Tear MicroRNA-Based Molecular Signature for Keratoconus Stratification: A Cross-Cohort Integration Study Authors: Shao-Hsuan Chang, Kuo-Hsuan Hung, Hsieh-Fu Tsai, Chung-Pei Ma, Hao-Chang Chiang, Lung-Kun Yeh Published: 2026 ABSTRACT Keratoconus (KC) is a progressive corneal ectatic disorder characterized by structural remodeling and immune-associated molecular alterations. However, whether tear-derived microRNAs reflect corneal molecular changes remains unclear. We analyzed three publicly available corneal transcriptomic datasets (N = 112) to identify immune-associated gene expression patterns using differential expression analysis and Gene Set Enrichment Analysis. In a prospective tear cohort (N = 20), tear miRNA profiling was performed, and candidate miRNAs were selected using LASSO regression. miRNA patterns were integrated with immune-related transcriptomic alterations, and an exploratory molecular-structural index was constructed using standardized hsa-miR-326 expression and keratometry (K2). We identified immune-associated transcriptional alterations in KC corneas and derived a five-miRNA tear signature. The signature showed discriminatory performance in an independent corneal epithelial miRNA dataset (AUC = 0.89, 95% CI: 0.69-1.00; N = 16) and was further evaluated in an additional tear cohort using RT-qPCR (N = 52). Among these candidates, hsa-miR-326 showed consistent downregulation across datasets. Network analysis identified associations between candidate miRNAs and immune-related genes, including IL1B, CD4, CXCL10, and TGFB1. The exploratory molecular-structural index demonstrated an inverse association with keratometric steepening (Spearman ρ = -0.86, p < 0.001). These findings suggest that tear-derived miRNAs may provide complementary molecular information associated with immune-related corneal alterations and warrant further investigation in KC characterization. 1. Introduction Keratoconus (KC; also abbreviated as KCN) is a progressive corneal disorder characterized by stromal thinning, protrusion, scarring, and varying degrees of visual distortion. Traditionally considered non-inflammatory, KC is a complex multifactorial disease influenced by genetic predisposition, environmental stimuli, oxidative stress, and chronic eye rubbing. 1 Altered expression of matrix metalloproteinases (MMPs) and pro-inflammatory mediators has also been reported.2,3 Histologically, KC involves disorganized collagen, reduced keratocyte density, and altered proteoglycan composition, leading to impaired tissue-level biomechanical properties and extracellular matrix (ECM) remodeling.4-6 Tear fluid is a promising source of non-invasive biomarkers, with specific protein and nucleic acid signatures reflecting disease status.7-9 Chronic eye rubbing may further exacerbate disease progression, potentially impacting tear composition.10-12 Despite advances in imaging-based diagnosis, current clinical parameters (e.g., keratometry) primarily reflect established structural changes and provide limited ability to identify early molecular-level of disease progression. Recent advances in multi-omics approaches have enabled the discovery of novel biomarkers and potential therapeutic targets in KC. 13 By comparing gene expression profiles, proteomic signatures, and metabolite levels between healthy and diseased corneas, key molecular features associated with disease onset and progression have been identified.14,15 MicroRNAs (miRNAs) have attracted particular attention as post-transcriptional regulators in corneal wound healing, inflammation, and neovascularization.14,16 Several studies have reported differential miRNA expression in KC patients. For instance, hsa-miR-143-3p, hsa-miR-182-5p, and hsa-miR-92a-3p are upregulated, while alterations in hsa-miR-151a-3p and hsa-miR-194-5p may distinguish disease stages.17,18 Nowak-Malczewska et al identified KC upregulated hsa-miR-184, hsa-miR-200a, hsa-miR-6728, hsa-miR-429, and hsa-miR-1299, alongside downregulated hsa-miR-6081, hsa-miR-27b, hsa-miR-23b, and hsa-miR-23a, suggesting their contribution to ECM disruption. 19 Given their stability in bodily fluids miRNAs are excellent biomarker candidates.20,21 However, cross-cohort variability remains substantial due to patient heterogeneity and differences in study design. Furthermore, limited integration of tear-derived miRNAs with corneal transcriptomic alterations has hindered the biological contextualization of candidate molecular signals. More importantly, the biological linkage between tear-derived molecular signals and corneal structural alterations remains poorly defined, limiting their translational interpretation in stratification. To address this gap, this study performed cross-cohort integration of three independent corneal transcriptomic datasets to define immune-related gene signatures while minimizing biological heterogeneity through consensus clustering. We then applied machine learning to prospectively collected tear miRNA profiles to identify candidate molecular features. Finally, tear-derived miRNA signatures were integrated with corneal structural parameters to explore molecular-structural coupling in KC, including an exploratory composite index linking miRNA expression with keratometric features. 2. Methods 2.1. Transcriptomic Dataset Collection The datasets were obtained from the NCBI database, including the Gene Expression Omnibus (GEO) and BioProject repositories (https://www.ncbi.nlm.nih.gov/geo/ and https://www.ncbi.nlm.nih.gov/bioproject/; accessed on 1st September 2024). Three independent RNA-seq datasets were collected for integrated transcriptomic analysis, including one GEO accession GSE77938, and three corresponding NCBI BioProject accessions PRJNA312169, PRJNA636666, and PRJNA1184491. Figure 1 illustrated the overall workflow of the analysis pipeline. RNA sequencing (RNA-Seq) data were generated using different Illumina platforms across the independent datasets. GSE77938 (BioProject: PRJNA312169) was generated using the Illumina HiSeq 1500 platform (Illumina, San Diego, CA, USA).19,22 The PRJNA1184491 dataset was generated using the Illumina HiSeq 2500 platform by the Federal Research and Clinical Center for Physical-Chemical Medicine (https://rcpcm.ru/), while the PRJNA636666 was also generated using the Illumina HiSeq 2500 platform. 23 All datasets employed paired-end sequencing strategies as described in their respective original studies. The transcriptomic samples were derived from corneas tissues, followed by total RNA extraction and sequencing. Detailed information regarding dataset accession numbers, sample numbers, sequencing platforms, and sequencing strategies was provided in Supplemental Table 1.Figure 1.The workflow of this study The workflow of this study 2.2. Data Stratification and Purification via Consensus K-Means Clustering To evaluate potential dataset-associated variation, principal component analysis (PCA) was first performed on the complete integrated cohort of 112 samples. Batch correction was performed using removeBatchEffect function from the limma package with dataset identity specified as the batch variable and disease status retained in the design matrix. 24 The resulting batch-adjusted expression matrix was subsequently used for consensus K-means clustering to identify transcriptionally homogeneous sample subgroups. The optimal number of clusters was determined using the elbow method. This stratification aimed to isolate representative keratoconus (KC) and healthy control samples with minimal intra-cluster variability for robust downstream analysis. The full dataset was subsampled at 80% per iteration (sample_rate = 0.8), and clustering was repeated with K ranging from 1 to 5 (k ∈ [1, 5]). Centroid initialization was repeated 100 times (n_init = 100), and a distinct random seed (random_state = i) was used for each iteration to enhance cluster stability and diversity. Based on cluster consistency and biological plausibility, we identified and retained two core clusters representing the most molecularly distinct KC and control subgroups, while excluding two intermediate, mixed-composition clusters. This purification process refined the cohort to 63 samples (18 controls and 45 KC). 