2.1. Data source and preprocessing

Data sets associated with SCI, namely GSE213240, GSE183591, GSE45006, and GSE45550, have been obtained from the Gene Expression Omnibus (GEO) repository (https://www.ncbi.nlm.nih.gov/geo/ [Accessed: March 15, 2023]). In particular, the GSE45550 collection, originating from Rattus norvegicus soleus muscle tissue, incorporates 18 specimens from SCI cases (at 3, 8, and 14 days following injury) along with 6 control specimens [19]. For GSE45006, isolation occurred from Rattus norvegicus spinal cord tissue, yielding 20 SCI specimens and 4 sham-operated controls across time points of 1 day, 3 days, 1 week, 2 weeks, and 8 weeks after injury. Similarly, GSE183591, drawn from Sprague Dawley (SD) rat spinal cord, features 16 SCI specimens and 4 sham-operated controls evaluated at 1 week, 2 weeks, 4 weeks, and 8 weeks post-injury [20]. The GSE213240 single-cell RNA sequencing (scRNA-seq) dataset (platform: GPL25947) consists of 8 spinal cord tissue samples from SCI patients, sequenced via high-throughput technology. Expression matrices, sample clinical information, and annotation files for the corresponding microarray platforms were obtained using the GEOquery package. Probe ID-to-gene name mappings were extracted from the annotation files, and probe IDs in the expression matrices were converted to gene names by merging the data. Gene name cleaning was then performed by splitting and removing redundant entries separated by spaces, followed by deduplication to retain unique expression records for each gene. Sample IDs and phenotypic data were extracted from the clinical information to construct group files, and the processed expression matrices and group information were saved as standardized files. To minimize biases due to small sample sizes, the “sva” R package was used to integrate the GSE45550 and GSE45006 datasets, which were subsequently designated as the training set. Batch effects were corrected using the ComBat algorithm (based on an empirical Bayesian framework) prior to merging, and post-merge principal component analysis (PCA) validated the complete removal of batch clustering (S1A, B Fig). GSE183591 served as the external validation set. Additionally, a total of 35 macrophage polarization-related genes (MPRGs) were retrieved from the published literature [21]. Three M1 macrophage marker genes (TNF-α, HLA-DRα, and iNOS) and five M2 macrophage marker genes (IL-10, CCL22, ARG-1, CD206, and CD163) were retrieved from published literature [22].

2.2. Identification of differentially expressed genes (DEGs) between SCI and control groups

After batch effect correction, DEGs distinguishing SCI cases from controls within the GSE45006 and GSE45550 datasets were detected via the “limma” R package (version 3.54.0), under thresholds of adjusted P < 0.05 and |log₂FC| > 0.5 [23]. Enrichment analysis for the biological functions of these DEGs (adjusted P < 0.05) was then conducted, employing the “clusterProfiler” R package (version 4.7.1) in conjunction with the “org.Rn.e.g.,db” annotation package, and leveraging resources from the Kyoto Encyclopedia of Genes and Genomes (KEGG, https://www.genome.jp/kegg/ [Accessed: March 20, 2023]) as well as Gene Ontology (GO, https://www.geneontology.org/ [Accessed: March 20, 2023]) databases [24,25].

2.3. Screening of the hub genes

DE-MPRGs were initially ascertained via the overlap between DEGs and MPRGs, computed with the aid of the “ggvenn” R package (version 0.1.9). Thereafter, selection of genes demonstrating substantial biological relevance was accomplished through application of extreme gradient boosting (XGBoost), random forest (RF), and least absolute shrinkage and selection operator (LASSO) techniques. Specifically, for the LASSO model, family = “binomial” was set to suit binary classification tasks, and type.measure = “auc” was designated as the performance evaluation metric. A regularization parameter (λ) was introduced, and lambda.min (the value yielding the minimum error) was selected as the optimal regularization parameter via 5-fold cross-validation, followed by extraction of genes with non-zero coefficients. For the RF model, optimal parameters were set as mtry = 3 and ntree = 1,000; ensemble learning across multiple decision trees reduced the variance of individual models, while built-in out-of-bag (OOB) sample validation was used instead of an independent test set to evaluate model generalization without additional sample consumption. For the XGBoost model, a reasonable learning rate (eta = 0.5), tree depth (max_depth = 5), and number of iterations (nround = 25) were configured to balance fitting ability and complexity, with gradient boosting mechanisms employed to iteratively correct errors. Hub genes were designated as those emerging from the common overlap among the three machine learning algorithms. To eliminate the potential impact of data pooling, DEGs of GSE45006 and GSE45550 were separately identified using the “limma” R package (version 3.54.0) (adjusted P < 0.05 and |log₂FC| > 0.5). Subsequently, the “RobustRankAggreg” (RRA) package (version 3.54.0) was applied to integrate the ranking of DEGs across the two datasets. The aggregateRanks function was utilized to calculate the integrated score for each gene, with a lower score indicating a higher ranking. This analysis generated a cross-dataset list of significantly integrated genes, and special attention was paid to the rankings of the core genes within this integrated list.

