2.1. Data Source

In this study, we accessed relevant microarray datasets from the database by searching for the term “periodontitis.” The dataset GSE10334 contains samples from 19 periodontitis patients and six normal [22].

2.2. Identification of DEGs

The GSE10334 dataset was first normalized. After preprocessing, differential expression analysis was performed using the “limma” package to identify DEGs between the periodontitis and control groups. The criteria for statistical significance were set at p  < 0.05 and |log₂ fold change (FC)| ≥ 1. To visualize the DEG results, heatmaps and volcano plots were created using the “pheatmap” and “ggplot2” packages in R.

2.3. WGCNA Analysis

WGCNA is used to explore co-expression patterns and biological networks in gene expression data. First, clustering analysis was performed to remove outliers. A scale-free co-expressed gene network was constructed using the one-step function and the soft threshold (β) was determined with “pickSoftThreshold” to optimize the adjacency matrix. The adjacency matrix was then transformed into a topological overlap matrix (TOM) and its dissimilarity (1-TOM) calculated. Modules were detected using hierarchical clustering and dynamic tree cutting. Finally, gene significance and module membership were computed to identify clinically relevant modules and the feature gene network was visualized.

2.4. Enrichment Analysis

After intersecting the genes from the WGCNA blue module with the DEGs, potential hub genes related to the mechanism of periodontitis were identified. These potential hub genes were subjected to Gene Ontology (GO) functional analysis and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis using the “clusterProfiler” package [23]. The significant threshold for enrichment was set at p  < 005.

2.5. Immune Cell Infiltration Analysis

To assess immune cell infiltration in gene expression profiles associated with periodontitis, we utilized the CIBERSORT algorithm. This approach calculates the relative proportions of 22 immune cell subtypes (LM22) in each sample, thereby providing a comprehensive evaluation of immune cell infiltration levels.

2.6. MR

To ensure the robustness and reliability of the findings, we followed the guidelines of the STROBE-MR checklist in conducting our MR analysis [24]. The MR analysis was performed using the TwoSampleMR package. Genome-wide significant single nucleotide polymorphisms (SNPs) with p < 5 × 10⁻⁸ were selected as IVs, excluding those associated with potential confounders. Linkage disequilibrium (LD) pruning (R² < 0.01, 10,000 kb window) ensured IV independence. Exposure data were sourced from expression quantitative trait loci (eQTLs) in the IEU Open genome-wide association study (GWAS) database (https://gwas.mrcieu.ac.uk/datasets/). Additionally, we accessed the periodontal disease cohort from this database, identified by the ID finn-b-K11_PERIODON_CHRON, which included 195,395 standard samples and 3046 samples diagnosed with periodontal disease. These samples are of European descent and provided whole genome data that served as our outcome data. The inverse variance weighted (IVW) method with a significance threshold of p  < 0.05 was used as the primary analysis. Subsequently, we computed the F-statistic to assess the strength of the association between IVs and exposure factors. SNPs with an F-value less than 10 were excluded to mitigate potential weak instrument bias. Cochran’s Q test assessed heterogeneity (p  < 0.05 indicating potential invalid instruments), while leave-one-out analysis validated robustness. MR-Egger regression tested for directional pleiotropy (intercept p  < 0.05 suggesting pleiotropic bias). Additionally, we employed the MR-PRESSO package to detect and correct for horizontal pleiotropy, thereby ensuring the robustness of the results [25]. This integrated approach combined multiomics evidence to validate causal relationships between biomarkers and periodontitis. To visualize the design of this study more, we drew the flow chart (Figure 1).

2.7. Dimensionality Reduction Analysis of Single-Cell Data and Identification of Cell Types

Data from the GSE152042 dataset underwent quality control, normalization, and dimensionality reduction using Seurat and MonacoImmuneData frameworks. The PercentageFeatureSet function calculated the percentage of molecules and rRNA, ensuring cells expressed >500 genes, <4000 genes, with <30% molecule content and at least 100 unique molecular identifiers (UMIs) per cell. Data normalization and identification of highly variable genes were performed using FindVariableFeatures. Cells were clustered (resolution = 0.1, dim = 30) using FindNeighbors and FindClusters, and visualized with t-SNE. Intercellular communication was inferred, analyzed, and visualized using the “CellChat” package in R, which constructs a neighborhood graph and estimates communication probabilities based on gene expression, signal ligands, receptors, and related factors. Cell type annotation was performed using the SingleR algorithm. The SingleR algorithm classifies each cell population based on known marker genes. To validate the annotations, manual annotation was also performed, relying on well-established cell-type marker genes (Supporting Information 1: Table S1). To further validate the accuracy of the annotations, a comparison was made with the PanglaoDB database, a comprehensive resource for human and mouse scRNA-seq data (Supporting Information 2: Table S2). During the comparison, marker genes and cell-type classifications were cross-checked to ensure the accuracy of our annotations.