2.3. Data Processing and Identifying Differentially Expressed Genes (DEGs) Differential gene expression (DGE) analysis was conducted to identify genes with significant expression differences between the refined KC group (N=45) and the healthy control group (N=18) identified through consensus clustering. The analysis was performed using the DESeq2 package (version 1.38.3, Bioconductor, USA), with filtering criteria set to adjusted p-value (Benjamini-Hochberg corrected p-value) < 0.05 and a |log2FoldChange| > 2. These thresholds were applied to ensure that the selected genes exhibited substantial expression differences and biological relevance. FDR is an important statistical measure used to control the proportion of false discoveries among identified significant genes, thereby reducing false-positive findings in large-scale transcriptomic analyses. 25 Raw count data were analyzed using the DESeq2 standard workflow, including size factor normalization based on the median-of-ratios method and dispersion estimation for differential expression analysis. Variance-stabilizing transformation (VST) was applied to normalized counts for visualization. 2.4. Gene Function Enrichment Analysis We employed gene set enrichment analysis (GSEA) to investigate differences in gene expression between keratoconus and normal phenotypes. Specifically, we focused on genes associated with Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways to identify biological processes potentially involved in disease development. Through this step, we identified gene sets related to KEGG pathways that may be involved in the development of KC.26-28 GSEA ranks genes based on their correlation with the phenotype and calculates an Enrichment Score (ES) for each gene set. A higher ES indicates a stronger association between the gene set and the phenotype. For this analysis, we selected the c2.cp.kegg.v11.0.symbols gene set from the Molecular Signatures Database (MSigDB), which contains genes linked to KEGG pathways. To improve the accuracy of the analysis, we applied a False Discovery Rate (FDR) threshold of < 0.05. Finally, we conducted GSEA using the R programming language to explore significantly enriched gene sets and their associations with keratoconus and normal corneas. Statistical significance was assessed with a p-value < 0.05, and enrichment curves were plotted to further validate the gene sets that were significantly enriched in the two phenotypes. 2.5. Tear Sample Collection The study protocol was conducted in accordance with the Declaration of Helsinki and was approved by the Institutional Review Board (IRB) of Linkou Chang Gung Memorial Hospital (IRB No. 202401867B0; approval date: 01/08/25). KC patients were clinically diagnosed by experienced corneal specialists based on slit-lamp examination and corneal tomography. Healthy controls had no history of ocular disease, ocular trauma, or ocular surgery and had a best-corrected visual acuity of at least 20/25. Patients with severe systemic diseases, previous ocular surgery or trauma, or conditions that could compromise tear sample quality were excluded. For the ocular surface disease (OSD) group, participants with recent topical corticosteroid or immunomodulatory treatment were excluded to minimize medication-related effects on tear biomarkers. To ensure the reliability and integrity of the transcriptomic data, samples were processed following standardized quality control (QC) protocols. Participants with insufficient tear volume (Schirmer test <10 mm) or those with inadequate library yield for high-throughput sequencing were excluded from the final analysis. Consequently, a refined cohort of 20 participants (10 KC and 10 controls) was analyzed via miRNA-sequencing. Clinical characteristics of the KC participants and keratoconus severity grading of this cohort was listed in Supplemental Table 2. Clinical parameters, including central corneal thickness (CCT), refractive error, keratometry (Avg K), and tear production (Schirmer test), were objectively measured using a Dual-Scheimpflug Analyzer (Galilei G4; Ziemer Ophthalmic Systems AG, Switzerland). KC severity was staged according to the topography-based ABCD grading system to ensure a well-characterized clinical cohort. Subsequently, an additional, independent validation cohort of 52 distinct participants (20 KC, 20 controls, 8 dry eye disease, 4 cataract) was clinically diagnosed and staged by experienced ophthalmologists. This cohort encompassed a representative spectrum of disease severity, ranging from mild-to-moderate to severe KC. Tear fluid was obtained using Schirmer’s test strips under standardized conditions, which were stored at -80°C after collection until further processing. 2.6. miRNA Extraction and Sequencing Total RNA, including small RNAs, were extracted from the tear samples using the miRNeasy Mini Kit (QIAGEN, USA, No. 217004) according to the manufacturer’s protocol. MiRNA-seq libraries were size-selected (∼170-200 bp) to enrich for adapter-ligated miRNAs and validated for quality using the Qsep400 system (Bioptic Inc., Taiwan) and Qubit 2.0 Fluorometer (Thermo Fisher Scientific, USA). Raw sequencing reads were processed using the QIAseq miRNA Analysis workflow (QIAGEN) according to the manufacturer’s recommended pipeline. The workflow included adapter trimming, unique molecular identifier (UMI)-based read collapsing, and miRNA quantification. miRNAs were annotated against miRBase release v22. UMI length was set to 12 nucleotides, and mature miRNA sequences between 18-25 nt were quantified. Sequence matching allowed up to two mismatches, and length-based annotation of isomiRs was enabled. Homo sapiens was specified as the prioritized reference species. Sequencing was performed on an Illumina NovaSeq platform to generate 75 bp single-end reads, enabling comprehensive profiling of expressed miRNAs. Sequencing and annotation metrics were obtained from the QIAseq miRNA Analysis report, including total read counts, UMI statistics, annotated miRNA reads, and ambiguous annotation rates for each sample. Across the 20 tear miRNA-seq libraries, the median sequencing depth was 4.30 million reads per sample (range, 3.13-7.92 million reads), and the median proportion of reads annotated to miRNA records was 4.15% (range, 1.32-11.86%). Sequencing quality was evaluated using the QIAseq analysis report, including total read counts, UMI statistics, and miRNA annotation summaries for each sample. Detailed sample-level quality metrics were summarized in Supplemental Table 3. 2.7. miRNA Feature Selection and miRNA-mRNA Network Construction To explore the regulatory interactions between miRNAs and mRNAs associated with KC, we integrated an external miRNA dataset with our transcriptomic data. Differential expression analysis was first performed using the limma package. Feature selection was then carried out using a LASSO logistic regression model implemented in the glmnet package. The LASSO model was constructed using miRNA expression profiles from the discovery tear sequencing cohort only. Clinical covariates were not incorporated into the feature selection process, as the objective was to identify discriminatory miRNA features based on molecular expression patterns rather than to develop a clinically adjusted prediction model. Ten-fold cross-validation was performed to determine the optimal regularization parameter. The parameter λ corresponding to the minimum cross-validation error (λmin = 0.0049) was selected, and miRNAs with non-zero coefficients were retained as candidate biomarkers. Candidate miRNAs identified by LASSO regression in the discovery tear sequencing cohort were subsequently used to construct a logistic regression classifier. The classifier was trained using the discovery tear cohort and evaluated on an independent external corneal epithelial miRNA dataset (PRJNA842431) without additional feature selection or model retraining. Model performance was assessed using receiver operating characteristic (ROC) curve analysis and the area under the curve (AUC). To elucidate the post-transcriptional regulatory landscape, a miRNA-mRNA interaction network was constructed by integrating tear-derived miRNA signatures with corneal stromal transcriptomic alterations. Potential targets of the validated miRNAs were first retrieved from the miRDB database using the database-provided target prediction results without additional filtering. These predicted targets were then intersected with the 174 core mRNAs previously identified through GSEA. Genes shared between miRNA target predictions and the stromal mRNA signature were retained to construct the integrated miRNA-mRNA regulatory network, representing potential regulatory interactions between tear miRNAs and corneal stromal gene alterations. 