Spatially resolved proteomics provides insights into the normal-dysplasia-tumor continuum in oral squamous cell carcinoma

Abstract

Background

Oral epithelial dysplasia is a recognized precursor lesion for oral squamous cell carcinoma (OSCC), yet the proteomic changes underlying the progression from normal epithelium to dysplasia and ultimately to OSCC remain incompletely characterized. This gap is partially due to the high interpatient heterogeneity of OSCC, the frequent reliance on cross-patient comparisons, and the omission of dysplasia as a critical intermediate stage.

Methods

A histology-guided, spatially resolved proteomics workflow was applied to formalin-fixed paraffin-embedded tissue blocks from six patients with primary OSCC. For each case, spatially contiguous normal (N), dysplasia (D), and tumor (T) regions from the same lesion-bearing block, yielding 18 regional proteomes. Proteins were profiled using data-independent acquisition liquid chromatography-tandem mass spectrometry. Downstream analyses included differential expression analysis, Gene Ontology and Kyoto Encyclopedia of Genes and Genomes enrichment, soft-clustering trajectory analysis, protein-protein interaction analysis, exploratory clinicopathological stratification, and immunohistochemistry validation.

Results

The spatially matched N–D–T dataset identified 1,683 proteins in N, 2,100 in D, and 1,966 in T. Under the predefined differential-expression threshold, the N vs. D comparison yielded more differentially expressed proteins than the D vs. T comparison, suggesting that threshold-detectable proteomic alterations may already be present at dysplasia. Functionally, the N-to-D transition was associated with cell-cycle activation, proliferative signaling, hypoxia-associated responses, and metabolic alterations, whereas the D-to-T transition was more closely related to cytoskeletal organization, motor activity, and structural remodeling. Across the overall N-to-T contrast, glycolysis, pyruvate metabolism, and broader metabolic reprogramming predominated. Trajectory analysis resolved Early Warning, Progression, and dysplasia-restricted transient modules, suggesting a non-linear architecture of proteomic evolution during OSCC progression. MYO1B showed progression-associated spatial increase from N to D and T regions, whereas HSP90AB1 was prioritized as a candidate invasion-associated hub supported by proteomic, protein-protein interaction network, and external transcriptomic analyses.

Conclusions

Spatially resolved proteomics of archived formalin-fixed paraffin-embedded tissues provides a morphology-anchored proteomic continuum of oral malignant transformation in a small spatially matched cohort. The data suggest that OSCC progression involves early-shift, progressive, and transient dysplasia-specific programs. These findings offer candidate proteins and biological hypotheses for further validation in larger oral cancer cohorts.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12903-026-09284-5.

Background

Oral squamous cell carcinoma (OSCC), the predominant subtype of head and neck squamous cell carcinoma (HNSCC), remains a major global health burden. Cancers of the lip and oral cavity accounted for 389,485 new cases and 188,230 deaths worldwide despite advances in surgery, radiotherapy, and immune checkpoint blockade [1, 2]. This persistent burden underscores the significance of identifying lesions at risk of progression at an earlier stage, rather than relying solely on treatment intensification to improve outcomes. In this context, oral potentially malignant disorders (OPMDs) are clinically important because a substantial proportion of OSCC arises from clinically or histologically altered mucosa [3, 4]. Histologically defined oral epithelial dysplasia (OED) within OPMDs is likewise associated with an increased risk of malignant transformation [5].

Morphological change is often preceded by molecular alteration, making it important to define biomarkers across the normal–dysplasia–tumor (N–D–T) continuum [6]. Although genomic, transcriptomic, and epigenetic studies have provided important insights into oral carcinogenesis [7–9], these molecular layers do not always capture the functional state of tissues during disease evolution. Proteins, by contrast, are the most proximal effectors of phenotype and represent a major class of actionable therapeutic targets [10, 11]. Proteomics therefore offers a more functionally informative framework for characterizing progression-associated molecular reprogramming. Large-scale proteogenomic studies in HNSCC have shown that proteomic and phosphoproteomic states can uncover pathway activity and therapeutic vulnerabilities that are not evident from genomic or transcriptomic data alone [12]. In oral cancer–related settings, both salivary and tissue-based proteomic studies have further demonstrated the feasibility of biomarker discovery [13–16].