2.8. Candidate Drug Prediction

Evaluating the interactions between proteins and drugs plays a crucial role in determining whether target genes can serve as potential drug targets. In this study, the drug signatures database will be utilized for drug prediction [26]. This database contains 22,527 gene sets and 17,389 distinct compounds spanning 19,531 genes, aiming to link drugs and chemicals to their corresponding target genes. By uploading the identified target genes to DSigDB, potential drug candidates can be predicted to assess the drug activity of these target genes.

2.9. MD

This study utilizes MD to evaluate the binding energy and interaction patterns between drug candidates and target proteins at the atomic level, exploring their potential as therapeutic targets. The two-dimensional structures of small-molecule ligand drugs were obtained from the PubChem database (PDB, https://www.rcsb.org/), while the three-dimensional structures of target proteins (receptors) were retrieved from the RCSB Protein Data Bank (PDB, https://www.rcsb.org/). AutodockVina 1.2.2 was used to perform docking on selected drugs and their corresponding target proteins.

2.10. Molecular Dynamics Simulation

Molecular dynamics simulations were performed using Gromacs 2022.3. For preprocessing of small molecules, AmberTools 22 was utilized to assign the GAFF force field, while Gaussian 16W was employed to hydrogenate the molecules and compute the RESP charge. The resulting potential data were incorporated into the system’s topology file. The simulation was conducted at a fixed temperature of 300 K and atmospheric pressure (1 bar). Amber99sb-ildn was selected as the force field, with Tip3p water model used as the solvent. The system’s total charge was neutralized by adding appropriate Na⁺ ions. Energy minimization was performed using the steepest descent method, followed by an NVT (isothermal and isovolumic) and NPT (isothermal and isobaric) equilibration for 100,000 steps each, with a coupling constant of 0.1 ps and a duration of 100 ps. The final step involved a production simulation with a total of 50,000,000 steps, each with a time step of 2 fs, for a total simulation time of 100 ns. Upon completion of the simulation, trajectory analysis was carried out using the built-in tools in the software. Metrics such as the root-mean-square deviation (RMSD), root-mean-square fluctuation (RMSF), solvent-accessible surface area (SASA), and the radius of gyration of the protein were calculated, alongside free energy calculations (MMGBSA), free energy landscapes, and other relevant data [27].

3.1. Identification of DEGs

We retrieved the periodontitis dataset (GSE10334) from the GEO database and performed differential expression analysis to identify DEGs in the periodontitis group. Compared to the normal group, we found 167 DEGs in the periodontitis group, with 104 upregulated and 63 downregulated (Figure 3A,B).

3.2. Construction of WGCNA Network

We then conducted WGCNA on the core genes from the GSE10334 dataset to explore gene expression patterns related to periodontitis (Figure 3C). Fourteen distinct modules were identified (Figure 3D). Correlation analysis revealed that the blue module exhibited the strongest association with periodontitis (Figure 3E).

3.3. Enrichment Analysis

We identified disease-related core genes and overlapping genes between DEGs using the WGCNA method. Ultimately, we selected 118 overlapping genes as potential core genes involved in the development and progression of periodontitis (Figure 3F). Next, we carried out GO and KEGG enrichment analyses to examine the biological roles of these genes. The results of the GO analysis revealed that the overlapping genes are closely related to several immune-related biological processes, including immune response-activating cell surface receptor signaling, leukocyte-mediated immunity, B cell receptor signaling, immune response regulation, and B cell activation (Figure 3G). The KEGG analysis indicated that these candidate genes are primarily involved in immune system functions, signaling molecule interactions, parasitic infectious diseases, and transport and metabolic pathways (Figure 3H).

3.4. Immune Cell Infiltration

The CIBERSORT algorithm was employed to analyze immune cell infiltration in both the periodontitis and control groups (Figure 4A,B). The results showed significant differences between the two groups. In both groups, the infiltration of immune cells such as plasma cells, memory B cells, CD4⁺ memory T cells, activated T cells, follicular helper T cells, regulatory T cells (Tregs), resting dendritic cells (DCs), and activated DCs showed marked differences (Figure 4C). These results suggest a close relationship between immune cell infiltration and gene expression differences.

