The transcriptome (GSE10575; GSE51588; GSE32317; GSE55457) of OA and normal cartilage tissues used for raw analysis were downloaded from the Gene Expression Omnibus (GEO, https://www.ncbi.nlm.nih.gov/geo/) platform. To reduce the error bias of individual samples, we used the “limma” and “SVA” R packages to combine and correct the GSE10575 and GSE51588 datasets as the training cohort for screening OA signature genes. We collected 62 samples, including 16 normal cartilage and 46 OA samples. Two independent datasets, GSE32317 and GSE55457, were used as validation cohorts. GSE32317 comprises early- and end-stage OA, and the GSE55457 dataset contained 10 normal and 23 OA samples. Additionally, differentially expressed genes (DEGs) were screened under the criterion of |log2 FC| ≥ 1 and FDR < 0.05 using the “limma” R package.

We used KEGG analysis to detect the signaling pathways associated with DEGs in OA. Similarly, we conducted GO functional enrichment analysis to explore the biological functions of identified DEGs, including molecular functions (MFs), cellular components (CCs), and biological processes (BPs). To investigate the signaling pathways and biological functions activated in OA, we conducted Gene Set Enrichment Analysis (GSEA) (13).

Three machine learning algorithms were used to identify significant markers to predict OA status. The LASSO regression analysis allows variable selection and complexity regularization to obtain accurate predictors. We performed the LASSO regression to identify genes with OA diagnostic significance using the “glmnet” R package. The SVM algorithm is based on supervised learning and binary data classification and has been widely used for classifying and subclassifying disease genomes (14). SVM’s powerful classification capabilities in disease genome learning can help discover new biomarkers and understand disease driver genes (15). The WGCNA (16) can identify gene modules with the highest relevance to diseases. Therefore, we used LASSO regression analysis, WGCNA, and the SVM-RFE algorithm to screen diagnostic genes in OA and the intersection of the three algorithms for further validation.

To evaluate the diagnostic accuracy of these signature genes for OA, receiver operating characteristics (ROC) curves were generated using the mRNA expression datasets from 16 normal cartilage and 46 OA samples of the training cohort. The area under the ROC curve (AUC) can accurately reflect the validity of the signature gene as a diagnostic marker. The results were further validated in the GSE32317 and GSE55457 cohorts. The screened diagnostic genes were also separately validated for differential expression in the two independent validation cohorts.

A PPI network was constructed using the online STRING database (https://cn.string-db.org/) (17), with the minimum required interaction score of 0.7, to explore PPIs between signature genes and the classical PI3K/AKT and MAPK inflammatory signaling pathway. The interrelationships between signature genes and inflammatory signaling pathways, chondrocyte autophagy, and apoptosis were further explored using the same approach to clarify the functions of signature genes in OA development.

As previously described (18, 19), we collected knee cartilages from Sprague–Dawley (SD) rats (3-7 days old) under aseptic conditions, then extracted and cultured the chondrocytes in vitro. Briefly, after cutting into small pieces and 30-min digestion using trypsin (Solabio, China) at 37°C, the mixture was incubated with collagenase type II (1 mg/mL; Gibco), and high Glucose-containing DMEM medium (Gibco, USA) for 6 h. Chondrocytes were collected after centrifugation at 1500 rpm/3 min and cultured in FBS (10%; Gibco, USA) and penicillin/streptomycin (Solarbio, China) DMEM medium at 37°C. Third-generation chondrocytes were used for further studies. All experiments were approved by the Medical Ethics Committee of the First Affiliated Hospital of Guangxi Medical University (Approval number: 2022-E320-01).

