Identification and experimental validation of biomarkers associated with exercise in intervertebral disc degeneration through bulk RNA and single-cell RNA sequencing analysis Authors: Hang Zhang, Yujie Wu, Zhiyi Fu Published: 2026 ABSTRACT Long-term running positively affects the intervertebral disc. This study aimed to identify and validate biomarkers associated with exercise-related genes (ERGs) in intervertebral disc degeneration (IDD), potentially guiding the development of targeted therapies. IDD-related datasets and ERGs were extracted from public databases. Biomarkers were identified through differential expression analysis, weighted gene co-expression network analysis (WGCNA), protein-protein interaction (PPI) networks, machine learning techniques, and gene expression analyses. A nomogram based on these biomarkers was developed and evaluated. Subsequently, functional enrichment, RNA methylation, drug prediction, molecular docking, and single-cell RNA sequencing (scRNA-seq) analyses were performed. Finally, reverse transcription quantitative PCR (RT-qPCR) was conducted to validate the expression of biomarkers. TNFAIP6 and CHI3L1 were identified as biomarkers for IDD, both exhibiting significantly higher expression in IDD samples. These findings were corroborated by RT-qPCR. A nomogram was developed and validated, with the calibration curve indicating its effectiveness in predicting IDD risk (P = 0.798 in the Hosmer-Lemeshow test). All biomarkers were significantly co-enriched in the “spliceosome” pathway. Additionally, compound 16 was found to target CHI3L1, while acetylsalicylic acid was predicted to target TNFAIP6. Molecular docking analysis revealed favorable binding energies between the compounds and biomarkers, including a binding energy of -9.1 kcal/mol between compound 16 and CHI3L1. Finally, scRNA-seq analysis identified homeostatic fibroblasts, progenitor cells, fibrochondrocytes, and regulatory chondrocytes as key cell types in IDD, with dynamic expression patterns of CHI3L1 observed during the differentiation of these key cells. Biomarkers associated with exercise, specifically TNFAIP6 and CHI3L1, were identified. These findings provide valuable insights for the development of targeted therapies for patients with IDD.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1038/s41598-026-57855-x. 1༎Introduction Degenerative disc disease is a prevalent condition closely associated with chronic lower back pain. Its incidence increases with age in adults, resulting in significant social and economic burdens1,2. The etiology of intervertebral disc degeneration (IDD) is multifactorial, involving genetic, biochemical, and mechanical influences. Continuous strain on the intervertebral disc, coupled with insufficient blood supply and nutrition, renders it susceptible to external factors, ultimately leading to IDD3. Research indicates that intervertebral discs begin to degenerate around the age of 20, and this degeneration becomes irreversible with advancing age, posing a serious threat to overall health4. Traditional treatments for IDD, including pharmacotherapy, physical therapy, and surgical interventions, primarily address symptoms rather than the underlying causes of degeneration. While drug therapies, such as anti-inflammatory medications, can alleviate symptoms like pain and inflammation, they do not prevent disease progression5. Thus, the identification of new biomarkers for IDD is essential for enhancing diagnosis and treatment strategies. Exercise therapy is defined as a structured physical activity program designed for specific therapeutic goals, aimed at restoring normal musculoskeletal function or alleviating pain associated with illness or injury. In the context of chronic low back pain, maintaining physical activity, restoring normal function, and engaging in exercise are recognized treatments, serving as effective self-management strategies for prompt return to work and daily activities, thereby reducing pain and disability6,7. Furthermore, intervertebral discs are integral components of the motor system, and research by Shu et al. has demonstrated that long-term running positively impacts intervertebral discs8. Additionally, exercise-induced FNDC5/irisin has been shown to protect nucleus pulposus cells from aging and apoptosis by activating autophagy9. However, the specific mechanisms by which exercise-related genes (ERGs) influence IDD remain unclear. Thus, this study aims to enhance understanding of the molecular mechanisms underlying IDD and provide new references for its diagnosis and treatment. In 2009, Tang et al. made a significant conceptual and technological advancement in the field of single-cell RNA sequencing (scRNA-seq) by sequencing the transcriptomes of individual blastomeres and oocytes10. With ongoing advancements in sequencing technology, single-cell sequencing has emerged as an effective tool for analyzing cell specificity and differences. This technology allows exploration of the collaborative dynamics between cells, investigation of tissue heterogeneity, accurate characterization of diseases based on cellular characteristics, precise measurement of gene expression levels, identification of rare cell types, and detection of changes in cellular components11–15. This study identified biomarkers related to motor genes in IDD based on transcriptome data and validated their expression in clinical samples. Additionally, the specific biological functions and processes associated with these biomarkers were analyzed. Furthermore, using single-cell data, the expression of biomarkers in key cell types was examined to provide new references for the treatment of clinical IDD. Materials and methods Data collection In this study, transcriptomic and scRNA-seq datasets related to IDD were retrieved from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/). Specifically, the GSE70362 dataset (GPL17810 platform) was utilized as the training set, comprising transcriptomic data from 16 IDD nucleus pulposus tissue samples (graded III, IV, and V) and 8 control nucleus pulposus tissue samples (graded I, II, and I-II). The GSE56081 dataset (GPL15314 platform) served as the validation set, providing transcriptomic data from 5 IDD nucleus pulposus tissue samples and 5 control samples. Additionally, the scRNA-seq dataset GSE230809 (GPL30173 platform) was employed, including 8 IDD and 3 control nucleus pulposus samples. A gene set of 206 exercise response-related genes was also obtained from reference16, derived from the Gene Set Enrichment Analysis (GSEA) database, encompassing various biological processes associated with exercise response (Supplementary Table S1). Differential expression analysis In the training set, differentially expressed genes (DEGs) between IDD and control samples were identified using the “limma” package (v 3.56.2)17 with criteria set at |log2 fold change (FC)| > 0.50 and P < 0.05. The identified DEGs were visualized using a volcano plot and heat map. The volcano plot, generated with the “ggplot2” package (v 3.5.1), highlighted the top 10 upregulated and downregulated genes, labeled and ranked by their P values in ascending order. The heat map, displaying the top 10 upregulated and downregulated genes, was created using the “ComplexHeatmap” package (v 2.16.0)18. WGCNA The ERG scores for the training