Exploration and experimental validation of oxidative stress-related diagnostic genes in interstitial cystitis based on transcriptomics

Abstract

Background

Interstitial cystitis (IC) is a chronic pain syndrome with an elusive diagnosis and poorly understood pathogenesis, in which oxidative stress (OS) is increasingly implicated. This study aimed to identify and validate OS-related diagnostic biomarkers for IC using an integrative computational and experimental approach.

Methods

We performed bioinformatics analysis on human bladder transcriptomic datasets (GSE11783, GSE57560) to identify OS-related differentially expressed genes (DEOSGs). Three machine learning algorithms [least absolute shrinkage and selection operator (LASSO), support vector machine-recursive feature elimination (SVM-RFE), random forest] were applied to screen for robust diagnostic markers. Immune cell infiltration was analyzed using Cell-type Identification By Estimating Relative Subsets Of RNA Transcripts (CIBERSORT). Putative therapeutic agents were identified through the Drug Signatures Database (DSigDB) and further assessed via molecular docking and molecular dynamics (MD) simulations. The top candidate genes were validated in a cyclophosphamide-induced rat model of IC via reverse transcription quantitative real-time polymerase chain reaction (RT-qPCR) and western blotting.

Results

We identified 58 DEOSGs in IC. Three machine learning methods consistently pinpointed S100A8 and TLR2 as key diagnostic genes. A nomogram model incorporating these genes showed high diagnostic accuracy [area under the receiver operating characteristic curve (AUC) >0.9]. Immune profiling revealed significant correlations between S100A8/TLR2 expression and specific CD4⁺ T cell subsets. In the IC rat model, both messenger RNA (mRNA) and protein levels of S100A8 and TLR2 were significantly upregulated. Drug repurposing analysis nominated simvastatin as a potential therapeutic agent modulating the TLR2 pathway.

Conclusions

Our study identifies S100A8 and TLR2 as novel diagnostic biomarkers for IC. The S100A8/TLR2 axis represents a promising therapeutic target, with simvastatin highlighted as a potential repurposing candidate, offering new strategies for precise diagnosis and mechanism-based treatment of IC.

Highlight box.

Key findings

• S100A8 and TLR2 are identified as oxidative stress (OS)-driven diagnostic biomarkers for interstitial cystitis (IC) through integrated bioinformatics and machine learning.

• Both genes are significantly upregulated in IC patients and validated in a cyclophosphamide-induced rat model at messenger RNA and protein levels.

• The S100A8/TLR2 axis correlates with distinct immune cell infiltration patterns, particularly CD4⁺ T cell subsets, linking OS to immune dysregulation.

• A nomogram incorporating these genes achieves excellent diagnostic accuracy (area under the receiver operating characteristic curve >0.9).

• Drug-repurposing analysis nominates simvastatin as a potential therapeutic agent targeting this pathway.

What is known and what is new?

• It is known that IC lacks objective diagnostic biomarkers and that OS contributes to its pathogenesis.

• This manuscript newly establishes S100A8 and TLR2 as functionally interconnected, OS-linked biomarkers, validated experimentally, and proposes their axis as a novel therapeutic target with drug-repurposing potential.

What is the implication, and what should change now?

• These findings provide a foundation for developing objective diagnostic tools and mechanism-based therapies for IC.

• Clinical validation in patient cohorts and preclinical evaluation of simvastatin targeting the S100A8/TLR2 axis should be prioritized to translate these discoveries into practice.

Introduction

Interstitial cystitis (IC)—a chronic condition featuring bladder pain, urgency, and frequency—poses diagnostic and etiological challenges (1). Currently, there is a lack of a unified definition and diagnostic methodology for the disease, and the prevalence of IC varies significantly across different regions, with current evidence estimating it to be between 0.01% and 6.5% (2,3). Moreover, earlier research indicated that the incidence rate in women is about five times that in men (4). In contrast, recent investigations have revealed no significant sex-based differences in the prevalence of IC (5). The exact pathogenic mechanisms underlying the disease are still not fully understood. Nevertheless, continued research efforts in recent years have led to the widespread acceptance of several factors as potential contributors to IC, including neurogenic inflammation, infection, autoimmune responses, mast cell activation, glycosaminoglycan (GAG) layer defects, and altered bladder epithelial permeability (2,6-9). Dysfunction of the urothelial barrier permeability, induced by local pathology or systemic inflammation, constitutes one of the pathogenic factors in IC. The degradation of the superficial GAG layer in the urothelium results in abnormally elevated permeability, enabling the infiltration of noxious urinary solutes (e.g., potassium ions) into the subepithelial tissues of the bladder wall (10). Such infiltration directly stimulates nerve and muscle fibers, inducing depolarization, and concurrently activates mast cells residing in the submucosa. Simultaneously, urothelial cells autonomously secrete an antiproliferative factor (APF) that suppresses cellular proliferation and disrupts the assembly of tight junction proteins (such as E-cadherin and zonula occludens-1). This profoundly compromises the epithelial regenerative capacity, promotes apoptotic cell death, and establishes a self-perpetuating cycle of continuous barrier function deterioration (11,12). The compromised urothelial barrier function is responsible for pain related to bladder filling, inflammatory processes and urinary urgency (13). Research indicates that oxidative stress (OS) and free radicals potentially contribute to IC pathogenesis by disrupting urothelial barrier integrity through alterations in permeability (14-17). Clinically, IC profoundly compromises physical and psychosocial functioning, underscoring the need for pathophysiology-informed management.