We used the Cell Counting Kit-8 (CCK-8, C0037, Beyotime, China) assay to determine the safe concentration of the PDK1 inhibitor BX795 (HY-10514, MCE) and the PDK1 activator PS48 (GC13920, GLPBIO) on chondrocytes. Briefly, BX795 and PS48 were dissolved in dimethyl sulfoxide (DMSO, Beyotime, China) as a storage solution and stored at -20°C. To exclude the effects of DMSO on chondrocytes, storage solutions were diluted with phosphate-buffered saline (PBS, Solarbio, China) or DMEM medium during the experiment. The solubility of BX795 and PS48 was increased by heating to 37°C before the experiment. Third-generation chondrocytes were cultured in a 96-well plate (8×10³ cells per well) for 24 h, then intervened using BX795 (0.1-32 μM) and PS48 (2.5-320 μM) for two days. After adding 10 μL CCK-8 into each well of chondrocytes and 2-h culture, the absorbance values at 460 nm were recorded. To evaluate the viability of chondrocytes, we employed a live/dead cell staining kit (40747ES76, YEASEN, China). Briefly, chondrocytes were treated with BX795 (1 μM) or PS48 (5 μM) in 6-well plates (n = 8×10⁴ cells) after two days of incubation. PBS was rinsed clean of cell culture medium, then incubated with 4.5-μM propidium iodide (PI) and 2-μM Calcein-AM solution at 37°C for 30 min in the dark. After washing with PBS, a fluorescence imaging microscope (Olympus, BX53) was used to view and record the results.

Third-generation primary chondrocytes were cultured in 6-well plates (8×10⁴ cells/well) for 48 h. The medium was aspirated and washed 3 times with PBS, followed by 4% paraformaldehyde (POM, Biosharp, China) soaking for 15-30 min. Finally, they were incubated with 0.2 mg/mL of L safranin-O staining solution (Solarbio, China) for 5 min in the dark.

We used the RNAeasy™ Plus Animal RNA Isolation Kit with Spin Column (Beyotime, R0032, China) to extract total RNA from chondrocytes. The PrimeScript™ RT reagent Kit with gDNA Eraser (Takara, China) was used to reversibly transcribe 1 mg of the extracted RNA into cDNA. The expressions of PDK1, apoptosis-related genes (Bcl-2, Bax, and Casp-3), and autophagy-related genes (ATG7, LC3 I/II, and Beclin-1) in chondrocytes were measured by qRT-PCR. Primer sequences were designed according to the GenBank database and are shown in Table 1 . A total reaction volume of 20 μL mix containing 10 μL SYBR Premix Ex Taq mixture (Thermo Fisher Scientific, USA), 0.8 μL of each primer, 3.4 μL sterile distilled water, and 5 μL cDNA was prepared before amplification. Target mRNA levels were compared to the control and normalized to GAPDH by the 2^-ΔΔCt method.

Total chondrocyte protein was extracted using a 1% phosphatase inhibitor cocktail (CW2383, CWBIO), protease inhibitor cocktail (HY-K0010, MCE), and phenylmethanesulfonyl fluoride (PMSF, BOSTER, China)-contained RIPA lysis buffer (Beyotime, China). After determining the concentration using a BCA protein assay (Beyotime, China) and mixing with loading buffer and 10-min boiling at 100°C, protein samples were separated by 10% polyacrylamide gels. After transferring the protein bands using polyvinylidene fluoride (PVDF) membranes, membranes were blocked with 5% nonfat milk and incubated for 2 h at room temperature. Subsequently, blocked membranes were incubated with primary antibodies against GAPDH (Abcam), PDK1(Thermo Fisher Scientific), Beclin-1(Proteintech, China), LC3 I/II (Abcam), Atg7 (ZENBIO, China), P38 (Abcam), p-P38 (Abcam), JNK (CST), p-JNK (CST), ERK (Abcam), p-ERK (Abcam), p‐PI3K (Abcam), PI3K (Abcam), p‐Akt (Abcam), and Akt (Abcam) at 4°C overnight with 1:1000 dilution. After three washes with PBST and 1 h of incubation at room temperature with Rabbit anti-Goat IgG (H+L) Secondary Antibody (Thermo Fisher Scientific), signals were visualized with enhanced chemiluminescence (ECL, Beyotime) and recorded by Odyssey Infrared Imaging System.

The donors agreed to use human-derived cartilage samples, and this study was approved by the Medical Ethics Committee of the First Affiliated Hospital of Guangxi Medical University (Approval number: 2022-E320-01). Cartilage sample sections were pretreated with 3% hydrogen peroxide, blocked with 10% goat serum (Gibco), and probed with PDK1 (1:400) primary antibody overnight at 4°C. Unbound antibodies were washed away with PBS and incubated with IgG secondary antibody (1:300, Thermo Fisher Scientific) for 2 h at room temperature. Next, diaminobenzidine (DAB) substrates (Boster, China) were used to develop the positively stained areas, and sections were counterstained with hematoxylin. Tunel staining (Promega) was performed to determine the level of apoptosis in sectioned bone tissues. Finally, sections were observed using an upright microscope (Olympus).

