Source of data

The Gene Expression Omnibus (GEO) database (http://www.ncbi.nlm.nih.gov/geo/) was utilized to obtain datasets related to HF. The GSE116250 dataset on the GPL16791 platform consisted of samples from 50 left ventricular tissues and 14 control samples [18]. The GSE57338 dataset (GPL11532) was used as the validation set, containing 177 samples of left ventricular tissue and 136 control samples [19]. the GSE5406 dataset (GPL96) was also used as the validation set, consisting of 194 samples of left ventricular tissue and 16 control samples [20]. Additionally, the GSE183852 dataset, based on the GPL24676 sequencing platform, comprised 5 transmural left ventricular apical tissue samples and 2 control samples [21]. To validate the comparability of GSE5406 with the model-development dataset (GSE116250) and initial validation dataset (GSE57338), we extracted baseline characteristics including species, tissue source, sample grouping, data type, sequencing platform, and sample size (Table 1). All three datasets were derived from human left ventricular tissues, with consistent grouping of heart failure (HF) and non-failing (NF) controls. GSE57338 and GSE5406 were microarray datasets normalized using the Robust Multi-Array Analysis (RMA) algorithm, while GSE116250 was a high-throughput sequencing dataset standardized uniformly. Batch effects were evaluated using the sva package in R, and no significant batch bias was observed after standardization, ensuring the three datasets were comparable and suitable for external validation [22].

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Standardization of the three datasets.

https://doi.org/10.1371/journal.pone.0352124.t001

The platelet activation-related pathway (REACTOME_PLATELET_ACTIVATION_SIGNALING_AND_AGGREGATION) was accessed by searching for the keyword “Platelet” in the Molecular Signatures Database (MSigDB) (https://www.gsea-msigdb.org/gsea/msigdb). The 261 genes within this pathway were designated as platelet activation-related genes (PCRGs).

Differential gene expression identification

The “DESeq2” package (v 1.38.0) [23] was employed to perform differential expression analysis on the GSE116250 dataset, with the aim of identifying differentially expressed genes (DEGs). Genes were filtered based on thresholds of P.adj < 0.05 and |log₂Fold Change (FC)| > 1. Volcano and heatmap plots were generated using the “ggplot2” (v 3.4.1) [24] and “ComplexHeatmap” (v 2.14.0) [25] packages, respectively. The top 10 upregulated and downregulated DEGs, selected based on their |log₂FC| values, were highlighted in both the volcano plot and the heatmap.

Weighted gene co-expression network analysis (WGCNA)

To identify gene modules co-expressed with NEK7 and platelet activation, we first intersected PCRGs with DEGs to pinpoint those PCRGs that were significantly differentially expressed between the HF group and the control group.

The “GSVA” package (v 1.46.0) [23] was then used to calculate the enrichment scores of these differentially expressed PCRGs. A Wilcoxon test confirmed their differential expression between the two groups (P < 0.05). A similar analysis was performed for NEK7 expression.

Based on these findings, the enrichment scores of differentially expressed PCRGs and NEK7 expression levels were used as phenotypes to construct a co-expression network using the “WGCNA” package (v 1.71) [26]. The method of constructing a signed network was adopted. During the calculation of adjacency relationships, only positive correlations (r > 0) were retained. By excluding negative correlations, this approach minimizes the interference from irrelevant reverse expression relationships related to functions and focuses the analysis on the synergistic regulatory modules related to disease progression. Based on the hclust function, we performed sample clustering to eliminate outliers and ensure data quality. In the signed network setting, the pickSoftThreshold function was used to select the optimal soft threshold power (β) within the range of 1–20. Dynamic tree cutting was used to identify co-expression modules (minimum size of 100 genes), merging those with a similarity greater than 0.4. The module that showed the strongest positive correlation with both PCRG scores and NEK7 expression was selected as the key module. Further analysis identified genes in this module that met dual thresholds (|MM| > 0.8, |GS| > 0.2) as key module genes [26].

Acquisition of candidate genes and their functional analysis

The “ggvenn” package (v 1.7.3) [27] was employed to intersect DEGs with key module genes to identify candidate genes. To explore the related biological functions and pathways, Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses were conducted on the candidate genes using the “clusterProfiler” package (v 4.2.2) [28], with significance set at P < 0.05. The GO analysis included categories such as cellular component (CC), molecular function (MF), and biological process (BP). Subsequently, these candidate genes were uploaded to the Search Tool for the Retrieval of Interacting Genes/Proteins (STRING) database (http://www.string-db.org/), employing an interaction score threshold of>0.15 to construct the protein-protein interaction (PPI) network.