OS, characterized by an imbalance favoring oxidants over antioxidants, results in the accumulation of reactive oxygen and nitrogen species (ROS/RNS) (18). This disrupts redox signaling and control and/or causes molecular damage, and it plays a role in the pathogenesis of various conditions, including cardiovascular, autoimmune, and neoplastic diseases, as well as neurological and chronic pain disorders (19). ROS/RNS are small, highly reactive molecules encompassing both free radicals and non-radical species, examples of which include the superoxide anion, hydroxyl radical, hydrogen peroxide, and nitric oxide. The production of ROS can be induced by both hypoxia and pro-inflammatory stimuli. Sustained exposure of tissues, cells, and biological macromolecules to these elevated levels of oxidants initiates a series of biochemical reactions that cause oxidative damage, thereby compromising normal cellular functions (20). In IC, OS is not merely a bystander but a central modulator influencing all recognized pathogenic pathways (21). The excessive ROS directly impairs the urothelial barrier, facilitating the infiltration of noxious urinary solutes that stimulate sensory nerves, thereby generating pain and urgency. Concurrently, ROS sensitizes bladder afferents by activating transient receptor potential (TRP) ion channels, markedly lowering the pain threshold. A self-perpetuating cycle is established wherein ROS drives pro-inflammatory cytokine release, which in turn amplifies ROS production, perpetuating tissue injury and neuronal sensitization (22). Additionally, OS may disrupt bladder contractility by interfering with efferent signaling and cholinergic transmission. Supporting a pathogenic role, animal studies demonstrate that timely antioxidant administration alleviates oxidative damage, pain, and inflammation in IC models, highlighting its therapeutic relevance (23-25).

Collectively, the evidence underscores OS as a cornerstone in the pathogenesis of IC, driving disease progression through interconnected mechanisms. Targeting this pathway consequently represents a rational and promising therapeutic strategy. Advances in bioinformatics now offer powerful, high-resolution tools for dissecting complex disease etiology and uncovering potential biomarkers. To this end, we integrated systematic bioinformatics analyses with machine learning algorithms to enhance the robustness and discriminatory power of biomarker discovery for IC. This computational approach identified a panel of diagnostic marker genes implicated in IC pathology. The biological and diagnostic relevance of these candidates was further substantiated through validation in independent animal models, strengthening the translational potential of our findings. We present this article in accordance with the TRIPOD and ARRIVE reporting checklists (available at https://tau.amegroups.com/article/view/10.21037/tau-2026-0255/rc).

Methods

Acquisition of dataset

This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. For this study, we obtained two gene expression microarray datasets from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/): GSE11783 and GSE57560 (Table 1). The GSE11783 dataset comprises transcriptomic data from 11 human samples, specifically 10 IC samples and 6 healthy control samples. The GSE57560 dataset includes 13 IC samples and 3 healthy control samples. During the data preprocessing phase, probe IDs were converted to official gene symbols using the platform-specific annotation files, and probes lacking gene symbol annotations were excluded. The two IC datasets (GSE11783 and GSE57560) were merged by common gene symbols, resulting in a combined expression matrix of 32 samples. Batch effect correction was subsequently performed using the ComBat algorithm from the sva R package (version 3.54.0), with the original dataset ID as the batch covariate and parametric prior adjustment enabled. The effectiveness of this batch-effect correction was evaluated by principal component analysis (PCA). Before correction, the first principal component (Dim1) accounted for 58.8% of the total variance and clearly separated samples by dataset of origin, indicating a strong batch effect. After ComBat correction, Dim1 explained only 17.5% of the variance, and the separation by dataset disappeared while the biological distinction between IC and control samples was preserved. Furthermore, OS-associated genes were identified from the GeneCards database and filtered using a relevance score cutoff of >7, yielding a final set of 1,878 genes (table available at https://cdn.amegroups.cn/static/public/tau-2026-0255-1.xlsx). All data were downloaded on May 8, 2025.

Table 1. Characteristics of the GEO datasets used in this study.

Dataset	Platform	Control (n)	IC (n)	Total (n)
GSE11783	GPL570	6	10	16
GSE57560	GPL6244	3	13	16
Combined	–	9	23	32

GEO, Gene Expression Omnibus; IC, interstitial cystitis.

Identification of DEGs

Differential gene expression analysis between IC samples and controls was performed using the “limma” package in R, with significance thresholds set at |log₂ fold change (FC)| >1 and an adjusted P value (adj.P) <0.05. Genes with a logFC >1 were defined as up-regulated, whereas those with a logFC <−1 were defined as down-regulated. The differentially expressed genes (DEGs) were visualized using a volcano plot generated with the “ggplot2” R package. A Venn diagram was constructed using the “Venn” package to identify the genes common to both IC and OS, termed OS-related DEGs (DEOSGs), which were subsequently visualized in a heatmap created with the “pheatmap” R package.

Enrichment analysis

To elucidate the putative biological functions and underlying mechanisms, we performed Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis on the OS-associated DEGs using the “clusterProfiler” package in R. GO terms, including biological process (BP), cellular component (CC), molecular function (MF), and KEGG pathways, with an adjusted P<0.05 for each category were considered statistically significant.

Construction of protein-protein interaction (PPI) network

The STRING database (“STRING v10: PPI networks, integrated over the tree of life”) (version 12.0, https://cn.string-db.org/) was utilized to construct a PPI network. Genes without interactions were removed to enhance network reliability. Subsequently, the Degree algorithm within the CytoHubba plugin in Cytoscape was employed to identify significant gene modules and pinpoint hub genes, which were then visualized.

Identification of diagnostic biomarkers using machine learning

For further refinement of candidate genes, we utilized three distinct machine learning algorithms—least absolute shrinkage and selection operator (LASSO), support vector machine-recursive feature elimination (SVM-RFE), and random forest (RF)—within a nested cross-validation (CV) framework to obtain unbiased performance estimates. Specifically, the dataset was partitioned using an outer 5-fold CV loop for performance evaluation, with an inner 5-fold CV loop embedded within each outer training fold for hyperparameter tuning. For LASSO logistic regression (implemented via the glmnet R package), the regularization parameter λ was tuned over a grid of 100 values; the optimal λ was selected as the value that minimized binomial deviance (lambda.min) in the inner CV (26). For SVM-RFE (implemented via the e1071 and caret R packages), recursive feature elimination was conducted simultaneously with a grid search over the cost parameter C (0.01, 0.1, 1, 10) and the radial basis function kernel parameter σ (0.001, 0.01, 0.1) (27). The feature elimination process was performed independently within each outer fold. For RF (implemented via the randomForest package with 500 decision trees), the mtry parameter (number of variables randomly sampled as candidates at each split) was tuned over a range from 1 to 10 via inner 5-fold CV (28). The final model performance was evaluated on the held-out outer test sets, and the cross-validated metrics—including the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and their 95% confidence intervals (CIs)—were averaged across the five outer folds. The optimal hyperparameters and cross-validated performance of each model are summarized in Table 2. Ultimately, the overlapping genes derived from all three algorithms were defined as the key hub genes.