set were calculated using the plageParam function from the “GSVA” package (v 1.53.28)19. Differences in ERG scores between IDD and control samples were assessed using the Wilcoxon test (P < 0.05). Subsequently, weighted gene co-expression network analysis (WGCNA) was conducted using the “WGCNA” package (v 1.73)20 to identify gene modules most associated with ERG scores. Initially, samples were clustered using the GoodSamplesGenes function from the “WGCNA” package to detect and remove outliers. An optimal soft threshold (power) was selected to achieve a scale-free topology with an R2 value exceeding 0.85 while maintaining mean connectivity close to zero. A co-expression matrix was constructed with a minimum module size of 100 genes and a module merging parameter (mergeCutHeight) of 0.4, resulting in distinct gene modules, each represented by a different color. Following module identification, the Pearson correlation test was employed to calculate correlation coefficients between ERG scores and each gene module (|cor| > 0.30 and P < 0.05). The modules exhibiting the highest positive and negative correlations with ERG scores were designated as key modules, with genes within these modules identified as key module genes for further detailed analysis. Identification and function analyses of candidate genes The intersection of DEGs and key module genes was identified as candidate genes for further analysis using the “ggvenn” package (v 0.1.10) (https://CRAN.R-project.org/package=ggvenn). To investigate the biological functions and pathways associated with these candidate genes, Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) analyses were performed using the “clusterProfiler” package (v 4.15.0.003)21(P < 0.05). The GO analysis encompassed categories such as biological process (BP), cellular component (CC), and molecular function (MF). The top five terms for BP, CC, and MF, as well as the top ten KEGG pathways, were displayed, ranked by P value from lowest to highest22–24. Additionally, to examine protein-protein interactions (PPIs) among the candidate genes, a PPI network was constructed using the STRING database (https://string-db.org/), with parameters set to a confidence score ≥ 0.40. The PPI network was visualized using Cytoscape (v 3.10.3)25. To further identify important genes within the network, the Degree, Maximum Neighborhood Component (MNC), Density of Maximum Neighborhood Component (DMNC), and Maximal Clique Centrality (MCC) algorithms were applied through the CytoHubba plug-in in Cytoscape (v 3.10.3). The top 20 genes from each algorithm were selected, and genes that appeared across all algorithms were classified as important genes. Determination of biomarkers In the training set, two machine learning algorithms, LASSO and Boruta, were employed to identify feature genes based on the important genes. LASSO analysis was conducted using the “glmnet” package (v 4.1.8)26. The results were confirmed at the minimum lambda value with coefficients ≠ 0, using 10-fold cross-validation. To evaluate the robustness and generalization capability of the predictive model under small-sample conditions, an internal validation method involving 1,000 Bootstrap resampling iterations with Out-of-Bag (OOB) assessment was implemented. In each resampling iteration, samples were drawn with replacement from the original training set (GSE70362) to create a Bootstrap training set, with unsampled samples serving as the OOB test set for that iteration. Within each Bootstrap training set, a LASSO-penalized logistic regression model was re-fitted using 10-fold cross-validation, and the optimal λ value was selected. This model was subsequently applied to the corresponding OOB test set to calculate predicted probabilities of IDD and the area under the receiver operating characteristic (ROC) curve (AUC). AUC values from the 1,000 iterations were summarized, and the mean, standard deviation, and 95% confidence interval were calculated. For Boruta analysis, the “Boruta” package (v 8.0.0)27 was utilized to select genes based on their importance. The intersection of genes identified by both LASSO and Boruta was determined using the “ggvenn” package (v 0.1.10) (https://CRAN.R-project.org/package=ggvenn), which facilitated the identification of feature genes. Subsequently, gene expression analyses of the feature genes were performed in both the training and validation sets using the Wilcoxon test (P < 0.05). Genes exhibiting significant differential expression between IDD and control samples (P < 0.05) and consistent expression trends across both datasets were considered biomarkers for further analysis. Construction and validation of nomogram In the training set, a nomogram was developed to predict the risk of IDD based on the identified biomarkers using the “rms” package (v 6.8.1)28. The accuracy of the nomogram was evaluated by generating a calibration curve with the same “rms” package. The results of the Hosmer-Lemeshow (HL) test were assessed, where a P value greater than 0.05 indicated good model fit. To construct the calibration curve, bias correction was applied using the Bootstrap internal validation method, implemented via the calibrate() function in the rms package. This involved setting method = “boot” and B = 1000. In each Bootstrap resample, samples were drawn with replacement from the original training set, matching the original sample size, and a logistic regression model was re-fitted to estimate optimism in model performance. Consequently, a bias-corrected calibration curve was generated. Additionally, ROC curves were produced using the “pROC” package (v 1.18)29, where an AUC greater than 0.70 indicated excellent predictive capability. Function analyses To assess potential interactions and functional relationships among the biomarkers and other genes, the GeneMANIA database (http://genemania.org/) was utilized. A gene-gene interaction (GGI) network was constructed, offering valuable insights into functional connections within the gene network. To further elucidate the biological functions and signaling pathways associated with the biomarkers, GSEA was conducted on the training set. First, Spearman correlation coefficients between the biomarkers and other genes were calculated and ranked from high to low using the “psych” package (v 2.4.6.26)30. The background gene set “c2.cp.kegg.v7.0.symbols.gmt” was obtained from the MSigDB (https://www.gsea-msigdb.org/gsea/msigdb). GSEA was then conducted using the “clusterProfiler” package (v 4.15.0.003) with significance thresholds set at |Normalized Enrichment Score (NES)| > 1, P < 0.05, and False Discovery Rate (FDR) < 0.25 (BH method). The top 10 enriched pathways, ranked by |NES| from high to low, were displayed. Subcellular localization and RNA methylation analyses To explore the subcellular localization of the biomarkers, relevant information was gathered from the GeneCards database (https://www.genecards.org/) with a confidence level of 3 to ensure high reliability. Results were visualized through a subcellular-biomarkers network created using Cytoscape (v 3.10.3), providing a clear and intuitive representation of the data. Additionally, to investigate the RNA methylation sites of the biomarkers in IDD, SRAMP (http://www.cuilab.cn/sramp/) was employed to identify m6A modification sites in RNA as well as their