Decoding regional keratinization in human oral mucosa through high-resolution spatial transcriptomics Authors: Noy Rosental Vazani, Michal Lusthaus, Ana Caetano, Rami Khosravi, Avital Sarusi Portuguez, Carlos E Nemcovsky, Omer Bender, Daniel Z Bar, Miron Weinreb, Evgeny Weinberg Published: 2026 ABSTRACT AbstractOral mucosa exhibits region-specific keratinization, essential for periodontal health, yet the spatial and molecular mechanisms driving these differences remain poorly understood. This study aimed to generate a high-resolution spatial transcriptomic atlas of the human oral mucosa around the mucogingival junction, to reveal stromal-epithelial interactions, that distinguish keratinized from non-keratinized programs. Formalin-fixed paraffin-embedded specimens from the mucogingival junction area of two healthy donors were analyzed with the 10 × Genomics Visium HD platform, yielding two keratinized and two non-keratinized regions. Spatial clustering, pseudotime trajectory inference, cell-type integration with a single-cell reference, and ligand-receptor network analysis were applied to delineate epithelial and stromal compartments. Sixteen reproducible clusters, recapitulating tissue architecture, were identified and revealed distinct transcriptional signatures, distinguishing gingiva from lining mucosa. Pseudotime analysis revealed bifurcating epithelial lineages, originating from a shared basal progenitor layer into keratinized and non-keratinized programs. Gingival keratinization was driven by stromal collagen ligands (COL1A1, COL1A2, COL6A1, COL6A2) engaging epithelial receptors (CD44, SDC1), further reinforced within the epithelium by desmosomal adhesion via DSG1-DSC2/3. Gingival keratinization emerges from integrated stromal collagen signaling and epithelial adhesion. This spatially resolved framework advances understanding of oral mucosal specialization and provides a foundation for biologically guided regenerative therapies.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1038/s41598-026-53276-y.Keywords: Oral mucosa, Keratinization, Spatial transcriptomics, Stromal-epithelial interactions, Collagen-desmosome signalingSubject terms: Biological techniques, Cell biology Introduction The oral mucosa is a dynamic barrier that supports mastication, speech, host defense, and wound healing, while maintaining the structure and esthetic integrity of the oral cavity1,2. It consists of stratified squamous epithelium supported by the lamina propria, and is classified into three types: masticatory mucosa (keratinized, in stress-bearing areas, such as the gingiva and hard palate), lining mucosa (non-keratinized), and specialized mucosa (on the dorsal tongue surface)3. Despite their shared stratified squamous architecture, these tissues differ in mechanical resilience, barrier function, immune responsiveness, and healing patterns4–6. Although initially attributed to functional adaptation to mechanical forces, landmark transplantation studies have demonstrated that keratinization patterns are primarily genetically determined and maintained by intrinsic molecular signals from the underlying connective tissue, underscoring its role in directing epithelial differentiation7,8. Subsequent research has revealed that epithelial-stromal interactions, mediated by growth factors, extracellular matrix components, and basement membrane (BM) constituents, are key regulators of epithelial differentiation and gene expression9–12. However, the molecular mechanisms driving oral mucosal keratinization remain unclear. Understanding oral mucosal keratinization has direct therapeutic relevance. In periodontics and implantology, sufficient keratinized tissue supports functional stability and esthetic outcomes2,13, whereas its deficiency predisposes to recession, root exposure, hypersensitivity, and esthetic problems14,15. Current solutions rely on invasive autologous grafting16,17, underscoring the need for biologically guided alternatives18. Previous bulk and single-cell RNA-seq studies have identified diverse stromal populations in oral mucosa, that may influence epithelial keratinization19,20. However, single-cell methods require tissue dissociation, disrupting spatial context essential for understanding stromal-epithelial communication. Spatial transcriptomics (ST) overcomes this by mapping gene expression in situ while preserving architecture, enabling high-resolution profiling21. ST has already advanced fields such as development and cancer22, and is now driving progress in oral and craniofacial research23,24. Despite recent advances, the spatial gene expression patterns, distinguishing (keratinized) gingiva from non-keratinized oral mucosa, remain poorly defined. In this study, we employed high-resolution ST using the 10 × Genomics Visium HD platform to comprehensively characterize the molecular architecture of healthy human oral mucosa around the mucogingival junction (MGJ). By analyzing tissue sections