Table 2. Cross-validated performance and optimal hyperparameters of the three machine learning models.

Algorithm	Optimal hyperparameters	AUC (95% CI)	Sensitivity (95% CI)	Specificity (95% CI)
LASSO	λ=0.0498	0.923 (0.831–1.000)	0.913 (0.818–1.000)	0.667 (0.333–1.000)
SVM-RFE	C=10.00, σ=0.010	0.879 (0.746–1.000)	0.957 (0.867–1.000)	0.333 (0.000–0.667)
RF	mtry =1.2 (mean)	0.915 (0.815–1.000)	0.826 (0.652–0.957)	0.778 (0.444–1.000)

Performance evaluated via nested 5-fold × 5-fold cross-validation. Sensitivity and specificity were calculated at a probability threshold of 0.5. AUC, area under the receiver operating characteristic curve; CI, confidence interval; LASSO, least absolute shrinkage and selection operator; RF, random forest; SVM-RFE, support vector machine-recursive feature elimination.

Nomogram construction and evaluation using receiver operating characteristic (ROC) curve

A nomogram model incorporating these genes was developed using the “rms” R package, providing a tool for clinical diagnostic applications. The predictive performance of individual genes and the composite nomogram model was subsequently assessed using ROC curve analysis, with the AUC calculated via the “pROC” package. Differential expression levels of the hub genes between IC samples and controls were visualized using boxplots created in GraphPad Prism (version 9.0.0).

Immune landscape and correlation analysis with genes

CIBERSORT, an acronym for cell-type identification by estimating relative subsets of RNA transcripts, is a sophisticated bioinformatic tool that employs a support vector regression-based algorithm to deconvolute gene expression profiles of heterogeneous tissues, enabling the inference of relative abundances of constituent cell types. The Spearman’s rank correlation between infiltrating immune cells and diagnostic genes was calculated. To account for multiple testing across the 22 immune cell types, raw P values were adjusted using the Benjamini-Hochberg procedure. Correlations with an adj.P<0.05 were considered statistically significant. The results were visualized as a correlation heatmap generated with the corrplot package in R.

Single-gene gene set enrichment analysis (GSEA)

GSEA for a single gene is performed by assessing the association between the target gene and genes in a dataset, identifying datasets with high correlation to the target gene, which are subsequently subjected to enrichment analysis. In our investigation, samples within the dataset were initially stratified into high-expression and low-expression groups according to the median expression level of an individual gene. Reference gene sets were acquired from the Molecular Signatures Database (MSigDB). We then conducted GSEA employing the “clusterProfiler” R package to delineate the biological pathways implicated by the diagnostic marker genes. The background gene set was c2.cp.kegg_medicus. v2025.1.Hs.symbols. Statistical significance was assessed using the Benjamini-Hochberg false discovery rate (FDR) adjustment; pathways with an adj.P<0.05 were considered significantly enriched.

Evaluation of applicant drugs

In order to investigate the relationships between pharmacological compounds and the critical hub genes, pertinent data were retrieved from the Drug Signatures Database (DSIGDB) (https://tanlab.ucdenver.edu/dsigdb).

Molecular docking

Molecular docking was performed to predict the binding mode between small molecules (ligands) and target proteins (receptors, such as enzymes or transcription factors). A semi-flexible docking strategy was adopted to simulate the formation of stable complexes, which is essential for elucidating mechanisms of action and for the identification of potential lead compounds. Docking calculations were conducted using AutoDock Vina version 1.2.7 (29,30), where simvastatin (Resveratrol, PubChem CID: 54454) was docked with TLR2 (PDB ID: 2Z7X). Protein structures were preprocessed with PyMOL version 2.4 by removing water molecules and non-essential ligands, followed by the addition of hydrogen atoms. The ligand structure was energy-minimized using ChemDraw version 20.0. PDBQT files were generated with AutoDock Tools version 1.5.6, and the docking grid box was defined to cover the entire protein structure, while all other parameters were kept at default values. For each docking run, nine conformations were generated, and the pose with the lowest binding energy and highest clustering frequency was selected as the most probable binding mode. Visualization of docking results was carried out using PyMOL version 2.4 to display ligand-protein binding conformations and molecular interactions. Detailed docking parameters, including grid box center coordinates and dimensions, are provided in Table S1.

Molecular dynamics (MD)

MD simulations were performed using GROMACS 2025.3 (31). The protein was modeled with the Amber99sb-ildn force field, solvated in SPC water, and neutralized with Na⁺ ions. The system was energy-minimized using the steepest descent algorithm, followed by equilibration under NVT and NPT ensembles for 1 ns each (1,000,000 steps, 2 fs time step, coupling constant 0.1 ps). Production MD simulations were then conducted for 100 ns (50,000,000 steps, 2 fs time step). Trajectories were analyzed using GROMACS tools to calculate root mean square deviation (RMSD) and radius of gyration (Rg).

Establishment of animal models for IC

Seven-week-old female Sprague-Dawley (SD) rats were purchased from Shandong Pengyue Laboratory Animal Technology Co., Ltd. The animals were maintained in the designated animal facility at the Affiliated Provincial Hospital of Shandong First Medical University under well-ventilated conditions, a 12-hour light/dark cycle, and an ambient temperature of 20–24 °C. A suitable housing density was maintained, with five rats per cage. All rats had free access to water and food. Rats were randomly allocated into two groups: a normal control group (n=8) and an IC model group (n=8) induced by cyclophosphamide (CYP; Sigma-Aldrich, USA) administered intraperitoneally at 75 mg/kg every three days for a total of three doses. After successful modeling, the rats exhibited symptoms such as frequent urination and bladder pain-like behaviors. Twenty-four hours prior to sacrifice, filling cystometry was performed under isoflurane inhalation anesthesia (induction 4%, maintenance 1.5–2.0% in oxygen). A polyethylene catheter (PE-50) was inserted into the bladder dome through a small abdominal incision and connected via a three-way stopcock to a pressure transducer and a microinjection pump. The bladder was continuously infused with sterile room-temperature saline at a rate of 0.04 mL/min. Intravesical pressure was recorded using Laboratory View 6.0 software (National Instruments, USA). The following urodynamic parameters were analyzed from at least three stable micturition cycles per animal: micturition interval (MI, seconds), voided volume (UV, mL), maximum voiding pressure (P_max, cmH₂O), and residual volume (RV, mL). n=8 per group. At 24 hours following the final treatment, all rats were sacrificed via carbon dioxide asphyxiation. Due to the need to allocate bladder tissue for multiple parallel assays [histology, reverse transcription-quantitative polymerase chain reaction (RT-qPCR), and western blot (WB)], the final number of biological replicates analyzed per assay was as follows: RT-qPCR, n=3 per group; WB, n=3 per group; hematoxylin and eosin (H&E) staining, n=2 per group (representative image). All available tissue was utilized; no data were excluded from analysis. Bladder tissues were harvested immediately upon confirmation of euthanasia. All animal experiments were performed under a project license (No. 2025-051) granted by the Animal Ethics Committee of the Affiliated Provincial Hospital of Shandong First Medical University, in compliance with the national guidelines for the care and use of laboratory animals in China.