positions within the RNA secondary structure. Regulatory network, drug prediction, and molecular docking analyses To further investigate the underlying mechanisms of IDD, regulatory network analyses were performed. Transcription factors (TFs) targeting the biomarkers were predicted using the KnockTF database (http://www.licpathway.net/KnockTF/index.html). A TF-mRNA network was subsequently constructed and visualized with Cytoscape (v 3.10.3). Additionally, microRNAs (miRNAs) targeting the biomarkers were predicted using the miRWalk (http://mirwalk.umm.uni-heidelberg.de/) and miRDB (https://mirdb.org/) databases. Key miRNAs were identified by overlapping predictions from both databases. Following this, lncRNAs targeting these key miRNAs were identified using the miRNet database (https://www.mirnet.ca/miRNet/home.xhtml). This methodology aimed to establish a comprehensive lncRNA-miRNA-mRNA network, which was visualized with Cytoscape (v 3.10.3). To explore potential pharmacological treatments for IDD based on the identified biomarkers, the DrugBank (https://go.drugbank.com/) and DGIdb databases (https://www.dgidb.org/) were extensively utilized. A drug-biomarker network was constructed and visualized using Cytoscape (v 3.10.3). Subsequently, to evaluate the binding affinity between the drug and the biomarkers, the drug with the highest interaction score was selected for molecular docking studies. The 3D molecular structures of the drugs (acting as ligands) were obtained from the PubChem database (https://pubchem.ncbi.nlm.nih.gov/), while the protein crystal structures of the biomarkers (acting as receptors) were sourced from UniProt (https://www.uniprot.org/) and the PDB databases (http://www.rcsb.org). Molecular docking analyses were conducted using AutoDock Vina (v 1.2.3) (https://github.com/ccsb-scripps/AutoDock-Vina), and binding energy was calculated. A binding energy of less than − 5 kcal/mol typically indicates a strong binding affinity between the drug and the biomarker, suggesting potential for effective molecular interaction. The scRNA-seq analysis The scRNA-seq data from GSE230809 were analyzed utilizing the “Seurat” package (v 5.1.0)17. To ensure high data quality for downstream analysis, quality control (QC) procedures were implemented to filter out low-quality cells and genes. Specifically, cells with fewer than 200 detected genes and genes expressed in fewer than 3 cells were excluded. Retained cells met the following criteria: 300 < nFeature_RNA (genes per cell) < 10,000, nCount_RNA (total RNA count per cell) < 22,000, and percent.mt (proportion of mitochondrial gene expression) < 5%. Following QC, the data were normalized using the NormalizeData function. The FindVariableFeatures function, along with the vst method, was employed to select 2,000 highly variable genes (HVGs). Principal component analysis (PCA) was performed on these genes using the RunPCA function. To mitigate technical batch effects between donors affecting clustering results, the Harmony algorithm was employed for data integration. Initially, data from each sample were merged, PCA was conducted based on the standardized expression matrix, and then the Harmony function was applied, with sample source designated as the batch variable, to construct an integrated low-dimensional representation (integrated.Harmony). Subsequent construction of the neighbor graph, unsupervised clustering, and UMAP visualization were performed based on the first 20 principal components (dims = 1:20) after Harmony correction (P < 0.05). Unsupervised clustering analysis was executed using the FindNeighbors and FindClusters functions. UMAP visualization was applied with a resolution of 0.4 to illustrate the clusters. Cell annotations were performed based on literature reviews and validated using the FindAllMarkers function31. The marker genes corresponding to each cell type are presented in Supplementary Table S2. To gain insights into the biological pathways and functions of annotated cells, functional enrichment analysis was conducted using ReactomeGSA (https://github.com/reactome/ReactomeGSA) on the GSE230809 dataset. The top 10 enrichment results were displayed using the “ComplexHeatmap” package (v 2.16.0). Furthermore, the distribution of annotated cells in IDD and control samples was examined. Expression differences of biomarkers between IDD and control samples within each annotated cell type were assessed using the Wilcoxon test (P < 0.05). Key cell types were defined based on the following criteria: (1) significant enrichment among the top 10 pathways in the Reactome pathway enrichment analysis (P < 0.05), and (2) significant expression differences in biomarkers between IDD and control samples within that cell type (Wilcoxon test, P < 0.05). Cell-to-cell communication, pseudo-temporal, and cell cycle regulation analyses To investigate interactions between key cell types and their relationships with other annotated cell types in the GSE230809 dataset, cell-to-cell communication analysis was conducted using the “CellChat” package (v 1.6.1)32. Cell-cell communication networks and receptor-ligand interaction bubble diagrams were generated to illustrate the strength of interactions and the involved receptor-ligand pairs in intercellular communication. For further analysis, secondary dimensional reduction on the key cell types was performed using the clusterCells function from the “Seurat” package (v 5.1.0), following the methods outlined in Sect. 2.10. The key cell types were subsequently subdivided into subclusters. Additionally, pseudo-temporal analysis was executed using the “Monocle2” package (v 2.28.0)33 to examine the differentiation states and trajectories of the key cell types, including analysis of biomarker expression changes throughout the pseudo-temporal progression. To explore the role of the cell cycle in regulating various biological processes in key cell types, cell cycle analysis was conducted utilizing the “Seurat” package (v 5.1.0). The CellCycleScoring function was employed to calculate differential grade scores for each key cell type, providing insights into cell cycle regulation within these populations. The RT-qPCR For validation of biomarker expression in clinical samples, RT-qPCR analysis was conducted. Ten intervertebral disc nucleus pulposus tissue samples (five controls and five cases) were obtained from the Ninth People’s Hospital. Informed consent was obtained from all participants, and clinical information is detailed in Supplementary Table S3. The study received approval from the ethics committee of the Ninth People’s Hospital (approval number: SH9H-2025-T435-1). Total RNA was extracted from the tissue samples using TRIzol reagent (Ambion, USA) according to the manufacturer’s protocol. RNA concentration was measured using a NanoPhotometer N50. cDNA synthesis was performed by reverse transcription with the SureScript First-Strand cDNA Synthesis Kit on a S1000™ Thermal Cycler (Bio-Rad, USA) under the following conditions: 50 °C for 10 min, 85 °C for 5 s, followed by a hold at 4 °C. The qPCR reaction was prepared as follows: 3 µL of cDNA, 5 µL of 2 × Universal Blue SYBR Green qPCR Master Mix, and 1 µL each of forward and reverse primers (10 µM). Primer sequences are listed in Supplementary Table S4. Amplification was carried out on a CFX Connect Real-Time