containing both keratinized gingiva and non-keratinized lining (alveolar) mucosa, we identified distinct epithelial and stromal compartments and delineated the signaling pathways that drive regional keratinization. These findings advance understanding of oral mucosal biology and lay the groundwork for developing targeted therapies to induce oral mucosal keratinization. Materials & methods Tissue collection and preparation of tissue blocks for spatial transcriptomics This study was approved by the Institutional Ethics Committee of Tel Aviv University (IEC No. 0007945–1) all participants provided written informed consent, and all methods were performed in accordance with relevant guidelines and regulations. Oral mucosal specimens were collected from the mucogingival junction (MGJ) area of the anterior mandible in two healthy Caucasian donors (a 34-year-old male and a 28-year-old female; Fig. 1A) during routine periodontal procedures. Exclusion criteria included smoking, systemic disease, pregnancy, lactation, or periodontal disease. (A) Schematic representation of the experimental design and workflow; (B) Two H&E-stained, FFPE oral mucosal tissue sections, obtained from different patients at the MGJ, analyzed using the 10 × Genomics Visium HD platform. Each sample includes the interface between gingiva and non-keratinized lining (alveolar) mucosa; (C) Spatial map showing 16 graph-based clusters identified using the 10 × Genomics Space Ranger pipeline and manually annotated; (D) UMAP plot showing each spatial spot as an individual point, with colors corresponding to cluster identity; (E) Spatial maps of both tissue specimens annotated by major tissue compartments, showing the distribution of gingival epithelium, lining mucosa epithelium, shared basal layer of epithelium, and stroma; (F) Detailed spatial annotation of epithelial layer clusters within one tissue specimen, showing the stratified organization of gingival epithelium and lining mucosa epithelium; (G) Schematic annotation of clusters based on histological localization and gene expression features. For each cluster, the top three up-regulated genes are indicated. Although Clusters 5, 3, and 1 show limited extension across the MGJ, they are illustrated schematically according to their predominant anatomical localization: Cluster 5 in gingival stroma, and Clusters 3 and 1 in lining mucosa stroma. Tissues were fixed in 4% formaldehyde, processed as formalin-fixed paraffin-embedded (FFPE) blocks, sectioned at 5 μm, and stained with hematoxylin and eosin (H&E). Histology confirmed that both specimens encompassed the transition zone between keratinized gingiva and non-keratinized lining (alveolar) mucosa. The gingival epithelium exhibited parakeratinization, versus the thicker non-cornified alveolar epithelium. The lamina propria varied in connective tissue density between gingiva and alveolar mucosa. To maximize spatial capture efficiency, the two original specimens were embedded together into a single 7 × 7 mm paraffin block, with careful orientation, to preserve anatomical integrity. This consolidation enabled simultaneous processing and spatial mapping of both samples within a single capture area of the 10 × Genomics Visium HD platform. RNA quality assessment in FFPE tissue block RNA was extracted from freshly sectioned material using the RNeasy FFPE Kit (Qiagen). Integrity was assessed with the Agilent TapeStation, measuring mean fragment length and the percentage of RNA fragments > 200 nucleotides (DV200). The sample yielded a DV200 of 72%, well above the 30% threshold recommended by 10 × Genomics, confirming suitability for ST library preparation. Spatial transcriptomics workflow Serial 5 μm sections were cut from the unified paraffin blocks and mounted on charged slides (Leica Biosystems). H&E staining followed the 10 × Genomics Visium HD protocol using a Leica ST5020 multistainer, and slides were imaged at 40 × with an Olympus VS200 scanner (Evident). Spatial RNA transfer to Visium HD slides was performed with the 10 × CytAssist, and cDNA synthesis and library preparation were carried out using the Visium HD Human Transcriptome Probe Kit v2 (CG0000685, Rev. B). In brief, mRNA molecules were captured with barcoded probes, reverse transcribed, and amplified. Final libraries were sequenced on an Illumina NextSeq 2000 following 10 × Genomics guidelines. Spatial transcriptomics analysis ST data were generated with the 10 × Genomics Visium HD platform. Raw FASTQ files were processed using Space Ranger (v3.1.2) with manual image alignment to match spatial data with H&E-stained sections for accurate tissue registration. Sequencing yielded ~ 286 million reads, with 97% mapping to the probe set and 92.6% saturation (Supporting Information 1). A total of 113,066 eight-µm bins were retained under tissue across both sections, detecting 17,480 unique genes. These metrics reflect deep coverage and high data quality for downstream spatial