2.8. Protein-Protein Interaction (PPI) Network and Pathway Analysis Genes within the integrated miRNA-mRNA regulatory network were further analyzed using Protein-Protein Interaction (PPI) network analysis. 29 The PPI network was constructed by mapping candidate genes to the STRING database (https://string-db.org/ (accessed on 18th Jan 2026), version 12.0). The default medium confidence interaction score (combined score ≥0.4) was applied for network construction. The resulting interaction network was analyzed using Cytoscape (version 3.10.2). Hub genes were identified using the cytoHubba plugin based on Degree centrality, which evaluates the number of direct interactions associated with each node. The top-ranked genes according to degree centrality were selected as hub genes for subsequent functional interpretation. Gene Ontology (GO) enrichment analysis was performed for the identified hub genes, with statistical significance defined as p-value < 0.05. 2.9. Reverse Transcription-Quantitative Polymerase Chain Reaction (RT-qPCR) To validate the expression patterns derived from the sequencing data and network analysis, we subsequently enrolled an independent validation cohort consisting of 52 tear samples, including 20 healthy controls and 20 patients with keratoconus (KC), 4 cataract patients, and 8 dry eye disease (DED) patients, which were distinct from the discovery cohort. This cohort was used for RT-qPCR validation of candidate miRNAs and evaluation of associations between miRNA expression and clinical parameters. Clinical metadata including age, sex, contact lens use, eye rubbing history, atopy/allergy status, and ocular surface-related parameters of the KC participants were collected and summarized in Supplemental Table 4. hsa-miR-326 (5′-CCU​CUG​GGC​CCU​UCC​UCC​AG-3′) was selected for further experimental verification via RT-qPCR, given its prominent position as the central regulatory hub with the highest connectivity in the miRNA-mRNA network. A miRNA-specific stem-loop reverse transcription strategy was applied for hsa-miR-326 detection. The stem-loop RT primer was designed according to the mature miRNA sequence, consisting of a universal stem-loop sequence fused with a target-specific complementary region. (5′-CTC​AAC​TGG​TGT​CGT​GGA​GTC​GGC​AAT​TCA​GTT​GAG​CTG​GAG​GAA​G-3′). For reverse transcription (RT), 1 µL of the extracted miRNA was reverse-transcribed into cDNA using specific RT-primers. Quantitative PCR was performed using the SYBR Green PCR Master Mix (Thermo Fisher Scientific, USA, No. A66732)) with the following specific primers: miRNA-forward (5′-CGG​CGG​CCT​CTG​GGC​CCT​TCC​T-3′) and universal reverse primer (5′-CTG​GTG​TCG​TGG​AGT​CGG​CAA​TTC-3′). The forward primer contained a GC-rich 5′ extension as part of the target-specific miRNA primer design strategy. PCR amplification was conducted in a total reaction volume of 20 µL, containing the cDNA template, 0.5 µM of each primer, and SYBR Green Master Mix. The expected amplicon size was approximately 57 bp. The expression levels of hsa-miR-326 were normalized to the small nuclear RNA U6 as an internal control. Relative quantification was calculated using the 2-ΔΔCt method, with the healthy control group serving as the calibrator. Statistical significance was determined with a p-value < 0.05 considered significant. 2.10. Statistical Analysis This study employed R software 4.5.2 (R Foundation for Statistical Computing, Vienna, Austria) and Python 3.10.11 for data processing, statistical analysis, and result visualization. Statistical significance was defined as a p-value < 0.05 and was used to evaluate all comparisons and analyses. The data processing steps included imputation of missing values, data filtering, standardization, and transformation. In R, differential expression analysis of tear miRNA-seq data was performed using the limma (v3.66.0) and edgeR (v4.8.2) packages, while differential gene expression analysis of the integrated mRNA datasets was conducted using DESeq2 (v1.38.3). Batch-effect correction was performed using the removeBatchEffect function implemented in limma. Gene set enrichment analysis (GSEA) was performed to analyze significant gene sets related to KEGG pathways and visualize the results. Feature selection was performed using LASSO regression implemented in the glmnet package (v4.1.10). In Python, data processing and additional statistical analyses were conducted using the pandas (v2.2.2) and scikit-learn (v1.6.1). Data visualization was performed using relevant R packages, including ggplot2 (v4.0.2) and ggrepel (v0.9.7). Clinical parameter analyses in the RT-qPCR validation cohort were performed using the Mann-Whitney U test for comparisons between binary clinical groups and Spearman rank correlation analysis for associations between miRNA expression and continuous corneal parameters. 3. Results 3.1. Clustering Analysis and Selection of Representative Groups Raw gene expression data for keratoconus were obtained from three independent RNA-seq datasets (PRJNA312169, PRJNA636666, and PRJNA1184491) downloaded from the GEO databases. After validating clustering within each dataset, we combined the three datasets into a pooled cohort for integrated analysis. Consensus K-means clustering was then applied to the integrated dataset to assess heterogeneity across samples. Using the elbow method to determine the optimal number of clusters, we identified four distinct subgroups. The distribution of samples among the identified clusters was shown in Figure 2A, while the elbow method used for determining the optimal cluster number was presented in Figure 2B. Each cluster’s composition was subsequently examined to identify biologically representative groups for downstream analysis. Among the four clusters, two showed clear biological relevance: one cluster consisted predominantly of control samples (94.7%; KC: 1; CON: 18), while another was enriched for keratoconus (KC) samples (80.3%; KC: 45; CON: 11). The remaining two clusters appeared to be mixed or heterogeneous groups without clear disease/control predominance. For downstream network analysis, we retained only the two dominant clusters, resulting in 45 KC and 18 control samples (N=63). This selection ensured that the analysis focused on representative samples with low intra-cluster heterogeneity, minimizing the influence of mixed or atypical samples. The sample identifiers included in the final 63-sample cohort were provided in Supplemental Table 5.Figure 2.Consensus clustering, batch-effect assessment, sample purification, and pathway enrichment analysis of integrated transcriptomic datasets. (A) Consensus matrix heatmaps for k = 2 to 5 generated from consensus clustering of the transcriptomic data. (B) Elbow analysis for k-means clustering selection. The plot represents the average sum of squared errors (SSE) across k = 1 to 5, with error bars representing the standard deviation from 100 independent clustering runs. (C) Principal component analysis (PCA) of the integrated cohort (N = 112) before batch-effect correction, colored according to dataset origin. (D) PCA of the integrated cohort (N = 112) after batch-effect correction, demonstrating reduced dataset-associated variation. (E) PCA of the retained representative cohort (N = 63) after k-means-based sample purification, showing improved separation between keratoconus and control samples. (F) GSEA identifying signaling pathways significantly enriched in the control group. (G) GSEA identifying signaling pathways significantly enriched in the KC group. Samples are colored by dataset origin: Dataset A (PRJNA312169), dataset B (PRJNA636666), and dataset C (PRJNA1184491) Consensus clustering, batch-effect assessment, sample purification, and