2.1. Preparation of the rabbit model and RNA sample collection

New Zealand rabbits, aged 6–7 months, weighing 2.5–3 kg, and exhibiting normal limb movement (assessed via visual observation, gait analysis, and functional tests), were selected for all in vivo experiments. The animals were sourced from Wangdu Tong Hui Animal Breeding Co. Ltd. (Baoding City, China; Animal Certificate number: 210509). The animal experimental protocol was approved by the Medical Ethics Committee of the Third Hospital of Hebei Medical University (Approval number: z2023-012–01). All experimental procedures involving animals were conducted in accordance with the ethical standards set forth by the Animal Experimentation Ethics Committee of the Third Hospital of Hebei Medical University, Shijiazhuang, China (reference number: z2023-012-1). The rabbits were maintained in standard laboratory conditions with a controlled temperature at 18–22°C, humidity at 50–60%, and a 12-hour light/dark cycle. The rabbits were fed a standard laboratory rabbit diet provided by Qingdao Kangda Aibo Biotechnology Co., Ltd. (Qingdao, China; License No.: SCXK(lu)20220004) and had access to water ad libitum. Rabbits were randomly allocated to four groups using computer-generated randomization (R v4.3.0; seed = 2024), which were sorted to assign animals to the Control, Model, PMMA, and ES-PMMA groups, ensuring unbiased distribution [38]: a blank control group (n = 5), a model group (n = 10), PMMA (n = 10), and ES-PMMA (n = 10) groups. For anesthesia and analgesia, animals were anesthetized using urethane (Shanghai Yuanye Bio-Technology Co., Ltd., Shanghai, China; Catalog No.: S30146) administered via marginal ear vein injection at a dose of 1.5 g/kg body weight. Postoperative pain was managed by the local injection of lidocaine at the surgical site. A blank group served as the control group, for which only the skin and subcutaneous tissue of the distal femur were incised without damaging the bone. Subsequently the wound was meticulously sutured layer-by-layer, and antibiotics were administered for 24 hours to prevent infection. Samples were obtained directly from the cancellous bone tissue of the distal femur. In the model group, a bone defect was created at the distal femur to simulate bone injury. A circular bone defect, with a diameter of 3 mm and a depth of 1 cm, was carefully formed by sequentially removing layers of skin, subcutaneous tissue, and periosteum at the lateral end of the femur. The incisions were then closed in layers. To minimize animal suffering, the following measures were implemented: (1) preoperative fasting for 6 hours to reduce stress during anesthesia; (2) daily monitoring of wound healing, mobility, and behavior post-surgery; and (3) prompt euthanasia upon completion of the study to avoid prolonged distress. After 10 days, and in accordance with the guidelines set by the Institutional Animal Care and Use Committee (IACUC), the rabbits were euthanized humanely via the intravenous injection of KCL (10% solution) following the completion of tissue collection. Cancellous bone tissue surrounding the bone defect in the distal femur was collected for analysis. Rabbits in the PMMA group were treated in a manner similar to the model group, with the exception that a PMMA mold was implanted into the bone defect to simulate its application for treating such defects. A cylindrical PMMA mold, with a diameter of 3 mm and a length of 1 cm, was inserted into the bone defect at the distal femur. Following a period of 10 days for healing and integration, the cancellous bone tissue surrounding the mold in the distal femur was harvested. For the ES-PMMA Group, we inserted an ES-PMMA implant into an artificially created bone defect to mimic its application to treat these defects. ES-PMMA was prepared as described previously [26–28]. In brief, enoxaparin sodium (ES) was incorporated into poly (methyl methacrylate) (PMMA) at a concentration of 5% w/w under controlled polymerization conditions (monomer-to-polymer ratio of 2:1, polymerization temperature of 37°C). The mixture was molded into cylindrical implants (3 mm diameter × 1 cm length) to ensure consistent drug release and mechanical strength. A cylindrical ES-PMMA implant (3 mm diameter × 1 cm length) was placed inside the manufactured bony lesion located at the distal femur. Ten days post-surgery, the cancellous bone tissue surrounding the molded region in the distal femur was harvested for analysis. All samples were placed in Trizol Reagent (Invitrogen, Carlsbad, CA, USA) and sent to Beijing Zhongke Shengxin Biotechnology Co. Ltd. for sequencing as shown in Fig 1, which depicts the steps from incision to implant insertion and sample harvesting.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Surgical procedure and tissue sample collection in rabbits. A. A lateral incision was made in the distal femur of the rabbit, and the skin and subcutaneous tissues were dissected layer by layer to fully expose the bone tissue. B. A drill was used to create a bone defect by drilling perpendicular to the bone surface. C. The bone defect was refined to achieve a diameter of 3 mm and a depth of 1 cm. D. According to the ratio established in our previous study (monomer-to-polymer ratio of 2:1), we prepared Polymethylmethacrylate (PMMA)/Enoxaparin Sodium-Polymethylmethacrylate (ES-PMMA) cylindrical implants with a diameter of 3 mm and a length of 1 cm. E. The cylindrical implant was inserted into the bone defect, and the subcutaneous tissues and skin were sutured layer-by-layer, followed by wound dressing. F. After allowing for a period of 10 days for healing and integration, we harvested cancellous bone tissue around the molded area in the distal femur.