Recognition of key genes

To identify key genes in the GSE116250 dataset, this study applied two machine learning algorithms to conduct feature selection and dimensionality reduction on previously selected candidate genes. First, the Least Absolute Shrinkage and Selection Operator (LASSO) regression analysis was utilized, employing the “glmnet” software package (v 4.1.4) (https://www.jstatsoft.org/v33/i01). A logistic regression model was constructed, and 10-fold cross-validation was used to determine the optimal penalty coefficient (lambda). The cross-validation aimed to minimize the deviation, resulting in the selection of lambda = 0.000160471. This model retained 10 feature genes with non-zero regression coefficients, effectively selecting features and controlling multicollinearity [29]. Second, the Support Vector Machine Recursive Feature Elimination (SVM-RFE) algorithm was concurrently used, executed through the “e1071” software package (v 4.7–1.1) [30]. This method also assessed the model performance using 10-fold cross-validation. By recursively eliminating features, it identified a combination of 5 feature genes corresponding to the lowest average error rate. The common candidate key genes selected by these algorithms were identified as the intersection and visualized using the “ggvenn” software package (v 1.7.3). To identify differentially expressed candidate biomarkers between the HF and control groups, Wilcoxon tests (P < 0.05) were performed on datasets GSE116250 and GSE57338.

Nomogram construction and evaluation

To mitigate overfitting, logistic regression with L2 regularization (ridge regression, penalty = 1) was performed using the “rms” package (v 6.5−0) [31] to construct the nomogram, with the maximum iteration number (maxit) set to 1000 to ensure convergence. The formula for calculating the line graph is: logit(P)=β₀ + β₁ × TIMP2 + β₂ × COL16A1+β₃ × MDK + β₄ × ISLR, where P is the probability of HF, β₀ is the intercept, and β₁-β₄ are the regression coefficients. For model stability assessment, 10 iterations of 10-fold repeated cross-validation were performed on the training set (GSE116250) using the “glmnet” package to calculate the mean AUC, sensitivity, and specificity. Bootstrap optimism correction with 1000 resamplings was conducted using the “rms” package to calculate the corrected AUC and calibration curves for the training and validation sets, correcting for overfitting bias. For cross-dataset calibration assessment, integrated calibration curves and stratified calibration curves were plotted by combining GSE116250, GSE57338, and GSE5406 datasets.To evaluate the discrimination of the nomogram, the “pROC” package (v 1.18.0) [32] was used to generate a Receiver Operating Characteristic (ROC) curve, and the area under the curve (AUC) was calculated to evaluate the nomogram’s predictive value (an AUC > 0.7, approaching 1, indicates higher diagnostic accuracy). Calibration curves were plotted, and Hosmer-Lemeshow test was performed (P > 0.05 indicates good calibration). For external validation sets (GSE57338, GSE5406), logistic calibration was used to correct for systematic bias in gene expression scales before plotting calibration curves. Decision curve analysis (DCA) was performed using the “rmda” package (v 1.6) [33] to evaluate the clinical net benefit of the nomogram. Furthermore, the expression levels of the biomarkers were verified in the additional external validation dataset GSE5406.

Analysis of the basic characteristics of key genes

To investigate the expression specificity of key genes, NEK7 and these genes were analyzed using the Human Protein Atlas (HPA) database (https://www.proteinatlas.org/) to display their expression across various tissues. Additionally, the chromosomal distribution of NEK7 and the key genes was visualized using the “RCircos” package (v 1.2.2) [34]. To determine their subcellular localization, NEK7 and the key genes were examined in the Universal Protein Resource (UniProt) database (https://www.uniprot.org/), which provided information on their subcellular distribution. Concurrently, Spearman correlation analysis (|correlation (cor)| > 0.3 and P < 0.05) was performed using the “psych” package (v 2.2.9) [35] to assess the correlations among NEK7 and key genes in both HF and control groups of the GSE116250 study.

Gene set enrichment analysis (GSEA)

GSEA was conducted via the “ClusterProfiler” package (v 4.2.2) to explore the biological functions of biomarkers in HF development. Spearman correlations between NEK7, key genes, and all genes in GSE116250 were calculated using the “psych” package (v 2.2.9). Genes were ranked by correlation coefficients to generate lists for NEK7 and each key gene, followed by GSEA with the “c2.cp.kegg_legacy.v2024.1.Hs.symbols” reference set from “org.Hs.e.g.,db” (v 3.16.0) [36] (P.adj < 0.05, |Normalized Enrichment Score (NES)| > 1). The top five signaling pathways, sorted in descending order according to P.adj value, of the NEK7 and key genes enrichment results were visualized.

Immune cell infiltration analysis

Immune cell infiltration analysis in GSE116250 examined differences in the immune microenvironment between the HF and control groups. Infiltration scores for 28 immune cells [37] were computed using the “GSVA” package (v 1.46.0). After filtering samples by confidence, Wilcoxon rank‑sum tests were performed to compare infiltration levels between the two groups. To control the false discovery rate (FDR) under multiple testing, the raw P values were adjusted using the Benjamini‑Hochberg (BH) procedure, and differential immune cells (DICs) were defined as those with adjusted P  .adj < 0.05. Spearman correlations were employed via the “psych” package (v 2.2.9) to explore associations among DICs and between NEK7/key genes and DICs (|cor| > 0.3, P < 0.05).

Construction of molecular regulatory networks

To investigate the regulatory mechanisms of NEK7 and key genes, their transcription factors (TFs) were predicted using the JASPAR database in NetworkAnalyst. miRNAs interacting with them were predicted from TarBase v9.0 (NetworkAnalyst) and miRDB, with the intersection of the results from these two databases considered.