Quantitative PCR System (Bio-Rad, USA) with the following protocol: initial denaturation at 95 °C for 1 min, followed by 40 cycles of 95 °C for 20 s, 55 °C for 20 s, and 72 °C for 30 s. Relative mRNA expression was quantified using the 2−ΔΔCT method. RT-qPCR results were exported to Excel and analyzed statistically with GraphPad Prism 10 (https://www.graphpad.com/). Statistical analysis Statistical analysis was performed using R (v 4.3.3). Comparisons between the two independent groups (IDD and control) were conducted using the Wilcoxon rank-sum test (Mann-Whitney U test) (P < 0.05), with all reported P values being two-sided. For the RT-qPCR data, normality was first assessed using the Shapiro-Wilk test; if the data followed a normal distribution, the t-test was used; alternatively, if the data did not meet normality, the Wilcoxon rank-sum test was employed instead (P < 0.05). Results Identification of DEGs and key module genes In the training set, a total of 515 DEGs were identified, comprising 206 genes that were upregulated and 309 genes that were downregulated in IDD samples (Fig. 1A-B). Significant differences in ERG scores were observed between IDD and control samples (P < 0.0001) (Fig. 1C). For WGCNA, no outlier samples were detected in the training set (Fig. 1D). The optimal power value was determined to be 7, yielding an R2 value greater than 0.85 (indicated by the red line) and achieving mean connectivity close to 0 (Fig. 1E). The co-expression matrix was subsequently used to merge similar modules, resulting in the identification of 16 gene modules, excluding the gray module, which represents unclassified genes (Fig. 1F). The MEblack module, containing 687 genes, exhibited a strong positive correlation with the ERG score (cor = 0.65, P = 5 × 10− 4), while the MEpink module, comprising 614 genes, also demonstrated a strong positive association with the ERG score (cor = -0.67, P = 4 × 10− 4) (Fig. 1G). As a result, the total of 1,301 genes derived from these two key modules were classified as key module genes. Fig. 1Identification of DEGs and key module genes. (A) Volcano plot illustrating differentially expressed genes between IDD and control samples; blue indicates downregulated genes, red indicates upregulated genes. (B) Heatmap of DEGs across IDD and control samples; blue indicates downregulation, red indicates upregulation. (C) Comparison of ERG scores between IDD and control groups; blue indicates the control group, red indicates the disease group. (D) Sample clustering and trait heatmap for the training set. (E) Soft-threshold power selection for WGCNA. (F) Gene cluster dendrogram and module assignment. (G) Module-trait correlation heatmap showing associations between gene modules and ERG scores. Identification of DEGs and key module genes. (A) Volcano plot illustrating differentially expressed genes between IDD and control samples; blue indicates downregulated genes, red indicates upregulated genes. (B) Heatmap of DEGs across IDD and control samples; blue indicates downregulation, red indicates upregulation. (C) Comparison of ERG scores between IDD and control groups; blue indicates the control group, red indicates the disease group. (D) Sample clustering and trait heatmap for the training set. (E) Soft-threshold power selection for WGCNA. (F) Gene cluster dendrogram and module assignment. (G) Module-trait correlation heatmap showing associations between gene modules and ERG scores. Function exploration of candidate genes and determination of 7 important genes A total of 126 candidate genes were identified by intersecting DEGs with key module genes (Fig. 2A). These candidate genes were significantly enriched in 405 GO terms (P < 0.05), including 25 BPs, 14 CCs, and 30 MFs. Noteworthy enriched GO terms included “regulation of angiogenesis” (BP), “collagen-containing extracellular matrix (ECM)” (CC), and “ECM structural constituent” (MF) (Fig. 2B, Supplementary Table S5). Furthermore, the candidate genes were significantly enriched in 16 KEGG pathways (P < 0.05), such as “complement and coagulation cascades,” “protein digestion and absorption,” and “fatty acid elongation” (Fig. 2C, Supplementary Table S5). To further explore interactions among these candidate genes, a PPI network was constructed (Fig. 2D). Notably, ACAN displayed strong interactions with several other genes, including TNFAIP6 and CNMD, suggesting that ACAN may serve as a critical regulator within the network. Subsequently, seven important genes were identified by overlapping the top 20 genes from each algorithm (Fig. 2E, Supplementary Fig. S1A-D). Fig. 2Functional exploration of candidate genes and screening of important genes. (A) Venn diagram identifying 126 candidate genes shared by DEGs and WGCNA-derived key module genes. (B) GO enrichment analysis of candidate genes across biological process, cellular component, and molecular function categories. (C) KEGG pathway enrichment analysis of candidate genes. (D) PPI network of candidate genes. (E) UpSet plot illustrating the intersection of top 20 candidate genes screened by four cytoHubba algorithms. Functional exploration of candidate genes and screening of important genes. (A) Venn diagram identifying 126 candidate genes shared by DEGs and WGCNA-derived key module genes. (B) GO enrichment analysis of candidate genes across biological process, cellular component, and molecular function categories. (C) KEGG pathway enrichment analysis of candidate genes. (D) PPI network of candidate genes. (E) UpSet plot illustrating the intersection of top 20 candidate genes screened by four cytoHubba algorithms. Identification of TNFAIP6 and CHI3L1 as biomarkers for IDD From an initial set of seven important genes, five genes (TNFAIP6, IGFBP3, CHI3L1, ACAN, CDH1) were identified through LASSO analysis [log(lambda.min) = -4.8077] (Fig. 3A). The Bootstrap OOB validation results demonstrated that the model achieved a mean AUC of 0.9923, with a standard deviation of 0.0299 and a 95% confidence interval of 0.8889-1.0000 over 1,000 resampling iterations. These results indicate that the LASSO-based logistic regression model exhibited high discriminatory capability and low performance variability during repeated resampling, reflecting strong stability within the current cohort (Supplementary Fig. S2). Additionally, seven genes were identified through Boruta analysis (Fig. 3B). By intersecting the five genes identified by LASSO and the seven genes from Boruta, five feature genes were confirmed (Fig. 3C). Subsequent gene expression analysis revealed that TNFAIP6 and CHI3L1 were significantly upregulated in IDD samples from both the training and validation sets (P < 0.05) (Fig. 3D). Therefore, TNFAIP6 and CHI3L1 were designated as biomarkers for IDD, offering potential utility for early diagnosis and understanding the disease’s underlying mechanisms. Further validation using RT-qPCR corroborated these findings, demonstrating significantly higher expression of TNFAIP6 and CHI3L1 in IDD samples (P < 0.01) (Fig. 3E-F). Fig. 3Machine learning and identification of biomarkers. (A) LASSO regression coefficient profiles and cross-validation curve. (B) Boruta feature-selection results. (C) Venn diagram showing the intersection of genes selected by LASSO and Boruta. (D) Expression validation of candidate genes in the training and validation