pathway enrichment analysis of integrated transcriptomic datasets. (A) Consensus matrix heatmaps for k = 2 to 5 generated from consensus clustering of the transcriptomic data. (B) Elbow analysis for k-means clustering selection. The plot represents the average sum of squared errors (SSE) across k = 1 to 5, with error bars representing the standard deviation from 100 independent clustering runs. (C) Principal component analysis (PCA) of the integrated cohort (N = 112) before batch-effect correction, colored according to dataset origin. (D) PCA of the integrated cohort (N = 112) after batch-effect correction, demonstrating reduced dataset-associated variation. (E) PCA of the retained representative cohort (N = 63) after k-means-based sample purification, showing improved separation between keratoconus and control samples. (F) GSEA identifying signaling pathways significantly enriched in the control group. (G) GSEA identifying signaling pathways significantly enriched in the KC group. Samples are colored by dataset origin: Dataset A (PRJNA312169), dataset B (PRJNA636666), and dataset C (PRJNA1184491) To evaluate the impact of dataset integration and sample purification, PCA was performed before and after batch-effect correction. Following correction, dataset-associated variation was reduced, although residual heterogeneity remained among the integrated cohorts (Figure 2C and D). Consensus K-means clustering was subsequently applied using the batch-adjusted expression matrix to identify transcriptionally coherent sample subgroups. After exclusion of mixed-composition clusters, the retained 63 samples demonstrated improved separation in PCA space, with cumulative variance explained by PC1 and PC2 increasing from 42.1% to 47.7% (Figure 2E). These findings suggested that consensus clustering reduced residual heterogeneity and enriched biologically representative KC and control populations for downstream analysis. 3.2. Differential Expression Analysis Between KC and Control Groups Following differential gene expression (DGE) analysis, gene set enrichment analysis (GSEA) was performed to explore the biological significance of the transcriptional differences between KC and control samples. A total of 5598 genes were identified as significantly differentially expressed. GSEA subsequently identified several metabolic pathways that were significantly enriched between the KC and control groups (Figure 2F and G). These included steroid hormone biosynthesis, mucin-type O-glycan biosynthesis, and retinol metabolism (Table 1). The involvement of these pathways suggested altered epithelial regulation and impaired ocular surface homeostasis in KC. Immune-related pathways, including chemokine signaling, TNF signaling, and IL-17 signaling, showed relatively higher enrichment scores in control samples under the current analytical framework. Rather than indicating increased inflammatory activity in healthy corneas, this pattern may reflect disruption or dysregulation of immune homeostasis and stress-response signaling in KC tissues. Together, these findings suggest that KC is characterized by coordinated alterations in epithelial, metabolic, and immune-associated signaling pathways, rather than representing a purely inflammatory disease process.Table 1.KEGG Pathway Enrichment Analysis Results for Control and Keratoconus GroupsCategoryIDKEGG pathwayNESGene %FDR q-valueControl (CON)hsa04657IL-17 signaling pathway-2.48235321.51E-13hsa04061Viral protein interaction with cytokine and cytokine receptor-2.34198414.68E-11hsa05134Legionellosis-2.16949234.34E-06hsa04062Chemokine signaling pathway-2.13966458.87E-09hsa04620Toll-like receptor signaling pathway-2.13757473.26E-06hsa05323Rheumatoid arthritis-2.09899432.04E-06hsa04668TNF signaling pathway-2.07242316.41E-06hsa05133Pertussis-2.06949321.90E-05hsa04936Alcoholic liver disease-2.04956451.09E-05hsa04060Cytokine-cytokine receptor interaction-2.04671472.50E-10Keratoconus (KC)hsa00140Steroid hormone biosynthesis2.223139507.96E-04hsa00512Mucin type O-glycan biosynthesis1.888838508.39E-03hsa00830Retinol metabolism1.654084343.54E-02hsa05204Chemical carcinogenesis - DNA adducts1.712496423.61E-02hsa00982Drug metabolism - cytochrome P4501.756272384.20E-02 Table 1.KEGG Pathway Enrichment Analysis Results for Control and Keratoconus GroupsCategoryIDKEGG pathwayNESGene %FDR q-valueControl (CON)hsa04657IL-17 signaling pathway-2.48235321.51E-13hsa04061Viral protein interaction with cytokine and cytokine receptor-2.34198414.68E-11hsa05134Legionellosis-2.16949234.34E-06hsa04062Chemokine signaling pathway-2.13966458.87E-09hsa04620Toll-like receptor signaling pathway-2.13757473.26E-06hsa05323Rheumatoid arthritis-2.09899432.04E-06hsa04668TNF signaling pathway-2.07242316.41E-06hsa05133Pertussis-2.06949321.90E-05hsa04936Alcoholic liver disease-2.04956451.09E-05hsa04060Cytokine-cytokine receptor interaction-2.04671472.50E-10Keratoconus (KC)hsa00140Steroid hormone biosynthesis2.223139507.96E-04hsa00512Mucin type O-glycan biosynthesis1.888838508.39E-03hsa00830Retinol metabolism1.654084343.54E-02hsa05204Chemical carcinogenesis - DNA adducts1.712496423.61E-02hsa00982Drug metabolism - cytochrome P4501.756272384.20E-02 KEGG Pathway Enrichment Analysis Results for Control and Keratoconus Groups 3.3. Identification of Candidate miRNA Signatures To identify potential tear miRNA for KC, we performed a systematic transcriptomic analysis. The initial screening via volcano plot (Figure 3A) identified a significant number of differentially expressed miRNAs. Notably, hsa-miR-1260a emerged as the most significantly down-regulated marker, while several others, such as hsa-miR-222-3p and hsa-miR-221-5p, showed pronounced up-regulation in KC samples. Following this, Figure 3B displayed the top 30 significant miRNAs through a hierarchical clustering heatmap. The expression patterns of these candidates successfully partitioned the cohort into two primary clusters. To further refine the diagnostic signature, we employed LASSO regression analysis, which narrowed the selection to 7 LASSO-derived candidate miRNAs (Figure 3C). The boxplots in Figure 3D detailed the expression levels and statistical significance (p-values) of these biomarkers, where hsa-miR-1260a (p<0.001), hsa-miR-320b (p<0.001), hsa-miR-326 (p<0.001), hsa-miR-30b-3p (p<0.001), hsa-miR-1291 (p<0.001), and hsa-miR-1290 (p=0.002) exhibited a consistent and significant decrease in expression within the KC group compared to the control, while hsa-miR-222-3p (p<0.001) showed a distinct increase in KC samples.Figure 3.Identification of differentially expressed miRNAs and construction of the miRNA-mRNA regulatory network in KC. (A) Volcano plot of miRNA expression profiles comparing KC and controls. (B) Heatmap depicting the expression patterns of the top 30 miRNAs across samples. (C) LASSO regression analysis yielding a sparse model of 7 tear-derived miRNAs. (D) Boxplots showing individual expression levels of the 7 tear-derived miRNAs. Significant differences were indicated for each miRNA. (E) Regulatory network between 5 candidate miRNAs and their interactions with 10 identified hub genes. (F) PPI network of 10 hub genes with functional enrichment categories, including cytokine-mediated signaling pathways, response to stress, and cell migration Identification of differentially expressed miRNAs and construction of the miRNA-mRNA regulatory network in KC. (A) Volcano plot of miRNA expression profiles comparing KC and controls. (B) Heatmap depicting the expression patterns of the top 30 miRNAs across samples. (C) LASSO regression analysis yielding a sparse model of 7 tear-derived miRNAs. (D) Boxplots showing individual expression levels of the 7 tear-derived miRNAs. Significant differences were indicated for each miRNA. (E) Regulatory network between 5 candidate miRNAs and their interactions with 10 identified hub genes. (F) PPI network of 10 hub genes with functional enrichment categories, including cytokine-mediated signaling pathways, response to stress, and cell migration 3.4. Integration of miRNA-mRNA Regulatory Network and Functional Characterization of Feature Genes We identified the transcriptomic signature of keratoconus from multi-cohort RNA-seq data derived from corneal stroma tissues, resulting