3.5. There Is a Causal Relationship Between HLA-C and the Risk of Periodontitis

The eQTL data were sourced from the IEU OpenGWAS project and periodontitis GWAS data were obtained from FinnGen R11. To investigate the genetic relationship between 118 potential genes and periodontal disease, we conducted a combined analysis of GWAS and MR, identifying seven core genes. Among them, HLA-C (IVW, nine SNPs, p = 0.03) was identified as a risk factor for periodontitis, while STAP1 (IVW, four SNPs, p = 0.01) and PTP4A3 (IVW, 10 SNPs, p = 0.04) were recognized as protective factors (Figure 4D).

We then further investigated the causal relationship between HLA-C and periodontitis. Using the IVW method, a significant association was observed, with an odds ratio (OR) of 1.0729 (95% CI: 1.0056–1.1448, p = 0.033). The MR-Egger test revealed no evidence of heterogeneity (p = 0.704; Figure 4D). Moreover, reverse MR analysis showed no evidence of reverse causality, suggesting that HLA-C is not inversely related to periodontitis (Figure 4E). The MR-Egger regression intercept also indicated no pleiotropy (p = 0.704), further supporting the robustness of our findings. Sensitivity analysis using the leave-one-out approach demonstrated that no individual polymorphism significantly influenced the results, confirming the reliability of the causal effect estimates (Figure 4F).

Additionally, correlation analysis with 22 types of immune cells revealed that HLA-C was positively correlated with plasma cells (r = 0.59), while it was negatively correlated with resting DCs (r = −0.50) and CD8⁺ T cells (r = −0.53; Figure 4G,H). These results suggest a close relationship between immune cell infiltration and gene expression differences.

3.6. Analysis of Single-Cell RNA Sequencing Data of Marker Gene

The annotations from SingleR were largely consistent with the manual annotations, particularly for B cell populations, where the clustering results were almost identical (Figure 5A,B). Minor discrepancies were observed in other cell types, such as fibroblasts and pericytes. The comparison with the PanglaoDB database showed strong consistency between our annotations and those in the database, especially in terms of key marker genes and cell-type classifications (Figure 5C). These results further reinforce the confidence in the accuracy of our cell-type annotations, particularly for B cell subsets and other major cell types (Figure 5D–F). In the GSE152042 dataset, uniform manifold approximation and projection (UMAP) clustering identified 14 distinct cell clusters (Figure 6A,B). Using SingleR, eight cell types were annotated across all samples, including CMP, monocytes, B cells, endothelial cells, tissue stem cells, epithelial cells, DCs, and T cells (Figure 6D,E). Dimensionality reduction and proportional representation analysis revealed a significant increase in the proportion of B cells in periodontal disease samples, indicating heightened B cell activity during the disease progression (Figure 6F). Further analysis of immune cell marker gene expression identified five B cell subtypes: switched memory B cells, plasmablasts, maive B cells, exhausted B cells, and low-density basophils. Corresponding dimensionality reduction and proportional representation plots were also generated. Notably, the number of plasmablasts was significantly higher in periodontal disease patients compared to healthy controls, suggesting a strong correlation between plasmablast levels and the severity of periodontal disease (Figure 6G–I). Additionally, the results showed elevated HLA-C expression in periodontal disease samples, particularly in plasmablasts (Figure 6J).

3.7. Comparative Analysis of Cell–Cell Interactions

We compared the interaction networks of different cell subsets between normal control and periodontal disease samples. Cell communication analysis revealed that, compared to the normal controls, the periodontal disease group exhibited a higher number of ligand–receptor interactions and greater communication intensity among the eight cell types analyzed (Figure 7A–C). Notably, in the normal control group, plasmablasts showed minimal or no interactions with other cell types (Figure 7B). In contrast, in periodontal disease samples, activated B cells facilitated significant interactions between plasmablasts and various other cell types (Figure 7D). This suggests that plasmablasts may play a pivotal role in the pathogenesis of periodontal disease through crosstalk with a range of immune cells.