https://doi.org/10.1371/journal.pone.0332041.g001

2.2. Quality control procedure for RNA sequencing

Total RNA was extracted using TRIzol reagent (Invitrogen), and RNA integrity was evaluated using an Agilent 2100 Bioanalyzer (Agilent Technologies). Only samples with an RNA Integrity Number (RIN) ≥ 7.0 were included for library construction. Ribosomal RNA contamination was assessed by aligning raw reads to the SILVA rRNA database, and samples with > 5% rRNA reads were subjected to ribosomal depletion. Raw sequencing data were pre-processed using Trimmomatic software (v0.39) (http://www.usadellab.org/cms/?page=trimmomatic) with default parameters to remove splice contaminants and low-quality reads. Filtered clean reads were then aligned to the human reference genome (GRCh37) using Hisat2 software (v2.2.1) (https://daehwankimlab.github.io/hisat2/), with the following key parameters: --outFilterMismatchNmax 2 (maximum 2 mismatches per read), --alignIntronMin 20 (minimum intron size 20 bp), --alignIntronMax 1000000 (maximum intron size 1 Mb), and --outFilterMultimapNmax 10 (allow up to 10 multi-mapping locations per read). The alignment between each sample and the reference genome were compiled into a table for comparative analysis. Reads with a mapping quality (MAPQ) < 30 were filtered out. Strand-specificity was confirmed using RSeQC. Subsequently, gene expression profiles were extracted using Feature Count software (v2.0.0) (https://sourceforge.net/projects/subread/), with default parameters.

2.3. Identification of differentially expressed genes (DEGs)

To model and identify sequence differences between the normal, model, PMMA, and ES-PMMA groups, principal component analysis (PCA) in R package (https://cran.r-project.org/) was first performed for data visualization and outlier detection, followed by DEG analysis. The screening thresholds for DEGs were as follows: (1) statistical significance (p < 0.05), (2) fold change (|log2FC| > 1), (3) adjusted p-value (Benjamini-Hochberg correction for multiple testing), (4) replication number (biological replicates ≥3 for reliability), (5) normalization (TMM/FPKM for library size correction), and (6) filtering (low-expression gene removal).

2.4. Differential expression analysis and the identification of key candidate genes associated with ES-PMMA

The DESeq2 package in R software was used to identify DEGs between the model and control groups, between the PMMA and model groups, and between the ES-PMMA and model groups, based on transcriptomic sequencing data. These gene sets were named DEG-1, DEG-2, and DEG-3, respectively. The screening threshold for DEGs was set to P < 0.05 and a |log2FC| > 0.5. A volcano plot and a heat map were then generated in the DESeq2 package in R to visualize the DEGs. We also generated a Venn diagram to identify the intersection between DEG-1 and DEG-2 to identify genes specifically associated with traditional PMMA. Subsequently, another Venn diagram was generated to identify the intersection between DEG-1 and DEG-3 to identify genes specifically associated with ES-PMMA. Finally, key candidate genes associated with ES-PMMA were obtained by excluding common genes between genes specifically associated with ES-PMMA and genes specifically associated with PMMA from the ES-PMMA gene set.

2.5. Identification of candidate genes by protein-protein interaction (PPI) network analysis and correlation analysis

To investigate the interaction between key candidate genes associated with ES-PMMA, we next constructed a PPI network using the STRING database (https://string-db.org/). STRING-derived interaction data were imported into Cytoscape (https://cytoscape.org/), in which the Degree, Closeness, and Betweenness algorithms were used to score the genes. “Degree” represents the number of proteins that gene nodes directly interact with in the PPI network. “Closeness” is used to measure the average length of the shortest path from this gene to all other genes in the network. “Betweenness” is used to measure the frequency of gene occurrence in all the shortest paths, which indicates the key regulatory factor. Three centrality algorithms quantify the influence of genes in the PPI network from different dimensions. The combined use can comprehensively identify the key regulatory factors under the background of differential expression. The top 20 genes identified by each algorithm were intersected, thus resulting in the identification of 12 candidate genes. Subsequently, Spearman’s correlation analysis was performed to investigate the relationship between candidate genes in the ES-PMMA–model dataset.