Drug prediction and molecular docking

We queried the DSigDB database through “enrichR” [38] to predict candidate drugs/compounds interacting with NEK7 and key genes (P < 0.05), thereby identifying potential HF-targeting drugs. For each candidate, we selected the drug with the highest binding score and lowest P-value for molecular docking. The 3D structures of target proteins were downloaded from the RCSB PDB, and ligand structures were obtained from PubChem. Molecular docking was conducted using the CB-Dock2 online server, and binding affinity was evaluated as the total score (absolute Vina score). A total score > 4.0 kcal/mol, > 5.0 kcal/mol, and >7.0 kcal/mol indicated moderate, good, and strong binding activity, respectively.

scRNA-seq data analysis

scRNA-seq data from GSE183852 were analyzed using the “Seurat” package [39]. Initial filtering excluded cells with fewer than 200 detected genes and genes present in fewer than 3 cells; retained cells/genes met the following criteria: 1,000–35,000 features per cell, fewer than 10,000 total reads per cell, and less than 10% mitochondrial reads. The filtered data were normalized using the LogNormalize method with a scale factor of 10,000 via the “NormalizeData” function, and the top 2,000 highly variable genes were identified via “FindVariableFeatures”. PCA was used to analyze their distribution in the HF and control groups, with “JackStraw” computing PC gene P-values to select significant PCs (P < 0.05). Unsupervised clustering (resolution = 0.4) via “FindNeighbors” and “FindClusters” was visualized by UMAP. Cell types were annotated using literature-derived marker genes [40], and marker gene expression in clusters was analyzed.

Acquisition of key cells

To screen for genes related to key cells, infiltration proportions of cell types in GSE183852 HF and control groups were visualized via “ggplot2” (v 3.4.1). NEK7 and key gene expression across cell types were analyzed, with Wilcoxon tests identifying differential expression between groups within each cell type (P < 0.05). Key cell populations were selected for high abundance in both groups, detectable NEK7/key gene expression, and significant group expression differences.

Pseudo-time and functional enrichment analysis

Key cell populations underwent secondary clustering using the aforementioned approach. In GSE183852, the Monocle algorithm analyzed their pseudo-time, visualizing NEK7 and key gene expression patterns across developmental stages. “FindAllMarkers” identified genes influencing differentiation in HF samples (retaining positive markers with min.pct = 0.25, logfc.threshold = 0.25). Critical cells were subgrouped by these markers, with a heatmap displaying the top 30 DEGs. ClusterProfiler performed functional enrichment (GO: BP, CC, MF; KEGG) on differentiation-related genes (P.adj < 0.05).

Cell communication and TF regulatory analysis

To analyze interactions among critical cell populations in GSE183852 for both HF and control groups, the CellChat package (v 1.6.1) [41] was utilized to construct cell communication networks, focusing on critical cell interactions to generate maps. Additionally, the viper package (v 1.32.0) [42] analyzed regulatory relationships within critical cells, inferring TF activities across states and their downstream regulatory impacts.

Animal model of HF

To investigate the functional roles of key genes identified from public databases, HF was induced in male C57BL/6 mice (8–10 weeks old) provided by Cyagen Biosciences Inc., Suzhou, China, via continuous subcutaneous infusion of angiotensin II (Ang II). All procedures were approved by the Animal Ethics Committee of Guizhou Medical University (Approval No. 2200536; Date: March 3, 2022). The experiments adhered strictly to institutional guidelines for the humane care and use of laboratory animals.

Animal husbandry and welfare

All mice were maintained under specific pathogen-free (SPF) conditions in the barrier facility of the Laboratory Animal Center at Guizhou Medical University. Environmental parameters were strictly controlled, including temperature (22 ± 1°C), humidity (50 ± 10%), and a 12-hour light/dark cycle. Mice were group-housed, with three to four animals per cage, and had ad libitum access to a standard breeding diet and water. Bedding was replaced twice weekly. All personnel involved in the study received standardized training in laboratory rodent handling, procedural techniques, and postoperative care.

Humane endpoints and health monitoring

Humane endpoint criteria were established in accordance with the guidelines from NC3Rs. To implement these endpoints, animal health was systematically assessed through daily observations of general appearance (posture, coat condition), activity levels, and behavior, in addition to weekly body weight measurements. Immediate euthanasia was mandated if any of the following predefined criteria were met: (1) a body weight loss exceeding 20% of baseline; (2) severe lethargy, prostration, or unresponsiveness; (3) overt signs of severe distress, such as labored breathing or hunched posture; or (4) any unforeseen severe clinical event, such as seizures or paralysis. No animals died prior to meeting these endpoint criteria.

Model induction and surgical procedure

Mice were randomly assigned to either the Ang II infusion group (or the normal control (NC) group. Angiotensin II (Ang II, Macklin, catalog #A920840) was dissolved in sterile saline to deliver 1.5 mg/kg/day via osmotic minipumps (Alzet Model 2004, catalog #2004) [43]. All mice underwent subcutaneous implantation of a minipump for continuous delivery over four weeks. Pumps in the Ang II group contained the Ang II solution, while those in the NC group contained only saline.