The CIBERSORT (http://cibersort.stanford.edu/) (20) algorithm was used to quantify the content of infiltrated immune cells between OA and normal cartilage. We used the “corrplot” R package to analyze and visualize the relationship of the 22 infiltrated immune cells. We applied the “vioplot” R package to represent the variations of infiltrated immune cells between normal and OA cartilage samples with violin plots. Spearman’s rank correlation was conducted to clarify the correlation between infiltrated immune cell levels and PDK1 expression in different samples. Finally, the “ggplot2” R package was applied to visualize the associations.

Bioinformatics analyses were implemented in R (x64 4.0.2) and RStudio. The results of in vitro experiments are presented as means ± standard deviations. One-way analysis of variance (ANOVA) was performed between groups, followed by Tukey’s and t-test to determine significant differences. The significance level was set at p < 0.05, and figure legends indicate sample sizes.

After eliminating the batch effect of the two cohorts (GSE10575 and GSE51588), the ‘limma’ R package was used to calculate the merged cohorts’ differences. After screening, we identified 286 DEGs, 161 downregulated and 125 upregulated ( Figures 1A, C ). The distribution of DEGs among different cohorts and samples was visualized using a heatmap ( Figure 1B ). Meanwhile, we conducted KEGG and GO enrichment analyses to clarify the signaling regulatory pathways and biological functions of DEGs. The associations of the DEGs with the ECM-receptor interaction and the pathways associated with PI3K/AKT and chemokine were also observed ( Figure 2A ). Inhibition of the PI3K/AKT signaling pathway can promote autophagy activity and reduce the inflammatory response in chondrocytes (21, 22). We also found that DEGs were associated with neutrophil degranulation, humoral immune response, and neutrophil migration in the biological processes (BPs), suggesting that DEGs might have a regulatory relationship with immune regulation in OA ( Figure 2B ). Collagen fibers are important components of the extracellular matrix (ECM) of cartilage and play an important role in maintaining the growth and development of cartilage cells. In the cellular components (CCs) module, DGEs were significantly enriched on collagen fibers, banded collagen, and complex of collagen, indicating a close association between DGEs and cartilage development ( Figure 2C ). In terms of molecular functions (MFs), we also observed that DEGs were associated with extracellular matrix structural constituent conferring tensile strength, suggesting that DEGs were involved in the production of cartilage ECM and were associated with chondrocyte growth and development ( Figure 2D ). The degradation and abnormal metabolism of cartilage ECM are pathological features of OA development. The inhibition of cartilage ECM degradation and loss is an effective way to treat OA (23).

The GSEA can be based on the entire gene expression dataset for biological functions and can more comprehensively and systematically reveal the different biological behavior and functional pathways between treatment and control groups. Herein, we found that the main biological behaviors enriched in the treatment group were related to skeletal system development, mainly bone development, cartilage development, chondrocyte differentiation, and collagen fibril organization ( Figure 3A ). Regarding the regulation of signaling pathways, the treatment group was mainly enriched in axon guidance, and ECM preceptor interaction ( Figure 3B ). Cartilage ECM plays a crucial role in cartilage development and differentiation. Increased cartilage ECM catabolism is a key factor in OA disease progression (24). The main enriched biological functions and signaling pathways in the control group were ATP syntheses coupled electron transport, cellular respiration ( Figure 3C ), cell cycle, and DNA replication ( Figure 3D ).