in the selection of 174 differentially expressed genes guided by KEGG pathway analysis. To further elucidate the regulatory mechanisms associated with the disease, we integrated these stroma-derived signatures with the 7 candidate miRNAs identified from tear fluid (hsa-miR-1260a, hsa-miR-222-3p, hsa-miR-30b-3p, hsa-miR-320b, hsa-miR-326, hsa-miR-1290, and hsa-miR-1291). This integrative analysis revealed 33 target genes that overlapped with the 174 stroma-derived signatures, suggesting a potential miRNA-mRNA regulatory cross-talk in the corneal stroma (Table 2). Among these targets, the top 10 hub genes were identified within the PPI module based on their degree centrality and their direct involvement in primary biological pathways associated with corneal remodeling (Table 3). Specifically, IL1B, CD4, CXCL10, and TGFB1 emerged as the core hub genes.Table 2.Integrated Regulatory Pairs of Identified Tear miRNAs and Their Corresponding Differentially Expressed Genes in Corneal Stromahsa-miR-326hsa-miR-30b-3phsa-miR-320bhsa-miR-1291hsa-miR-222-3phsa-miR-1290hsa-miR-1260a S100A7A SULT1E1 CREB5 (3)IL1B (2) GALNT18 MAP2K7 NOS1 (2) IL10RA CYP1A2 (2)CYP1A2 (2) CHRFAM7A SOCS3 GALNT15 RELT (3) PDE1B NOS1 (2)RELT (3)RELT (3)CD4 (2)CD4 (2)GRK3 (2) TNFRSF8 (2) JAK3 GRIN2A (2) GALNT16 CXCL11 TNFRSF8 (2)​ HPGDS CREB5 (3) DGKB TNFRSF1B GCNT1 CREB5 (3)​ CCR1 GRK3 (2) HLA-DPA1 ​​​​ CCL26 CXCL10 ​​​​​ TGFB1 CXCL16 ​​​​​ GRIN2A (2) ​​​​​​ PIK3R5 ​​​​​​ IL1B (2) ​​​​​​ MMP1 ​​​​​​Table 3.Top 10 Hub Genes Identified Within the PPI Module Based on Degree Centrality and Their Associated Primary Biological PathwaysGeneDegreeBetweennessPrimary biological pathway IL1B 130.1389Pro-inflammatory cytokine, stress response, and stromal remodeling CD4 130.0564Immune cell signaling and T-cell mediated response CXCL10 120.0394Chemokine signaling and inflammatory cell recruitment TGFB1 120.0358Fibrosis, extracellular matrix (ECM) synthesis, and remodeling CCR1 110.0712Chemokine receptor activity and cell migration IL10RA 100.0224Regulation of anti-inflammatory signaling SOCS3 90.1157Regulation of cytokine signaling CXCL11 90.0129Immune response and chemotactic activity CXCL16 70.0018Cell adhesion and chemotaxis MMP1 70.0004Matrix metalloproteinase, collagen degradation, and ECM turnover Table 2.Integrated Regulatory Pairs of Identified Tear miRNAs and Their Corresponding Differentially Expressed Genes in Corneal Stromahsa-miR-326hsa-miR-30b-3phsa-miR-320bhsa-miR-1291hsa-miR-222-3phsa-miR-1290hsa-miR-1260a S100A7A SULT1E1 CREB5 (3)IL1B (2) GALNT18 MAP2K7 NOS1 (2) IL10RA CYP1A2 (2)CYP1A2 (2) CHRFAM7A SOCS3 GALNT15 RELT (3) PDE1B NOS1 (2)RELT (3)RELT (3)CD4 (2)CD4 (2)GRK3 (2) TNFRSF8 (2) JAK3 GRIN2A (2) GALNT16 CXCL11 TNFRSF8 (2)​ HPGDS CREB5 (3) DGKB TNFRSF1B GCNT1 CREB5 (3)​ CCR1 GRK3 (2) HLA-DPA1 ​​​​ CCL26 CXCL10 ​​​​​ TGFB1 CXCL16 ​​​​​ GRIN2A (2) ​​​​​​ PIK3R5 ​​​​​​ IL1B (2) ​​​​​​ MMP1 ​​​​​​ Integrated Regulatory Pairs of Identified Tear miRNAs and Their Corresponding Differentially Expressed Genes in Corneal Stroma Table 3.Top 10 Hub Genes Identified Within the PPI Module Based on Degree Centrality and Their Associated Primary Biological PathwaysGeneDegreeBetweennessPrimary biological pathway IL1B 130.1389Pro-inflammatory cytokine, stress response, and stromal remodeling CD4 130.0564Immune cell signaling and T-cell mediated response CXCL10 120.0394Chemokine signaling and inflammatory cell recruitment TGFB1 120.0358Fibrosis, extracellular matrix (ECM) synthesis, and remodeling CCR1 110.0712Chemokine receptor activity and cell migration IL10RA 100.0224Regulation of anti-inflammatory signaling SOCS3 90.1157Regulation of cytokine signaling CXCL11 90.0129Immune response and chemotactic activity CXCL16 70.0018Cell adhesion and chemotaxis MMP1 70.0004Matrix metalloproteinase, collagen degradation, and ECM turnover Top 10 Hub Genes Identified Within the PPI Module Based on Degree Centrality and Their Associated Primary Biological Pathways To further elucidate the regulatory architecture, a miRNA-hub gene interaction network was constructed. While the initial panel comprised 7 miRNAs, this refined network (Figure 3E) highlighted 5 core miRNAs (hsa-miR-326, hsa-miR-1291, hsa-miR-1290, hsa-miR-222-3p, and hsa-miR-30b-3p) that directly govern the identified hub genes. The findings underscored a functional interplay where individual hub genes, such as IL1B and CD4, are simultaneously governed by multiple miRNAs from the tear-derived panel. Notably, hsa-miR-326 identified as the most highly connected node within the regulatory network, potentially regulating multiple inflammatory and signaling genes such as IL1B, TGFB1 and MMP1. These results suggested potential upstream regulatory mechanisms whereby altered tear miRNA profiles may reflect or influence the molecular landscape of the corneal stroma in KC pathology, thereby driving the coordinated dysregulation of key inflammatory and remodeling pathways. By combining these multi-omics layers, our study provided a more comprehensive understanding of the molecular landscape underlying keratoconus. The interaction network among the feature genes was illustrated in Figure 3F, where each node represented a specific gene and the connecting lines indicate known or predicted interactions. Functional enrichment analysis of this network revealed that these hub genes were involved in several critical biological domains including cytokine-mediated signaling pathways, response to stress, cellular response to mechanical stimulus, and cell migration. Specifically, genes such as IL1B, and CXCL10 were central to cytokine signaling, while others like TGFB1, CCR1, and CXCL11 facilitated the cell migration and biological response to stress. The presence of IL1B and TGFB1 further highlighted the importance of the cellular response to mechanical stimulus in the disease context. The identified nodes, encompassing pro-inflammatory cytokines (IL1B), immune markers (CD4), and key chemokines (CXCL10, CXCL11, and CCR1), characterized a coordinated post-transcriptional response associated with chronic tissue inflammation and mechanical stress responses in progressive keratoconus. Consistent with the KEGG enrichment analysis of stromal transcriptomic signatures, functional characterization of the PPI hub network indicated that these hub genes were mainly involved in immune-related processes, including cytokine signaling, chemokine activity, inflammatory regulation, and immune cell recruitment, with additional contributions from extracellular matrix remodeling-related processes. 3.5. External Validation of Tear miRNAs in Corneal Epithelium To validate the miRNAs discovered in the tear cohort, we performed a cross-tissue validation using an independent corneal epithelial miRNA dataset (PRJNA842431) (Supplemental Table 6). Volcano plot and hierarchical clustering analyses of the validation cohort revealed a distinct miRNA expression profile, with the top 30 significant miRNAs stratifying the KC and control groups (Figure 4A and B). Further evaluation of specific miRNAs (Figure 4C) confirmed that hsa-miR-1291 (p = 0.009), hsa-miR-326 (p = 0.014), and hsa-miR-30b-3p (p = 0.014) were significantly downregulated in the KC group, displaying high consistency with the tear findings. In contrast, hsa-miR-1290 (p = 0.529) and hsa-miR-222-3p (p = 0.894) exhibited no statistically significant differential expression.Figure 4.External validation of miRNA expression profiles in an independent corneal epithelial dataset (PRJNA842431). (A) Volcano plot showing differentially expressed miRNAs between KC and control samples. (B) Heatmap depicting the expression patterns of the top 30 miRNAs across samples. (C) Boxplots showing individual expression levels of the 5 epithelium-derived miRNAs. (D) ROC curves illustrating the discriminatory potential of these miRNAs in epithelial tissue. (E) The blue line highlighted the diagnostic efficacy of the combined five-miRNA signature panel External validation of miRNA expression profiles in an independent corneal epithelial dataset (PRJNA842431). (A) Volcano plot showing differentially expressed miRNAs between KC and control samples. (B) Heatmap depicting the expression patterns of the top 30 miRNAs across samples. (C) Boxplots showing individual expression levels of the 5 epithelium-derived miRNAs. (D) ROC curves illustrating the discriminatory potential of these miRNAs in epithelial tissue. (E) The blue line highlighted the diagnostic efficacy of the combined five-miRNA signature panel ROC curve analysis was employed to assess the diagnostic efficacy of the selected miRNAs (Figure 4D). Given the limited discriminatory performance of individual markers, a five-miRNA panel was developed to enhance predictive performance (Figure 4E). To test its cross-tissue generalizability, the model was trained on the tear profiles and subsequently projected onto the corneal epithelial dataset. The multi-marker signature yielded a superior diagnostic performance, with an AUC of 0.89 (95% CI: 0.69-1.00) in the independent external validation cohort (N = 16). This cross-tissue validation underscored the biological relevance of the identified tear miRNAs and supported the five-miRNA signature in distinguishing KC from controls across tissue contexts. 