RNA isolation and RT-qPCR

Following tissue processing, total RNA was isolated from the tissue samples of both groups using the SteadyPure Universal RNA Extraction Kit (Aikerui Biotech, Changsha, China), adhering strictly to the manufacturer’s instructions. The concentration and purity of the extracted total RNA were assessed using a UV spectrophotometer (NanoDrop2000, Thermo Fisher Scientific). Reverse transcription and RT-qPCR analysis were performed using the Evo M-MLV RT Mix Kit with gDNA Clean for qPCR Ver.2 (Aikerui Biotech, Changsha, China). The reverse transcription reaction conditions included incubation at 37 °C for 15 min, followed by 85 °C for 5 s, and storage at 4 °C. Quantitative real-time PCR was carried out using a LightCycler^® 480 II instrument. The PCR conditions included denaturation at 95 °C for 30 s, followed by 40 cycles of 95 °C for 5 s and 60 °C for 30 s for real-time plate reading. A melt curve analysis was performed subsequently, comprising 95 °C for 15 seconds, 60 °C for 1 minute, and a final ramp to 95 °C for 1 second. Messenger RNA (mRNA) expression levels were normalized using β-actin as an internal reference gene. Relative quantification was calculated using the 2^−ΔΔCq method. All samples were prepared and analyzed in triplicate. Data are expressed as the mean ± standard error of mean (SEM) for subsequent statistical comparisons. Differences were evaluated for statistical significance using the t-test, with a P<0.05 deemed significant. The nucleotide sequences of primers used in this study are listed below: β-actin (forward: 5’-GGAGATTACTGCCCTGGCTCCTA-3’, reverse: 5’-GACTCATCGTACTCCTGCTTGCTG-3’), TLR2 (forward: 5’-GCCCTCAGTCTTGGAGTGTC-3’, reverse: 5’-GCGCCTAAGAGCAGGATCAA-3’), S100A8 (forward: 5’-GGGATGACTTCAGGAAAATGGTC-3’, reverse: 5’-CTGTCTTTATGAGCTGCCACG-3’) (Table S2).

WB analysis

Cryogenically ground bladder tissues were added to protein lysis buffer containing protease and phosphatase inhibitors, incubated on ice for 10 minutes to facilitate lysis, followed by sonication. Subsequently, the lysates were centrifuged at 12,000 rpm for 10 minutes at 4 °C, and the resulting supernatant was collected for protein quantification. Proteins were resolved using gel electrophoresis and subsequently transferred to polyvinylidene fluoride (PVDF) membranes. Membranes were blocked with 5% non-fat dry milk for 2 hours and then probed with specific primary antibodies overnight at 4 °C. Primary antibodies employed were: rabbit anti-S100A8 (15792-1-AP, Proteintech; 1:1,000) and rabbit anti-TLR2 (CY5102, Abways; 1:1,000). Following washes with TBST, membranes were incubated for 1 hour at room temperature with a horseradish peroxidase-conjugated goat anti-rabbit IgG secondary antibody (RGAR001, Proteintech; 1:10,000). Protein bands were visualized using an enhanced chemiluminescence (ECL) substrate (PK10002, Proteintech) with high sensitivity. Semi-quantitative analysis of the WBs was performed using ImageJ and Photoshop software.

Statistical analysis

Effect size and statistical power

Effect sizes are reported as Cohen’s d with 95% CIs, calculated using the pooled standard deviation. Post hoc power analysis was conducted using G*Power 3.1 based on the observed effect sizes and actual sample sizes for each assay (RT-qPCR: n=3 per group; WB: n=3 per group). For RT-qPCR, achieved power was 0.976 (S100A8) and 0.864 (TLR2). For WB, achieved power was 0.570 for both proteins. All effect sizes were large (Cohen’s d>2.3); however, the protein-level comparisons are underpowered and should be interpreted as supportive rather than definitive evidence (Table S3).

General statistical procedures

All statistical analyses were conducted using GraphPad Prism (version 9.0.0) and R (version 4.4.2). For two-group comparisons, the two-tailed independent t-test was used. Significance levels are denoted as follows: *, P<0.05; **, P<0.01; ***, P<0.001.

Results

Identification of DEGs

The overall study workflow is depicted in the flowchart presented in Figure 1. Gene expression levels from the GSE11783 and GSE57560 datasets were normalized, with the effects of this normalization illustrated in Figure 2A,2B. Application of the statistical thresholds (adj.P<0.05 and |log₂ FC| >1) yielded 1,181 DEGs, with 643 up-regulated and 538 down-regulated. The DEGs between the IC group and controls were then graphically represented in a volcano plot (Figure 2C). From the GeneCards database, 1,065 genes associated with OS were identified by applying a relevance score cutoff of >7. The overlap between these two gene lists resulted in 58 DEOSGs (Figure 2D). A heatmap generated with the “pheatmap” R package was used to display the expression patterns of these 58 DEOSGs (Figure 2E).

Figure 1.

Workflow of this study. DEG, Differentially Expressed Gene; GO, Gene Ontology; GSEA, gene set enrichment analysis; KEGG, Kyoto Encyclopedia of Genes and Genomes; LASSO, least absolute shrinkage and selection operator; ROC, receiver operating characteristic; SVM-RFE, support vector machine-recursive feature elimination.

Figure 2.