Nevertheless, the proteomic landscape spanning the continuous transition from OPMD to OSCC remains insufficiently characterized. Most available studies compare tumor and normal tissues across different individuals or focus only on extreme states, such as tumor versus adjacent normal tissue, while omitting the intermediate dysplastic state [13, 17, 18]. This limitation is compounded by the marked inter-patient and intra-lesional heterogeneity of OSCC and its precursor lesions [19, 20], which reduces the consistency and interpretability of currently proposed molecular markers. Moreover, although spatially resolved mass spectrometry (MS) combined with histology-guided microdissection enables precise protein profiling of specific formalin-fixed paraffin-embedded (FFPE) lesional regions [21], it has mainly been used to dissect biologically distinct tumor compartments rather than to reconstruct the sequential trajectory of malignant transformation [22]. Therefore, a histology-anchored, within-patient proteomic trajectory that incorporates dysplasia as the bridge between normal mucosa and carcinoma is needed.

In this study, we applied spatially resolved proteomics to FFPE tissue slides from OSCC patients and systematically sampled morphologically N, D, and T regions that were spatially contiguous within the same lesion-bearing tissue block. This design enabled delineation of a morphology-anchored proteomic trajectory of OSCC progression while reducing inter-patient confounding. It also allowed the identification of candidate molecules with potential for early risk stratification, early detection, and therapeutic targeting.

Materials and methods

Human participants and sample preparation

Six patients who underwent primary surgical resection for OSCC at the National Center of Stomatology in 2023 were included. Inclusion criteria were as follows: (i) histopathologically confirmed OSCC contained spatially contiguous N, D, and T regions within the same FFPE tissue block by experienced pathologists; and (ii) sufficient tissues available for serial sectioning and spatial sampling. Patients who had received neoadjuvant chemotherapy and/or radiotherapy before surgery were excluded. Detailed clinicopathological characteristics, including age, sex, lesion site, tobacco smoking, alcohol consumption, TNM stage, clinical stage, pathological grade, follow-up duration, and disease status at the last follow-up, are summarized in Table 1. Recurrence was defined as clinically, radiologically, or pathologically confirmed tumor relapse during follow-up. No evidence of disease (NED) was defined as the absence of clinically detectable recurrence at the last follow-up.

Table 1.

Clinicopathological characteristics of the patients

	P1	P2	P3	P4	P5	P6
Age (years)	47	58	59	53	63	60
Sex (M/F)	M	F	M	F	F	M
Site	tongue	gum	tongue	tongue	tongue	tongue
Tobacco smoking	0	0	1	0	0	1
Alcohol consumption	0	0	1	0	1	1
T stage	4	2	4	1	2	2
N stage	2	2	0	0	1	0
M stage	0	0	0	0	0	0
Clinical Stage	IV	IV	IV	I	III	II
Pathological grade	II	II	II	I	II	I-II
Follow-up duration (months)	39	38	11	39	39	38
Disease status at last follow-up	NED	NED	Recurrence	NED	NED	NED

Recurrence was defined as clinically, radiologically, or pathologically confirmed tumor relapse during follow-up. NED was defined as the absence of clinically detectable recurrence at the last follow-up

M male, F female, NED no evidence of disease

Tobacco smoking: 1, smoker, 0, non-smoker; Alcohol consumption: 1, alcohol user, 0, non-user

For each FFPE block, regions of interest were annotated under pathological guidance before sampling. Epithelial areas were preferentially selected to improve regional specificity. Obvious non-target areas, including necrotic, stromal-rich, and inflammatory regions, were avoided where possible.

Sample preparation followed the previously described ProteomEx workflow [23], including protein anchoring, hydrogel embedding, ~ 6× linear expansion, and Coomassie staining. Briefly, regions of interest were first identified on spatially contiguous tissue sections and marked using laser capture microdissection (LCM)-guided burning marks before tissue expansion. After expansion and re-imaging, one 1-mm punch sample was excised from each representative N, D, and T region, yielding a total of 18 spatially resolved samples for liquid chromatography–tandem mass spectrometry (LC–MS/MS) analysis. The overall study workflow is shown in Fig. 1A. This study was approved by the Ethics Committee of Peking University School and Hospital of Stomatology. All samples and clinical data were obtained with informed consent, and all procedures were conducted in accordance with the Declaration of Helsinki.