datasets. **P < 0.01, ***P < 0.001, ns, not significant. (E-F) RT-qPCR validation of TNFAIP6 and CHI3L1 expression in IDD and control samples (t-test). (E) TNFAIP6. (F) CHI3L1. Machine learning and identification of biomarkers. (A) LASSO regression coefficient profiles and cross-validation curve. (B) Boruta feature-selection results. (C) Venn diagram showing the intersection of genes selected by LASSO and Boruta. (D) Expression validation of candidate genes in the training and validation datasets. **P < 0.01, ***P < 0.001, ns, not significant. (E-F) RT-qPCR validation of TNFAIP6 and CHI3L1 expression in IDD and control samples (t-test). (E) TNFAIP6. (F) CHI3L1. Establishment of well-performing nomogram A nomogram model was developed based on the identified biomarkers, with specific point values assigned to each gene. Higher total points in the model correlated with an increased risk of IDD (Fig. 4A). Bias correction of the calibration curve was performed using 1,000 Bootstrap resamples, and the HL test yielded a P value of 0.798, confirming the model’s accuracy (Fig. 4B). ROC analysis revealed an AUC for the nomogram model of 0.8516 (DeLong method, 95% CI: 0.6917-1.0000), indicating strong discriminatory ability within the training set (Fig. 4C). Overall, these results exhibited good predictive potential during internal evaluation of the training set; however, the generalizability of the model requires further validation in larger, independent cohorts. Fig. 4Construction and evaluation of the nomogram. (A) Nomogram constructed using CHI3L1 and TNFAIP6 to predict IDD risk. (B) Calibration curve for the nomogram, including the HL test result. (C) ROC curve evaluating the diagnostic performance of the nomogram in the training set. Construction and evaluation of the nomogram. (A) Nomogram constructed using CHI3L1 and TNFAIP6 to predict IDD risk. (B) Calibration curve for the nomogram, including the HL test result. (C) ROC curve evaluating the diagnostic performance of the nomogram in the training set. Investigation of functions and subcellular localization of biomarkers Using GeneMANIA, 20 genes associated with the biomarkers were identified. For instance, CHI3L1 was linked with CHI3L2, while TNFAIP6 interacted with IL1B. These gene interactions were implicated in key processes such as “tertiary granule,” “regulation of inflammatory response,” and “amino sugar catabolic process” (Fig. 5A), suggesting that these genes play essential roles in the cellular and molecular mechanisms underlying the disease. GSE revealed that both biomarkers were significantly enriched in multiple pathways (|NES| > 1, FDR < 0.25, P < 0.05). Specifically, TNFAIP6 showed significant enrichment in the “hematopoietic cell lineage” and “mismatch repair (MMR)” pathways, whereas CHI3L1 was notably enriched in the “ECM-receptor interaction” and “arachidonic acid metabolism” pathways (Fig. 5B). Notably, the “spliceosome” pathway was identified among the top three enriched pathways for both TNFAIP6 and CHI3L1 (Fig. 5C), highlighting the potential involvement of these biomarkers in fundamental cellular processes such as RNA splicing. Furthermore, subcellular localization analysis indicated that TNFAIP6 was present in the nucleus, cytosol, and plasma membrane, while CHI3L1 localized to the cytoskeleton and endoplasmic reticulum (Fig. 5D). Notably, both TNFAIP6 and CHI3L1 were found in the extracellular space. This localization suggests that these biomarkers may be involved in intercellular communication and signaling, which are essential for elucidating the disease’s pathophysiology and developing targeted therapies. Fig. 5Functional analysis of biomarkers. (A) GeneMANIA interaction network of biomarker-related genes. (B) GSEA plots showing enriched pathways associated with CHI3L1 and TNFAIP6. (C) Correlation analysis of the top three enriched pathways for each biomarker. (D) Subcellular localization network of CHI3L1 and TNFAIP6. Functional analysis of biomarkers. (A) GeneMANIA interaction network of biomarker-related genes. (B) GSEA plots showing enriched pathways associated with CHI3L1 and TNFAIP6. (C) Correlation analysis of the top three enriched pathways for each biomarker. (D) Subcellular localization network of CHI3L1 and TNFAIP6. Comprehensive regulatory network analysis of biomarkers TNFAIP6 and CHI3L1 RNA methylation analysis revealed that the CHI3L1 gene had six predicted m6A methylation sites, two of which were categorized as highly confident (Fig. 6A). In contrast, the TNFAIP6 gene had two predicted m6A methylation sites, with one exhibiting medium confidence (Fig. 6B). These methylation sites may play a critical role in the modulation of gene expression and function. Regulatory network analysis further identified 119 TFs targeting the biomarkers. Notably, several TFs, such as MYB, MYC, and TFAP2C, were found to co-target both TNFAIP6 and CHI3L1 (Fig. 6C), suggesting potential coordinated regulation of these biomarkers by multiple TFs, which could enhance the understanding of their functional roles in the cellular context. Additionally, 35 key miRNAs were identified through the overlap of predictions from miRWalk and miRDB (Fig. 6D). Subsequently, lncRNAs targeting these key miRNAs were identified, leading to the construction of a lncRNA-miRNA-mRNA network. For example, the interaction between TNFAIP6, hsa-miR-181b-5p, and MALAT1 was elucidated (Fig. 6E), providing insights into the complex regulatory interactions involving the biomarkers and highlighting potential mechanisms for their post-transcriptional regulation. Fig. 6RNA methylation and regulatory network analysis. (A) Predicted m6A modification sites of CHI3L1. (B) Predicted m6A modification sites of TNFAIP6. (C) TF-mRNA regulatory network targeting the biomarkers. (D) Venn diagram identifying overlapping miRNA-target pairs predicted by miRWalk and miRDB. (E) lncRNA-miRNA-mRNA regulatory network related to the biomarkers. RNA methylation and regulatory network analysis. (A) Predicted m6A modification sites of CHI3L1. (B) Predicted m6A modification sites of TNFAIP6. (C) TF-mRNA regulatory network targeting the biomarkers. (D) Venn diagram identifying overlapping miRNA-target pairs predicted by miRWalk and miRDB. (E) lncRNA-miRNA-mRNA regulatory network related to the biomarkers. Detection of potential therapeutic drugs for IDD In the pursuit of effective targeted therapies, five drugs were predicted to interact with the biomarkers. Specifically, hyaluronic acid, acetylsalicylic acid, and donepezil were identified as potential interactants for TNFAIP6, while recombinant chemokine and compound 1634 were associated with CHI3L1 (Fig. 7A). To further investigate these interactions, acetylsalicylic acid and compound 16 were selected for molecular docking analysis with TNFAIP6 and CHI3L1, respectively. The docking results indicated that the binding energy between CHI3L1 and compound 16 was − 9.1 kcal/mol, suggesting a strong binding affinity (Fig. 7B). In contrast, the binding energy between TNFAIP6 and acetylsalicylic acid was − 3.9 kcal/mol, falling short of the threshold for strong binding affinity (< -5 kcal/mol) established in this study, indicating a potentially weak interaction (Fig. 7C). Notably the drug predictions and molecular docking analyses were preliminary screenings based on computational biology, merely