An Integrative Bioinformatics Framework Nominates Candidate Limbal Stem-Cell Exosome Cargo for Keratoconus by Coupling Corneal Transcriptomics, Disease-Gene Evidence and Extracellular-Vesicle Repositories

1. Introduction

Keratoconus is the most common corneal ectasia and a leading indication for corneal transplantation in young adults; it is defined by progressive stromal thinning, biomechanical weakening and disruption of collagen and other ECM components [1,2]. Although historically regarded as a non-inflammatory condition, accumulating tear-film and tissue evidence now supports a low-grade inflammatory and immune contribution to disease progression [3,4,5]. Current management—rigid contact lenses, corneal cross-linking and, ultimately, keratoplasty—addresses biomechanics and optics but does not target the underlying molecular pathology, leaving a clear unmet need for disease-modifying, regenerative approaches.

The corneal epithelium is continuously renewed by limbal epithelial stem cells residing in the limbal niche, whose dysfunction underlies limbal stem-cell deficiency and impaired corneal surface homeostasis [6,7,8]. Stem-cell-derived exosomes—nanoscale extracellular vesicles (EVs) carrying proteins, mRNA and microRNA—have emerged as a cell-free regenerative modality, and limbal-niche- and mesenchymal stem-cell (MSC)-derived EVs accelerate corneal epithelial repair in preclinical models [9,10,11,12,13,14]. A central translational question, however, is which molecular cargo a therapeutic LSC exosome should carry to correct keratoconus pathology. No curated LSC-exosome cargo dataset is available in the major EV repositories, which precludes a purely data-driven cargo definition and instead motivates an integrative, evidence-ranked nomination strategy.

Here we develop a reproducible computational framework that couples real keratoconus corneal transcriptomics with disease-gene evidence, PPI topology and documented EV-cargo competence to prioritise candidate LSC-exosome cargo. We deliberately treat the absence of LSC- and induced-pluripotent-stem-cell (iPSC)-EV cargo data as an explicit, declared limitation rather than imputing it, and we provide a transparent Cargo Prioritization Score whose every component is traceable to a public dataset, public database or stated scoring rule.

2. Materials and Methods

The overall analytical workflow, from public-data acquisition through differential expression, functional and network analysis, multi-source evidence integration and Cargo Prioritization Scoring, is summarised in Figure 1.

2.1. Dataset and Preprocessing

Processed gene-level RNA-sequencing counts for human keratoconus (KTCN) and non-keratoconus (non-KTCN) corneas were obtained from Gene Expression Omnibus accession GSE77938 [16,17], which comprises an independent discovery cohort (8 KTCN versus 8 control) and a replication cohort (17 KTCN versus 17 control), all corneal tissue (50 samples in total; Table 1). Genes with at least 10 counts in a number of samples equal to the smaller group were retained.

2.2. Differential Expression

Differential expression was performed independently in each cohort with DESeq2 [18] using the model ~group (KTCN versus Control); log2 fold changes were shrunk with apeglm [19]. Ensembl identifiers were mapped to gene symbols and Entrez IDs (org.Hs.eg.db). DEGs were defined as adjusted P < 0.05 (Benjamini-Hochberg) and |log2 fold change| > 1. A replicated (consensus) DEG set was defined as DEGs significant in discovery that were concordant in sign and nominally significant (P < 0.05) in the replication cohort.

2.3. Functional Enrichment

Gene Ontology Biological Process, KEGG and Reactome over-representation analyses of the consensus DEGs were performed with clusterProfiler [20] and ReactomePA [21] against the set of all tested genes as background (Benjamini-Hochberg P < 0.05; redundant GO terms were simplified at a semantic-similarity cutoff of 0.7).

2.4. Protein-Protein Interaction Network and Hub Genes

The 400 most significant consensus DEGs (by discovery adjusted P) were submitted to the STRING database [22]; only high-confidence interactions (combined score >= 700) were retained. The network was analysed in NetworkX, and hub genes were ranked by degree and betweenness centrality (igraph/NetworkX) [26,27].