Identification of DEOSGs. (A,B) PCA plots before and after batch correction. (C) Volcano plot of DEGs between IC and control. (D) Venn diagram of DEOSGs. (E) Heatmap of DEOSGs expression. DEG, Differentially Expressed Gene; DEOSGs, oxidative stress-related differentially expressed genes; Dim1, the first principal component; IC, interstitial cystitis; OS, oxidative stress; PCA, principal component analysis.

GO and KEGG enrichment analysis

GO and KEGG pathway analyses elucidated the biological properties and functional pathways enriched among the DEOSGs. The analysis yielded 489 significantly enriched GO terms and 98 KEGG pathways (table available at https://cdn.amegroups.cn/static/public/tau-2026-0255-2.xlsx; https://cdn.amegroups.cn/static/public/tau-2026-0255-3.xlsx). Among the GO terms, 407 belonged to BP, 45 to MF, and 37 to CC. The top five significantly enriched GO terms for each category are presented in bubble plots (Figure 3A). Within the BP category, the key DEGs showed significant enrichment in processes including leukocyte migration, phagocytosis, superoxide anion generation, superoxide metabolic process, and respiratory burst. For the CC category, the key DEGs were associated with vesicles and oxidase complexes. In the MF subgroup, enrichment was observed for terms related to NADPH oxidase (NOX) activity. KEGG pathway analysis identified significant enrichment for pathways including leukocyte transendothelial migration, neutrophil extracellular trap formation, phagosome, cell adhesion molecules, IL-17 signaling pathway, natural killer cell mediated cytotoxicity, chemokine signaling pathway, Fc gamma R-mediated phagocytosis, NF-kappa B signaling pathway, TNF signaling pathway, Rap1 signaling pathway, ferroptosis, and Fc epsilon RI signaling pathway (Figure 3B).

Figure 3.

Functional enrichment of DEOSGs. (A) Top GO terms in BP, CC, and MF. (B) Top enriched KEGG pathways. BP, biological process; CC, cellular component; DEOSGs, oxidative stress-related differentially expressed genes; GO, Gene Ontology; KEGG, Kyoto Encyclopedia of Genes and Genomes; MF, molecular function.

PPI network construction and hub gene identification

A PPI network based on the DEOSGs was constructed using the STRING database to elucidate their interactions. The generated network comprised 58 nodes and 267 edges (Figure 4A). The top 18 hub genes were then identified and visualized using the Degree algorithm within the Cytoscape software (Figure 4B).

Figure 4.

PPI network and hub gene screening. (A) PPI network of DEOSGs. (B) Top 18 hub genes identified by the degree algorithm. DEOSGs, oxidative stress-related differentially expressed genes; PPI, protein-protein interaction.

Screening of candidate diagnostic biomarkers via machine learning approaches

In order to accurately pinpoint key genes possessing the greatest diagnostic potential, feature selection was performed using three machine learning algorithms—LASSO, SVM-RFE, and RF—to isolate the most biologically relevant markers. LASSO logistic regression analysis selected 5 candidate genes from the hub gene set (Figure 5A), SVM-RFE analysis pinpointed 4 genes (Figure 5B), and RF analysis identified 5 genes (Figure 5C). The intersection of results from all three algorithms, visualized via a Venn diagram, yielded two final consensus genes: S100A8 and TLR2 (Figure 5D). To subsequently evaluate the diagnostic accuracy of S100A8 and TLR2 as biomarkers for IC, a nomogram prediction model was developed (Figure 6A), which utilizes a composite score to estimate the probability of an IC diagnosis. ROC curve analysis based on expression data was employed to assess the diagnostic efficacy of the individual genes and the combined nomogram. The AUC values for both genes and the nomogram model were all greater than 0.9 (Figure 6B). The expression levels of these two genes were compared between IC samples and controls within the dataset, revealing significantly elevated expression in the IC group (Figure 6C). Furthermore, calibration curves showed that the nomogram model predicted IC very well (Figure 7A), and decision curve analysis (DCA) indicated that the model had high clinical application value (Figure 7B). These results demonstrate the strong discriminatory power of these two genes for identifying IC. The nested cross-validated performance of each individual machine learning model is reported in Table 2, with all models achieving AUC values above 0.85, further confirming the robustness of the selected biomarkers. Together, these findings offer a viable strategy for the diagnosis and potential therapeutic targeting of IC patients.

Figure 5.

Feature selection by machine learning. (A) LASSO coefficient profiles. (B) SVM-RFE accuracy over feature subset sizes. (C) RF variable importance. (D) Overlap of genes selected by the three algorithms. CV, cross-validation; LASSO, least absolute shrinkage and selection operator; RF, random forest; SVM-RFE, support vector machine-recursive feature elimination.

Figure 6.

Nomogram and diagnostic performance. (A) Nomogram for predicting IC based on S100A8 and TLR2. (B) ROC curves of S100A8, TLR2, and the nomogram model. (C) Expression levels of S100A8 and TLR2 in the combined dataset (IC vs. Control). ****, P<0.0001. AUC, area under the curve; CI, confidence interval; IC, interstitial cystitis; ROC, receiver operating characteristic.

Figure 7.

Calibration and clinical utility of the nomogram. (A) Calibration curve. (B) Decision curve analysis.

Immune landscape and correlation with diagnostic marker genes

The results from the earlier functional enrichment analysis indicated that immune-related pathways play a pivotal role in the pathogenesis of IC. The relative proportions of 22 immune cell types within each sample are presented as bar plots (Figure 8A). Immune infiltration analysis performed via the CIBERSORT algorithm revealed significant differences between control and IC groups in the abundances of T cells CD8, T cells CD4 memory resting, T cells CD4 memory activated, NK cells activated, Macrophages M0, Mast cells resting, and Neutrophils (Figure 8B). We then conducted a correlation analysis to evaluate the associations between the two genes of interest and the infiltrating immune cell subsets. Following Benjamini-Hochberg correction, both S100A8 and TLR2 remained significantly positively correlated with Macrophages M0 (S100A8: Spearman rho =0.52, adj.P =0.01; TLR2: rho =0.47, adj.P =0.03) and significantly negatively correlated with Mast cells resting (S100A8: rho =−0.48, adj.P =0.02; TLR2: rho =−0.43, adj.P =0.045) (Figure 8C, Table S4).