Surgery was performed under aseptic conditions. Anesthesia was induced and maintained with inhaled isoflurane (4% for induction, 1.5–2% for maintenance). A small interscapular incision was made to create a subcutaneous pocket where a primed minipump was inserted. The incision was closed with absorbable sutures. Postoperative analgesia was administered via subcutaneous injection of buprenorphine (every 12 hours) for at least 48 hours.

At the experimental endpoint, four weeks post-implantation, cardiac function was evaluated by echocardiography. Following assessment, animals were euthanized by cervical dislocation under deep anesthesia (5% isoflurane), with death confirmed by the absence of pedal and corneal reflexes.

Echocardiography and heart morphology

Transthoracic echocardiography was performed on isoflurane-anesthetized (0.5–1%) mice prior to sacrifice using a SiliconWave 60 system equipped with an L38-22K3 transducer (30 MHz). M-mode images were analyzed by a blinded operator to determine functional and structural parameters, including left ventricular ejection fraction (LVEF), left ventricular fractional shortening (LVFS), left ventricular end-diastolic diameter (LVEDD), left ventricular end-systolic diameter (LVESD), diastolic and systolic left ventricular posterior wall thickness (LVPWTd, LVPWTs), and diastolic and systolic interventricular septal thickness (IVSd, IVSs). Following sacrifice, hearts were excised, photographed, and weighed to calculate the heart weight to body weight (HW/BW) ratio.

Reverse transcription quantitative polymerase chain reaction (RT-qPCR)

Total RNA was extracted from heart tissues utilizing TRIzol reagent (Vazyme, catalog #R401-01). Reverse transcription was conducted using the HP All-in-one qRT Master Mix II RT203-Ver.1 kit (Kunming Yungeng Biotechnology, catalog #24Y0124). Quantitative PCR was performed, with normalization to GAPDH across three biological replicates per group and three technical replicates per sample. Relative gene expression levels were determined using the 2 ⁻ ^ΔΔCt method (P < 0.05) [44]. This section represents a preliminary mRNA expression detection with a limited animal sample size (n = 3 per group).

Statistical analysis

Bioinformatic analyses were performed using R software (v4.2.2). Statistical significance was established at P < 0.05. Data from animal experiments were analyzed through GraphPad Prism 10.4.2, employing unpaired t-tests or Wilcoxon rank-sum tests as appropriate. The results are expressed as the mean ± SEM, with statistical significance set at P < 0.05.

Acquisition of candidate genes

In the dataset GSE116250, 1,073 DEGs were identified between the HF and control groups. Among these, 744 genes were up-regulated and 329 were down-regulated (P.adj < 0.05, |log₂FC| > 1) (Fig 1A,B, S1 Table).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Differential Expression Analysis of Genes in HF versus Control Groups.

(A) Volcano plot showing DEGs between the HF and control groups. The x-axis represents log₂ FC, and the y-axis shows -log10(P.adj). Significant up-regulated genes are highlighted in orange, down-regulated genes in cyan, and non-significant genes in gray. (B) Heatmap of top 10 up- and down-regulated genes based on |log₂FC|. The heatmap presents relative expression levels using a color scale, while the density distribution above the heatmap indicates the percentage of gene expression within each group.

https://doi.org/10.1371/journal.pone.0352124.g001

Acquisition of key module genes

DEGs were intersected with 261 PCRGs (S2 Table), identifying 14 differential PCRGs (APOA1, ISLR, LEFTY2, GNB3, F2R, TGFB2, SERPINA3, F2RL2, KNG1, GP5, F5, LCK, F2RL3, GNG4) (Fig 2A). In the HF versus control groups, ssGSEA scores of these 14 PCRGs (P = 0.00001) (Fig 2B, S3 Table) and NEK7 expression (P = 0.00562) were higher in HF (Fig 2C). In the training set, one outlier was removed before performing WGCNA (Fig 2D). The optimal soft threshold (β = 10) was determined when R² = 0.8, approximating a scale-free distribution (Fig 2E). Merging modules at height 0.4 yielded 11 modules (excluding gray) (Fig 2F). The turquoise module, which showed the highest correlation with both the ssGSEA scores of the 14 PCRGs and NEK7 expression, was identified as key (Fig 2G). Filtering by |MM| > 0.8 and |GS| > 0.2 identified 1,121 key module genes (Fig 2H, S4 Table).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Identification and Analysis of DEGs and PCRGs in HF.

(A) Venn diagram showing the intersection between DEGs and PCRGs, identifying 14 differentially expressed PCRGs. (B) Comparison of ssGSEA scores for the 14 PCRGs, revealing significant differences between the HF and control groups (P = 0.00001). (C) NEK7 expression levels were significantly higher in HF compared to controls (P = 0.00562), suggesting its association with HF. (D) Sample clustering dendrogram after removing an outlier from the training set. (E) Determination of the optimal soft threshold (β = 10) for WGCNA analysis based on R² and mean connectivity. (F) Gene module cluster dendrogram merged at a height of 0.4, resulting in 11 modules. (G) Heatmap highlighting the turquoise module as key due to its highest correlation with ssGSEA scores of PCRGs and NEK7 expression. (H) Scatter plot selecting 1,121 key module genes based on |MM| > 0.8 and |GS| > 0.2 criteria.

https://doi.org/10.1371/journal.pone.0352124.g002

The functions and pathways of candidate genes

DEGs were intersected with key module genes, resulting in the identification of 26 candidate genes (MDK, ISLR, TIMP2, COL16A1, SSC5D, PODN, COL14A1, LUM, OXER1, SCARF2, SLC43A2, F2R, QPCT, AEBP1, ACHE, ARHGAP22, OGN, MRC2, CTSK, CLEC11A, CPEB1, HAPLN3, ECM2, EPHA3, CHSY3, and FAP) (Fig 3A).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Functional and Pathway Analysis of Candidate Genes in HF.