3.6. Experimental Validation of hsa-miR-326 via RT-qPCR Representative findings from a severe KC case (Case # KC1) were shown in Figure 5A–C, demonstrating characteristic corneal ectatic changes identified by slit-lamp examination and Dual-Scheimpflug imaging. To validate the bioinformatic-derived findings, RT-qPCR analysis was performed in an independent clinical cohort (N = 52), including KC patients (n = 20), healthy controls (n = 20), dry eye disease (DED) patients (n = 8), and cataract patients (n = 4). As shown in Figure 5D, hsa-miR-326 was significantly downregulated in tear samples from KC patients compared with healthy controls (p < 0.05). Among the examined groups, KC patients exhibited the lowest hsa-miR-326 expression, whereas no significant differences were observed among cataract, dry eye disease (DED), and healthy control groups. The absence of similar alterations in DED and cataract groups suggested that reduced tear hsa-miR-326 expression may reflect KC-associated molecular changes rather than non-specific ocular surface inflammation alone. This finding was consistent with the expression pattern observed in the discovery tear cohort.Figure 5.Clinical characterization and experimental validation of hsa-miR-326 in KC. (A) External eye and slit-lamp photography of a representative KC patient (case #KC1). (B) Dual-Scheimpflug corneal topography and pachymetric analysis of the right eye (OD) and (C) left eye (OS), demonstrating characteristic corneal steepening and thinning associated with KC. (D) RT-qPCR validation of hsa-miR-326 expression in tear samples from KC patients and healthy controls (N=52, 20 KC, 20 controls, 8 DED, 4 cataract). (E) Correlation analysis between the MSI and Max K2 severity in the KC cohort (n=20) Clinical characterization and experimental validation of hsa-miR-326 in KC. (A) External eye and slit-lamp photography of a representative KC patient (case #KC1). (B) Dual-Scheimpflug corneal topography and pachymetric analysis of the right eye (OD) and (C) left eye (OS), demonstrating characteristic corneal steepening and thinning associated with KC. (D) RT-qPCR validation of hsa-miR-326 expression in tear samples from KC patients and healthy controls (N=52, 20 KC, 20 controls, 8 DED, 4 cataract). (E) Correlation analysis between the MSI and Max K2 severity in the KC cohort (n=20) We further explored the potential coordinated relationships between molecular alterations and structural features in KC. Standardized z-scores were calculated for hsa-miR-326 expression and keratometric measurements, and an exploratory molecular–structural index (MSI), defined as z(miR-326) − z(K2), was generated to represent the relative balance between molecular and structural alterations within individuals. The MSI demonstrated a significant inverse association with Max K2 in the KC cohort (Spearman ρ = -0.86, p<0.001, n = 20) (Figure 5E), suggesting a potential relationship between tear-derived miRNA alterations and corneal structural changes. 3.7. Association of hsa-miR-326 With Clinical Characteristics To further evaluate potential clinical confounding factors, we examined associations between tear hsa-miR-326 expression and ocular surface-related characteristics, including atopy, meibomian gland dysfunction (MGD), itching, and contact lens use. No significant associations were observed between these factors and hsa-miR-326 expression (Supplemental Table 7). In addition, correlation analyses with corneal structural parameters, including Max K2, minimum pachymetry, and ABCD max score, showed no significant association with hsa-miR-326 expression alone. The MSI showed a strong correlation with Max K2 but was not significantly associated with minimum pachymetry or ABCD max score (Supplemental Table 8). Because steroid hormone-related pathways were identified in transcriptomic enrichment analysis, we therefore evaluated potential sex-associated differences within the KC cohort. No significant differences were observed between female and male participants in tear hsa-miR-326 expression (median: 0.37 vs. 0.19, p = 0.31), MSI (median: 0.13 vs. -0.41, p = 0.40), Max K2 (median: 46.0 vs. 47.5 D, p = 0.94), or minimum pachymetry (median: 514 vs. 483 μm, p = 0.47) (Supplemental Table 9). These findings suggested that the identified molecular-structural associations were unlikely to be primarily attributable to sex distribution within the KC validation cohort. 4. Discussion Following k-means clustering to reduce biological heterogeneity, transcriptomic analysis identified 5,598 genes that were differentially expressed in keratoconus (KC) compared with controls. Subsequent pathway analysis (KEGG and GSEA) prioritized 174 genes enriched in immune-related pathways, suggesting that immune signaling represents a dominant transcriptional axis within the refined corneal cohorts. In parallel, tear miRNA profiling combined with LASSO regression identified a small set of regulatory candidates, including hsa-miR-222-3p, hsa-miR-30b-3p, hsa-miR-1260a, hsa-miR-320b, hsa-miR-326, hsa-miR-1290, and hsa-miR-1291. Rather than acting as isolated markers, these miRNAs converge on shared immune- and stress-associated gene modules, indicating coordinated regulatory structure across tear fluid and corneal transcriptomic signals. Network analysis further highlighted a tightly connected set of immune-associated genes, including IL1B, CXCL10, CXCL11, CCR1, TGFB1, and CD4, which were enriched in cytokine-mediated signaling, mechanical stress response, and inflammatory activation pathways. These results suggest that KC-related molecular alterations are organized into interdependent immune-stress response modules rather than independent gene changes. External validation using an independent corneal epithelial miRNA dataset demonstrated partial reproducibility, with hsa-miR-1291, hsa-miR-326, and hsa-miR-30b-3p showing concordant expression patterns. The remaining candidates showed tissue- or compartment-dependent variability, which may reflect differences between intracellular epithelial regulation and extracellular tear secretion dynamics. Overall, these findings indicate that tear-derived miRNA profiles capture biologically coherent signals that reflect, at least partially, corneal immune-associated transcriptional alterations in KC. Within the framework of non-coding RNA research in keratoconus, previous studies have reported dysregulation of multiple miRNAs and lncRNAs in corneal tissues.19,30 However, most prior work has relied on single-cohort designs and isolated molecular layers, limiting cross-study comparability and translational interpretation. In contrast, the present study incorporates cross-cohort transcriptomic filtering combined with tear-based profiling, enabling comparison between corneal intracellular transcriptional changes and extracellular RNA signals. Notably, hsa-miR-1260a has been previously implicated in stromal tissue alterations and inflammatory signaling pathways involving IL1B, TGFB1, and CCR1. 