Fig. 1.

Spatially-resolved proteome profiling of histopathologically N, D, and T regions in OSCC. A Study workflow. Spatially contiguous N-D-T regions were collected from OSCC cases using LCM-guided marking followed by expansion-assisted punch sampling. Proteomes were measured by data-independent acquisition-based LC–MS/MS. B Venn diagram showing the overlap of proteins identified in the N (n = 1683), D (n = 2100), and T (n = 1966) regions. C Pearson correlation matrix. D Heatmap of protein expression across the three regions. E-G Cumulative protein abundance distributions in the N (E), D (F), and T (G), together with the number of proteins within each abundance quantile (Q1: 0–25%, Q2: 25–50%, Q3: 50–75%, Q4: 75–100%). OSCC, oral squamous cell carcinoma; OED, oral epithelial dysplasia; LC–MS/MS, Liquid Chromatography-Tandem Mass Spectrometry; N, normal epithelium; D, dysplasia; T, tumor

Spatial proteomics sequencing and DIA data processing

Peptides obtained from spatially resolved punch samples were analyzed on a nanoElute^® system coupled online to a timsTOF Pro mass spectrometer (Bruker Daltonics, Bremen, Germany) equipped with a CaptiveSpray nano-electrospray ion source. All 18 samples were prepared and analyzed within the same experimental batch using the same LC–MS/MS platform and matched chromatographic gradients. Mobile phase A consisted of water with 0.1% formic acid (FA), and mobile phase B consisted of acetonitrile with 0.1% FA; all solvents were LC–MS grade. Data were acquired in data-independent acquisition (DIA) mode using Bruker otofControl v6.2 and HyStar v5.1 [24]. The ion mobility range was scanned from 0.7 to 1.3 1/K0 with a ramp time of 166 ms. MS1 and MS2 spectra were acquired over an m/z range of 100–1700.

A project-specific spectral library was generated from pooled peptide digests prepared from whole serial tissue sections adjacent to those used for spatial sampling from the same patient specimens. Total proteins were extracted from the whole-slide tissue sections, quantified, denatured, reduced, alkylated, and digested with trypsin. The resulting peptides were pooled and fractionated by high-pH reversed-phase chromatography into 30 fractions. These fractions were then analyzed by DIA on the same timsTOF Pro mass spectrometer to generate the spectral library.

Spectral library construction and DIA data quantification were performed using DIA-NN (v1.8.1) against the UniProt human reference proteome database (UP000005640, release 2023-07-25; 20,456 reviewed canonical proteins). A separate contaminant FASTA file containing common laboratory protein contaminants was appended to the human database for searching. Library-based matching was enabled, and match-between-runs was activated to improve peptide identification consistency across the DIA runs. Precursor- and protein-level false discovery rates were controlled at 1%. All remaining DIA-NN analytical parameters were retained at the software default settings. The resulting project-specific library contained 8,965 protein isoforms, 8,965 protein groups, and 102,528 precursors in 88,559 elution groups. Protein-level quantitative matrices exported from DIA-NN were used for downstream bioinformatic analyses. Quantitative values were log2-transformed and standardized by z-score normalization before differential expression and trend-based analyses. The mass spectrometry proteomics data have been deposited to the ProteomeXchange Consortium via the iProX partner repository [25, 26] with the dataset identifier PXD078064.

Statistical and bioinformatic analysis

Statistical and bioinformatic analyses were performed in R (v4.0). Differential expression analysis was conducted for N vs. D, D vs. T, and N vs. T using the limma package, with significance thresholds of P < 0.05 and |fold change| > 1.5. Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses were performed using the clusterProfiler package. One-way ANOVA was used to identify differentially expressed proteins (DEPs) across the three groups. The Mfuzz package was then used to cluster DEPs according to their expression trends. The resulting Mfuzz clusters were further summarized into trajectory modules based on their quantitative cluster-center profiles across the three stages.