(A) Venn diagram depicting the intersection between DEGs and key module genes from WGCNA, identifying 26 candidate genes. (B) Circular plot illustrating GO enrichment analysis of the 26 candidate genes, with significant enrichment in 237 GO terms (BPs, CCs, MFs). (C) KEGG pathway analysis indicating notable enrichment in protein digestion/absorption and glycosaminoglycan biosynthesis pathways. (D) PPI network analysis highlighting LUM, SSC5D, TIMP2, OGN, PODN, MRC2, and ISLR as potential key regulatory factors due to high interaction scores.

https://doi.org/10.1371/journal.pone.0352124.g003

Functional enrichment analysis characterized the biological functions and pathways of these candidate genes. They were significantly enriched in 237 GO entries (173 BPs, 27 CCs, 37 MFs; P < 0.05), including collagen metabolic process, extracellular structure organization, and ECM structural constituent for compression resistance (Fig 3B, S5 Table). Additionally, two KEGG pathways were enriched: protein digestion and absorption, and glycosaminoglycan biosynthesis, chondroitin sulfate/dermatan sulfate (P < 0.05) (Fig 3C, S6 Table). PPI analysis identified LUM, SSC5D, TIMP2, OGN, PODN, MRC2, and ISLR as key network regulators with more interactions (score>0.15) (Fig 3D).

Screened key genes: TIMP2, COL16A1, MDK and ISLR

LASSO regression identified 10 characteristic genes (MDK, ISLR, TIMP2, COL16A1, SLC43A2, ACHE, OGN, ECM2, EPHA3, FAP) with non-zero coefficients at a lambda value of 0.000160471 (Fig 4A, B). SVM-RFE analysis achieved the highest prediction accuracy using five genes: TIMP2, COL16A1, MDK, ISLR, PODN (Fig 4C). The overlap of these analyses yielded four candidate key genes: TIMP2, COL16A1, MDK, ISLR (Fig 4D).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Identification of Key Genes in HF.

(A-B) LASSO regression analysis identifying 10 characteristic genes with regression coefficients not penalized to zero at a lambda value of 0.000160471. (C) SVM-RFE analysis determining the highest model prediction accuracy with five characteristic genes: TIMP2, COL16A1, MDK, ISLR, and PODN. (D) Overlapping the characteristic genes from LASSO and SVM-RFE analyses resulted in four candidate key genes: TIMP2, COL16A1, MDK, and ISLR. (E-F) Expression analysis confirmed that these four candidate key genes were significantly up-regulated in both datasets compared to control groups, identifying them as key genes (P < 0.05).

https://doi.org/10.1371/journal.pone.0352124.g004

Expression analysis demonstrated that these four genes were significantly up-regulated in the HF versus control groups across datasets with consistent trends. Thus, these genes were identified as key genes (P < 0.05) (Fig 4E, F).

Candidate nomogram predictive performance validation and reliability assessment

First, the comparability of the three datasets was verified: baseline characteristics including species, tissue source, sample grouping and sequencing platform were highly consistent, and batch effect examination using the sva package confirmed no significant bias after unified standardization (Table 1). The candidate diagnostic nomogram was constructed incorporating L2 regularization (ridge regression) to effectively prevent overfitting (Fig 5A). 10 iterations of 10‑fold repeated cross‑validation on the training set GSE116250 yielded a mean AUC of 0.989, mean sensitivity of 0.972 and mean specificity of 0.965, verifying robust model stability.Bootstrap optimism correction with 1000 resamplings generated a corrected AUC of 0.985 for the training set; the calibration curves closely fitted the ideal diagonal, and the Hosmer–Lemeshow test yielded P = 0.928, indicating excellent calibration (Fig 5B). ROC analysis showed the nomogram achieved an AUC of 0.990 (95% CI: 0.980–1.000) in the training set GSE116250, 0.916 in the initial validation set GSE57338, and 0.908 in the additional external validation set GSE5406, with all values exceeding 0.9 and demonstrating superior discriminative ability across datasets (Fig 5C). Integrated cross‑dataset calibration confirmed that predicted probabilities were highly consistent with actual observed probabilities without significant deviation. For the two external validation sets, logistic calibration corrected systematic gene‑expression scaling bias; the recalibrated curves closely approximated the ideal line, with Hosmer–Lemeshow P = 0.179 for GSE57338 and P = 0.836 for GSE5406 (Fig 5D). Decision curve analysis (DCA) across GSE116250, GSE57338 and GSE5406 showed that the nomogram provided a significantly higher net benefit than any single‑gene model across most high‑risk threshold intervals (Fig 5E), supporting its reliable predictive performance in bioinformatic analysis. The detailed regression coefficients of the nomogram model are listed in Table 2.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Nomogram model coefficients.

https://doi.org/10.1371/journal.pone.0352124.t002

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Development and comprehensive validation of the four-gene signature-based candidate diagnostic nomogram for heart failure.