indicating possible binding interactions that do not equate to in vivo pharmacological activity or therapeutic efficacy. These findings offer initial insights for future research into targeted therapies but necessitate rigorous validation through cellular, biochemical, and animal studies. Fig. 7Drug prediction and molecular docking. (A) Predicted interaction network between biomarkers and potential therapeutic compounds. (B) Molecular docking results for CHI3L1 and compound 16, showing favorable binding energy and key interaction residues. (C) Molecular docking results for TNFAIP6 and acetylsalicylic acid. In the larger left figure, blue represents the molecular structure of the key gene receptor protein, and yellow represents the small molecular structure of the drug ligand. In the enlarged inset, blue indicates specific bonding interactions, and yellow represents the small molecular structure of the drug ligand. Drug prediction and molecular docking. (A) Predicted interaction network between biomarkers and potential therapeutic compounds. (B) Molecular docking results for CHI3L1 and compound 16, showing favorable binding energy and key interaction residues. (C) Molecular docking results for TNFAIP6 and acetylsalicylic acid. In the larger left figure, blue represents the molecular structure of the key gene receptor protein, and yellow represents the small molecular structure of the drug ligand. In the enlarged inset, blue indicates specific bonding interactions, and yellow represents the small molecular structure of the drug ligand. Identification of four key cell types in IDD After filtering out ineligible cells and genes, a total of 30,607 cells and 24,346 genes were retained (Supplementary Fig. S3A). A set of 2,000 HVGs was identified, with the top 10 HVGs highlighted (Supplementary Fig. S3B). PCA was conducted, leading to the selection of the top 20 principal components for further analysis (P < 0.05) (Supplementary Fig. S3C). Using Uniform Manifold Approximation and Projection (UMAP), cells were classified into 12 distinct clusters (resolution = 0.4) (Fig. 8A). These clusters were annotated as fibrochondrocytes, regulatory chondrocytes, stem cells, fibroblasts, blood vessel cells, and progenitor cells (Fig. 8B-C). This clustering and annotation facilitated the identification of distinct cell populations and their potential roles within the biological context. Functional analysis uncovered specific gene expression patterns across different cell types: homeostatic fibroblasts exhibited significantly high expression related to ALKBH2-mediated reversal of alkylation damage; progenitor cells demonstrated high expression in the synthesis of hepoxillins (HX) and trioxillins (TrX); fibrochondrocytes showed elevated expression in the metabolism of ingested MeSeO2H into MeSeH; regulatory chondrocytes had heightened expression of MGMT-mediated DNA damage reversal; and blood vessel cells displayed increased expression in fructose biosynthesis (Fig. 8D). These findings emphasized the distinct functional roles of each cell type and their potential contributions to overall biological processes. Additionally, the distribution of annotated cells is shown in Fig. 8E. Wilcoxon test results indicated significant expression differences for the two biomarkers between IDD and control samples in most annotated cells, with the exception of blood vessel cells (Fig. 8F-G). This significant differential expression suggests that these biomarkers may play key roles in the pathogenesis of IDD. Consequently, homeostatic fibroblasts, progenitor cells, fibrochondrocytes, and regulatory chondrocytes were identified as key cell types for IDD. These cell types and their associated biomarkers could serve as potential therapeutic targets and diagnostic markers for IDD. Fig. 8Single-cell analysis and key cell-type identification. (A) UMAP clustering of single-cell transcriptomes. (B) Heatmap of marker-gene expression across annotated cell clusters. (C) Cell-cluster annotation results. (D) Pathway activity analysis across different cell types. (E) Distribution of annotated cell types in IDD and control samples; purple indicates low expression, green indicates high expression. (F) Expression of CHI3L1 and TNFAIP6 across annotated cell types. (G) UMAP distribution of CHI3L1 and TNFAIP6 expression, with color intensity representing gene expression levels. Single-cell analysis and key cell-type identification. (A) UMAP clustering of single-cell transcriptomes. (B) Heatmap of marker-gene expression across annotated cell clusters. (C) Cell-cluster annotation results. (D) Pathway activity analysis across different cell types. (E) Distribution of annotated cell types in IDD and control samples; purple indicates low expression, green indicates high expression. (F) Expression of CHI3L1 and TNFAIP6 across annotated cell types. (G) UMAP distribution of CHI3L1 and TNFAIP6 expression, with color intensity representing gene expression levels. Exploration interactions and differentiation trajectories of key cell types Cell-cell communication analysis provided significant insights into the interactions among key cell types and annotated cells. Regulatory chondrocytes were observed to engage with all other key cell types, underscoring their pivotal role in cellular communication within the context of IDD (Fig. 9A). For example, the primary signaling pathway identified between regulatory chondrocytes and homeostatic fibroblasts was MDK-NCL, suggesting its potential importance in facilitating cellular interactions in IDD (Fig. 9B). Moreover, analysis indicated that both regulatory chondrocytes and fibrochondrocytes frequently communicated with fibroblasts (Fig. 9C-D), highlighting a potentially significant interplay among these cell types. Pseudo-time analysis was performed to trace the differentiation trajectories of the four key cell types. For fibrochondrocytes, the differentiation process progressed from dark blue to dark red, categorizing cells into 11 distinct stages (stages 1–11) and 5 subtypes. The expression of CHI3L1 demonstrated an increasing trend throughout fibrochondrocyte development (Fig. 10A). Likewise, for regulatory chondrocytes, differentiation also progressed from dark blue to dark red, with cells classified into 7 stages (stages 1–7) and 4 subtypes. Notably, CHI3L1 expression initially declined slowly before experiencing a rapid decrease in the later stages of regulatory chondrocyte differentiation (Fig. 10B). In the case of homeostatic fibroblasts, the differentiation trajectory from dark blue to dark red resulted in 7 stages (stages 1–7) and 5 subtypes. CHI3L1 expression first decreased before gradually increasing in the latter stages of homeostatic fibroblast differentiation (Fig. 10C). For progenitor cells, differentiation also progressed from dark blue to dark red, resulting in 3 stages (stages 1–3) and 5 subtypes, with CHI3L1 expression initially increasing and then decreasing in the later stages of progenitor cell differentiation (Fig. 10D). These findings provide a comprehensive understanding of the dynamic gene expression changes during the differentiation of key cell types, which may be crucial for elucidating the underlying mechanisms of IDD. Lastly, cell cycle regulation analysis indicated that progenitor cells predominantly resided in the G2M phase (Fig. 