2.5. Database Integration and Cargo Prioritization Score

Keratoconus disease-gene association scores were retrieved from the Open Targets Platform (disease MONDO_0015486) [23]. Documented human EV-cargo competence and a mesenchymal stem-cell (MSC) EV reference proteome were derived from ExoCarta [24] and Vesiclepedia [25]; EV nomenclature follows MISEV2023 [15]. For each replicated DEG, five components were scaled to [0,1] or set as indicators: differential-expression evidence (S_DE, scaled |log2 fold change|), disease association (S_disease, Open Targets score), PPI hub centrality (S_hub, scaled degree), documented EV-cargo competence (S_EVcargo) and MSC-EV cargo membership (S_MSC). To integrate these complementary evidence sources into a single ranking metric, we empirically assigned weights that emphasized transcriptomic dysregulation and disease relevance while retaining contributions from network topology and extracellular-vesicle evidence. The resulting Cargo Prioritization Score (CPS) was calculated as: CPS = 0.30·S_DE + 0.25·S_disease + 0.20·S_hub + 0.15·S_EVcargo + 0.10·S_MSC. A disease-gene-cargo-pathway network was assembled for the top-ranked candidates.

3. Results

3.1. Differentially Expressed Genes in Keratoconus Cornea

Across 27,830 testable genes in the discovery cohort, 1,677 genes were differentially expressed (152 up- and 1525 down-regulated in KTCN), a pronounced down-regulation bias consistent with loss of corneal structural transcripts (Figure 2). Of these, 1,380 were replicated in the independent cohort and were carried forward as the consensus DEG set (Table 2; Figure 3).

Table 3. Top 15 replicated DEGs ranked by discovery adjusted P.

Gene	log2FC (disc)	padj (disc)	log2FC (rep)	padj (rep)
CCN2	-3.71	1.64e-21	-3.27	5.17e-13
PCDH10	-5.67	1.17e-17	-2.9	1.28e-05
RNASE1	-6.89	2.33e-17	-6.24	2.37e-17
CCN1	-2.51	3.04e-17	-2.63	2.98e-15
MATN3	-4.34	1.09e-16	-3.9	2.24e-17
C1QB	-6.45	1.46e-16	-5.63	7.33e-19
MEF2C	-4.1	3.44e-16	-3.24	1.15e-14
GFRA1	-7.72	6.49e-16	-3.9	5.12e-08
GUCY1A1	-5.7	1.9e-15	-4.86	1.18e-24
CYBB	-5.2	4.59e-15	-5.03	3.62e-14
MRC1	-6.39	6.6e-15	-5.34	1.07e-15
LRRC32	-6.28	7.49e-15	-4.67	1.11e-19
TMEM119	-5.76	4.51e-14	-4.34	9.05e-17
FBN1	-4.5	4.84e-14	-4.21	1.94e-25
INHBA	-6.63	5.43e-14	-5.24	3.03e-14

Table 3. Top 15 replicated DEGs ranked by discovery adjusted P.

Gene	log2FC (disc)	padj (disc)	log2FC (rep)	padj (rep)
CCN2	-3.71	1.64e-21	-3.27	5.17e-13
PCDH10	-5.67	1.17e-17	-2.9	1.28e-05
RNASE1	-6.89	2.33e-17	-6.24	2.37e-17
CCN1	-2.51	3.04e-17	-2.63	2.98e-15
MATN3	-4.34	1.09e-16	-3.9	2.24e-17
C1QB	-6.45	1.46e-16	-5.63	7.33e-19
MEF2C	-4.1	3.44e-16	-3.24	1.15e-14
GFRA1	-7.72	6.49e-16	-3.9	5.12e-08
GUCY1A1	-5.7	1.9e-15	-4.86	1.18e-24
CYBB	-5.2	4.59e-15	-5.03	3.62e-14
MRC1	-6.39	6.6e-15	-5.34	1.07e-15
LRRC32	-6.28	7.49e-15	-4.67	1.11e-19
TMEM119	-5.76	4.51e-14	-4.34	9.05e-17
FBN1	-4.5	4.84e-14	-4.21	1.94e-25
INHBA	-6.63	5.43e-14	-5.24	3.03e-14

3.2. Functional Enrichment

The consensus DEGs were strongly enriched for ECM organisation and degradation, collagen and integrin biology, and immune/cytokine signalling across GO, KEGG and Reactome (Figure 4; Table 4). The leading GO term was “extracellular matrix organization” (adjusted P = 4.5e-27, 85 genes), and Reactome was led by “Extracellular matrix organization” (adjusted P = 3.6e-33). This dual ECM-and-immune signature is concordant with the contemporary view of keratoconus pathology [2,5].

3.3. PPI Network and Hub Genes

A high-confidence STRING network of the consensus DEGs contained 200 connected proteins and 621 edges (Figure 5). The highest-degree hubs—FN1, ITGAM, PTPRC, ITGB2, TYROBP, FCGR3A, CCL2, CD163—comprised both ECM/collagen proteins and immune effectors, all down-regulated in keratoconus (Table 5), pinpointing co-ordinately suppressed modules as rational intervention nodes.