and functional analyses. Raw sequencing data were aligned, quantified, and subjected to graph-based clustering using the 10 × Genomics Space Ranger pipeline. Marker identification and heatmap visualization Filtered feature-barcode matrices from Space Ranger were imported into Seurat (v5). Clustering metadata from Space Ranger’s output were integrated into the Seurat object via barcode matching for spatial annotation. Data were normalized to correct technical variation. Cluster-specific markers were identified based on log₂FC > 0.1, expression in ≥ 10% of cells, and adjusted p < 0.05. Mitochondrial genes (defined by the “MT-” prefix) were excluded to avoid metabolic bias. The top five markers per cluster (ranked by log₂FC) were visualized using a heatmap. Gene ontology enrichment (GO) was performed on top marker genes. Notably, some marker genes from the basal epithelial cluster (4) and subepithelial stromal cluster (10) appeared associated with the adjacent compartment, such as COL17A1 in basal epithelium and KRT15 in stroma. Given the close proximity of these layers and the 8-μm bin size of Visium HD, this likely reflects partial signal mixing, as also reported in tomo-seq studies25. To preserve interpretive accuracy, we omitted non-representative genes from downstream analyses and focused on biologically coherent marker sets. Pseudotime trajectory analysis To explore epithelial lineage trajectories, we applied Slingshot (v2.7.0) to a subset of clusters corresponding to the epithelial compartment. Dimensionality reduction was re-performed on this subset using Seurat’s RunPCA and RunUMAP functions. The resulting UMAP embeddings served as input for Slingshot to infer pseudotemporal trajectories across spatially defined epithelial states. Integration with single-cell reference Cell-type annotation of ST bins was performed using SingleR (v2.6.0), referencing a publicly available scRNA-seq dataset from human oral mucosa (GSE16424126). Annotations were assigned to each bin based on the highest-scoring reference match, facilitating the interpretation of spatial cellular heterogeneity. In total, 110,971 cells were mapped to 29 distinct cell types. Cell–cell communication analysis To investigate intercellular signaling within spatially resolved oral mucosa, we applied the CellChat R package (v.1.6.1) to infer ligand-receptor networks from ST data. The raw count matrix was filtered and normalized, and cell identity labels were assigned based on unsupervised clustering of ST profiles. Using CellChatDB human database, we identified key ligand-receptor pairs, estimated communication probabilities between epithelial and stromal clusters, and reconstructed putative signaling networks across the tissue. Results Spatial transcriptomic profiling of oral mucosa To comprehensively characterize the molecular architecture of healthy gingiva and non-keratinized lining oral mucosa, ST profiling was performed using the 10 × Genomics Visium HD platform. H&E-stained FFPE sections were prepared from oral mucosal tissue samples harvested around the MGJ of two individuals (Figs. 1A-B). Each section encompassed keratinized gingiva and non-keratinized lining mucosa, yielding four regions for analysis – two keratinized and two non-keratinized. Space Ranger analysis identified 16 spatially consistent clusters across both tissue samples (Figs. 1C-D, G), with cluster-specific differential gene expression results available in Table 1 and Supporting Information 2. For Seurat-identified cluster markers, a heatmap of top genes per cluster, and GO enrichment results, see Supporting Information 3–5. Table 1Top 10 enriched genes for each graph-based cluster.ClusterAnnotationTop 10 enriched genes1Deep stroma (lining mucosa)IGKC, COL1A1, COL1A2, VIM, COL3A1, TNXB, MFAP4, DCN, IGHA1, CFD2Immune-cell-rich stromaIGKC, IGHG1, IGHA1, COL3A1, COL1A1, COL1A2, CD74, IGHG3, VIM, IGLC13Superficial stroma (lining mucosa)VIM, TNXB, MFAP4, COL3A1, COL6A2, DCN, COL1A2, CFH, APOD, TMSB4X4Basal epithelial layerKRT5, COL17A1, KRT15, EEF2, CD44, DST, MT1X, DSC3, COL7A1, CXCL145Gingival stromaCOL1A1, COL1A2, COL3A1, SPARC, IGKC, COL6A2, COL6A1, AEBP1, MMP2, COL5A16Stratum granulosum (gingiva)S100A8, S100A9, KRTDAP, SPRR1B, KRT16, KRT1, PERP, SBSN, DSC2, KLK137Superficial layer (lining mucosa)SPRR3, MAL, EMP1, S100A9, CRCT1, KRT78, CRNN, ERO1A, CNFN, ECM18Upper stratum spinosum (lining mucosa)KRT4, S100A8, S100A9, SPRR3, RHCG, CRNN, ANXA1, MAL, CSTA, EMP19Lower stratum spinosum (lining mucosa)KRT4, S100A8, S100A9, RHCG, SPRR3, CSTA, ANXA1, SERPINB1, KRT13, SLPI10Basement membrane-adjacent stromaCD74, VIM, COL3A1, TMSB4X, COL6A2, APCDD1, COL6A1, COL17A1, LYZ, DCN11Suprabasal layer (lining mucosa)KRT4, GBP6, KRT13, S100A2, RHCG, CD9, LY6D, CSTA, CRABP2, TRIM2912Stratum intermedium (lining mucosa)SPRR3, MAL, EMP1, S100A9, CRNN, S100A8, RHCG, CNFN, TMPRSS11B, ANXA113Suprabasal layer (gingiva)KRT5, KRT16, DSP, AQP3, KRT6A, LY6D, DSC3, DSG1, CD44, FABP514Stratum