After that, correlations between these hub genes were investigated through Spearman correlation analysis. Employing the GENEMANIA database (http://genemania.org/ [Accessed: April 27, 2023]), an interaction network linking genes for the hub genes was established. Ultimately, incorporation of the hub genes into a nomogram was achieved via the “rms” R package (version 6.2.0). Reliability of this nomogram was assessed by means of receiver operating characteristic (ROC) curves and decision curve analysis (DCA).

2.4. Gene set enrichment analysis (GSEA) of the hub genes

Training set specimens were categorized into cohorts displaying elevated and reduced expression utilizing the median values from hub genes, prompting subsequent differential expression evaluation. Genes in the entire collection received ordering according to log₂FC metrics. Application of the “ClusterProfiler” R package (version 4.7.1) enabled GSEA implementation under the cutoff of adjusted P < 0.05 [26].

2.5. Ingenuity pathway analysis (IPA) of the hub genes

To explore the biological functions of the hub genes, IPA was performed based on the QIAGEN Knowledge Base (https://digitalinsights.qiagen.com/products-overview/qiagen-knowledge-base/ [Accessed: April 16, 2023]). Subsequently, we analyzed the diseases and functional annotations to which the hub genes are enriched.

2.6. Immune analysis

Gene sets pertaining to immune responses were sourced from the ImmPort repository (https://www.immport.org/ [Accessed: April 21, 2023]). Through single-sample gene set enrichment analysis (ssGSEA) executed within the “GSVA” R package (version 1.46.0), enrichment metrics for such gene sets were derived for control and SCI cohorts [27]. Variances in these enrichment metrics across the control and SCI cohorts underwent evaluation. Associations linking differentially expressed immune response gene sets to hub genes were probed via Spearman’s rank correlation technique. Moreover, marker genes indicative of human M2 and M1 macrophages were mapped to their rat orthologs employing the “homologene” R package. Scrutiny extended to the relationships connecting hub genes with these rat M1/M2 macrophage marker genes (stemming from human orthologs).

2.7. Evaluation of immune infiltration

Upon procurement of gene expression profiles, immune cell infiltration abundances were estimated employing the CIBERSORT method. Relative abundances among 22 immune cell subtypes per specimen were illustrated through stacked bar charts, alongside evaluation of infiltration variances across cohorts. Moreover, associations linking differentially infiltrated immune cell subtypes to one another, plus those connecting hub genes to such subtypes, were examined via Spearman’s rank correlation.

2.8. Creation of micro RNA (miRNAs)-mRNA network and transcription factors (TFs)-mRNA network

Orthologs corresponding to hub genes in humans were mapped through the “homologene” R package. miRNet repository (https://www.mirnet.ca/miRNet/home.xhtml [Accessed: April 29, 2023]) facilitated the forecasting of miRNAs directed toward hub genes. Concurrently, transcription factors (TFs) aimed at these genes underwent prediction leveraging the NetworkAnalyst tool (https://www.networkanalyst.ca/ [Accessed: April 30, 2023]). Interactions in TF-mRNA and miRNA-mRNA regulatory frameworks were depicted employing Cytoscape software (version 3.8.0) [28].

2.9. Drug prediction of the hub genes

Conversion to human orthologs for hub genes was executed by means of the “homologene” R package. Queries in the Drug-Gene Interaction Database (DGIdb; https://dgidb.genome.wustl.edu/ [Accessed: April 27, 2023]) relied on individual hub genes functioning as search keywords to pinpoint pharmaceuticals that engage with them.