(A) Nomogram constructed using L2 regularization (ridge regression) to prevent overfitting, incorporating TIMP2, COL16A1, MDK, and ISLR for HF risk prediction. (B) Calibration curve of the nomogram in the training set (GSE116250) after 1000-times bootstrap optimism correction; Hosmer–Lemeshow test P = 0.928 indicates excellent calibration. (C) Combined receiver operating characteristic (ROC) curves of the nomogram in the training set (GSE116250, AUC = 0.990), initial validation set (GSE57338, AUC = 0.916), and additional external validation set (GSE5406, AUC = 0.908). (D) Calibration curves of the nomogram in two external validation sets (GSE57338 and GSE5406) after logistic calibration to correct systematic expression bias; Hosmer–Lemeshow test P = 0.179 for GSE57338 and P = 0.836 for GSE5406. (E) Decision curve analysis (DCA) of the nomogram and single genes in GSE116250, GSE57338, and GSE5406, showing superior net benefit of the nomogram across most high-risk thresholds. The model was validated by 10 iterations of 10-fold repeated cross-validation, 1000-times bootstrap optimism correction, and cross-dataset calibration to ensure stability and generalizability.

https://doi.org/10.1371/journal.pone.0352124.g005

Exploration of basic properties in key genes

NEK7 exhibited the highest expression in myocardial tissue (Fig 6A). Chromosomal localization identified NEK7 and COL16A1 on chromosome 1, MDK on chromosome 11, ISLR on chromosome 15, and TIMP2 on chromosome 17 (Fig 6B). Subcellular localization findings were as follows: TIMP2, MDK, and ISLR were located outside the cell membrane; NEK7 was found in the cytoplasm; COL16A1 was located in the ECM (Fig 6C). Correlation analysis (|cor| > 0.3, P < 0.05) revealed positive correlations among all genes, with COL16A1 and ISLR showing the strongest correlation (cor = 0.92, P < 0.001) (Fig 6D).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Exploration of Basic Properties in Key Genes.

(A) Expression levels of key genes (NEK7, TIMP2, COL16A1, MDK, and ISLR) in myocardial tissue. (B) Chromosomal localization of the key genes: NEK7 and COL16A1 on chromosome 1, MDK on chromosome 11, ISLR on chromosome 15, and TIMP2 on chromosome 17. (C) Subcellular localization of the key genes: TIMP2, MDK, and ISLR localized outside the cell membrane, NEK7 primarily in the cytoplasm, and COL16A1 in the ECM. (D) Correlation analysis indicating positive correlations among all key genes, with the highest correlation between COL16A1 and ISLR (cor = 0.92, P < 0.001).

https://doi.org/10.1371/journal.pone.0352124.g006

Key genes shared enriched signaling pathways

The results of the GSEA revealed that the top five pathways significantly enriched by the biomarkers were as follows (P.adj < 0.05, |NES| > 1). NEK7 showed potential correlative enrichment in 20 pathways, including the TGFβ signaling pathway and WNT signaling pathway (Fig 7A, S7 Table). TIMP2 was associated with enrichment in 10 pathways (Fig 7B, S8 Table), such as the TGFβ signaling pathway and the regulation of the actin cytoskeleton. COL16A1 was involved in six pathways (Fig 7C, S9 Table), including the TGFβ signaling pathway and ECM receptor interaction. MDK showed enrichment in seven pathways (Fig 7D, S10 Table), including ECM receptor interaction and the TGFβ signaling pathway. ISLR was enriched in 12 pathways (Fig 7E, S11 Table), involving the TGFβ signaling pathway, Notch signaling pathway, and ECM receptor interaction.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. GSEA of Key Genes.

(A) GSEA results for NEK7, showing enrichment in 20 pathways including TGFβ signaling pathway and WNT signaling pathway. (B) GSEA results for TIMP2, showing enrichment in 10 pathways including TGFβ signaling pathway and regulation of actin cytoskeleton. (C) GSEA results for COL16A1, showing enrichment in six pathways including TGFβ signaling pathway and ECM receptor interaction. (D) GSEA results for MDK, showing enrichment in seven pathways including ECM receptor interaction and TGFβ signaling pathway. (E) GSEA results for ISLR, showing enrichment in 12 pathways (e.g., TGFβ signaling pathway, Notch signaling pathway, and ECM receptor interaction).

https://doi.org/10.1371/journal.pone.0352124.g007

Key genes were implicated in the immune microenvironment of HF

The heatmap (Fig 8A) displayed the abundances of 28 types of immune cell infiltration in the HF and control groups from dataset GSE116250. With a significance level of P.adj < 0.05, 9 DICs were identified, including CD56dim natural killer (NK) cells, central memory CD8 T cells, and effector memory CD8 T cells and so on (Fig 8B). The correlation between different immune cells is such that NK cells and Effector memory CD8 T cells show the highest positive correlation (cor = 0.73, P < 0.001; |cor| > 0.3, P.adj < 0.05) (Fig 8C). The analysis of the correlation between key genes and different immune cell subsets revealed that NEK7, COL16A1, and ISLR were all significantly positively correlated with CD56dim natural killer cells, with correlation coefficients of 0.46 (P < 0.001), 0.69 (P < 0.001), and 0.77 (P < 0.001), respectively (Fig 8D-F, S12 Table); TIMP2 was significantly negatively correlated with type 17 helper T cells (r = −0.39, P < 0.001) (Fig 8G, S12 Table); and MDK was significantly negatively correlated with effector memory CD8 T cells (r = −0.66, P < 0.001) (Fig 8H, S12 Table).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 8. Analysis of Immune Microenvironment in HF.