2.6. Identification of key genes by machine learning

Three machine learning algorithms were used to identify key genes from the candidate gene set in the ES-PMMA dataset: LASSO regression, support vector machine recursive feature elimination (SVM-RFE), and Boruta. The analysis object in the machine learning stage was the gene expression data of the model group (Model, n = 10) and the ES-PMMA group (Heparin, n = 10). No independent training set and test set division was conducted; instead, 10-fold cross-validation was used to evaluate the stability of the model. Ten-fold cross-validation was implemented during model training to optimize hyper parameters and prevent overfitting. Model performance was quantified using the Area Under the Receiver Operating Characteristic Curve (AUC-ROC) and precision-recall curves. SVM-RFE is a recursive feature elimination algorithm based on Support Vector Machine (SVM) that removes features with the lowest contribution iteratively, gradually selecting the most critical genes for classification (between the model group and the ES-PMMA group). Our analysis involved a linear kernel and a feature elimination step size set so that 10% of the lowest-ranked features were removed in each iteration, thereby gradually reducing the feature set to enhance the generalization ability of the model. Boruta was used to identify potential key genes most closely related to the therapeutic effects of ES-PMMA. The number of iterations was 1000 and Z-score (importance score) was calculated for each real gene and shadow feature. When the feature Z value was> than the maximum shadow feature Z value and p was < 0.01, this condition was marked as “confirmed” (a significant important feature). Next, the glmnet package in R was used to implement LASSO analysis. The optimal λ that minimized the error of the validation set was selected by 10-fold cross-validation. The regularization path comprised 100 exponentially decreasing λ values with an elastic net parameter α of 1 (pure LASSO). Features with non-zero coefficients (|β| > 0) were retained. Finally, the genes identified by the three machine learning algorithms were intersected.

2.7. ROC analysis, nomogram construction, and expression analysis of biomarkers

The pROC package in RStudio (https://www.rstudio.com/) was used to generate ROC curves to evaluate the ability of key genes to distinguish between the ES-PMMA group and the model group; then, the area under the ROC curve (AUC) was calculated with 95% confidence intervals. An AUC value > 0.7 was considered to indicate that a given key gene had good discrimination ability. Additionally, precision-recall curves (PRC) were generated using the PRROC package (https://cran.r-project.org/web/packages/PRROC/) to assess performance under class imbalance. The area under the precision-recall curve (AUPRC) was computed, with AUPRC values > 0.5 (above random chance) considered clinically meaningful. For the selected key genes, their ability to distinguish between the model group and the ES-PMMA group was evaluated. The dataset used was the same as that in the machine learning stage. The nomogram was constructed based on these key genes, and a risk prediction model was built in the same dataset (the model group and the ES-PMMA group), without introducing a new dataset division. The Hosmer-Lemeshow test was used for calibration metrics. Finally, we constructed a transcription factor (TF)-biomarker-microRNA (miRNA) regulatory network. We used this network to investigate whether a given biomarker was regulated by a TF and identify the interaction between biomarkers and miRNA, based on the GhEA3 predictive biomarkers of TFs (https://maayanlab.cloud/chea3/) database.

2.8. RNA isolation and real-time quantitative PCR (RT-qPCR)

To further verify the expression levels of the key candidate genes associated with ES-PMMA, an additional cohort of animals (n = 6 per group) was subjected to the same surgical and sampling procedures as previously described in Section 2.1. All additional procedures were approved by the Third Hospital of Hebei Medical University (Approval number: z2023-012–01) and complied with the ARRIVE guidelines. Bone tissues were collected, and total RNA was extracted using TRIzol reagent (Invitrogen, USA) in accordance with the manufacturer’s protocol. RNA concentration and purity were assessed spectrophotometrically (NanoDrop, Thermo Fisher Scientific). cDNA was synthesized from 1 µg of total RNA using a SureScript First-Strand cDNA Synthesis Kit (GeneCopoeia, USA) in accordance with the manufacturer’s instructions.