10E). Furthermore, progenitor cells exhibited higher S and G2M scores compared to other key cell types (Fig. 10F). This suggests that progenitor cells might possess a greater proliferative capacity, which could be relevant to their roles in tissue repair and regeneration processes in IDD. Fig. 9Cell-cell communication analysis. (A) Cellular interaction network among key cell types; different colors represent distinct key cells, line thickness indicates interaction strength, and circles denote ligand-receptor pairs. (B) Bubble plot of ligand-receptor pair interactions between key cell types. (C) Interaction network between key cell types and other annotated cells. (D) Heatmap illustrating ligand-receptor interaction patterns among various cell types; horizontal and vertical axes represent different cell types, with the vertical axis indicating signal-sending cells. Color intensity in the central annotated plot reflects the number of interactions, while bar plots above and to the right depict interaction counts for each cell type. Cell-cell communication analysis. (A) Cellular interaction network among key cell types; different colors represent distinct key cells, line thickness indicates interaction strength, and circles denote ligand-receptor pairs. (B) Bubble plot of ligand-receptor pair interactions between key cell types. (C) Interaction network between key cell types and other annotated cells. (D) Heatmap illustrating ligand-receptor interaction patterns among various cell types; horizontal and vertical axes represent different cell types, with the vertical axis indicating signal-sending cells. Color intensity in the central annotated plot reflects the number of interactions, while bar plots above and to the right depict interaction counts for each cell type. Fig. 10Pseudotemporal trajectory and cell-cycle analysis. (A-D) Pseudotemporal trajectory analysis for fibrochondrocytes, regulatory chondrocytes, homeostatic fibroblasts, and progenitor cells, depicting cell-state progression and dynamics of CHI3L1 expression. (E) Proportions of key cell types across cell-cycle phases; different colors signify various cell types, with the vertical axis representing the cell cycle and the horizontal axis indicating proportions of different key cells. (F) S-phase and G2/M-phase scores of key cell types; the horizontal axis denotes the cell cycle, while the vertical axis shows the respective cell cycle scores. Pseudotemporal trajectory and cell-cycle analysis. (A-D) Pseudotemporal trajectory analysis for fibrochondrocytes, regulatory chondrocytes, homeostatic fibroblasts, and progenitor cells, depicting cell-state progression and dynamics of CHI3L1 expression. (E) Proportions of key cell types across cell-cycle phases; different colors signify various cell types, with the vertical axis representing the cell cycle and the horizontal axis indicating proportions of different key cells. (F) S-phase and G2/M-phase scores of key cell types; the horizontal axis denotes the cell cycle, while the vertical axis shows the respective cell cycle scores. Discussion This study identified TNFAIP6 and CHI3L1 as biomarkers for IDD linked to ERGs through the integration of transcriptomic and single-cell transcriptomic data. Both genes demonstrated significantly elevated expression in degenerated nucleus pulposus tissue, and this differential expression was validated in independent clinical samples using RT-qPCR. The nomogram model constructed based on these two biomarkers exhibited strong predictive efficacy for IDD risk (AUC = 0.8516, HL test P = 0.798), serving as a preliminary tool for disease risk assessment. The ensuing discussion examines the biological functions of these biomarkers, the potential pathways involved, and their expression characteristics at the single-cell level, aiming to provide novel insights into the molecular mechanisms and therapeutic targets for IDD. TNFAIP6 (Tumor Necrosis Factor Alpha Induced Protein 6) is an inflammation-related protein whose expression can be induced by pro-inflammatory cytokines such as tumor necrosis factor alpha (TNF-α) and interleukin-1 (IL-1)35,36. TNFAIP6 plays a pivotal role in tissue repair by regulating the stability of the ECM and facilitating cell migration, thus promoting wound healing and tissue regeneration37. CHI3L1 (Chitinase-3-like protein 1) is a secreted glycoprotein that mediates various biological processes, including inflammation, macrophage polarization, apoptosis, and carcinogenesis38. The CHI3L1 signal is essential for growth, proliferation, invasion, metastasis, angiogenesis, and the activation of tumor-associated macrophages, as well as Th2 polarization of CD4 + T cells in cancer39. The inflammation-sensitive CHI3L1 protects intervertebral disc chondrocytes during IDD through AKT3 signaling40. The biomarkers identified in this study are both associated with inflammation, a key contributor to the chronic inflammation and cell death characteristic of IDD41. Notably, TNFAIP6 has been previously reported as a biomarker for IDD42, and our study confirms that TNFAIP6 is significantly overexpressed in IDD samples (P < 0.05). Furthermore, literature supports the notion that CHI3L1 acts as a biomarker for IDD43, indicating significantly elevated levels in IDD samples as well. These findings suggest that TNFAIP6 and CHI3L1 may be integral to the initiation and progression of IDD, although their precise roles as biomarkers warrant further validation. The evaluation and validation of the nomogram model based on these biomarkers demonstrate its robust predictive capacity for IDD, providing a substantial theoretical foundation for risk prediction in this context. Additionally, GSEA revealed that TNFAIP6 was significantly enriched in pathways such as “hematopoietic cell lineage” and “MMR,” while CHI3L1 showed significant enrichment in pathways like “ECM-receptor interaction” and “arachidonic acid metabolism.” These findings highlight that these functions and associated pathways may play critical roles in the pathophysiology of IDD. Hematopoietic cell lineages encompass various blood cells and their precursor cells that differentiate from hematopoietic stem cells (HSCs). This includes myeloid cells (e.g., granulocytes, monocytes, macrophages, red blood cells) and lymphoid cells (e.g., T cells, B cells, NK cells)44. The dynamic balance of these lineages relies on the self-renewal and directed differentiation of HSCs. Core mechanisms underlying IDD involve chronic inflammation, ECM degradation, and aging of nucleus pulposus cells45. Hematopoietic cells, particularly myeloid cells, are thought to play a pivotal role in IDD through immune inflammatory responses. Their pro-inflammatory subsets, such as M1 macrophages and Th1/Th17 cells, are key effector cells in degeneration, while imbalances in anti-inflammatory subsets, like M2 macrophages and regulatory T cells (Tregs), can exacerbate the pathological process. Regulating hematopoietic cell function may represent a novel approach for IDD intervention46. MMR is a highly conserved DNA repair mechanism that corrects base mismatches (e.g., G-T or A-C mismatches) and insertion/deletion loops (IDLs) generated during DNA replication, thereby maintaining genomic stability. This process is essential for preventing tumorigenesis and regulating the cell cycle47. Abnormalities in