spinosum (gingiva)KRT16, AQP3, DSP, KRT6B, KRT6A, KRT1, PERP, DSG1, FABP5, SFN15Endothelial compartmentVIM, AQP1, TMSB4X, A2M, CLU, CD74, SPARCL1, EGFL7, CCL14, PECAM116Stratum corneum (gingiva)LCE3D, RPTN, LOR, SPRR2G, SPRR2D, SPRR2E, CDSN, KRT80, CNFN, SPINK7Genes were selected from the graph-based cluster differential-expression output using positive log2 fold-change and adjusted p < 0.05, excluding mitochondrial genes. Top 10 enriched genes for each graph-based cluster. Genes were selected from the graph-based cluster differential-expression output using positive log2 fold-change and adjusted p < 0.05, excluding mitochondrial genes. Sixteen graph-based clusters, identified by Space Ranger, were manually annotated using histology and gene expression. Guided by tissue architecture and marker genes, clusters were classified as epithelial or stromal (Fig. 1E) and further subdivided into sub-compartments and layers (Fig. 1F), revealing clear separation of gingival and lining epithelia at the MGJ. Epithelial clusters were annotated based on widely accepted masticatory and lining mucosal layers3, with additional strata introduced to capture the finer resolution revealed by our spatial data (Fig. 1G). Cluster 4, adjacent to the stroma, was shared across gingiva and lining mucosa. Its markers included KRT5 and KRT15, common markers of basal epithelial cells24,27, and TOP2A (proliferation marker), supporting its annotation as the basal layer. Gingiva-specific epithelial clusters included Cluster 13 (suprabasal), Cluster 14 (stratum spinosum), Cluster 6 (stratum granulosum), and Cluster 16 (stratum corneum). Lining mucosa-specific epithelial clusters included Cluster 11 (suprabasal), Cluster 9 (lower spinosum), Cluster 8 (upper spinosum), Cluster 12 (stratum intermedium), and Cluster 7 (superficial layer). The stromal compartment was also subdivided into spatial zones. Cluster 10, the BM-adjacent stroma, was shared across gingiva and lining mucosa, and showed upregulation of COL17A1, a hemidesmosomal component associated with the epithelial-stromal interface28. Cluster 5, predominantly localized to the gingival lamina propria, was enriched in structural collagens (COL1A1, COL1A2) and annotated as gingival stroma. Clusters 1 and 3 represented deep and superficial stroma of non-keratinized mucosa, expressing VIM, LMNA, and TMSB4X. Cluster 2, which was present within the stromal compartment of both gingival and lining mucosa, was enriched in immunoglobulin transcripts, suggesting an immune-cell-rich stromal niche. Cluster 15, characterized by ENG, VWF, and PECAM1, represented vascular endothelium. Full annotations and distributions are shown in Fig. 1G. Although certain stromal clusters were classified as compartment-specific – gingival stroma (Cluster 5) and lining mucosa (Clusters 3 and 1) – our spatial data reveal that they are not strictly confined to one region. Cluster 5, while mainly gingival, extended into the lining mucosa, whereas lining-associated clusters 3 and 1 reached the adjacent gingiva. Unlike the sharply demarcated epithelial boundaries, stromal compartments thus display spatial overlap across the MGJ. Pseudotime trajectory analysis Next, to further dissect transcriptomic differences between lining mucosa and gingiva, we reconstructed epithelial trajectories using pseudotime analysis of spatially annotated clusters. Pseudotime revealed a sequential maturation process (Fig. 2A) originating from the shared basal layer (Cluster 4), consistent with its progenitor role. In gingiva, differentiation progressed through suprabasal, spinous, granular and corneum layers, with the granular layer (Cluster 6) showing heterogeneity in terminal differentiation. In the lining mucosa, maturation followed a smooth and continuous trajectory through suprabasal, spinous, intermedium, and superficial layers, reflecting gradual and coordinated non-keratinized differentiation. UMAP-based trajectory inference further demonstrated a clear bifurcation from the shared basal epithelium into two distinct lineages (Fig. 2B), indicating early divergence of keratinized and non-keratinized programs. Pseudotime analysis of epithelial differentiation in the oral mucosa. (A) Boxplots show the distribution of UMAP-based pseudotime values across spatially annotated epithelial clusters, inferred using the Slingshot algorithm. Tissue types are color-coded: green (shared basal), blue (gingiva), and red (lining mucosa). (B) UMAP visualization of Slingshot-inferred trajectories reveals bifurcated differentiation pathways originating from the shared basal cluster (green) toward gingival (blue) and lining mucosa (red) lineages. Spatial distribution of canonical keratins in gingival and lining epithelium We analyzed the spatial distribution of canonical keratins, used in histological classification of oral mucosa29. In gingival epithelium, KRT1 and KRT10 increased along the differentiation axis, peaking in the