Figure 8.

Immune cell infiltration and correlation with hub genes. (A) Stacked bar plot of 22 immune cell fractions in each sample. (B) Comparison of immune cell abundances between IC and control groups (*, P<0.05; **, P<0.01; ***, P<0.001). (C) Spearman correlation heatmap between S100A8/TLR2 and 22 immune cell types. Asterisks indicate significant correlations after Benjamini-Hochberg correction (*, adj.P <0.05; **, adj.P <0.01; ***, adj.P <0.001). IC, interstitial cystitis.

Single-GSEA analysis

GSEA with FDR correction (adj.P< 0.05) identified several significantly enriched pathways in the high-expression groups of S100A8 and TLR2, notably the CCR_CXCR_GNB_PI3K_RAC cascade and the PRNP_PI3K_NOX2 signaling axis (Figure 9A,9B). These results indicate that S100A8 and TLR2 are functionally coupled to OS-related signaling.

Figure 9.

Single-gene GSEA of S100A8 and TLR2. (A) Enriched pathways for S100A8. (B) Enriched pathways for TLR2. Enrichment plots showing pathways with FDR q-value <0.05. FDR, false discovery rate; GSEA, gene set enrichment analysis.

Identification of drug candidates

Potential therapeutic compounds targeting the two genes of interest were identified via the DSigDB module in the EnrichR database. Evaluation based on adj.P and combined scores allowed us to pinpoint the most promising drug candidates, which are likely to modulate pathways relevant to IC therapy and merit further investigation. The top 10 candidate drugs are listed in Table 3: simvastatin (CTD 00007319), adapalene (CTD 00002588), trimellitic anhydride (CTD 00000808), alfacalcidol (CTD 00000387), maltotetraose (BOSS), muramyl dipeptide (CTD 00005307), 24939-16-0 (CTD 00001660), celecoxib (MCF7 UP), hexadecanal (BOSS), and imiquimod (BOSS).

Table 3. Top 10 gene-targeted drugs for IC.

Term	adj.P	Combined score	Genes
Simvastatin CTD 00007319	0.03	329,955.2	S100A8; TLR2
ADAPALENE CTD 00002588	0.03	11,068.43	TLR2
TRIMELLITIC ANHYDRIDE CTD 00000808	0.03	9,282.083	TLR2
Alfacalcidol CTD 00000387	0.03	9,282.083	TLR2
Maltotetraose BOSS	0.03	6,159.553	TLR2
Muramyl Dipeptide CTD 00005307	0.03	5,821.717	TLR2
24939-16-0 CTD 00001660	0.03	5,821.717	TLR2
Celecoxib MCF7 UP	0.03	5,239.424	S100A8
Hexadecanal BOSS	0.03	4,348.373	TLR2
IMIQUIMOD BOSS	0.03	4,167.854	TLR2

adj.P, adjusted P value; IC, interstitial cystitis.

Molecular docking and dynamics

Molecular docking analysis revealed a strong binding affinity between simvastatin and TLR2, with a calculated binding energy of −8.643 kcal/mol. The most favorable docking conformation is illustrated in Figure 10A. To further substantiate the docking results, a 100 ns MD simulation was conducted on the simvastatin-TLR2 complex. Analysis of the RMSD trajectory for the simvastatin-TLR2 complex indicated that the protein-ligand association remained stable over the entire simulation course (Figure 10B). The Rg of the complex, reflecting structural compactness, also remained stable during the simulation (Figure 10C). The MD simulations collectively suggest that simvastatin and TLR2 maintain a highly stable binding interaction. Molecular docking of simvastatin with S100A8 resulted in a weak binding affinity (ΔG =−4.16 kcal/mol), suggesting that simvastatin is unlikely to form a stable complex with S100A8. Consequently, further MD simulation and free energy analysis were not performed for this target.

Figure 10.

Molecular docking and molecular dynamics simulation of simvastatin with TLR2. (A) Docking poses of simvastatin with TLR2. (B) RMSD trajectories of the simvastatin-TLR2 complex over 100 ns. (C) Rg of the complex. Simvastatin docking with S100A8 yielded weak binding affinity (ΔG =−4.16 kcal/mol); therefore, further MD simulation was not performed for that target. MD, molecular dynamics; ns, nanosecond; Rg, radius of gyration; RMSD, root mean square deviation.

Independent validation of the identified marker genes

Filling cystometry confirmed pronounced bladder dysfunction in the CYP-induced IC model (n=8 per group). Compared to normal controls, the IC group exhibited a significantly shortened MI (81.8±10.7 vs. 302.9±20.8 s, P<0.001), markedly reduced UV (0.044±0.008 vs. 0.191±0.011 mL, P<0.001), elevated P_max (44.5±0.9 vs. 38.4±1.7 cmH₂O, P<0.001), and increased RV (0.061±0.001 vs. 0.023±0.002 mL, P<0.001) (Figure 11A-11D; Table S5). These urodynamic changes are characteristic of bladder overactivity with impaired emptying, consistent with the clinical manifestation of IC. Representative pressure-time tracings from all animals are shown in Figure S1.

Figure 11.

In vivo validation in the CYP-induced IC rat model. (A-D) Urodynamic parameters measured by filling cystometry (n=8 per group): (A) MI, (B) UV, (C) P_max, (D) RV. (E) Representative H&E staining (×200). (F) RT-qPCR analysis of S100A8 and TLR2 mRNA (n=3 per group). (G,H) WB and densitometric quantification (n=3 per group). Data are mean ± SEM. *, P<0.05; ****, P<0.0001. CYP, cyclophosphamide; H&E, hematoxylin and eosin; IC, interstitial cystitis; MI, micturition interval; mRNA, messenger RNA; P_max, maximum voiding pressure; RT-qPCR, reverse transcription-quantitative polymerase chain reaction; RV, residual volume; SEM, standard error of the mean; UV, voided volume; WB, western blot.

A cyclophosphamide-induced IC rat model was successfully generated. Histopathological examination via H&E staining of representative bladder sections demonstrated that the IC model group exhibited characteristic IC-like pathologies, including marked tissue edema, diffuse inflammatory cell infiltration, and disruption of the bladder urothelium (Figure 11E).