To complement differential-expression analysis, intra-patient root mean square (RMS) distances were calculated using all quantified proteins between matched N-D, D-T, and N-T proteomic profile. This analysis provided an effect-size-oriented assessment of overall proteomic remodeling across the N-D-T continuum. Together with principal component analysis (PCA) trajectory visualization, it was further applied in an exploratory stratified analysis according to clinical stage (Stage I–II vs. Stage III–IV) and lymph node (LN) status (LN0 vs. LN+). These analyses were intended to assess potential associations between clinicopathological features and the magnitude or timing of proteomic remodeling.

For the secondary analysis focused on invasion status, all preprocessing steps, including missing-value imputation, were performed across the full sample set before the construction of the combined non-invasive epithelial reference group (ND; N combined D). This avoided group-specific preprocessing between ND and T. The subsequent limma-based differential analysis accommodated unequal group sizes between ND group (n = 12) and T group (n = 6). Protein–protein interaction (PPI) networks were constructed using STRING (v12.0) and visualized in Cytoscape (v3.10.4). Hub proteins were identified using cytoHubba (MCC, top 10) and MCODE with default settings.

Immunohistochemistry validation

Immunohistochemistry (IHC) was performed on 3-µm FFPE sections from five independent cases containing spatially contiguous morphologically N, D, and T regions to assess the spatial expression patterns of MYO1B. Corresponding hematoxylin and eosin (H&E)-stained serial sections were reviewed to confirm the histopathological regions.

Sections were deparaffinized, rehydrated, and subjected to antigen retrieval in EDTA buffer (pH 9.0) using a pressure cooker for 2–3 min. Sections were incubated overnight at 4 °C with a rabbit monoclonal anti-MYO1B antibody (Abcam, ab194356, clone EPR16223; 1:500 dilution). Immunodetection was performed using the GTVision™ Detection System III (including DAB)/Mo&Rb kit (GK500705) according to the manufacturer’s instructions. DAB was used as the chromogen for 30–40 s, and hematoxylin was used for counterstaining. Negative controls were prepared by omitting the primary antibody.

MYO1B staining was evaluated by pathologists based on staining localization, intensity, and distribution across N, D, and T regions. To complement pathological assessment, semi-quantitative ImageJ (version 1.54 g) analysis was performed using integrated optical density (IOD), which captures both staining intensity and positive staining area. For each case, three non-overlapping 20× fields were analyzed per region under identical threshold settings, and the total IOD of the entire field was recorded. The mean IOD of the three fields was used as the case-level value for each region. Differences among paired N, D, and T regions were assessed using the Friedman test.

Results

Spatially-resolved proteome profiling across histopathologically defined N, D, and T regions in OSCC

Spatially contiguous regions were collected from six OSCC cases. In each case, three adjacent regions, histopathologically defined as N, D, and T, were selected under pathological guidance and collected using LCM-guided marking followed by expansion-assisted punch sampling, yielding 18 regional proteomes. Representative images showing the annotated H&E-stained regions, LCM-guided burning marks, and sampled regions after tissue expansion are provided in Supplementary Figure S1. Proteomes were then measured by DIA-based LC–MS/MS to generate spatially resolved profiles across disease progression (Fig. 1A). Across all cases, 1683 proteins were identified in N, 2100 in D, and 1966 in T (Fig. 1B).

Pearson correlation analysis and heatmap visualization showed higher inter-sample consistency in N, whereas D and T displayed greater heterogeneity. The dispersion was particularly evident in D, indicating possible molecular instability in this stage (Fig. 1C, D). Proteins were then ranked by abundance contribution and stratified into quartiles (Q1: 0–25%, Q2: 25–50%, Q3: 50–75%, Q4: 75–100%) (Fig. 1E–G). High-abundance proteins (Q1–Q2) were dominated by epithelial keratins, including KRT1, KRT9, KRT6A, KRT14, KRT10 across N, D, and T. Lower-abundance proteins (Q3–Q4) included inflammatory proteins (S100A8 and S100A9), desmosome-related factors (DSC3, PKP1, and SERPINB3), the secreted antimicrobial protein (DCD), cytoskeleton-, adhesion-, and motility- related proteins (ACTB, FLNB, and PCDHGC4), nuclear proteins (LMNA, H2AC8, H2BC12, and H1-4), translation-related proteins (RPL8), and metabolic enzymes such as LDHA. These region-resolved identification, correlation, and abundance patterns supported the analytical quality of the dataset for downstream analyses.