granular layer (Cluster 6) (Fig. 3A). Unexpectedly, KRT4 and KRT13 were also expressed in gingiva, at levels exceeding KRT10 and KRT1, indicating that these traditionally defined lining markers are not strictly restricted to non-keratinized mucosa. In contrast, lining epithelium showed striking KRT4 dominance across suprabasal layers, with lower KRT13 and negligible KRT1/10 (Fig. 3B), establishing KRT4 as the principal marker of the non-keratinized phenotype. Notably, this keratinization profile appears specific to gingiva and may not extend to other keratinized sites, such as the hard palate. Spatial expression patterns of canonical keratins across gingival and lining epithelial layers. Line plots display the mean expression levels of KRT1, KRT10 (markers of keratinized epithelium), and KRT4, KRT13 (markers of non-keratinized epithelium) across manually annotated epithelial layers in gingival (keratinized) (A) and lining (alveolar, non keratinized) (B) oral mucosa. Single-cell integration reveals regional epithelial and stromal heterogeneity To establish a high-resolution transcriptional framework, we leveraged a public single-cell RNA sequencing (scRNA-seq) dataset of human oral mucosa (GSE164241), published by Williams et al.26. A total of 110,971 cells were mapped across 29 cell types (Supporting Information 6). Bin-level annotations are provided in Supporting Information 7, and spatial distributions in Supporting Information 8. Williams’ dataset defined three epithelial subtypes: H.Epi 1.1 (basal keratinocytes), H.Epi 1.2 (gingiva-specific), and H.Epi 1.3 (expressing keratinization markers). As expected, our spatial analysis revealed that the shared basal layer (Cluster 4) in both gingiva and lining mucosa, corresponded to H.Epi 1.1, marked by KRT5, KRT14, and KRT15 (Figs. 4A-B). In contrast, H.Epi 1.3, despite its keratinization-associated signature, aligned more closely with non-keratinized lining mucosa (Figs. 4A–B). Contrary to the original classification of H.Epi 1.2 as “gingiva-specific,” our data reveal a 3.9-fold enrichment of this subtype in the suprabasal compartment of the lining mucosa (Fig. 4B), highlighting how ST reveals microenvironmental heterogeneity that may be obscured in dissociated single-cell analyses. (A) Spatial distribution of P.Epi 1–3 subtypes across the tissue sections; (B) Distribution of H.Epi 1.1–1.3 subtypes across epithelial graph-based clusters; (C) Distribution of H.Fib 1.1–1.5 subtypes across stromal clusters, separated by gingival and lining mucosa compartments. Stacked bars represent total counts with percentages of each H.Fib subtype. The unexpected detection of epithelial gene signatures within stromal clusters (Fig. 4A) likely reflect limited spatial resolution, with signal contamination from adjacent epithelium. In the reference scRNA-seq dataset by Williams (2021), fibroblasts were classified into five transcriptionally distinct subtypes (H.Fib 1.1–1.5), each associated with putative functional roles. Applying this classification to our spatial data revealed distinct subtype distributions across six stromal clusters, demarcated by gingival and lining compartments (Supporting Information 9). H.Fib 1.1 was consistently rare (< 10%), while H.Fib 1.4 and 1.5 dominated most regions (60–80%) (Fig. 4C). H.Fib 1.4 was strongly enriched in gingival stroma (78.7% in Cluster 5 vs. 23.9% in lining), while lining clusters such as Cluster 1 (deep stroma) and Cluster 3 (superficial stroma) were enriched in H.Fib 1.5 and H.Fib 1.3, respectively. These spatial patterns only partly aligned with predicted functions, suggesting anatomical context modulates fibroblast behavior beyond transcriptional identity. Cell–cell communication based on tissue compartments and graph-based clusters Inspired by classical transplant studies highlighting stromal regulation of epithelial fate7,8, we applied CellChat to explore how stromal-epithelial interactions shape keratinization mechanisms. For this, we performed a compartment-based analysis by subdividing each sample into four regions: (1) gingival epithelium, (2) gingival lamina propria, (3) lining mucosa epithelium, and (4) lining mucosa lamina propria (Supporting Information 10), to detect dominant signaling pathways across these structural domains. The first analysis revealed that only collagen and desmosome signaling networks significantly contributed to intercellular communication (Fig. 5A). Collagen acted in a paracrine fashion from gingival stroma (sender) to epithelium (receiver) (Fig. 5B), while desmosome signaling was confined to intra-epithelial interactions, where epithelial cells acted as both senders and receivers (Fig. 5C). Comparison of mean intercellular communication probability and tissue-compartment contributions for COLLAGEN and DESMOSOME signaling pathways, as computed by CellChat analysis. (A) Bar plot displaying the average communication probability for COLLAGEN (red) and DESMOSOME (cyan) pathways