RT-qPCR was performed to determine the expression levels of SLA2, RAG1 and S100B, using a Bio-Rad iQ5 Real-Time PCR system (Bio-Rad, USA) with 2 × SYBR Green qPCR MasterMix (Applied Biosystems, USA). The reaction mixture (20 µL) contained 10 µL MasterMix, 1 µL each of forward and reverse primers (10 µM), 2 µL of cDNA template, and 6 µL of nuclease-free water. The thermal cycling conditions were as follows: 95°C for 10 min (initial denaturation), followed by 40 cycles of 95°C for 15 sec, 55°C for 30 sec, and 72°C for 30 sec. GAPDH was used as a reference gene for normalization. All reactions were performed in triplicate, and no-template controls were included. Primer sequences for target genes (SLA2, RAG1, S100B) and GAPDH were designed and synthesized by Wuhan Jinkaruo Biological Engineering Co., Ltd (Table 1). Relative mRNA expression levels were calculated using the 2 − ΔΔCT method [36].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Primer sequences for target genes in rabbit (Oryctolagus cuniculus).

https://doi.org/10.1371/journal.pone.0332041.t001

2.9. Statistical analysis

Statistical analysis was conducted using SPSS 26.0 (IBM Corp., Armonk, NY, USA). One-way ANOVA was used to conduct an overall difference test on multiple groups (including the Control, Model, PMMA, and ES-PMMA groups), and then pairwise comparisons between groups were performed using Tukey’s post-hoc test (such as a direct comparison between the ES-PMMA group and the Model group). To control for the false positives caused by multiple tests, the FDR (False Discovery Rate) correction (i.e., the Benjamini-Hochberg correction) was adopted to ensure the reliability of the statistical results in the paper. A corrected p-value < 0.05 was set as the significance threshold. For RT-qPCR analysis, all quantitative PCR data were normalized to the reference genes GAPDH using the 2^(-ΔΔCt) method. Statistical analyses were performed using one-way ANOVA followed by Tukey’s multiple comparisons test for multiple comparisons after confirming normality (Shapiro-Wilk test, all p > 0.05) and homogeneity of variance (Levene’s test, all p > 0.1). All analysis were conducted using GraphPad Prism 9.0 with statistical significance set at α = 0.05 (two-tailed).

3.1. Quality control of sequencing data

All RNA samples exhibited high integrity (RIN: 8.6 ± 0.5, range 7.8–9.5). Sequencing generated an average of 25 million paired-end 150-bp reads per sample, with post-trimming read lengths of 145 ± 3 bp and adapter contamination rates < 0.5%. Analysis revealed that all genes (100%) had quality scores > 35, with error rates < 0.1% (Fig 2A, B). Alignments between each sample and the reference genome was > 90% (except for C1 at 86.38%) (Table 2). These metrics confirmed the suitability of the data for downstream analysis. PCA was then used to model and visualize the dispersion of samples among the normal, model, PMMA, and ES-PMMA groups based on transcriptomic sequencing data. As shown in Fig 2C, no strong outliers were identified. PC1 and PC2 explained 45.22% and 8.85% of the total variance, respectively, with a cumulative contribution of 54.07%. These results indicated that the sequencing data of all samples were of good quality.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Control group (n = 5): Only the skin and subcutaneous tissue of the distal femur were incised and sutured.

https://doi.org/10.1371/journal.pone.0332041.t002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Quality control. (A) Distribution of quality scores for base positions. The x-axis represents the relative position of each base in the reads, whereas the y-axis represents the sequencing quality scores (higher scores indicate better base quality, with a 1% error rate at a quality score of 20 and a 0.1% error rate at a quality score of 30); (B) Nucleotide composition plot showing the position of reads on the x-axis and the proportion of single bases on the y-axis. Different colors indicate various types of bases; (C) Principal Component Analysis (PCA) distribution map; yellow, blue, green, and purple indicate the Polymethylmethacrylate (PMMA), control, Enoxaparin Sodium-Polymethylmethacrylate (ES-PMMA), and model groups, respectively.