3.4. Cargo Prioritization

Integrating the five evidence layers, the Cargo Prioritization Score ranked 1,253 replicated DEGs. 499 keratoconus-associated targets from Open Targets contributed disease-gene weight, including established keratoconus genes among the top candidates. The highest-scoring candidate cargo were COL1A1, FN1, COL4A1, COL3A1, COL5A1, MMP1, VCAN, PTPRC, INHBA, ITGAM (Figure 6; Table 6)—dominated by ECM and collagen constituents that are down-regulated in keratoconus, documented as EV cargo, and present in the MSC-EV reference proteome, and therefore represent biologically coherent restoration cargo for a limbal stem-cell exosome. Because keratoconus is characterised by ECM loss, these candidates are framed as restoration (replacement) cargo, whereas up-regulated immune candidates would instead represent suppression targets requiring a different cargo modality.

4. Discussion

Using only real, publicly available data, this study converts a fragmented evidence landscape into a single, ranked and fully traceable shortlist of candidate limbal stem-cell exosome cargo for keratoconus. The convergence of the differential-expression, network and disease-gene layers on extracellular-matrix and collagen biology is reassuring: it recapitulates the established molecular hallmark of keratoconus and indicates that the prioritisation is anchored in disease-relevant biology rather than analytical artefact.

The framework reframes a data gap as a hypothesis-generating opportunity. Because no LSC-exosome cargo repository exists, we do not claim to describe the LSC-EV proteome; instead we ask which disease-relevant molecules a therapeutic LSC exosome would ideally carry, and we restrict candidates to those for which EV delivery is at least plausible (documented EV cargo) and precedented in a stem-cell EV context (MSC-EV reference). The mechanistic hypothesis—that restoring ECM and collagen cargo, while modulating the immune component, could slow keratoconus progression—is consistent with preclinical evidence that limbal-niche- and MSC-derived EVs promote corneal epithelial and stromal repair, and is depicted in Figure 7.

The CPS is intentionally simple and transparent: all weights are fixed a priori, every component is traceable to a named public source, and the score can be recomputed or reweighted by any user. This favours interpretability and auditability over black-box optimisation, which we regard as appropriate for a hypothesis-prioritisation tool intended to guide, not replace, experimental validation.

4.1. Limitations

Several limitations must be stated explicitly. First, no curated LSC-exosome (or iPSC-exosome) cargo dataset exists in ExoCarta, Vesiclepedia or EVpedia (zero limbal/corneal EV experiments were found); consequently the LSC-EV cargo is nominated, not measured, and the MSC-EV proteome is used only as the closest available stem-cell EV reference. Second, the transcriptomic evidence derives from a single keratoconus dataset (GSE77938); although it includes an internal replication cohort, external validation in additional cohorts and ocular-surface diseases is required. Third, the public EV repositories are biased towards well-studied, highly abundant proteins, so EV-cargo competence is a permissive rather than a stringent filter. Fourth, gene-symbol harmonisation between transcriptomic and EV-repository nomenclature is imperfect and may under-count matches. Fifth, the analysis is entirely computational: the prioritised cargo are hypotheses that require proteomic confirmation in authentic LSC exosomes and functional validation in corneal models. Finally, the CPS weights, while transparent, are expert-chosen and not learned from outcome data.

4.2. Translational Outlook

Beyond target nomination, the candidate cargo identified here define a concrete translational path. The in-silico shortlist (Figure 8, stage 1) would first require confirmation that the prioritised molecules are genuinely present in authentic limbal stem-cell exosomes, followed by functional and preclinical validation, regulatory interactions and early-phase clinical evaluation of a cell-free exosome product. A structured validation programme spanning cargo confirmation, functional assays, disease models and EV engineering/delivery is outlined in Figure 9. We emphasise that all stages downstream of the present computational nomination remain to be performed; the framework is intended to focus and de-risk, not to substitute for, that experimental work.

5. Conclusions

We present a reproducible, integrity-constrained bioinformatics framework that nominates and ranks candidate limbal stem-cell exosome cargo for keratoconus by integrating real corneal transcriptomics, disease-gene evidence, PPI topology and extracellular-vesicle repositories. The approach yields an evidence-traceable shortlist led by ECM/collagen restoration cargo, an explicit map of current data gaps, and a transparent scoring scheme that is directly testable. It provides a rational starting point for designing engineered therapeutic exosomes for corneal ectasia and is readily generalisable to other ocular-surface diseases as suitable datasets become available.