across all sender–receiver compartments in the tissue. The y-axis indicates the mean probability, reflecting the overall strength or frequency of predicted intercellular signaling for each pathway. (B–C) Bar plots depicting the relative contributions of individual tissue compartments, either as senders or receivers, to the COLLAGEN (B) and DESMOSOME (C) signaling networks. To refine spatial resolution and identify cellular sources and targets of these pathways, a second CellChat analysis was performed using the 16 graph-based clusters from the Space Ranger pipeline. Collagen signaling was predominantly active in gingival tissue (Figs. 6A-B), driven by stromal Cluster 5 expressing COL1A1, COL1A2, COL6A1, and COL6A2, which targeted epithelial receptors CD44 and SDC1 (Figs. 6C, G). COL1A1/2 → CD44 were the main interactions, while SDC1 played a minor role. Notably, the BM-adjacent stroma (Cluster 10) acted as a relay, transmitting signals from deeper stroma to the epithelium. COLLAGEN and DESMOSOME signaling across graph-based clusters, identified using the 10 × Genomics Space Ranger pipeline and annotated based on histological localization and gene expression profiles. (A, D) Chord diagrams visualizing intercellular communications mediated by COLLAGEN (A) and DESMOSOME (D) pathways, as inferred by CellChat from spatial transcriptomic data of the oral mucosa. Outer segments represent annotated cell clusters, while ribbons (chords) between segments denote significant ligand–receptor interactions. The direction of each ribbon indicates signal flow from sender to receiver, and ribbon width reflects the predicted signaling strength. (B, E) Schematic representations of COLLAGEN (B) and DESMOSOME (E) signaling in keratinized gingiva (left) and non-keratinized lining mucosa (right). Arrows indicate the direction and relative magnitude of the most significant signaling pathways, highlighting tissue-specific differences between gingiva and lining mucosa. (C, F) Relative contributions of the top ligand-receptor pairs in COLLAGEN (C) and DESMOSOME (F) signaling. Horizontal bars represent the strength of each ligand-receptor interaction. (G) Dot plot summarizing significant ligand–receptor interactions between source and target cell clusters, as inferred from CellChat analysis. The y-axis indicates the target cluster, while the x-axis shows the ligand–receptor pairs. Dot size reflects the interaction probability (prob), and dot color indicates the source cluster identity. The eight ligand-receptor interactions on the left correspond to COL1A1, COL1A2, and COL6A1/COL6A2 signaling through CD44 and SDC1 receptors (COLLAGEN signaling), while the two interactions on the right correspond to desmosomal adhesion pathways DSG1–DSC2 and DSG1–DSC3 (DESMOSOME signaling). In lining mucosa, collagen signaling was weak, limited to faint COL6A1-SDC1 interactions, with no evidence of stromal induction. Desmosome signaling (Figs. 6D-E) was robustly detected within the gingival epithelium, predominantly mediated by DSC2 → DSG1 and, to a lesser extent, DSG1 → DSC3 interactions (Figs. 6F-G). These ligand-receptor pairs drove strong communication across the suprabasal, spinosum, and granulosum clusters, supporting both vertical (top-down and bottom-up) and lateral (intra-layer) communication, thereby reinforcing epithelial polarity and structural adhesion. In contrast, the alveolar mucosa exhibited only weak desmosomal signaling, limited to clusters 9 (lower stratum spinosum) and 11 (suprabasal layer), lacking the organized stratification observed in gingiva. Discussion This study applied the 10 × Genomics Visium HD platform to generate a high-resolution ST atlas of human oral mucosa. Analysis revealed a layer-resolved architecture, where basal epithelium diverges into keratinized and non-keratinized programs, with gingival keratinization instructed by stromal collagen → CD44/SDC1 signaling and reinforced by desmosomal adhesion (DSG1-DSC2/3). These results align with a recent review, noting that periodontal phenotypes emerge from interconnected molecular and cellular networks rather than single genes or cell types30. Our Visium HD atlas of FFPE MGJ sections from two donors resolves 16 reproducible clusters that recapitulate oral mucosa architecture with finer detail of epithelial sublayers and stromal zones. The clear separation of gingival and lining epithelia, contrasted with graded, yet partially overlapping, stromal fields, supports a model of sharp epithelial boundaries within a more continuous mesenchymal landscape31. The subepithelial stroma (Cluster 10), enriched for BM-associated genes such as COL17A1, highlights the role of BM components in regulating epithelial organization. Indeed, specific BM collagens like COL4A6 and COL18A1 have been shown to promote oral keratinization11,12, underscoring the epithelial-stromal interface as a key signal-transmission hub. Within this niche, growth factors have been shown to