https://doi.org/10.1371/journal.pone.0332041.g002

3.2. GSVA revealed the top 10 pathways that were activated or inhibited

GSVA was performed to evaluate pathway activity between groups. The reference gene set was the “h.all.v7.4.symbols.gmt” gene set obtained from the MSigDB database. Permutation number was 100 times and FDR-adjusted p-values < 0.05 were considered significant pathways. The FPKM data of gene expression was first converted from the original counts to FPKM values through the count ToFpkm function. Subsequently, a logarithmic transformation (log2 (FPKM + 1)) was applied for normalization prior to GSVA analysis. Pathways with |NES| > 1 were presented. When comparing the model group and the control group, 20 pathways were activated, and 11 pathways were inhibited. When comparing the PMMA and model groups, 17 pathways were activated, and 24 pathways were inhibited. When comparing the PMMA and control groups, 23 pathways were activated, and 28 pathways were inhibited. Therefore, we selected the top 10 pathways for further analysis. When comparing the ES-PMMA and model groups, analysis revealed that a range of inflammation-related pathways, among which TNF-α signaling via NF-KB (adj.p = 0.017, NES = −0.39) and inflammatory response (adj.p = 0.014, NES = −0.42), were the most significant pathways, thus indicating that ES-PMMA successfully alleviated inflammation. In contrast, the E2F target pathway (adj.p = 0.07, NES = 0.55) regulates cell cycle progression, the G2M checkpoint pathway (adj.p = 0.11, NES = 0.47) controls mitosis, and the HEME metabolism pathway (adj.p = 0.16, NES = 0.37) governs heme biosynthesis. Metabolism-related pathways included spermatogenesis (adj.p = 0.18, NES = 0.19) and estrogen response late (adj.p = 0.18, NES = 0.14). All of these pathways are critical for cell proliferation and differentiation. Although the adjusted p-value exceeded 0.05, these findings still suggest/indicate the potential of ES-PMMA to promote cell proliferation and differentiation (Fig 3A, B). When comparing the PMMA and model groups, the inflammation-related pathways, including the IFN-α response (adj.p = 0.64) and IFN-γ response (adj.p = 0.64), showed no statistically significant differences. This suggests that, at the selected time point (10 days post-implantation), PMMA did not induce a significant inflammatory perturbation in the bone defect tissue compared to the model group (Fig 3C, D). When comparing the model and control group, epithelial-mesenchymal transition (EMT) (adj.p = 7.52e-05, NES = 0.39), angiogenesis (adj.p = 0.0002, NES = 0.20), Notch signaling (adj.p = 0.0004,NES = 0.11), Kras signaling-up (adj.p = 0.0005, NES = 0.21), coagulation myogenesis (adj.p = 0.0005, NES = 0.19), TGF-β signaling (adj.p = 0.0008, NES = 0.8), TNF-α signaling via NF-κβ (adj.p = 0.0018, NES = 0.32) and the WNT-β-catenin (adj.p = 0.0018, NES = −0.04) signaling pathways were activated. However, the Mtorc1 signaling (adj.p = 0.0306, NES = −0.08), Myc target v2 (adj.p = 0.0221, NES = −0.24), mitotic spindle (adj.p = 0.0090, NES = −0.34), reactive oxygen species (ROS) (adj.p = 0.0062, NES = −0.10), DNA repair (adj.p = 0.0044, NES = −0.20), Myc target v1 (adj.p = 0.0018, NES = −0.26), HEME metabolism (adj.p = 0.0008, NES = −0.12), spermatogenesis (adj.p = 0.0007, NES = −0.12), E2F target (adj.p = 0.0005, NES = −0.25), and G2M checkpoint (adj.p = 0.0005, NES = −0.27) pathways were all inhibited, thus indicating that cell proliferation and differentiation had been inhibited (Fig 3E, F). Collectively, these results indicated that cell proliferation and differentiation were restricted during inflammation, at least to some extent.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. GSVA (Gene Set Variation Analysis) analysis of pathway activity. (A-B) Model vs. Control: Top 10 pathways show upregulated inflammation and suppressed proliferation. (C-D) PMMA vs. Model: Upregulated inflammation and mixed proliferation effects. (E-F) ES-PMMA vs. Model: Suppressed inflammation and enhanced proliferation. Barplots show enrichment t-scores; heatmaps display pathway enrichment. Blue representing inhibition and orange representing activation.

https://doi.org/10.1371/journal.pone.0332041.g003

3.3. Inter-group differences and the identification of key candidate genes associated with ES-PMMA

3.3.1. Screening for DEGS.

A total of 6637 DEGs, including 3871 up-regulated and 2766 down-regulated genes, were identified between the model and control groups. A total of 1071 DEGs, including 713 up-regulated genes and 358 down-regulated genes, were identified between the PMMA and model groups. A total of 2059 DEGs, including 1239 up-regulated and 820 down-regulated genes, were identified between the model and ES-PMMA groups. The DEGs identified between the model and control groups, the PMMA and model groups, and the ES-PMMA and model groups, were named DEG-1, DEG-2, and DEG-3, respectively. Volcano plots demonstrating these DEGs are shown in Fig 4A–C with a screening threshold of adjusted p-value (FDR) < 0.05 and |log2FC| > 0.5.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Differential gene expression between groups. (A) A total of 6637 DEGs, including 3871 up-regulated and 2766 down-regulated genes, were identified between the model and control groups. (B) A total of 1071 DEGs, including 713 up-regulated genes and 358 down-regulated genes, were identified between the PMMA and model groups. (C) A total of 2059 DEGs, including 1239 up-regulated and 820 down-regulated genes, were identified between the ES-PMMA and model groups. Green representing inhibition and red representing activation.