2.10. Processing and annotation of scRNA-seq data

Analysis at the single-cell resolution scrutinized the interplay linking hub genes to macrophage polarization. Sequencing profiles derived from GSE213240 initially underwent filtration through the “Seurat” R package (version 5.0.1) [29]. In detail, quality assurance imposed three thresholds: mitochondrial gene percentage confined below 10%, detected features per cell (nFeature) bounded from 200 to 2,500, and counts per cell (nCount) capped under 6,000. Thereafter, selection of the foremost 2,000 highly variable genes occurred via the FindVariableFeatures tool, relying on the association of expression averages with dispersions; these selections persisted for ongoing evaluations. Dataset specimens subsequently experienced PCA. The JackStraw function was employed to conduct a permutation test for the null distribution, aiming to quantify the percentage of variance explained by the top 30 principal components (PCs). Prominent variance-explaining principal components (PCs) underwent selection through examination of the elbow curve diagram, whereupon cellular unsupervised grouping ensued via invocation of FindClusters and FindNeighbors functions (resolution = 0.2). Labeling and contrasting of cell identities proceeded with deployment of the “SingleR” R package (version 2.0.0) integrated alongside the CellMarker repository (http://117.50.127.228/CellMarker/ [Accessed: April 22, 2023]). Ultimately, disparities in hub gene expression across M2 and M1 macrophage populations received scrutiny (adjusted P < 0.05; Bonferroni correction).

2.11. Expression analysis of the hub genes

For additional corroboration of the robustness inherent in our results, disparities in hub gene expression across control and SCI cohorts were assessed via the Wilcoxon rank-sum test (P < 0.05) within the training cohort together with the GSE183591 collection.

2.12. Spinal cord injury rat model

Female Wistar rats in good health (Rattus norvegicus; 200 g body weight, aged 8 weeks, sourced from Janvier Labs, France) underwent random allocation into one sham-operated group and three SCI groups (3 days, 7 days, and 14 days post-injury). The initial number of animals in each SCI group was 8 rats, whereas the sham-operated group consisted of 7 rats. Compliance with the NIH Guidelines for the Care and Use of Laboratory Animals (8th edition, revised 2010, NIH) guided the investigation, wherein the fewest possible animals per cohort sufficed to yield data robust enough for statistical inference, aligning with core ethical tenets; experimental designs secured endorsement from the Anhui Medical University Animal Ethics Committee (approval LLSC20231977, Hefei, China), while every operation conformed to prevailing ethical norms and followed the ARRIVE Checklist (S1 File). Proper husbandry attended to all subjects, who occupied enclosures limited to at most three individuals apiece, maintained within specific pathogen-free (SPF) environments featuring regulated ambient temperature (22 ± 1 °C), humidity levels (50 ± 1%), alongside a 12 h light/12 h dark photoperiod. Ad libitum access prevailed for both feed and hydration.

As indicated beforehand, induction of the contusion/compression SCI paradigm relied on a customized 28 g aneurysm clip (Fehlings Laboratory, Canada). Concisely, subjects received a 1:1 blend of N₂O and O₂ alongside isoflurane (1.5–3%) for eliciting systemic anesthesia. Subsequently, posterior integument incision occurred at the T10–T8 cord segment, trailed by T9 lamina resection to unveil the neural tissue; the clip then encircled the cord and persisted for 60 s, culminating in a compression-contusion lesion [30–32]. Rats, upon emergence from anesthetic effects, exhibited trailing of hindlimbs. Laminotomy at T9, sparing neural integrity, constituted the sham cohort. Application of ocular salve averted corneal aridity, whereas a thermal mat preserved subject thermoregulation. The animals received comprehensive post-surgical care. Immediately following surgery, they were administered a subcutaneous injection of 4 ml saline. The bladder was manually emptied two to three times daily (based on signs of wet abdomen and bladder distension) until spontaneous urination was re-established. Preventive antibiotics, including 2.5% Baytril (enrofloxacin, Bayer AG, Leverkusen, Germany), meloxicam (2.0 mg/kg), and buprenorphine (0.05 mg/kg subcutaneously), were administered daily. All analgesics and antibiotics were administered daily after surgery and continued for up to seven consecutive days, or until the day of sacrifice for animals in the 3-day group. Treatment was discontinued once rats regained spontaneous urination and showed no obvious signs of pain or infection. Throughout the experiment, trained personnel monitored the animals daily for signs of pain, stress, or other discomfort. The animals’ general condition was assessed using a scoring system, and appropriate interventions were performed based on the scores to minimize suffering. If the predetermined humane endpoint criteria were met, the animals were euthanized immediately. Euthanasia was performed by placing the animals in an induction chamber and administering 5% isoflurane by inhalation until they reached a state of complete unconsciousness, as indicated by the absence of righting reflex, respiration, and response to noxious stimuli. Once unconsciousness was confirmed, subsequent experimental procedures were conducted. Animals from the experimental group were sacrificed on days 3, 7, and 14 post-SCI for spinal cord tissue collection. A total of 7 rats per group were included in the subsequent experiments (1 rat died in each SCI group during surgery or the experimental period, while all 7 rats in the sham-operated group survived). This approach ensured that the animals remained unaware during euthanasia and complied with internationally recognized guidelines, including the AVMA Guidelines for the Euthanasia of Animals (2020 Edition) and Directive 2010/63/EU of the European Parliament. Additional information is provided in the Supporting Information, including the Inclusivity in global research questionnaire (S2 File) and the PLOS ONE humane endpoints checklist (S3 File).