31 These pathways overlap with those identified in our network analysis, supporting consistency between independent datasets. Among the identified candidates, hsa-miR-326 is particularly notable. Although its role in keratoconus remains insufficiently characterized, prior studies in other disease contexts have demonstrated its ability to regulate IL1B and modulate TGFB1-associated inflammatory signaling.32,33 Its central network position in our analysis suggests potential relevance in immune-related regulation in KC. Increasing evidence indicates that KC is associated with chronic immune dysregulation rather than representing a purely non-inflammatory ectatic disorder.34,35 Consistent with this evolving concept, our enrichment analysis demonstrated that the immune-related genes retained after transcriptomic refinement were predominantly associated with cytokine-mediated signaling, inflammatory responses, cellular responses to stress, and immune regulatory processes. These enriched biological processes suggest that immune activation in KC is organized into coordinated signaling networks that may influence corneal homeostasis. Persistent cytokine signaling has been implicated in disrupting the balance between extracellular matrix synthesis and degradation, thereby promoting stromal thinning and progressive biomechanical weakening.34-36 The simultaneous enrichment of stress-response pathways further supports the concept that inflammatory signaling and cellular stress are biologically interconnected processes in KC rather than independent pathogenic events. Recent single-cell transcriptomic studies further support the presence of immune-associated dysregulation at the cellular level in KC. While bulk transcriptomic analyses provide an integrated view of tissue-level molecular alterations, single-cell approaches have revealed that disease-associated changes may arise from specific corneal cell populations. Dou et al demonstrated cell-type-specific transcriptional alterations in corneal epithelial cells, stromal cells, and stromal immune cells, identifying abnormal expression of immune-related genes including IL23A, CXCL1, and IL1RN, as well as alterations in genes associated with extracellular matrix regulation. 37 These findings provide important biological context for the immune-related pathways identified in our bulk transcriptomic analysis and suggest that inflammatory dysregulation in KC represents a coordinated alteration across multiple cellular compartments rather than a consequence of bulk tissue averaging alone. Our enrichment results also provide biological context for the hub genes identified in the integrated miRNA–mRNA network. Key genes including IL1B, CXCL10, CXCL11, CCR1, TGFB1, and CD4 were not interpreted as isolated differentially expressed genes but rather as components of interconnected immune signaling pathways. IL1B represents a major upstream inflammatory cytokine capable of inducing downstream chemokine production and matrix remodeling, whereas CXCL10 and CXCL11 participate in chemokine-mediated communication that regulates immune-cell recruitment and local inflammatory responses. 38 TGFB1 provides a mechanistic link between immune activation and extracellular matrix remodeling through its established role in stromal repair and fibrotic regulation.39,40 Together with CD4-related immune activity and SOCS3-associated signaling regulation, these molecules may reflect modulation of inflammatory balance under chronic ocular surface stress. 41 Although direct mechanistic confirmation in corneal tissue remains limited, these components collectively suggest a coordinated inflammatory signaling environment associated with disease activity. It should be noted that the network was constructed by anchoring tear-derived miRNAs to a pre-selected set of immune-related corneal genes. The identified miRNA signature may represent a molecular pattern associated with immune signaling and tissue remodeling rather than isolated miRNA-gene interactions. The observed cross-compartment concordance provides biological support for the hypothesis that extracellular tear miRNAs may capture, at least partially, immune-associated molecular alterations occurring in the cornea. However, pathway enrichment findings should be interpreted in the context of both the biological signals captured by the dataset and the analytical framework used for feature selection. Therefore, these results should be considered hypothesis-generating associations linking tear miRNAs with corneal transcriptomic alterations, rather than evidence of definitive regulatory mechanisms. The integrative analysis of tear-derived miRNAs and stroma-derived mRNAs revealed a five-miRNA signature that may represent upstream regulatory associations of inflammatory and structural pathways in KC. Among these, hsa-miR-326 emerged as a central hub linking tear-based molecular signals with corneal transcriptomic alterations. This regulatory module provides a framework for understanding how extracellular molecular changes may relate to corneal structural remodeling processes.1,42-44 In advanced KC, keratocyte density is reduced due to apoptosis, leading to decreased cellular metabolic and transcriptional activity.13,45-48 Therefore, observed reductions in mRNA expression may partly reflect cellular loss rather than direct suppression of biological processes. The transcriptional patterns observed in the integrated stromal datasets (Supplemental Figure 1) further support coordinated dysregulation of immune-related pathways rather than isolated miRNA-mRNA pairwise regulation. Overall, these findings suggest that tear-derived miRNAs may reflect upstream molecular states associated with corneal structural alterations, including stromal thinning and extracellular matrix remodeling. Functional validation will be required to determine direct regulatory relationships. It is important to note that while our primary research focus was centered on immune-related genes, we also conducted a comprehensive analysis involving the entire set of DEGs. Although the initial candidate genes spanned a wide range of biological processes, the high-confidence PPI network and topological filtering (targeting the top 20 hub genes as shown in Supplemental Table 10) yielded a convergence toward specific functional pathways (Supplemental Figure 2). This analytical consistency supports the interpretation that immune-related signaling represents a prominent biological axis within the selected regulatory network. From a translational perspective, this study provides a framework for linking tear-derived molecular signals with corneal transcriptomic alterations in KC. Given the accessibility and non-invasive nature of tear sampling, tear miRNA profiles may provide complementary molecular information that supplements conventional structural assessment, rather than serving as standalone diagnostic biomarkers. Importantly, the integrative design of this study suggests that keratoconus is better interpreted as a biologically heterogeneous condition, in which molecular signals may vary across patients despite similar clinical classification. In this context, tear-derived miRNA patterns may contribute to future approaches aimed at refining disease characterization when combined with established structural imaging and clinical parameters. Recent studies investigating tear-derived molecular biomarkers have highlighted the potential of extracellular RNAs as accessible indicators of ocular disease-associated biological alterations. 49 However, candidate miRNAs have varied considerably across studies, likely reflecting differences in biological compartments, disease stages, and analytical platforms. 17 In the present study, the identified five-miRNA signature complements previous biomarker efforts by linking tear-derived molecular patterns with immune-associated corneal transcriptomic alterations. Among these candidates, hsa-miR-326 is of particular translational interest because it emerged as the central component of the integrated tear miRNA-corneal transcriptomic network. Although direct evidence supporting hsa-miR-326 as a KC-specific ocular biomarker remains limited, its network position and association in TGF-β-associated signaling pathways support further evaluation in biomarker-oriented studies. 