https://doi.org/10.1371/journal.pone.0332041.g004

3.3.2. Screening of key candidate genes in ES-PMMA bone cement.

A Venn diagram was used to identify genes with opposing expression trends in DEGs-1 and DEGs-2. We took the intersection of the up-regulated genes in DEGs-1 and the down-regulated genes in DEGs-2, and the intersection of the down-regulated genes in DEGs-1 and the up-regulated genes in DEGs-2, and then identified the union of the two intersections. We obtained 81 genes which were defined as bone cement-related DEGs (Fig 5A, B). Similarly, a Venn diagram was used to identify genes with opposing expression trends in DEGs-1 and DEGs-3 by taking the intersection of the up-regulated genes in DEGs-1 and the down-regulated genes in DEGs-3, and the intersection of the down-regulated genes in DEGs-1 and the up-regulated genes in DEGs-3, and then identifying the union of the two intersections. We identified 332 genes which were defined as heparin bone cement-related DEGs (Fig 5C, D). A 28-gene overlap subset was identified between heparin-cement co-regulated DEGs and cement-associated DEGs. Subtraction of this overlapping subset from the 332 heparin-cement co-regulated DEGs yielded 304 heparin-cement signature genes (Fig 5E).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Screening of key candidate genes associated with ES-PMMA. (A, B) Identification of bone cement-associated differentially expressed genes (DEGs). Venn diagram was used to obtain the genes with opposite expression trends (81 genes) in DEGs-1 and DEGs-2, which was defined as bone cement-associated DEGs. (C, D) Identification of heparin-cement co-regulated DEGs. Venn diagram was used again to obtain the genes with opposite expression trends in DEGs-1 and DEGs-3(332 genes in total), which was defined as heparin-cement co-regulated DEGs. (E) Selection of heparin-cement signature genes. A 28-gene overlap subset was identified between heparin-cement co-regulated DEGs and cement-associated DEGs. Subtraction of this overlapping subset from the 332 heparin-cement co-regulated DEGs yielded 304 heparin-cement signature genes.

https://doi.org/10.1371/journal.pone.0332041.g005

3.4. Enrichment analysis

3.4.1. Gene Ontology (GO) pathway enrichment analysis.

GO analysis revealed that the main enriched terms (FDR < 0.05) included response to lipopolysaccharides, response to molecules of bacterial origin, and cell-substrate adhesion (cell-matrix adhesion). The key candidate genes associated with ES-PMMA were enriched in 468 terms; of these, 350 terms were related to biological processes, 55 terms were related to cellular components, and 63 terms were related to molecular functions. The main enriched terms included tertiary granule (12/193), secretory granule membrane (14/193) and specific granule (10/193) (Fig 6A, B).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses of key candidate genes associated with ES-PMMA. (A, B). Results of GO enrichment analysis; (C, D). Results of KEGG enrichment analysis. Each module represents each pathway, the size of the module represents the level of gene expression in the pathway, and the color represents the P-value.

https://doi.org/10.1371/journal.pone.0332041.g006

3.5. Candidate genes identified by PPI network analysis

The PPI network was visualized by Cytoscape software and featured 103 nodes and 108 edges, representing the intricate web of interactions between these proteins within a biological system (Fig 7A). Data acquired from the STRING database were imported into Cytoscape, and genes were scored by the Degree, Closeness and Betweenness algorithms in the CytoHubba plug-in. The top 20 genes identified by each algorithm were intersected, resulting in the identification of 12 candidate genes (Fig 7B). Spearman’s correlation analysis was then performed to investigate the relationship between candidate genes in the transcriptomic sequencing dataset of the ES-PMMA and the model groups (|cor| > 0.3 and P < 0.05). Analysis revealed that Talin 1 (TLN1) and NCK-associated protein 1-like (NCKAP1L) exhibited the most significant positive correlation, whereas LIX1 and WDFY family member 4 (WDFY4) exhibited the most significant negative correlation (Fig 7C). TLN1, which mediates cytoskeletal linkage, and NCKAP1L, which regulates immune cell motility, may synergize to enhance immune cell adhesion and migration, potentially supporting the immunomodulatory effects of ES-PMMA during bone repair.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Screening of candidate genes via protein–protein interaction (PPI) network analysis and subsequent correlation analysis. A. A PPI network was constructed to investigate the interaction between the key candidate genes associated with Enoxaparin Sodium Polymethyl methacrylate (ES-PMMA); B. The top 20 genes identified using the Degree, Closeness, and Betweenness algorithms each were intersected, and 12 candidate genes were selected; C. Correlation analysis of the 12 candidate genes. Blue represents a positive correlation, and red represents a negative correlation. The larger the correlation, the larger the circle and the darker the color. ‘*’ represents statistical significance.