2.13. Reverse transcription-quantitative PCR (RT-qPCR)

T9 spinal cords were collected at 3, 7, and 14 days post-SCI or after laminectomy, immediately frozen, and stored at −80°C. Total RNA was then extracted using the NucleoSpin® RNA Isolation Kit (Macherey-Nagel GmbH & Co., Düren, Germany). Primers for each gene are listed in Table 1.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Primers for RT-qPCR.

https://doi.org/10.1371/journal.pone.0347599.t001

Relative expression levels of specific genes/markers in RT-qPCR were assessed using the Luna® Universal RT-qPCR Master Mix system (New England Biolabs, Massachusetts, USA). RT-qPCR was performed using the StepOne Plus system (Applied Biosystems, Massachusetts, USA). Duplicate cycle thresholds (CTs) were averaged, and data were normalized to the expression of β-actin mRNA using the 2^–ΔΔCt method [33].

2.14. Statistical analysis

R software (https://www.r-project.org/) enabled the implementation of all statistical computations. Disparities across cohorts received evaluation through Wilcoxon scrutiny, signifying meaningful divergence when P < 0.05. For RT-qPCR evaluations, Gaussian-form data invoked Student’s t-test application, while skewed distributions prompted Wilcoxon rank-sum deployment, with thresholds at P < 0.05 marking evidentiary weight. Multiplicity safeguards, where applicable, entailed P-value recalibration via the Benjamini-Hochberg (BH) procedure to mitigate false discovery rate (FDR).

3.1. Acquisition and enrichment analysis of DEGs between SCI and control groups

Between SCI and control cohorts, a cohort of 1,396 DEGs surfaced, featuring 820 instances of upregulation coupled with 576 cases of downregulation (Fig 1A; S1 Table). Hierarchical clustering via heatmap illustrated the premier 20 upregulated alongside downregulated genes (Fig 1B). Herein exhibited stand the foremost 10 enriched designations from GO alongside routes from KEGG. Across 1,297 GO designations, DEGs manifested enrichment (S2 Table), partitioned into 1,094 entries for biological process (BP), 137 for cellular component (CC), and 66 for molecular function (MF). As illustrations, engagement by DEGs encompassed membrane microdomain, membrane raft, synaptic membrane, positive regulation of response to external stimulus, immune response-regulating signaling pathway, myeloid leukocyte activation, glycosaminoglycan binding, cell adhesion molecule binding, and proteoglycan binding (Fig 1C). Among the KEGG pathways, 126 were significantly enriched (S3 Table), such as osteoclast differentiation, NF-κB signaling pathway, and pertussis (Fig 1D).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Differentially expressed genes analysis and enrichment analysis of DEGs.