RT-qPCR analysis confirmed that both S100A8 and TLR2 mRNA were substantially upregulated in the IC group compared to controls (S100A8: fold-change ≈31.1, Cohen’s d=4.44, 95% CI: 1.86–7.01, P=0.003; TLR2: fold-change ≈22.6, Cohen’s d=3.38, 95% CI: 1.22–5.53, P=0.01; n=3 per group; Figure 11F). These effect sizes were extremely large, and the achieved statistical power was high (0.976 and 0.864, respectively; Table S3), providing robust evidence for transcriptional upregulation of both genes in the IC bladder.

WB analysis showed a consistent trend at the protein level (S100A8: fold-change ≈1.68, Cohen’s d=2.38, 95% CI: 0.29–4.48, P=0.04; TLR2: fold-change ≈1.46, Cohen’s d=2.31, 95% CI: 0.24–4.37, P=0.047; n=3 per group; Figure 11G,11H). Although the protein-level effect sizes remained large (d>2.3), the small sample size limited statistical power to 57% for both comparisons (Table S3). Consequently, these protein-level results should be regarded as preliminary confirmatory evidence that requires validation in larger, independent cohorts.

Taken together, the convergent upregulation of S100A8 and TLR2 at both mRNA and protein levels, combined with the characteristic histopathological changes and the independent bioinformatic machine learning predictions, provides strong multi-dimensional evidence supporting the involvement of S100A8 and TLR2 in IC pathogenesis.

Discussion

Taken together, by applying a multi-algorithm machine learning framework to transcriptomic data, we have identified and validated S100A8 and TLR2 as robust biomarkers linked to OS in IC. This study not only corroborates their pathogenic involvement but also underscores their functional synergy in propagating inflammation and oxidative damage. The confirmation of their upregulation in vivo solidifies their relevance to IC pathology. Given their critical role, S100A8 and TLR2 emerge as high-priority candidates for the development of novel treatment strategies aimed at controlling OS and interrupting the progression of IC.

IC is a clinical syndrome with an elusive etiology, whose hallmark symptoms include urinary frequency, urgency, and suprapubic/pelvic pain that worsens with bladder filling and improves upon urination (32). IC imposes a substantial burden on quality of life, often co-occurs with anxiety or depression, and is associated with a suicide rate fourfold higher than that of the general population (33). The diagnosis of IC remains challenging due to the current lack of sensitive and specific biomarkers. The diagnosis of IC is confirmed through a process of exclusion and the identification of characteristic IC symptoms. Establishing a clinical diagnosis of IC necessitates a detailed history, physical examination, and laboratory investigations to document its fundamental symptoms and exclude infectious or other mimickers (1). Therefore, advancing the understanding of IC pathogenesis and discovering objective diagnostic tools are pressing priorities (34,35). In this context, the role of OS is well-substantiated (36,37). Studies have reported significantly increased levels of the OS marker 8-isoprostane in both urine and blood samples from IC patients (38-40). Additionally, Ener et al. demonstrated a significantly reduced total antioxidant capacity in the serum of IC patients compared to healthy controls (41). This established link makes the OS pathway a salient focus for the identification of novel diagnostic biomarkers (34,42).

Our integrated analysis of GEO and GeneCards databases identified 58 DEOSGs in IC. Enrichment analysis revealed these genes are significantly involved in immune-inflammatory responses and cell signaling pathways. This suggests that IC pathogenesis may involve a synergistic network where immune-inflammatory processes act as primary instigators, with OS serving as a critical downstream effector. Major cellular sources of ROS include mitochondrial oxidases, uncoupled endothelial nitric oxide synthase (eNOS), and the NOX family (43). The NOX enzyme family is of particular interest due to its dedicated role in generating ROS, specifically superoxide anion (O²⁻) and hydrogen peroxide (H₂O₂) (44). Studies in animal models have demonstrated a critical role for the NOX2 isoform in the pathogenesis of IC (41). Pharmacological inhibition of NOX2 significantly attenuated cyclophosphamide-induced bladder tissue damage, dysfunctional micturition patterns, and pain-related behaviors. Moreover, evidence suggests that elevated peripheral TNF levels can activate the NF-κB pathway in central microglia, triggering the release of neurotoxic mediators including pro-inflammatory cytokines and ROS, thereby inducing widespread OS (45,46). Research has shown that bladder tissues from a protamine sulfate and LPS-induced rat model of IC exhibit substantial upregulation of inflammatory markers (IL-6, TNF, NF-κB) alongside the induction of antioxidant enzymes such as heme oxygenase-1 (HO-1) and NAD(P)H quinone oxidoreductase 1 (NQO1). Given these insights, therapeutic strategies aimed at mitigating OS to disrupt the deleterious feedback loop between inflammation and oxidative damage are of considerable importance and therapeutic potential.

Building on our bioinformatic screening, the consensus of three machine learning algorithms pinpointed S100A8 and TLR2 as diagnostic markers for IC, a finding supported by robust ROC curve analysis and a predictive nomogram. Experimental validation in a CYP-induced rat model confirmed their significant upregulation in IC tissues at both mRNA and protein levels, underscoring their pathogenic relevance. S100A8, a damage-associated molecular pattern (DAMP) predominantly released from myeloid cells, exerts its biological effects largely through dimerization with S100A9 (47-49). This S100A8/A9 complex is a potent driver of inflammation and OS, promoting ROS generation, NLRP3 inflammasome activation, and pro-inflammatory cytokine release. Its extracellular effects are largely mediated via Toll-like receptors, leading to NF-κB and MAPK pathway activation and a cascade of cytokines such as tumor necrosis factor-alpha (TNF-α), interleukin-6 (IL-6), interleukin-8 (IL-8) (50-52). This cascade recruits and activates immune cells, establishing a self-perpetuating inflammatory feedback loop. The sustained inflammation drives key IC pathologies: vascular endothelial growth factor (VEGF)-mediated abnormal neovascularization and increased vascular permeability, and transforming growth factor beta 1 (TGF-β1)-driven fibroblast-to-myofibroblast transdifferentiation leading to bladder fibrosis, reduced compliance, and capacity. Critically, our co-identification of TLR2 alongside S100A8 suggests a novel, less-explored axis. TLR2 is a key innate immune sensor linked to chronic pain and autoimmunity (53,54). Evidence indicates S100A8 can signal through TLR2 (55). Thus, while S100A8/TLR4 interactions are known, the S100A8/TLR2 axis may represent a distinct mechanism contributing to immune dysregulation and pain sensitization in IC, warranting further investigation for its therapeutic potential.