https://doi.org/10.1371/journal.pone.0332041.g007

3.6. Machine learning

Next, we selected key genes from the 12 candidate genes associated with the ES-PMMA and model groups identified in Section 3.5. We employed a triple machine learning algorithm to screen key genes, as follows. SVM-RFE was performed using the R package “e1071” and parameter performance was evaluated by 10-fold cross-validation. The recursive elimination process was initiated with all candidate genes and involved three core steps: (1) training SVM models with optimal parameters, (2) computing feature weights, (3) eliminating the lowest ranked 10% of features. Termination criteria required a > 0.005 increase in average cross-validation error for three consecutive iterations. To enhance stability, external 3-fold cross-validation was repeated 10 times. Genes with a selection frequency > 0.8 (selected occurrences/total runs) were retained, yielding six feature genes (Fig 8A).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 8. Candidate genes were analyzed by three machine learning algorithms: LASSO regression, support vector machine recursive feature elimination (SVM-RFE), and Boruta. (A) SVM-RFE analysis identified six signature genes; (B) Boruta analysis identified five signature genes; (C, D) LASSO analysis identified five signature genes (lambda. min = 0.0162); E. Three key genes (RAG1, SLA2, and S100B) were identified after the three gene sets were intersected.

https://doi.org/10.1371/journal.pone.0332041.g008

Boruta analysis was employed to perform random forest-based permutation tests and retain genes with importance scores significantly higher than synthetic noise features. We calculated the Z-score (importance score) for each real gene and shadow feature. Iterations continued until all features attained definitive classifications. When the feature Z value was greater than the maximum shadow feature Z value and p < 0.01, the condition was marked as “Confirmed” (a significant important feature) resulting in five confirmed feature genes (Fig 8B).

Next, LASSO logistic regression was implemented via the “glmnet” package. The regularization path comprised 100 exponentially decreasing λ values with an elastic net parameter α = 1 (pure LASSO). The optimal λ was selected by 10-fold cross-validation (λ = 0.0162). Features with non-zero coefficients (|β| > 0) were retained, identifying five genes (Fig 8C, D). Integration of the three machine learning outcomes followed strict criteria: only genes concurrently selected by SVM-RFE, Boruta, and LASSO were retained. Three key genes (RAG1, SLA2, S100B) were ultimately identified (Fig 8E).

3.7. ROC analysis and evaluation

Our comprehensive analysis demonstrated outstanding predictive performance across all evaluation metrics. The machine learning models achieved exceptional classification: LASSO (accuracy = 1.00, recall = 1.00, AUC-ROC = 1.00), SVM (accuracy = 0.909, recall = 1.00, AUC-ROC = 1.00), and Boruta (accuracy = 1.00, recall = 1.00, AUC-ROC = 1.00) all showed perfect or near-perfect discrimination (Figure S1 A-F in S1 File). Next, the pROC package in RStudio was used to generate ROC curves for the three key genes using the 70:30 training: test split established in the machine learning phase. The AUC confidence intervals were as follows: RAG1 (0.6784–1), SLA2 (0.562–1) and S100B (0.5677–1), all significantly >0.7 (DeLong’s test, all p < 0.001), indicating their robust discriminatory ability between the ES-PMMA and model groups (Fig 9A). These genes were consequently selected as potential biomarkers. ROC analysis of the nomogram demonstrated perfect discrimination (AUC = 1.00, 95% CI: 1.00–1.00) in predicting risk stratification between groups (Fig 9B). In terms of calibration metrics, the Hosmer-Leme test (10 bins) (χ² = 40,043,602.04, df = 4, p < 2e-16, p ≤ 0.05) indicated significant calibration bias) while calibration curve analysis (slope = 0.98, 95% CI: 0.95–1.01) confirmed excellent agreement between predicted and observed outcomes (Fig 9C, D), thus validating the predictive accuracy of the nomogram.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 9. Assessment of key genes. A. The receiver operating characteristic analysis (pROC). package was used to generate ROC curves for the three key genes. Area Under the Curve (AUC) values of > 0.7 indicated that the genes had good discrimination performance; B. The rms package was used to construct a nomogram model for risk prediction in the Enoxaparin Sodium-Loaded Polymethylmethacrylate (ES-PMMA) versus the model group based on the three biomarkers; C. ROC analysis (AUC > 0.7) was performed to evaluate the predictive performance of the nomogram, with a result of 1, indicating good performance; In this curve, the abscissa represented the predicted probability, whereas the ordinate represented the actual probability. D. Calibration curve of the nomogram. The slope of the curve was close to 1, validating the predictive accuracy of the nomogram.