(A) Volcano plot of DEGs. (B) Heatmap of differentially expressed genes (Top 20 DEGs). (C) GO enrichment analysis of DEGs. (D) KEGG enrichment analysis of DEGs.

https://doi.org/10.1371/journal.pone.0347599.g001

3.2. Acquisition of the hub genes

Six differentially expressed macrophage polarization-related genes (DE-MPRGs) were identified by intersecting DEGs with MPRGs: Fgd2, Soat1, Comt, Myo1f, Maf, and Sh2b2 (Fig 2A). The least absolute shrinkage and selection operator (LASSO) algorithm selected three characteristic genes: Myo1f, Soat1, and Comt (Fig 2B). The extreme gradient boosting (XGBoost) algorithm identified five characteristic genes: Fgd2, Soat1, Comt, Myo1f, and Sh2b2 (Fig 2C). The random forest (RF) algorithm identified four characteristic genes: Fgd2, Soat1, Comt, and Myo1f (Fig 2D-E). The intersection of characteristic genes from the three algorithms yielded three hub genes: Soat1, Comt, and Myo1f (Fig 2F). Meta-analysis using the RRA package revealed that three genes were stably identified in the integrated results. Specifically, Soat1 was ranked 1,336th (top 8.84%, score = 0.0936), Myo1f was ranked 1,605th (top 10.62%, score = 0.1139), and Comt was ranked 2,972nd (top 19.66%, score = 0.2258) (S4 Table). Correlation analysis revealed significant positive correlations among the hub genes (Fig 2G), with the strongest correlation between Soat1 and Myo1f (R = 0.901, P = 5.77e-18) (Fig 2H). To explore interactions among the three hub genes, a gene-gene interaction network was constructed (Fig 2I). These genes formed a complex interaction network, with the following interaction types: co-expression (39.02%), physical interactions (25.61%), predicted interactions (21.99%), pathway associations (8.15%), co-localization (2.86%), and shared protein domains (2.37%). A nomogram incorporating the three hub genes was developed to predict SCI progression (Fig 3A). The nomogram’s ROC curve yielded an area under the curve (AUC) of 0.866 (Fig 3B). The DCA curve demonstrated that the nomogram provided a higher net benefit (Fig 3C). These findings indicate that the nomogram exhibits high predictive accuracy for SCI progression.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Acquisition of the hub genes.

(A) Venn plot of DEGs and macrophage polarization related genes (MPRGs). (B) LASSO algorithm. (C) XGBoost algorithm. (D) The number of RF trees correlates with model errors. (E) Average minimum Gini coefficient diagram. (F) Venn plot of three machine learning. (G) Hub gene correlation analysis. (H) Scatter plots of the two genes with the highest positive correlation. (I) PPI protein interaction network.

https://doi.org/10.1371/journal.pone.0347599.g002

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Construction and evaluation of nomogram model.

(A) The nomogram containing the three hub genes. (B) ROC curve. (C) DCA curve.

https://doi.org/10.1371/journal.pone.0347599.g003

3.3. Enrichment analysis of the hub genes

To further elucidate the roles of hub genes in SCI pathogenesis, GSEA was performed. The top five enriched KEGG pathways are presented. In the high-expression group, all hub genes were enriched in natural killer cell-mediated cytotoxicity, systemic lupus erythematosus, lysosome, and complement and coagulation cascades (Fig 4A-C). Additionally, both Soat1 and Myo1f were enriched in the cytokine-cytokine receptor interaction pathway in the high-expression group (Fig 4B-C).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. GSEA analysis of the hub genes.

(A) GSEA analysis of Comt. (B) GSEA analysis of Soat1. (C) GSEA analysis of Myo1f.

https://doi.org/10.1371/journal.pone.0347599.g004

3.4. IPA of the hub genes

A total of 582 canonical pathways were significantly enriched for DEGs between the SCI and control groups, with the top 20 enriched pathways presented (Fig 5A). The three hub genes were involved in four canonical pathways: dopamine receptor signaling, L-DOPA degradation, noradrenaline and adrenaline degradation, and dopamine degradation. Among these, the dopamine receptor signaling pathway had the highest |z-score| and is therefore illustrated in Fig 5B. In terms of disease and functional annotations, the DEGs were primarily associated with organismal injury and abnormalities (Fig 5C).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. IPA of the hub genes.