influence epithelial fate. Transforming growth factor-beta (TGF-β) proteins regulate keratinization through the opposing actions of latent TGF-β binding proteins (LTBPs)32. Bone morphogenetic protein-2 (BMP-2) also emerges as a key regulator, contributing to oral mucosal keratinization9. In contrast, stromal mechanics can suppress keratinization: elastin-rich stroma promotes a non-keratinized fate, while elastase treatment in lining mucosa induces keratinization10. Pseudotime analysis identified two distinct epithelial lineages originating from a shared basal progenitor population. This early bifurcation suggests that regional fate decisions are made near the BM and not solely determined at later stages. These findings align with murine tomo-seq data from Seubert et al., which demonstrated that dorsal and ventral tongue epithelia, despite sharing a common transition zone, follow site-specific differentiation programs with distinct transcriptional profiles25. Our spatial analysis of canonical keratins refines the classic keratinized versus non-keratinized oral epithelium model29. As expected, KRT1/10 increased with differentiation in gingiva, but KRT4/13 were also detected, challenging their restriction to lining mucosa. In contrast, lining epithelium showed strong KRT4 dominance with minimal KRT1/10 and lower KRT13, confirming KRT4 as a hallmark of non-keratinized phenotype. These findings suggest keratin expression is more nuanced than a binary model, highlighting the need to study other keratinized sites, such as the hard palate. Although spatial transcriptomics revealed unexpected KRT4 transcript expression within keratinized gingival epithelium, protein-level validation is required to confirm this observation. Integration with the Williams et al. scRNA-seq reference revealed both concordance and divergence with intact tissue architecture26. Basal keratinocytes (H.Epi 1.1) mapped consistently to the basal layer, whereas H.Epi 1.3 aligned with non-keratinized mucosa and H.Epi 1.2, previously described as gingiva-specific, was enriched in the lining epithelium. These results show that transcriptomic identity alone cannot reliably predict keratinization, highlighting the need for spatial context, as noted in recent interactomics studies25,27,33. Fibroblast subtypes also showed distinct regional distributions, with H.Fib 1.4 enriched in gingiva and H.Fib 1.5 in lining mucosa, underscoring how anatomical context shapes stromal and epithelial specialization. Similar spatial heterogeneity was noted by Caetano et al. in periodontitis, where stromal states depended on anatomical localization and immune context34. Importantly, while Williams et al. provided a foundational single-cell atlas of human oral mucosa, their study primarily focused on cellular heterogeneity and immune-stromal interactions related to periodontal inflammation. In contrast, the present study focuses on the MGJ as an anatomical transition zone between keratinized gingiva and non-keratinized lining mucosa. By applying high-resolution spatial transcriptomics to intact tissue sections, we preserved the spatial relationships between epithelial layers and underlying stromal compartments, enabling direct assessment of regional keratinization programs and epithelial–connective tissue crosstalk. Thus, the Williams et al. epithelial subtypes were used here as an external reference framework, whereas our principal conclusions arise from spatially resolved clustering, histology-guided annotation, pseudotime trajectory analysis, keratin expression mapping, and ligand–receptor inference. Our cell–cell communication analysis revealed a coordinated mechanism of gingival keratinization. Collagen was the dominant stromal signal, with COL1A1, COL1A2, COL6A1, and COL6A2 engaging CD44 and SDC1 in basal and early suprabasal layers. Within the epithelium, a desmosome interaction network (DSG1-DSC2/3) dominated, peaking in suprabasal and granular layers. Together, collagen signaling provides the stromal instruction, while the desmosomal network functions as the execution module, stabilizing epithelial architecture to permit the accumulation of late keratins and cornified-envelope proteins. While Seubert et al. (2024) identified site-specific fibroblast growth factor (FGF) ligands as stromal drivers of epithelial differentiation in the murine tongue25, our human gingival data point to a distinct, yet complementary mechanism, suggesting that FGF signals may act in parallel with the collagen axis. Previous studies propose a model in which BM components and an epithelium-adjacent stroma, delivering BMP-2 and modulating TGFβ, initiate keratinization, whereas elastin-rich stroma in lining mucosa counteracts it9–12,32. Integration with our data suggests that basal cells are first licensed to differentiate, a process reinforced by stromal collagen signaling and stabilized by intraepithelial desmosome adhesion within the gingival layers. Recent Visium-HD