals and DCM patients. A value of p < 0.05 was considered statistically significant.

We analyzed the differential expression of genes in the combined dataset using the “LIMMA” program package to better understand the changes at the gene level in DCM. The merged dataset identified 322 DEGs, comprising 164 upregulated and 158 downregulated genes. The heatmap and volcano plot of DEGs are shown in Supplementary Figures S1A,B, respectively. Genes exhibiting variance greater than 25% in the merged dataset were imported to the WGCNA. The “pickSoftThreshold” function from the “WGCNA” package was used to calculate the power parameter in the range of 1–20. To ensure the reliability of the scale-free network β = 7 (scale-free R² = 0.85) was considered as a soft threshold in our study (Supplementary Figure S2B). Additionally, the threshold was set to 0.25, and the green module was merged into the blue module (Supplementary Figure S2A). Figure 1A illustrates a total of six modules, each comprising genes with comparable co-expression characteristics. A random selection of colors was made to distinguish the different modules. The signature genes from the blue module (r = 0.82; p = 4 × 10⁻⁶⁷) showed the highest positive correlation and significant association with DCM when compared to other modules (Figure 1B). As a result, the blue module containing 971 genes was identified as being crucial for DCM and hence utilized for subsequent analyses. Moreover, correlations between gene module memberships (MMs) and gene significances (GSs) were evaluated in the blue module. In the blue module, there were significant positive correlations (r = 0.91, p < 1 × 10⁻²⁰⁰) between the MM and GS of genes (Figure 1C). CGs were obtained by utilizing the intersection of the DEGs screened by the LIMMA package, the genes in the blue module, and the genes in IGs (Figure 1D).

Functional enrichment analysis revealed that the CGs were primarily implicated in calcium-related pathways, such as calcium ion homeostasis, cellular calcium ion homeostasis, positive regulation of cytosolic calcium ion concentration, calcium ion binding, and calcium signaling pathway (Figure 2A). Calcium is essential for the normal physiological function of the heart. These findings suggest that calcium-related pathways may affect the onset and progression of DCM.

The PPI network obtained by STRING was visualized using Cytoscape. Following the removal of the isolated nodes, a PPI network with 37 nodes and 74 edges, containing 28 upregulated and nine downregulated genes was established (Figure 2B). Three CytoHubba plug-in algorithms (MCC, MNC, and EPC) were then applied to screen the hub genes of the PPI network. The top eight genes selected from each algorithm were used for Venn diagram analysis (Supplementary Figure S3). Thereafter, LTBP2, ASPN, ISLR, MFAP4, FNDC1, PODN, HTRA1, and FAP were identified as hub genes. Our study indicates that these hub genes might function as potential biomarkers, leading to the development of novel therapeutic options for the diseases being studied.

An in vivo model of DCM was created by repeatedly injecting mice with DOX. As shown in previous studies (24, 25), the successful establishment of the DCM model was marked by exacerbated cardiac dysfunction and ventricular dilatation in mice. Mice administered with cumulative doses of 20 mg/kg DOX exhibited impaired heart function (Supplementary Figures S4A,B; p < 0.001), dilated left ventricles (Supplementary Figures S4C,D; p < 0.001), and thin ventricular walls (Supplementary Figures S4E,F; p < 0.001), similar to the pathophysiological symptoms of human DCM (26). Using RT-qPCR, the levels of mRNA expression of the following genes were assessed in mice hearts: LTBP2, ASPN, ISLR, MFAP4, FNDC1, PODN, HTRA1, and FAP (DCM group, n = 5; control group, n = 5). Unfortunately, a sample from the DCM group had to be discarded due to a processing error. The current findings reveal that the levels of mRNA expression of hub genes (ASPN, MFAP4, PODN, HTRA1, and FAP) were significantly higher in the hearts of DOX treated mice as compared to controls (Figures 3A–E; p < 0.05), whereas there were no significant differences in the expression levels of the remaining genes (LTBP2, FNDC1, and ISLR; Figures 3F–H; p > 0.05). The in vivo changes in the expression of these hub genes provided additional support for the microarray findings, suggesting a possible link between these genes and DCM.

Using the independent predictors mentioned above, a risk prediction model was developed and presented as a nomogram (Figure 4A). The calibration curve for the DCM risk nomogram, which is used to predict DCM risk, was shown to have a high degree of agreement in this group (Figure 4B). The ROC curve of the risk prediction model depicted an area under the curve (AUC) of 0.948 (Figure 4C), suggesting a high accuracy and sensitivity of the model. These genes were then used to build a neural network model, with the scores and weights of the feature genes comprising the input layer of this model. The extracted feature information from the input layer was used to create the implicit layer. The output layer determined if the samples belonged to either the normal or DCM group (Figure 5A). Results of the model in the training group showed that the control group had 112 correct predictions out of 144 samples (77.8% accuracy), while the DCM group had 98 accurate predictions out of 129 samples (76.0% accuracy). The AUC value of the ROC curve was calculated to be 0.821 (Figure 5B). To validate the prediction accuracy of these two models, additional models were built and ROC curves were plotted in the test group (GSE116250) using the same methods. The AUC values calculated from the ROC curve plotted by the risk prediction model (Figure 4D) and the neural network model (Figure 5C) were 1 and 0.842, respectively.

In both the combined and validation datasets, the expression level of five hub genes (ASPN, MFAP4, PODN, HTRA1, and FAP) was significantly higher in DCM patients compared to healthy individuals (Supplementary Figure S5; Figure 6). In addition, ROC curve results in the train group revealed that these genes exhibited a better diagnostic ability to differentiate DCM from healthy individuals (Supplementary Figure S6). Moreover, the diagnostic efficiency of these genes was validated in the GSE116250, and the AUC of each gene was found to be greater than 0.7 in both the combined and validation datasets (Figure 7).

The infiltration of immune cells in the myocardium can have detrimental effects on cardiac function (27). Therefore, we sought to determine the infiltration status of immune cells in DCM and normal samples, as well as identified the relationship between the infiltration status of immune cells and the expression level of genes. Between normal and DCM samples, we observed abnormal regulatory levels exhibited by gamma delta T cells, activated CD8 T cells, activated dendritic cells, CD56 bright natural killer cells, eosinophils, activated CD4 T cells, immature dendritic cells, MDSC, macrophages, mast cells, natural killer T cells, neutrophils, plasmacytoid dendritic cells, regulatory T cells, T follicular helper cells, type 1 T helper cells, type 17 T helper cells, and type 2 T helper cells (Figure 8A). Additionally, the association of 23 different types of immune cells was evaluated. The findings showed that the majority of immune cells had a strong synergistic effect, with MDSC and activated dendritic cells having the strongest synergistic effect (Supplementary Figure S7). Furthermore, ASPN and HTRA1 showed a higher correlation with most of the immune cells, as revealed by the relationship between validated hub genes and immune cells (Figure 8B). Lollipop diagrams highlighted the in-depth association between ASPN and immune cells as well as HTRA1 and immune cells (Figures 8C,D). Overall, our results indicate that the five validated hub genes (ASPN, MFAP4, PODN, HTRA1, and FAP) may contribute to the progression of DCM by regulating multiple immune cells.