(A) Heatmap showing immune cell infiltration levels in the HF versus control groups. (B) Boxplot comparing 9 DIC types between the HF and control groups. (C) Heatmap showing the correlation between different immune cells. Red indicates positive correlation, purple indicates negative correlation, and the darker the color, the stronger the correlation. (D) Analysis of the Correlation between NEK7 and divergent immune cells. The x-axis represents the correlation coefficient, and the y-axis represents the different immune cells. Lines pointing to the left indicate negative correlation, and lines pointing to the right indicate positive correlation. (E) Analysis of the Correlation between COL16A1 and divergent immune cells. (F) Analysis of the Correlation between ISLR and divergent immune cells. (G) Analysis of the Correlation between TIMP2 and divergent immune cells. (H) Analysis of the Correlation between MDK and divergent immune cells.

https://doi.org/10.1371/journal.pone.0352124.g008

Molecular regulatory network construction, drug prediction, and molecular docking

As depicted in S1A Fig, in the TF-mRNA-miRNA regulatory network, NEK7 was predicted to interact with nine TFs and 38 miRNAs, as indicated in S13 and S14 Tables, respectively. Similarly, COL16A1 was associated with 13 TFs and three miRNAs, MDK with 11 TFs and four miRNAs, ISLR with six TFs and five miRNAs, and TIMP2 with nine TFs and 35 miRNAs. Furthermore, NEK7 was associated with 16 potential drugs, COL16A1 with one, MDK with 15, ISLR with two, and TIMP2 with 25 potential drugs (S1B Fig, S15 Table). Additionally, molecular docking results demonstrated strong binding activities between WZ4002 and NEK7, beta-Heparin and MDK, and Oroxylin A and TIMP2. Carmustine was found to show certain binding activity with ISLR (S1C Fig).

Annotation of 6 cell types via scRNA-seq data analysis

Regarding the single-cell dataset GSE183852, prior to quality control, it comprised 49,042 cells and 36,517 genes (S2A Fig). After quality control, 48,392 cells remained with the gene count unchanged (S2B Fig). PCA identified the top 10 highly variable genes, including SPP1 and GNLY (S2C Fig). Analysis of 50 PCs revealed a stable contribution after the top 30; thus, these 30 were selected (P < 0.05; S2D-E Fig). Cell clustering categorized the cells into 16 clusters (Fig 9A-B), with marker gene expression illustrated in Fig 9C (S16 Table). The annotation process identified six cell types: T cells, pericytes, neural cells, macrophages, fibroblasts, and endothelial cells (Fig 9D-F).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 9. Cell Clustering and Annotation via scRNA-seq Data Analysis.

(A) UMAP plot showing the clustering of cells into 16 distinct clusters. (B) UMAP plots comparing cell clusters in the control and HF groups. (C) Dot plot displaying the expression levels of marker genes for each cell type. (D) UMAP plot illustrating the annotation of six cell types: T cells, pericytes, neural cells, macrophages, fibroblasts, and endothelial cells. € UMAP plots comparing the distribution of annotated cell types between the control and HF groups. (F) Dot plot showing the identity scores of the annotated cell types.

https://doi.org/10.1371/journal.pone.0352124.g009

Key cell identification

The infiltration proportions of various cell types between the HF group and the control group indicated that fibroblasts constituted a significant proportion in both groups (Fig 10A). The analysis of expression levels of NEK7 and key genes across different cell types revealed high expression in fibroblasts and neural cells (Fig 10B). Further analysis of expression differences for NEK7 and key genes between the two groups across different cell types demonstrated inter-group differences in fibroblasts (Fig 10C). Consequently, fibroblasts were identified as key cells.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 10. Identification of Key Cells in HF.

(A) Proportions of various cell types in the HF and control groups, showing a high proportion of fibroblasts in both groups. (B) Expression levels of NEK7 and key genes across different cell types, highlighting their high expression in fibroblasts and neural cells. (C) Differential expression analysis of NEK7 and key genes between the HF and control groups in various cell types, demonstrating significant inter-group differences in fibroblasts.