We used three algorithms to identify OA signature genes more accurately: WGCNA, LASSO regression, and SVM-RFE. WGCNA can classify and visualize genes into different modules according to their different relevance to the disease ( Figure 4A ). We identified the genes in the MElighcyan module as the most relevant gene set for OA based on the highest absolute value of correlation ( Figure 4B ). Thus, we selected the MElightcyan gene set with the highest correlation coefficient for subsequent analysis, containing 224 genes ( Table S1 ). Besides, 15 genes were identified as OA diagnostic markers via LASSO regression ( Figure 4C ). Thirty-one OA signature genes were identified by the SVM-RFE algorithm ( Figure 4D ). Furthermore, the intersection of the feature genes obtained by the three algorithms was retrieved, and two (PDK1, CA2) were confirmed as OA diagnostic signature genes ( Figure 4E ). Next, we constructed ROC curves to evaluate the diagnostic ability of these two biomarkers to discriminate OA and control samples. The AUC was 0.898 for PDK1 and 0.856 for CA2 (Figure S1A). These results were further validated in the GSE32317 and GSE55457 cohorts. In the GSE32317 cohort, PDK1 and CA2 showed favorable diagnostic values with AUC of 0.875 and 0.903, respectively ( Figure S1B ). Similarly, in the GSE55457 cohort, PDK1 and CA2 also showed good diagnostic efficacy with AUC values of 0.796 and 0.661, respectively ( Figure S1C ). Furthermore, the model accuracy of the signature genes PDK1 and CA2 for predicting disease was further evaluated with an AUC of 0.918, indicating that the signature genes have a very credible value for predicting OA disease ( Figure 5A ).

Subsequently, we validated the differential expression of PDK1 and CA2 in OA using the GSE32317 and GSE55457 cohorts. PDK1 expression was higher in end-stage OA than in early-stage and higher in OA tissue than in normal cartilage ( Figures 5B, C ), suggesting that as OA disease progresses and the degree of inflammation increases, the expression level of PDK1 increases. This hypothesis was further confirmed by in vitro cellular experiments. We used different concentrations (0.5 and 1μg/mL) of LPS-induced OA chondrocyte models and characterized the survival of chondrocytes by calcein-AM/PI staining ( Figures 5D, E ). We observed that the growth activity of chondrocytes decreased with the increasing of LPS concentration, and the same result was observed in the safranin-o staining ( Figure 5F ) of the chondrocytes. The growth activity of the chondrocytes in the control group was significantly better than in the LPS group. Additionally, a positive correlation of the PDK1 expression with the dose of LPS was observed in both qRT-PCR ( Figure 5G ) and Western blot ( Figure 5H ). In contrast, there was no difference in CA2 expression between OA and normal cartilage ( Figure S1E ), except in early and end-stage OA samples ( Figure S1D ).

Based on the insights into the enrichment of DEGs in the PI3K/AKT signaling pathway ( Figure 2A ), we constructed PPI networks of diagnostic markers PDK1, CA2 with classical inflammatory signaling pathways (PI3K/AKT and MAPK), autophagy-related factors, and apoptosis-related factors to explore the potential relationship between diagnostic markers and inflammation regulation, autophagy, and apoptosis in OA ( Figure 6A ). Only PDK1 interacted with PI3K/AKT and MAPK pathways, autophagy-related factors, and apoptosis-related factors, suggesting that PDK1 might be involved in the autophagy and apoptosis of chondrocytes via PI3K/AKT and MAPK pathways. Therefore, we excluded CA2 from the subsequent experiments and used only PDK1. The CCK-8 kit was used to assess the potential cytotoxicity of the PDK1 inhibitor BX795 and the PDK1 activator PS48. Chondrocytes were cultured with different concentrations of BX795 (0.1-32 μM) and PS48 (2.5-320 μM). BX795 did not present cytotoxicity to chondrocytes between 0.1-1 μM compared to controls ( Figure 6B ), as well as PS48 between 2.5-10 μM ( Figure 6C ). Therefore, 1 μM of BX795 and 5 μM of PS48 were used for subsequent experiments.

Furthermore, we used 1 μg/mL LPS (Solarbio, China) to induce OA cell models in vitro. We divided the chondrocytes into four groups: (1) the control group, no treatment was given to the cells; (2) the OA group, chondrocytes were cultured with LPS (1 μg/mL); (3) OA+BX795 group, chondrocytes were induced by LPS for 24 h, then 1 μM of BX795 was added for culture for 24 h; (4) the OA+PS48 group, chondrocytes were induced by LPS for 24 h, then 5μM of PS48 for culture for 24 h. Our observations showed that most viable cells were found in the control group, followed by the BX795, OA, and PS48 groups ( Figures 6D, F ), and the same result was observed in the safranin-o staining ( Figure 6E ). Moreover, to understand the effect of PDK1 inhibitor BX795 and agonist PS48 on PDK1 protein expression, western blot was used to detect PDK1 protein expression between different groups, and the results showed that PDK1 protein expression in chondrocytes was decreased under the influence of BX795 ( Figure 6G ).