data from Zang et al. (2025) provide a disease-state comparator to our healthy gingival map24. Zang’s study reveals that smoking amplifies keratinization programs, including up-regulation of DSG1, which aligns directly with our desmosomal observations, suggesting that environmental stressors may drive aberrant, hyperkeratinizing programs. While this study leverages state-of-the-art ST, several limitations should be acknowledged. The analysis is constrained by the small number of specimens and by the resolution of Visium HD, which, although advanced, does not achieve true single-cell resolution. In addition, functional validation of the inferred collagen and desmosome pathways remains essential. Larger cohorts, higher-resolution imaging, and organoid models will be needed to confirm these findings. In summary, this study provides a high-resolution spatial transcriptomic analysis of human oral mucosa across the MGJ, revealing distinct yet spatially connected epithelial and stromal compartments in keratinized gingiva and non-keratinized lining mucosa. Pseudotime analysis demonstrated that epithelial differentiation diverges from a shared basal layer into separate keratinized and non-keratinized trajectories. Spatial mapping of canonical keratins further showed that oral epithelial keratin expression is more nuanced than the traditional binary classification, including unexpected KRT4 transcript expression within keratinized gingival epithelium. Finally, ligand-receptor analysis identified stromal collagen signaling and intraepithelial desmosome interactions as prominent features of the keratinized gingival compartment. Together, these findings refine our understanding of regional oral mucosal specialization and provide a spatial framework for future studies aimed at validating and therapeutically modulating gingival keratinization. Supplementary Information Supplementary Information 1: Space Ranger Web summary. Supplementary Information 2: Differential gene expression between Graph Based clusters. Supplementary Information 3: Seurat cluster markers. Supplementary Information 4: Heatmap displaying the expression patterns of the top positively enriched marker genes identified for each cluster. Marker gene selection was based on results from differential expression analysis using Seurat. Supplementary Information 5: DotPlot presenting the top 5 enriched GO Biological Process terms across the graph-based spatial transcriptomic clusters (x-axis). Dot size indicates the number of genes associated with each term, while color represents the adjusted p-values (Benjamini-Hochberg corrected, log₁₀ scale). Supplementary Information 6: Cell type distribution of spatial transcriptomic bins annotated using a single-cell reference. Spatial bins were annotated using SingleR and a reference human oral mucosa scRNA-seq dataset (GSE164241; Williams et al., 2021). Each bar represents a distinct cell type, ordered by frequency and color-coded to match the accompanying legend. The legend includes the absolute number of bins and their percentage of the total dataset. Supplementary Information 7: Bin Cell Type Annotations. Supplementary Information 8: Spatial mapping of annotated cell types in oral mucosa. Spatial localization of 29 distinct cell populations was inferred by integrating spatial transcriptomic data with a reference single-cell RNA-sequencing dataset of human oral mucosa (GSE164241; Williams et al., 2021). Cell types were annotated using the SingleR algorithm and visualized across spatial tissue sections. Each panel represents the spatial distribution of a single annotated cell population, with yellow overlay indicating predicted localization based on transcriptomic similarity. Supplementary Information 9: Detailed H.Fib1.1-1.5 distribution. Supplementary Information 10: Manual annotation of spatial transcriptomic sections into epithelial and lamina propria regions based on histological features and graph-based clustering, resulting in four anatomical compartments: (1) gingival epithelium, (2) gingival lamina propria, (3) lining mucosa epithelium, and (4) lining mucosa lamina propria. Author contributions All the authors contributed substantially to the manuscript. Data availability Spatial transcriptomics data have been submitted to EMBL-EBI BioStudies ( [https://www.ebi.ac.uk/biostudies/studies/S-BSST929] (https://www.ebi.ac.uk/biostudies/studies/S-BSST929)) and will be made publicly available upon publication. Declarations Competing interests The authors declare no competing interests. Ethical approval This study was approved by the Institutional Ethics Committee of Tel Aviv University (IEC No. 0007945–1). Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement Spatial transcriptomics data have been submitted to EMBL-EBI BioStudies ( [https://www.ebi.ac.uk/biostudies/studies/S-BSST929] (https://www.ebi.ac.uk/biostudies/studies/S-BSST929)) and will be made publicly available upon publication.