An exploratory stratified analysis by clinical stage and LN status was further performed to explore the clinical relevance of overall spatial proteomic changes (Supplementary Figure S2). The intra-patient RMS distance was calculated using all quantified proteins to estimate the magnitude of proteomic shifts across the N-D, D-T, and N-T intervals (Supplementary Table S1). PCA trajectory visualization and RMS distance analysis suggested possible differences in the timing of proteomic remodeling across clinical subgroups (Supplementary Figure S2A, B). Patients with advanced-stage disease (Stage III–IV) and LN+ status tended to show larger proteomic shifts during the N-to-D transition, whereas early-stage (Stage I–II) and LN0 cases appeared to show relatively greater remodeling during the D-to-T transition (Supplementary Figure S2C, D). These observations suggest that the timing and magnitude of proteomic remodeling along the N-D-T continuum may vary across clinical subgroups. However, because of the limited sample size, these findings should be interpreted as exploratory and require validation in larger independent cohorts.

Stage-resolved differential proteomics delineates early detectable alteration and later transformation-associated remodeling

Stage-resolved pairwise comparisons across N, D, and T samples were performed to characterize proteomic alterations during progression (Fig. 2 and Supplementary Figure S3A). In the N vs. D comparison, 46 DEPs were identified, including 41 upregulated and five downregulated proteins in D relative to N (Fig. 2A). Upregulated proteins included RNA-binding factors such as RBMXL2, ELAVL1, and KHSRP, together with mitochondrial metabolic enzymes such as NDUFS2 and PDHB. The differentiation-associated protein CRNN was among the downregulated proteins. Functional enrichment of the N-to-D DEP set highlighted terms related to proliferative regulation and hypoxia-associated signaling, including “positive regulation of cell cycle G1/S phase transition” and “HIF-1 signaling pathway”. “Pyruvate metabolism” and “central carbon metabolism in cancer” were also enriched (Fig. 2D, E).

Fig. 2.

Stage-resolved pairwise comparisons across normal epithelium (N), dysplasia (D), and tumor (T) regions. A-C Volcano plots showing differentially expressed proteins in the comparisons of N vs. D (A), D vs. T (B), and N vs. T (C). Red and blue dots indicate significantly upregulated and downregulated proteins, respectively, and selected proteins are labeled. D, F, H Top enriched GO terms, including biological process (BP), cellular component (CC), and molecular function (MF), for each comparison. E, G, I Top enriched KEGG pathways in each comparison. GO, Gene Ontology; KEGG, Kyoto Encyclopedia of Genes and Genomes

The D vs. T comparison showed a more focused DEP profile. 12 DEPs were identified, including six upregulated and six downregulated proteins in T relative to D (Fig. 2B). Upregulated proteins included MYL6B and ATP6V1D, whereas ADH1B was among the downregulated proteins. Enrichment analyses prioritized cytoskeleton-associated categories, including “cytoskeletal motor activity” and “myosin complex” (Fig. 2F, G).

The overall N vs. T comparison identified 85 DEPs, with 84 upregulated and one downregulated protein in T relative to N (Fig. 2C). Upregulated proteins included metabolism-associated factors such as ACOT7 and ATP6V0A1, as well as motility- and epithelial-mesenchymal transition - related proteins including FLOT1, MARCKS, and PGRMC2. KRT13 was the sole downregulated protein in this comparison. Enrichment analyses highlighted central carbon metabolism-related terms and pathways, including “pyruvate metabolic process” and “glycolysis/gluconeogenesis” (Fig. 2H, I).

Overall, stage-resolved comparisons delineated a progression pattern in which the N-to-D transition was associated with cell-cycle activation, hypoxia-associated signaling, and metabolic change. The D-to-T transition was characterized by cytoskeletal and transporter-related functions, whereas the global N-to-T contrast was dominated by glycolytic and pyruvate-associated reprogramming, together with reduced KRT13 expression. Complete lists of DEPs and GO/KEGG enrichment results are provided in Supplementary Table S2-S4.