(A) Top 20 classical IPA pathways (B) Interactive mapping of dopamine receptor signaling pathways. (C) Statistical analysis of disease and functional enrichment.

https://doi.org/10.1371/journal.pone.0347599.g005

3.5. Immune analysis between SCI and control groups

In SCI cohorts compared to controls, eleven gene sets associated with immune responses displayed substantial differential expression, spanning tumor necrosis factor (TNF) family member receptors, T cell receptor (TCR) signaling pathway, transforming growth factor-beta (TGF-β) family member receptors, interferon receptors, natural killer cells, interleukin receptors, B cell receptor (BCR) signaling pathway, cytokine receptors, antimicrobials, chemokine receptors, and antigen processing and presentation (Fig 6A; S5 Table). Scrutiny of correlations uncovered strong positive linkages across all hub genes and every differentially expressed immune response gene set, with the peak association evident between Myo1f and the B cell receptor (BCR) signaling pathway (R = 0.925, P = 5.44e-22) (Fig 6B). The marker genes of M1 and M2 macrophages from humans were converted to their orthologous genes in rats, namely RT1-Da, Tnf, Cd163, Mrc1, Arg1, Ccl22, and Il10. The relevant analysis indicated that Mrc1 was markedly positively associated with all hub genes (Fig 6C). RT1-Da and Cd163 were markedly positively associated with Myo1f and Soat1. Additionally, Comt was negatively associated with Tnf and Ccl22.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Immune analysis between SCI and control groups.

(A) There were 11 immune response gene sets that were significantly differential between SCI and control groups. (B) Relevant analysis between hub genes and immune response. (C) The relevant analysis between hub genes and macrophage polarization markers.

https://doi.org/10.1371/journal.pone.0347599.g006

3.6. Analysis of immune cell infiltration

The CIBERSORT algorithm facilitated quantification of infiltration fractions for 22 immune cell varieties within specimens, underscoring the pivotal influence of alterations in the immune milieu and associated signaling cascades on SCI advancement and recovery (Fig 7A). Elevated abundances of CD8 ⁺ T cells and resting dendritic cells marked the SCI cohort relative to controls, as our observations revealed. In opposition, the SCI ensemble manifested reduced incidences of memory CD4 ⁺ T cells, M0 macrophages, and follicular helper T cells (Tfh cells) when juxtaposed against the control ensemble (Fig 7B). Associations among immune infiltrates in these specimens underwent further interrogation (Fig 7C). Plasma cells forged a marked affirmative bond with resting natural killer (NK) cells (R = 0.63), while memory B cells evinced pronounced discord with naive B cells (R = −0.83). Associations spanning the triad of hub genes and immune infiltrates received additional examination (Fig 7D). Inverse linkages tied the three hub genes to resting memory CD4 ⁺ T cells, follicular helper T cells, M0 macrophages, and plasma cells. Affirmative ties linked every one of the three hub genes to resting dendritic cells and CD8 ⁺ T cells, mirroring the intergroup disparities noted earlier.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. Analysis of immune cell infiltration.

(A) The percentage stacked column plot illustrates the relative distribution of 22 immune cell types across each sample. (B) Box plots illustrate the levels of immune cell infiltration in the SCI and control groups. (C) The correlation heatmap illustrates the relationships between different immune cell compositions. (D) Heatmap of hub genes correlation with immune cells.

https://doi.org/10.1371/journal.pone.0347599.g007

3.7. The TF-mRNA-miRNA networks

Elucidation of the mechanistic underpinnings dictating hub gene operations in SCI necessitated assembly of regulatory architectures encompassing miRNA-mRNA and TF-mRNA interactions. The miRNA-mRNA construct, encompassing 273 miRNAs alongside 3 mRNAs (specifically, the set of three hub genes), emerged (Fig 8A). Illustrative regulatory pairings spanned MYO1F-hsa-miR-27b-3p, SOAT1-hsa-miR-485-5p, COMT-hsa-miR-637, and similar associations. Complementing this, the TF-mRNA framework integrated 30 TFs with the 3 mRNAs (i.e., the three hub genes) (Fig 8B). For instance, the TF TEAD1 could regulate SOAT1 and COMT. Three TFs (TFAP2C, NFKB1, and CREB1) were common TFs regulating both COMT and MYO1F.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 8. The networks of hub genes.