50 Nevertheless, differences among biomarker studies highlight the importance of considering disease stage, biological compartment, sample collection procedures, storage conditions, RNA processing methods, analytical platforms, and normalization strategies when interpreting tear miRNA signatures. To further explore the relationship between molecular and structural dimensions of disease, we constructed a proof-of-concept molecular-structural composite index integrating hsa-miR-326 expression and keratometry (K2). The observed association with disease severity suggests that combined molecular-structural representations may better capture disease complexity than either component alone. However, this index should be interpreted strictly as a hypothesis-generating clinical correlate rather than a validated predictive tool. Overall, these findings support the concept that tear-derived microRNAs may provide accessible molecular information that complements corneal imaging, offering a potential bridge between non-invasive sampling and tissue-level pathophysiology in keratoconus. Further validation in larger and more clinically stratified cohorts will be required to determine the stability and practical utility of these observations. This framework may provide a foundation for investigating molecular-based stratification and longitudinal disease monitoring in KC. 5. Limitation Several limitations should be considered when interpreting our findings. First, the integrated transcriptomic datasets were generated using different sequencing platforms, library preparation protocols, and normalization strategies, which may introduce technical variability despite the application of clustering methods to reduce heterogeneity. K-means clustering was employed to improve subgroup consistency. However, residual inter-study and inter-individual variability cannot be completely excluded. Second, the sample size of the prospective tear cohort was relatively small, which may limit statistical power and generalizability. In addition, certain subgroups, such as mild KC cases, did not yield a sufficient number of differentially expressed genes for robust network construction, limiting stage-specific analyses. Corneal tomographic parameters, including Max K2, minimum pachymetry, and ABCD grading, were available for evaluating structural disease severity. However, visual acuity outcomes were not systematically analyzed due to variability in data availability and potential influences from refractive correction and other optical factors. Contact lens use was also recorded and no significant association with tear hsa-miR-326 expression was observed. Other potential biological confounders should also be considered. Steroid hormone-associated pathways were identified in the transcriptomic enrichment analysis, raising the possibility that sex-related biological factors may influence immune regulation and miRNA expression in KC. Exploratory sex-stratified analyses in the present cohort revealed no significant differences in tear hsa-miR-326 expression, MSI, or major corneal structural parameters between female and male participants. However, given the limited sample size, subtle sex-dependent effects cannot be completely excluded and should be further evaluated in larger, balanced cohorts. Third, the epithelial miRNA dataset used for supplementary analysis does not represent an independent clinical validation cohort. Differences observed between tear-derived and tissue-level miRNA expression patterns may reflect compartment-specific regulation, selective secretion, or methodological differences. Therefore, the observed overlap should be interpreted as supportive biological evidence rather than formal validation. Furthermore, the predicted miRNA-mRNA regulatory relationships were inferred from computational integration and pathway-level associations. The absence of strict inverse expression trends between certain miRNAs and their target genes suggests that regulatory dynamics in KC may operate at a systems or pathway level rather than through simple one-to-one interactions. Experimental validation will be required to confirm direct mechanistic relationships. While the resulting five-miRNA panel demonstrated promising discrimination between KC and control, the limited sample size raises the possibility of overfitting. To evaluate the statistical reliability of this cohort, we performed a post-hoc power analysis. Under a highly stringent all miRNA screening threshold (alpha = 0.001) to minimize false positives during unbiased search, the effect size was estimated using the median Cohen’s d of the top 50 differentially expressed miRNAs (d = 2.09). Under these conditions, the statistical power for our sample size (N = 20) was approximately 0.75, which is close to the conventional 0.80 benchmark. Critically, when evaluating the predefined five-miRNA panel (alpha = 0.05), the median Cohen’s d increased to 2.59 when calculated specifically from the five-miRNA panel. Consequently, the statistical power was estimated to exceed 0.99, suggesting that the observed signal is unlikely to be explained solely by random variation within this cohort (Supplemental Figure 3). Furthermore, hsa-miR-1260a showed a significant up-regulation consistent with corneal epithelial data, its expression trend contrasted with the initial miRNA-sequencing results from tear samples. These discrepancies highlighted the complexity of miRNA secretory dynamics in KC. Consequently, these findings should be interpreted cautiously. In addition, although the RT-qPCR assay was consistently applied across all clinical samples using the same primer set and amplification conditions, comprehensive analytical validation, including standard curve-based amplification efficiency assessment, linearity testing, and independent specificity evaluation, was not performed in this study. Future studies incorporating formal assay validation will be required to establish the analytical performance of the proposed miRNA detection strategy. Furthermore, the cross-sectional design of the present study limits the ability to determine temporal relationships between tear miRNA alterations and KC progression. Therefore, whether the identified miRNA signatures represent early molecular changes preceding structural deterioration, consequences of disease progression, or dynamic responses to corneal stress remains unclear. Prospective longitudinal studies involving larger, independently validated cohorts and serial tear sampling will be required to determine the stability of these signatures and to evaluate their potential utility for monitoring disease progression or treatment response. 6. Conclusions This study integrated tear-derived miRNA profiles with corneal transcriptomic datasets to identify immune-related molecular signatures in KC. By combining consensus clustering to reduce biological heterogeneity, differential gene expression analysis, pathway enrichment, and LASSO-based feature selection, we identified five tear-derived miRNAs and 174 immune-related mRNAs associated with the molecular patterns distinguishing KC and control clusters. Key genes including IL1B, CXCL10, CXCL11, TGFB1, and CD4, formed a tightly interconnected network involved in cytokine signaling, mechanical stress response, and extracellular matrix remodeling. Several tear-derived miRNAs (including hsa-miR-326, hsa-miR-320b, and hsa-miR-30b-3p) were also supported by external validation in an independent corneal epithelial dataset, suggesting consistent biological associations across datasets. Among these, hsa-miR-326 emerged as a prioritized candidate within the integrated network, warranting further investigation in future functional studies. Overall, this integrative analytical framework provides a basis for linking tear-derived molecular signals with corneal transcriptomic alterations in KC, and may provide a foundation for future studies investigating molecular characterization of disease heterogeneity. Supplemental Material Supplemental Material - Tear MicroRNA-Based Molecular Signature for Keratoconus Stratification: A Cross-Cohort Integration Study Supplemental Material for Tear MicroRNA-Based Molecular Signature for Keratoconus Stratification: A Cross-Cohort Integration Study by Shao-Hsuan Chang, Kuo-Hsuan Hung, Hsieh-Fu Tsai, Chung-Pei Ma and Hao-Chang Chiang and Lung-Kun Yeh in Bioinformatics and Biology Insights.