Additionally, because DEP counts may be influenced by statistical thresholds, effect-size distribution, and inter-sample heterogeneity, patient-level proteomic magnitude was further examined using intra-patient RMS distances (Supplementary Table S1). Although the N vs. D comparison yielded more DEPs than the D vs. T comparison under the predefined threshold, RMS distances showed comparable and heterogeneous proteomic distances between the two transitions. Three patients showed a larger N-to-D magnitude, whereas the remaining three showed a larger D-to-T magnitude. At the cohort level, the mean RMS distance was 5.616 for N-to-D, 5.992 for D-to-T, and 5.280 for N-to-T, with corresponding median values of 5.571, 6.232, and 5.356, respectively. Thus, DEP number alone should not be taken as a direct measure of transition strength.

Proteomic dynamics reveal distinct evolutionary trajectories during OSCC progression

To capture dynamic trends across the N-D-T continuum, one-way ANOVA was performed across the three stages. This analysis identified 59 proteins with significant differential expression across the total proteome, defining a trajectory-informative protein set for pattern discovery (Supplementary Table S5). The corresponding heatmap showed separation across the N, D, and T phenotypes (Fig. 3A).

Fig. 3.

Distinct evolutionary trajectories during oral squamous cell carcinoma progression across the N-D-T continuum. A Heatmap of the 59 proteins significantly differentially expressed across the three regions identified by one-way ANOVA. B Mfuzz soft clustering identified eight distinct clusters within the trajectory-informative protein set. C Four trajectory modules grouped based on their Mfuzz cluster-center profiles and quantitative N–D–T trajectories. Left, representative expression trends of each module across N, D, and T; colored lines indicate individual proteins, and the black line indicates the module centroid. Middle, heatmaps showing module-specific protein expression patterns across the three groups. Right, bar plots showing relative expression changes within each module, calculated as log2(T/N) for Pg and log2(D/N) for the remaining modules. N, normal epithelium; D, dysplasia; T, tumor; Pg, Progression Module; EW, Early Warning Module; DsP, Dysplasia-Specific Peak Module; DsV, Dysplasia-Specific Valley Module

Mfuzz soft clustering was then applied to these proteins to capture progression patterns beyond static pairwise comparisons. Eight clusters were identified, containing 8, 13, 12, 3, 10, 2, 9, and 2 proteins, respectively, each with a distinct trajectory along the N-D-T axis (Fig. 3B and Supplementary Figure S3B-F). To improve biological interpretability, clusters with similar dynamic patterns were grouped into trajectory modules based on their Mfuzz cluster-center profiles and quantitative N–D–T trajectories (Fig. 3C and Supplementary Table S6). Clusters 1, 3, 4, and 5 showed increasing cluster-center trajectories from N to D and then to T and were grouped as the Progression Module (Pg, n = 33). This pattern was consistent with progressive accumulation of lesion-associated proteomic alterations. Cluster 2 showed a marked increase at D, with the cluster center remaining elevated in T relative to N, and was defined as the Early Warning Module (EW, n = 13). Dysplasia-restricted transient dynamics were represented by the Dysplasia-Specific Peak Module (DsP, cluster 7, n = 9), whose cluster center peaked at D, and the Dysplasia-Specific Valley Module (DsV, clusters 6 and 8, n = 4), whose cluster centers reached their lowest values at D.

MYO1B was selected as a representative progression-associated candidate for IHC validation because it was assigned to cluster 5 and the Pg module, together with its relevance to cytoskeletal remodeling, adhesion-related processes, and OSCC biology. Previous work has suggested potential involvement of MYO1B in oral carcinogenesis [8], but protein-level spatial evidence across the N–D–T transition remained limited. Results showed absent or very weak MYO1B staining in normal epithelium, increased staining in dysplasia, and more extensive positivity in tumor regions (Fig. 4A-C). Staining in dysplastic epithelium was heterogeneous, whereas tumor regions showed broader staining across epithelial tumor components. To further display regional staining patterns, representative high-magnification paired H&E and MYO1B IHC images of N, D, and T regions from two independent cases are provided in Supplementary Figure S4. Semi-quantitative ImageJ-based IOD analysis was further performed. Consistent with the proteomic trajectory, MYO1B IOD values increased from N to D and T regions. Overall differences among paired N, D, and T regions were significant by Friedman test (χ² = 10.000, P = 0.0067) (Fig. 4D). These findings support the association between MYO1B protein expression and histopathological progression across spatially contiguous N-D-T regions, although further validation in larger cohorts is required to determine its clinical utility.