Subsequently, we investigated the cell–cell communication and metabolic pathways of plasmablasts between HLA-C positive and negative groups. The results showed that plasmablasts in the HLA-C positive group exhibited significantly stronger communication intensity compared to the negative group, indicating extensive interactions with various cell types (Figure 7E). Compared to the HLA-C negative Tissue_stem_cell group, the HLA-C positive group also demonstrated enhanced MDK-NCL and CXCL12-CXCR4 signaling pathways in cell communication (Figure 7F). At the metabolic level, differential expression was concentrated in pathways such as drug metabolism—other enzymes, drug metabolism—cytochrome P450, glycosaminoglycan degradation, lysosphingolipid biosynthesis—ganglio series, terpenoid backbone biosynthesis, and porphyrin and chlorophyl metabolism (Figure 7G). Notably, the HLA-C positive group showed higher expression levels and stronger correlations in the drug metabolism—other enzymes and Glycosphingolipid biosynthesis—ganglio series pathways (Figure 7G).

3.8. Prediction of Potential Drug Candidates and MD Analysis

The DSigDB database was utilized in this study to predict potential therapeutic drugs. Based on adjusted p-values, the top 15 potential compounds were identified (Figure 8A). Among these, metronidazole (Rhein TTD 00010611/CTD 00001002) was determined to be the most significant drug associated with HLA-C (Figure 8B). To evaluate the binding affinity of the candidate drugs to their targets and assess the druggability of these targets, MD analysis was conducted.

Four candidate drugs with potential binding sites on HLA-C were identified, yielding four effective docking results (Figure 8C–F). Each candidate drug exhibited clear hydrogen bonds and strong electrostatic interactions with its protein target. Furthermore, the binding pockets of HLA-C were successfully occupied by all four candidate drugs. Notably, HLA-C and metronidazole displayed a binding energy of −5.00 kcal/mol, indicating a relatively stable interaction (Figure 8C).

3.9. Molecular Dynamics Simulation

To further confirm the binding of Metronidazole with HLA-C, molecular dynamics simulations were conducted and the RMSD, RMSF, H-bonds, SASA, and Rg of the HLA-C–metronidazole complex were calculated. The changes in the HLA-C and metronidazole binding complex before and after molecular dynamics simulation are shown in Figure 9A. The RMSD curve of metronidazole (red line) remains stable at approximately 0.13 nm (Figure 9A). Additionally, the RMSD curve of HLA-C (black line) is stable between 0.24 and 0.27 nm from 20 to 100 ns. The RMSD curve of the HLA-C–metronidazole complex (blue line) remains around 0.28–0.32 nm from 20 to 100 ns, with slight fluctuations, indicating that the binding of metronidazole to HLA-C is relatively stable.

Residues 10–40, 140–160, 180–190, and 210–230 on Chain A of HLA-C show greater residual flexibility, while residues 10–40 on Chain B also exhibit similar characteristics (Figure 9B). The Rg value of HLA-C remains stable between 30 and 50 ns and 60 and 90 ns, with an average Rg value ranging from 2.375 to 2.400 nm (Figure 9C,D). The SASA value of HLA-C is larger before the binding with metronidazole and decreases after binding, indicating that metronidazole binding leads to a reduction in the surface area of HLA-C (Figure 9E). Simulation results show that the HLA-C–metronidazole complex has stable hydrogen bonds, with up to six hydrogen bonds formed (Figure 9F). The formation of hydrogen bonds between the ligand and receptor contributes to the stability of the complex. Further analysis of the secondary structure of HLA-C–metronidazole showed minimal changes in secondary structure over time, indicating excellent protein stability. In the HLA-C–metronidazole complex, the EMM values of Tyr9 and Lys66 are the highest, contributing most significantly to the residues (Figure 9G). The 3D and 2D Gibbs free energy profiles of the HLA-C–metronidazole complex show that when the Rg value is between 2.32 and 2.40 nm, and the RMSD value is between 0.13 and 0.35 nm, the Gibbs free energy of the complex is lower (Figure 9H–K). These results suggest that HLA-C is an important target for metronidazole.

To further investigate the binding stability of the HLA-C–metronidazole complex, the binding free energy of the HLA-C–metronidazole complex was calculated using the MM/PBSA method from the final 10 ns stable RMSD trajectory. As shown in Table 1, the total binding free energy of the HLA-C–metronidazole complex is −19.48 kJ/mol, with van der Waals energy of −18.14 kJ/mol, electrostatic energy of −19.20 kJ/mol, and gas phase energy of −37.34 kJ/mol, which favors the stability of the HLA-C–metronidazole complex system (Table 1).