Consistent with established literature, our analysis reaffirms the involvement of diverse immune cells, such as T cells and macrophages, in IC (56,57). Leveraging the CIBERSORT algorithm, we characterized the infiltration landscape of 22 immune cell subsets, revealing distinct patterns in IC compared to controls. Notably, the expression of our identified biomarker genes, S100A8 and TLR2, showed significant correlations with the abundance of specific CD4⁺ T cell subsets—namely, resting and activated memory T cells. CD4⁺ T cells exhibit considerable heterogeneity and are broadly categorized into six principal subsets based on their functional profiles and cytokine secretion: T helper 1 (Th1), Th2, Th17, Th22, regulatory T (Treg), and follicular helper T (Tfh) cells (58). These subsets are crucial for pathogen response, immune regulation, and the maintenance of immune homeostasis (59). This correlation is functionally pertinent, as CD4⁺ T cells are pivotal modulators of the bladder immune niche in IC, comprising a substantial proportion of infiltrating cells. Their dysfunction, characterized by a Th1-skewed response and impaired regulatory capacity, actively perpetuates a chronic inflammatory microenvironment. Thus, our findings link OS-associated genes to specific adaptive immune perturbations, suggesting that S100A8 and TLR2 may influence or reflect the dysregulated CD4⁺ T cell responses that drive IC pathology, highlighting potential nodes for therapeutic intervention (60).

Beyond its conventional use in hyperlipidemia, simvastatin—an HMG-CoA reductase inhibitor—shows promise for repurposing in IC. Preclinical evidence indicates that simvastatin alleviates bladder inflammation and voiding dysfunction in cyclophosphamide-induced cystitis models (61). Notably, its mechanism may involve remodeling the TLR2 interactome, thereby suppressing TLR2-mediated signaling and the downstream production of pro-inflammatory cytokines such as TNF-α, IL-6, and IL-8 (62). This anti-inflammatory action directly targets a core pathogenic pathway in IC. Therefore, simvastatin represents a mechanistically rational candidate for therapeutic exploration in IC. Furthermore, its potential to modulate the TLR2 axis suggests that combinatorial strategies targeting the S100A8/TLR2 pathway with simvastatin could offer novel approaches for disrupting the chronic inflammatory cycle in IC.

The upregulation of S100A8 and TLR2 establishes them as strong candidate biomarkers for the early detection of IC, potentially enabling identification before symptomatic onset to improve diagnostic precision. Quantifying these markers could further guide personalized therapeutic strategies aimed at the core pathogenic pathways of OS and immune dysregulation. Additionally, longitudinal assessment of their levels may provide an objective means to monitor disease activity and treatment response. Building on this, we propose that the S100A8/TLR2 axis itself constitutes a novel therapeutic target. The potential repurposing of simvastatin to modulate this axis offers a promising strategy to disrupt the chronic inflammatory cycle in IC. In summary, our findings advance S100A8 and TLR2 from diagnostic indicators to actionable therapeutic targets, directly addressing a critical unmet need in IC management.

Several considerations merit acknowledgment when interpreting our findings. First, although eight rats were initially allocated to each experimental group, bladder tissue was divided across multiple downstream assays (histology, RT-qPCR, and WB), resulting in small analytic sample sizes for individual molecular analyses (RT-qPCR: n=3 per group; WB: n=3 per group; H&E: n=2 representative section per group). Consequently, the protein-level comparisons were underpowered (power =0.57), and the associated 95% CIs for Cohen’s d were wide. Although the observed effect sizes were consistently large across all endpoints (all d>2.3), the WB results should be considered preliminary and require confirmation in larger, independent cohorts with pre-specified sample sizes. Second, the bioinformatic analyses are based on retrospective public datasets, and technical heterogeneity in their generation could influence the results. The modest total sample size of the merged GEO dataset (n=32) also limits the precision of the machine learning performance estimates, as reflected in the wide CIs for specificity in the nested CV. Third, while we establish a robust association between S100A8/TLR2 and IC, the precise molecular mechanisms by which they influence OS and disease pathology remain to be fully mapped. Further mechanistic studies are needed to delineate the specific signaling cascades involved. Fourth, although supported by our animal model, the translational relevance of these biomarkers for diagnosis or therapy in human patients requires prospective clinical confirmation in well-characterized IC patient cohorts. Fifth, although the present study includes comprehensive urodynamic confirmation of bladder dysfunction (MI, UV, P_max, RV; n=8 per group), behavioral assessments such as spontaneous micturition frequency in home cages and direct pain-related behavioral tests (e.g., von Frey-evoked visceral sensitivity) were not performed. Incorporating these complementary behavioral endpoints in future studies would further strengthen the link between molecular biomarkers and specific aspects of bladder pain and dysfunction. Notwithstanding these constraints, this study provides substantive insights into IC pathogenesis, nominates S100A8 and TLR2 as concrete molecular targets, and establishes a comprehensive analytical framework for subsequent validation and mechanistic exploration, contributing to the long-term goal of refining IC diagnosis and management.

Conclusions

Through integrated bioinformatics and experimental validation, we conclusively identify S100A8 and TLR2 as co-diagnostic biomarkers and functionally linked therapeutic targets in IC. Their significant co-upregulation is conserved across patient datasets and animal models. We propose that the S100A8/TLR2 axis represents a promising therapeutic target, with simvastatin highlighted as a potential repurposing candidate drug. These findings provide new directions for the precise diagnosis and targeted therapy of IC, and subsequent research warrants in-depth exploration of their clinical translational potential.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. All animal experiments were performed under a project license (No. 2025-051) granted by the Animal Ethics Committee of the Affiliated Provincial Hospital of Shandong First Medical University, in compliance with the national guidelines for the care and use of laboratory animals in China.

Data Sharing Statement

Available at https://tau.amegroups.com/article/view/10.21037/tau-2026-0255/dss

tau-15-06-204-dss.pdf ^{(91.3KB, pdf)}

DOI: 10.21037/tau-2026-0255