(A) The miRNA-mRNA network of hub genes. (B) The TF-mRNA network of hub genes. (C) The hub gene-drug interaction network.

https://doi.org/10.1371/journal.pone.0347599.g008

3.8. Drug prediction

The hub gene-drug interaction network was constructed using DGIdb. The results revealed that a total of 37 small-molecule drugs were predicted by the database (Fig 8C). Five small-molecule drugs were associated with SOAT1: nifedipine, dexamethasone, testosterone, atorvastatin, and lovastatin. For COMT, 32 small-molecule drugs were predicted, including fluvoxamine, tolcapone, and entacapone. However, no small-molecule drugs were predicted for MYO1F as the target gene.

3.9. Differential expression of hub genes between different macrophage polarization states

In GSE213240, a total of 66,093 cells and 19,404 genes met the inclusion criteria (S2A Fig, B). Subsequent analysis identified 2,000 highly variable genes with the greatest variance, such as Igfbp5 (Fig 9A). Among these, the top 30 principal components (PCs) contributed the most to the variation (Fig 9B-9C). Clustering based on the top 30 PCs resulted in 17 cell clusters (Fig 9D). These clusters were classified into nine cell types, including T cells, B cells, M2 macrophages, M1 macrophages, oligodendrocytes, NK cells, endothelial cells, granulocytes, and monocytes. In SCI samples, granulocytes represented the largest proportion, followed by M1 macrophages, indicating an inflammatory microenvironment in SCI (Fig 9E). Notably, Soat1, Comt, and Myo1f showed significant differential expression between M1 and M2 macrophages, all being highly expressed in M1 macrophages (adjusted P < 0.0001, log₂FC > 1.50) (Fig 9F). These findings further confirm the association of these hub genes with the inflammatory microenvironment of SCI.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 9. Significant differences exist in the expression of hub genes between M1 and M2 macrophages.

(A) Visualization of highly variable genes. Red dots indicate highly variable genes, while black dots represent non-highly variable genes. (B) Jackstraw plot for PCA. The horizontal axis shows empirical values calculated by the Jackstraw function, the vertical axis displays computed theoretical values, and the legend on the right indicates the P-value for each principal component. (C) Number of available dimensions in the fragment plot. The horizontal axis represents the number of PCs, while the vertical axis shows the standard deviation. (D) UMAP plot of cell clusters. The x-axis represents the first principal component axis after UMAP dimensionality reduction, while the y-axis represents the second principal component axis. Points of different colors denote distinct cell clusters. (E) UMAP plot of annotation results for 9 cell types across all samples. (F) Violin plots showing the expression levels of Soat1, Comt, and Myo1f in M1 and M2 macrophages.The overlaid box plot (white) illustrates the median and interquartile range, while the scatter plot displays the expression values of individual cells. Between-group comparisons were performed using the Wilcoxon rank-sum test with Bonferroni correction for multiple testing.

https://doi.org/10.1371/journal.pone.0347599.g009

3.10. The expression levels of the hub genes

The expression of hub genes between the SCI and control groups was further analyzed in the training set and the GSE183591 dataset using the Wilcoxon test (Fig 10A, 10B). The results demonstrated a significant difference in the expression of all hub genes between the SCI and control groups in the training set, with higher expression levels observed in the SCI group (Fig 10A). In the GSE183591 dataset, the expression trends of all hub genes were consistent with those in the training set (Fig 10B). These findings suggest that all hub genes exhibit strong diagnostic potential for SCI.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 10. The expression levels of the hub genes.

(A) The expression of all hub genes of SCI group was markedly higher in the training set. (B) In GSE183591, the expression of all hub genes of SCI group was markedly higher.

https://doi.org/10.1371/journal.pone.0347599.g010

3.11. Validation of hub genes by RT-qPCR

RT-qPCR was performed on rat spinal cord tissue samples collected at 3, 7, and 14 days post-SCI to validate the reliability of the hub genes identified through database analysis. Consistent with the findings from the GEO databases, the expression levels of Soat1, Comt, and Myo1f were significantly elevated in the SCI group compared to the control group at all time points (Fig. 11, S6 Table). These results corroborated the data mining outcomes, further reinforcing the reliability of our findings.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 11. Verification of differentially expressed hub gene mRNAs.

Relative mRNA expression levels of the Comt, Myo1f, and Soat1 genes were assessed in both injured spinal cord and sham spinal cord samples by RT-qPCR at different time points. Data are expressed as the mean ± SD (n = 7). *P < 0.05, **P < 0.01, ***P < 0.001 (Student’s t-test).

https://doi.org/10.1371/journal.pone.0347599.g011