https://doi.org/10.1371/journal.pone.0332041.g009

3.8. Differential expression of candidate biomarkers

Quantitative analysis revealed statistically significant alterations in RAG1, S100B, and SLA2 expression across experimental groups. One-way ANOVA was used to conduct an overall difference test on multiple groups, and then pairwise comparisons between groups were performed using Tukey’s post-hoc test. (FDR-adjusted p < 0.05 for all comparisons). Compared to the control group, the expression of RAG1 and S100B in the model group was higher but with no significant difference, while SLA2 was significantly reduced (p = 0.007), thus suggesting that the inflammatory response period following bone injury persisted and had gradually stabilized (Fig 10A). Compared to the model group, the expression of RAG1 in PMMA group decreased mildly (p = 0.0232). Furthermore, the expression level of RAG1 was very low in both groups (FPKM<1), and there was no statistically significant difference in the expression levels of S100B and SLA2. Therefore, it can be concluded that PMMA had no significant effect on the inflammatory response after bone injury (Fig 10B). When compared to the model group, the expression levels of RAG1 and S100B in the ES-PMMA group were significantly lower (p = 0.0067 and p = 0.0288), while the expression levels of SLA2 were significantly higher (p = 0.0288). These directional reversals suggest that ES-mediated anti-inflammatory effects occurred through the suppression of RAG1/ S100B and the activation of SLA2 (Fig 10C).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 10. Gene expression analysis of Src-like adaptor 2 (SLA2), Recombination activating gene 1 (RAG1) and S100 calcium-binding protein B(S100B). A. Control Group vs. Model Group: RAG1 and S100B showed high expression, while SLA2 showed low expression. B. PMMA Group vs. Model Group: RAG1 exhibited low expression, whereas S100B and SLA2 showed no significant differences in expression levels. C. ES-PMMA Group vs. Model Group: RAG1 and S100B demonstrated low expression, while SLA2 showed high expression. (Mann-Whitney U test, * P < 0.5; **** P < 0.001).

https://doi.org/10.1371/journal.pone.0332041.g010

3.9. TF prediction by the three key genes

Computational interrogation of upstream regulatory mechanisms identified both transcription factors (TFs) and miRNAs governing key biomarkers. Using stringent ChEA3 analysis (integrating ENCODE ChIP-seq, ReMAP, and de novo motif screening; NES > 3.0 threshold), we predicted high-confidence TFs for: RAG1 (2 TFs), S100B (1 TF), and SLA2 (7 TFs). Parallel miRDB analysis (Target score ≥ 85, 80% validation accuracy) revealed miRNA regulators: 35 targeting RAG1, 5 for S100B, and 10 for SLA2. Integration of these predictions generated a TF-biomarker-miRNA network (Cytoscape 3.9.1) comprising 63 nodes and 60 confidence-weighted edges, highlighting SP1 (RAG1-associated TF) and miR-26b-5p (top-scoring RAG1 miRNA; score = 97) as central regulatory hubs (Fig 11).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 11. Computational interrogation identified upstream TFs (predicted via ChEA3 using ENCODE ChIP-seq, ReMAP, and de novo motif screening; NES > 3.0) and miRNAs (miRDB; Target score ≥ 85, 80% validation accuracy). The integrated network (Cytoscape 3.9.1) comprised 63 nodes (TFs, biomarkers, miRNAs) and 60 confidence-weighted edges, with SP1 (RAG1-associated TF) and miR-26b-5p (top RAG1 miRNA, score = 97) highlighted as central hubs.

https://doi.org/10.1371/journal.pone.0332041.g011

3.10. RT-qPCR Validation for Key Genes (SLA2, RAG1, S100B)

For S100B expression, the ES-PMMA Groups showed significantly reduced expression versus both PMMA (mean difference 0.27, [0.1412, 0.3988], p < 0.0001) and Model groups (mean difference 0.2000, [0.07123, 0.3288], p = 0.0016), thus confirming their anti-inflammatory potency (Figure S2A in S1 File). For SLA2 expression, both the model group and PMMA group showed marked suppression when compared to the control group (mean difference 0.3700, [0.1474,0.5926], p = 0.0008), (mean difference 0.4300, [0.2074,0.6526], p = 0.0001), while ES-PMMA partially reversed this effect (PMMA vs ES-PMMA: −0.2900, [−0.5126, −0.06740], p = 0.008) (Figure S2B in S1 File). RAG1 expression patterns further supported the immunomodulatory properties of ES-PMMA, with significant reductions when compared to the PMMA and Model groups (mean difference 0.1500, [0.008430,0.2916], p = 0.0354), (mean difference 0.1800, [0.3843,0.3216], p = 0.0098) (Figure S2C in S1 File). Consistent statistical significance across multiple comparison methods confirmed the anti-inflammatory benefits of ES conjugation to PMMA.