Through bioinformatic analysis and PPI construction, we have identified that PDK1 is involved in OA disease process and has protein interactions with mediators related to MAPK, PI3K/AKT signaling pathway. The phosphorylation of PI3K/AKT mediated by LPS was significantly attenuated after the PDK1 inhibitor was used in chondrocytes ( Figures 7A-C ). Similarly, we performed a Western blot analysis of the MAPK pathway and found that PDK1 downregulation using a PDK1 inhibitor significantly reduced the phosphorylation levels of MAPK inflammatory signaling pathway-related mediators (ERK, P38, and JNK) ( Figures 7D-G ). In summary, the phosphorylation of MAPK and PI3K/AKT inflammatory signaling pathways was attenuated by inhibiting PDK1 expression in OA cartilage, providing novel insights into inflammation inhibition and treatment in OA.

The apoptosis and autophagy of the chondrocytes are significantly connected to OA progression. Autophagy is a cellular activity that maintains cellular homeostasis and cellular activity by scavenging free oxygen radicals and fragmented organelles (25). Restoration of dysregulated autophagy in OA can increase cartilage protection and attenuates cartilage matrix degradation. Conversely, increased chondrocyte apoptosis exacerbates the degradation of the cartilage ECM and OA (25, 26). Here, we have found that PDK1 interacted with apoptosis and autophagy in chondrocytes through the MAPK and PI3K/AKT signaling pathways ( Figure 6A ). The qRT-PCR results confirmed this hypothesis ( Figures 8A-C ). The mRNA expression of autophagy-related markers (Beclin-1, Lc3I/II, and Atg7) was elevated in chondrocytes with suppressed PDK1 expression by BX795 compared to the OA group. The same results were further validated at the protein expression level via western blot analysis ( Figure 8D ). We collected cartilage from OA patients with their corresponding adjacent normal cartilage for Tunel staining in the OA chondrocyte apoptosis analysis. We detected more apoptotic cells in the OA cartilage than in normal cartilage ( Figure 8E ), suggesting that inhibition of apoptosis in chondrocytes is a desirable option for treating OA progression. The qRT-PCR results revealed that PDK1 inhibition downregulated various proteins associated with cell apoptosis, such as Bax, Bcl-2, and Casp-3. These results indicated that PDK1 inhibition rescued OA chondrocyte apoptosis ( Figures 8F-H ).

The immune system is divided into adaptive and innate immunity and is related to the body’s defense. Innate immune activation has been observed in cartilage damage in OA (27). To further understand the correlation between PDK1 and 22 immune cell infiltration in OA. First, we collected human cartilage tissues and used immunohistochemical analysis to confirm once again that PDK1 was overexpressed in OA cartilage tissues ( Figure 9A ). Next, we used the CIBERSORT algorithm to systematically analyze the association of PDK1 expression with the infiltration of 22 immune cells. The abundance of 22 immune cell infiltrates between normal and OA samples is shown in Figure S2A and Figure S2B . In addition, we observed differences in different types of immune cell infiltration between normal and OA samples, including T cells follicular helper, NK cells, Monocytes, Macrophages M1, Dendritic cells, Eosinophils, and Neutrophils ( Figure S2C ). PDK1 expression was positively associated with activated CD4 memory T cells (r = 0.29, p = 0.021), Plasma cells (r = 0.3, p = 0.016), Eosinophils (r = 0.36, p = 0.0044), and Neutrophils (r = 0.48, p<0.0001), while negatively associated with follicular helper T cells (r = -0.3, p = 0.016), and naïve CD4 T cells (r = -0.32, p = 0.01) ( Figure 9B ). The association of PDK1 with 22 immune cells can be visualized in Figure 9C .