Fig. 4.

MYO1B immunohistochemical staining and module-associated functional activities across the N–D–T continuum. A-C Representative high-magnification MYO1B immunohistochemical staining of N (A), D (B), and T (C) regions from an independent case containing spatially contiguous N–D–T tissues. The image was captured at 20× magnification. Scale bar, 50 μm. D Semi-quantitative ImageJ-based IOD analysis. Each point represents one case-level value, and lines connect paired regions from the same case. E Heatmap showing the activities of module-associated functional terms across samples along the N–D–T continuum. Functional terms were assigned to four modules, including Early Warning (EW), Progression (Pg), Dysplasia-Specific Peak (DsP), and Dysplasia-Specific Valley (DsV). For each term, an activity score was calculated as the mean protein-wise Z-score of proteins within that term and displayed across all samples. Terms were derived from GO biological process enrichment for protein-rich modules (Pg and EW, P < 0.05) and complemented by descriptive protein-level annotations for low-count modules (DsP and DsV); all terms were consolidated into 11 functional themes. N, normal epithelium; D, dysplasia; T, tumor

Functional enrichment reveals module-linked programs along the N–D–T continuum

To characterize functional changes across the N–D–T continuum, enriched biological terms and protein-level annotations were mapped to the four modules and visualized across subregions (Fig. 4E and Supplementary Tables S7, 8). For Pg and EW, which contained sufficient protein numbers, functional interpretation was based on enrichment analysis. For the smaller DsP and DsV modules, formal GO enrichment was considered underpowered. These modules were therefore interpreted using descriptive protein-level annotation based on database and literature evidence, with the annotation basis provided in Supplementary Table S8.

EW was characterized by an early functional shift that separated normal tissue from lesional regions (Fig. 4E, blue). This pattern involved coordinated changes in RNA metabolism and translational control and cytokine signaling, including terms related to translational initiation and cellular response to interleukin-7. Early changes in metabolic reprogramming and genome organization and maintenance further suggested departure from epithelial homeostasis at the dysplasia stage.

Proteins assigned to the Pg followed a stepwise increase along the N–D–T axis (Fig. 4E, red). This trajectory was associated with broader metabolic reprogramming and genome organization and maintenance, including glucose metabolic process, hexose metabolic process, and nuclear envelope organization. These changes occurred alongside vesicle-mediated transport, represented by Golgi vesicle transport, and innate immune defense and host–microbe interaction, represented by defense response to Gram-negative bacterium. Structural programs related to cell adhesion, junction remodeling, and invasion further contributed to the Pg phenotype.

DsP captured a dysplasia-associated peak-like pattern (Fig. 4E, yellow). Protein-level annotation suggested that DsP included factors related to growth-associated transcription/translation, RNA regulation, cytoskeletal remodeling, and stress-adaptive or pro-survival responses. These included proteins involved in MYC-related regulation, RNA binding/stability, spindle orientation, and anti-apoptotic stress responses.

DsV delineated a dysplasia-associated valley pattern (Fig. 4E, purple). Proteins in DsV were related to epithelial differentiation, polarity, cytoskeletal architecture, and barrier-associated functions, suggesting transient attenuation of epithelial structural integrity during dysplasia.

Proteomic features distinguishing non-invasive epithelium from invasive OSCC

As a secondary analysis focused on invasion status, invasive carcinoma samples (T) were compared with a combined non-invasive epithelial reference group (ND), consisting of histologically normal epithelium (N) and dysplasia (D). This comparison was based on the absence or presence of histological invasion, and was not intended to treat N and D as biologically equivalent stages. The stage-wise N–D and D–T analyses remained the primary framework for identifying sequential biological changes during tumor development.

A total of 96 DEPs were identified, including 68 upregulated and 28 downregulated proteins in T relative to ND (Fig. 5A; Supplementary Table S9). Representative upregulated proteins included HSP90AB1, EIF4G1, TOP1, PFKP, and PFKL, whereas representative downregulated proteins included CTBP2 and SHMT1. The full DEP list was then subjected to KEGG enrichment and PPI network analyses. KEGG enrichment highlighted pathways related to carbon metabolism and one-carbon/folate metabolism, including glycolysis/gluconeogenesis (PFKP/PFKL), the pentose phosphate pathway, central carbon metabolism in cancer, and one carbon pool by folate (SHMT1) (Fig. 5B). Complete enrichment outputs are provided in Supplementary Table S10.