DNA repair mechanisms can lead to various degenerative diseases, including cancer, neurological disorders, and premature aging. For instance, defects in the base excision repair (BER) and MMR pathways regarding oxidized guanine residues can induce trinucleotide amplification, which is implicated in Huntington’s disease, a severe neurodegenerative disorder48. ECM receptor interaction refers to the biological process where cells bind to ECM components (e.g., collagen, fibronectin, laminin, proteoglycans) via surface receptors (e.g., integrins, CD44, hyaluronic acid receptors). This interaction mediates cell adhesion, migration, proliferation, differentiation, and signal transduction49. The ECM of intervertebral discs, particularly the nucleus pulposus, is primarily composed of type II collagen and aggrecan, with its homeostasis dependent on the dynamic balance of ECM receptors. Dysfunction of ECM receptor interactions can lead to cell apoptosis, inflammatory responses, and ECM degradation, representing one of the fundamental mechanisms of IDD50. ECM receptor interactions can regulate the expression and release of inflammatory factors, thereby influencing the inflammatory microenvironment within intervertebral discs51. Thus, targeting ECM receptor interactions—such as enhancing adhesion signaling and inhibiting degradation pathways—may serve as a promising strategy for delaying IDD52. Arachidonic acid is an omega-6 polyunsaturated fatty acid found in cell membrane phospholipids. Upon cellular stimulation (e.g., inflammation or injury), phospholipase A2 (PLA2) releases arachidonic acid from these membranes. Arachidonic acid metabolism generates pro-inflammatory or anti-inflammatory mediators through pathways like cyclooxygenase (COX) and lipoxygenase (LOX), which drive inflammation, ECM degradation, and pain in IDD53. Elevated COX-2 expression has been observed in degenerated intervertebral discs, with prostaglandin E2 (PGE2) promoting the release of cytokines such as IL-6 and TNF-α, activating matrix metalloproteinases (MMPs) like MMP-3/13, and accelerating ECM degradation54. Therefore, targeting key enzymes involved in arachidonic acid metabolism (e.g., COX-2, 5-LOX) or regulating the metabolite ratios may offer therapeutic avenues for IDD. Additionally, computational predictions in this study indicated that Compound 1634 may bind to CHI3L1, with molecular docking results showing a binding energy of -9.1 kcal/mol, suggesting a significant potential for structural interaction between the two. Compound 16, derived from phloridzin, functions as a nucleotide transporter inhibitor and has been synthesized and evaluated as an inhibitor of the human concentrated nucleotide transporter 3 (hCNT3)55. hCNT3 indirectly impacts intracellular and extracellular adenosine concentrations by facilitating the uptake of adenosine precursors, such as adenosine nucleotides, and also regulates ATP/ADP/AMP metabolism, thereby affecting P2X/P2Y receptor-mediated inflammatory responses. During early stages of IDD, mechanical injury or oxidative stress leads to ATP release, and hCNT3-mediated adenosine production may temporarily suppress inflammation. A decline in hCNT3 function could result in compromised anti-inflammatory mechanisms, accelerating the progression of IDD56. Based on this speculation, Compound 16 may influence the pathological progression of IDD through the aforementioned pathways, offering a valuable research avenue for targeted interventions. However, as this analysis is currently limited to computational biology, the binding of Compound 16 to CHI3L1 and its therapeutic potential for IDD require systematic validation through in vitro and in vivo studies. This study identified four key cell populations through single-cell correlation analysis: homeostatic fibroblasts, progenitor cells, fibrochondrocytes, and regulatory chondrocytes. Among these, progenitor cells play a pivotal role in IDD. These cells, derived from stem cells, possess limited self-renewal capacity and differentiation potential57. Progenitor cells are essential for tissue development, maintenance, and repair, contributing new cells to tissues and aiding in damage repair and tissue homeostasis58. They may facilitate the repair of damaged intervertebral disc tissue by differentiating into new disc cells59,60. Pseudotemporal analysis revealed the heterogeneity of progenitor cell populations and the dynamic changes in gene expression during their differentiation process. Notably, TNFAIP6 expression increased in the later pseudotemporal stages, indicating the activation of inflammatory signaling during differentiation. Similarly, CHI3L1 expression rose throughout the pseudotemporal progression, potentially reflecting its role in ECM remodeling and highlighting its influence on key cell functions. This study focused on exercise-related genes (ERGs) and conducted a multidimensional analysis of the expression characteristics of TNFAIP6 and CHI3L1 in IDD through the integration of bulk and single-cell transcriptomic data. The single-cell analysis provided insights into the dynamic expression changes of CHI3L1 along the differentiation trajectories of key cell types, offering spatiotemporal information that previous bulk-level studies did not capture. Additionally, the nomogram model developed from these two biomarkers served as a preliminary tool for IDD risk prediction. Functional enrichment, regulatory network, and drug prediction analyses further provided valuable reference points for future investigations. However, this study has certain limitations. First, the control group in the public dataset included Pfirrmann Grade I, Grade II, and Grade I–II samples, which do not represent strictly normal intervertebral discs. This may have led to an underestimation of intergroup differences, potentially impacting the sensitivity and specificity of the biomarker screening. Second, RT-qPCR validation was conducted on a limited number of samples and assessed only at the mRNA level, lacking confirmation at the protein level and functional experiments to clarify the potential pathological mechanisms through which TNFAIP6 and CHI3L1 may be involved in IDD. Moreover, the ERG set was derived from literature integration and still requires direct validation through animal exercise intervention experiments. Lastly, the nomogram was evaluated solely within the training set and has not yet been validated in an independent external cohort; thus, its diagnostic utility should be interpreted with caution. To address these limitations, future studies should incorporate clearly non-degenerated normal samples and utilize multi-center independent cohorts. Additionally, confirming biomarker expression at the protein level, evaluating nomogram performance, and integrating in vitro cell models and in vivo animal models for functional experiments and exercise interventions will be essential. These steps are necessary to elucidate the mechanisms of action for these two genes in IDD and their relationship with exercise regulation. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1 Supplementary Material 2 Supplementary Material 2 Supplementary Material 3 Supplementary Material 3 Supplementary Material 4 Supplementary Material 4 Supplementary Material 5 Supplementary Material 5 Supplementary Material 6 Supplementary Material 6