https://doi.org/10.1371/journal.pone.0352124.g010

Functional enrichment analysis combined with pseudotime analysis

Secondary clustering of fibroblasts, based on the top 30 principal components (S3A-B Fig), resolved the cells into nine distinct subclusters (S3C-D Fig). Pseudotime analysis identified 15 developmental stages: clusters 0, 2, 4, 6, and 8 dominated the early stages, while clusters 1, 3, 5, 7 dominated the late stages (S3E Fig). Pseudotime expression of key genes in fibroblasts demonstrated that MDK/ISLR levels increased in early stages and decreased in late stages; COL16A1/NEK7 initially decreased then increased; TIMP2 declined continuously, linking these changes to disease development (S3F-G Fig). The FindAllMarkers function identified DEGs in HF samples, subgrouping key cells. The expression of the top 30 DEGs (S3H Fig) showed BTNL9/VWF/CD36 higher in late stages, and GJA4/ITGA7/NPY1R higher in early stages. GO/KEGG enrichment of these 30 genes (S3I Fig) revealed 40 GO entries (28 BPs, 7 CCs, 5 MFs; P.adj < 0.05) including blood-brain barrier maintenance and endothelium development (S17 Table), and 80 KEGG pathways (P.adj < 0.05) such as cell adhesion molecules and ECM receptor interaction (S18 Table).

Cell communication and TF regulatory analysis

Cell communication analysis indicated that fibroblasts, as key cells, were the most abundant in both groups. In the HF group, they engaged in 15 ligand-receptor pairs with the strongest interaction with macrophages (S4A Fig); in controls, they engaged in 13 pairs with the strongest macrophage interaction (S4B Fig). The probability of fibroblasts-to-macrophage communication was highest in both groups, with MIF-(CD74 + CD44) as the primary ligand-receptor pair (S4C-D Fig). TF activity analysis (S4E Fig) revealed that HNF4A, ATF3, EBF1 exhibited higher activity in HF, while SIX2, GATA6, JUN were more active in controls, suggesting distinct regulatory roles under physiological and pathological conditions.

Validation of the Ang II-induced HF Model

Cardiac function and structure were evaluated using echocardiography and gross morphological assessments. Mice subjected to Ang II infusion demonstrated cardiac enlargement and a significantly increased HW/BW (P < 0.01, Fig 11A, B). Analysis of cardiac function revealed systolic dysfunction, characterized by a significant decrease in LVEF (P < 0.001, Fig 11D). Additionally, structural remodeling was evident, as indicated by significant increases in both LVEDD and LVESD (both P < 0.001, Fig 11F, G). The interventricular septum displayed a tendency towards hypertrophy, marked by significant increases in IVSd (P < 0.0001, Fig 11J) and IVSs (P < 0.05, Fig 11K), whereas the LVPWTd remained unchanged (Fig 11H). Collectively, these functional and structural changes confirm the successful establishment of the Ang II-induced murine HF model.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 11. Successful establishment of the Ang II-induced HF model in mice.

(A) Representative gross heart images showing cardiac enlargement. (B) HW/BW. (C) Representative M-mode echocardiograms. (D-G) Key functional and structural parameters: left ventricular ejection fraction (LVEF, D), fractional shortening (LVFS, E), end-diastolic diameter (LVEDD, F), and end-systolic diameter (LVESD, G). (H-K) Wall thickness parameters: diastolic left ventricular posterior wall (LVPWTd, H), systolic left ventricular posterior wall (LVPWTs, I), diastolic interventricular septum (IVSd, J) systolic interventricular septum (IVSs, K). (*P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001).

https://doi.org/10.1371/journal.pone.0352124.g011

Key genes expression

Following the validation of the HF model (Fig 11), we evaluated the transcriptional changes in myocardial tissue. The results from RT-qPCR indicated that, in comparison to the NC group, the Ang II infusion group exhibited significantly higher expression levels of Nek7 (Table 3, Fig 12A), Timp2 (Fig 12B), C0l16a1 (Fig 12C), and Mdk (Fig 12D) (P < 0.05). However, the expression of Islr (Fig 12E), although elevated, did not reach statistical significance (primer sequences are provided in S19 Table). These expression patterns of the key genes in the HF group were consistent with the findings from the analysis of public databases.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Differential gene expression in Ang II-induced HF.

https://doi.org/10.1371/journal.pone.0352124.t003

Download:

• PNG
  larger image
• TIFF
  original image

Fig 12. Relative mRNA expression levels of key genes in Ang II-induced HF model.

The relative mRNA expression levels of Nek7 (A), Timp2 (B), Col16a1 (C), Mdk (D), and Islr (E) were quantified by qPCR, normalized against GAPDH, and presented as mean±SEM. The results show significant upregulation of Nek7 (P < 0.01), Timp2 (P < 0.001), Col16a1 (P < 0.05), and Mdk (P < 0.01) in the Ang II infusion group compared to the controls. The expression of Islr displayed an upward trend but was not statistically significant (ns, P = 0.5691). (*P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001). Note: Animal experiments in this section had a sample size of n = 3 per group, and only mRNA-level detection was performed; the results are preliminary findings.

https://doi.org/10.1371/journal.pone.0352124.g012

Preliminary detection of mRNA expression for key genes in the murine model aligns with trends observed in transcriptomes from human HF tissues (NEK7, TIMP2, COL16A1, MDK, ISLR), providing preliminary support for the pivotal roles these genes play in the pathogenesis of HF. The lack of significant difference in Islr expression, however, warrants further investigation. Nevertheless, the animal experimental results provide preliminary support for the conclusions drawn from bioinformatics analyses, emphasizing the critical roles of these key genes in the development of HF.