Fig. 5.

Proteomic features distinguishing non-invasive epithelium from invasive oral squamous cell carcinoma. A Volcano plot of differentially expressed proteins (DEPs) between invasive carcinoma (T) and non-invasive epithelium (ND; combined N and D). B KEGG enrichment analysis of DEPs. Bubble size indicates protein count, and color indicates significance. C Protein–protein interaction network of DEPs, showing HSP90AB1 as a centrally connected node. D Hub subnetwork identified by topological analysis, with HSP90AB1 consistently ranked as a key hub. E Core module identified by MCODE analysis, comprising EGFR, XRCC6, CTBP2, TOP1, ATIC, and HSP90AB1

The PPI network constructed from these DEPs comprised 66 nodes and 151 edges (Fig. 5C). Within the global network, HSP90AB1 showed extensive interactions with several highly connected proteins, including EGFR, ILF3, EIF4G1, and XRCC6. To prioritize candidate key nodes, topological and module-based analyses were integrated. HSP90AB1 was consistently ranked as a central candidate node by degree, MCC, and MCODE analyses (Fig. 5C–E). MCODE also identified a core module comprising EGFR, XRCC6, CTBP2, TOP1, ATIC, and HSP90AB1 (Fig. 5E). These findings support HSP90AB1 as a cross-method prioritized candidate hub in the invasion-focused ND–T analysis. External transcriptomic datasets further showed that HSP90AB1 was significantly upregulated in tumor tissues in GSE37991 (40 N vs. 40 T, P < 0.01) and GSE85195 (15 D vs. 34 T, P < 0.01) (Supplementary Figure S5A, B). Because the present six-patient discovery cohort had no death events and only one recurrence event during follow-up, formal survival analysis within this cohort was not statistically appropriate. To explore the potential clinical relevance of HSP90AB1 in an external dataset, overall survival analysis was performed using available HNSCC data through the Kaplan–Meier Plotter platform. Higher HSP90AB1 expression was associated with poorer prognosis in HNSCC patients (HR = 1.56, 95% CI: 1.18–2.06; log-rank P = 0.0016), with the high-expression group showing lower survival probability than those in the low-expression group (Supplementary Figure S5C). These external transcriptomic and survival data provide supportive evidence for the clinical relevance of HSP90AB1, but they should be interpreted as corroborative rather than as direct validation within the present proteomic discovery cohort.

To evaluate whether including N in the non-invasive reference group altered the direction of protein changes observed during the direct D-to–T transition, we performed a directional concordance analysis after standardizing all log2 fold changes so that positive values indicated higher abundance in invasive carcinoma. Among the 96 proteins significantly altered in the ND–T comparison, 95 showed concordant directionality in the D–T comparison, corresponding to a concordance rate of 98.96% (Supplementary Table S11). Concordance was observed for both T-upregulated proteins (67 of 68, 98.53%) and T-downregulated proteins (28 of 28, 100%). The log2 fold changes in the ND–T and D–T comparisons were also strongly correlated (Spearman’s rho = 0.920, P < 0.01; Supplementary Figure S5D). Representative ND–T hub proteins showed tumor-dominant or progression-like expression patterns across patient-matched N, D, and T regions (Supplementary Figure S5E-H).

Discussion

This study used spatially resolved proteomics to characterize protein-level changes across histologically defined N, D, and T regions within OSCC. Compared with genomics and transcriptomics, proteomics offers a more direct functional layer by capturing post-transcriptional and post-translational layers that define tumor behavior [11, 27, 28]. Previous salivary and tissue proteomics studies in OSCC have identified proteins associated with tumor progression, immune modulation, and microenvironmental remodeling, highlighting the promise of protein-level alterations as disease indicators [17, 29, 30]. However, most previous studies have focused on cross-patient tumor–normal comparisons or mixed-morphology tissue samples, with dysplasia absent or not explicitly resolved [13, 31, 32]. By using FFPE material blocks from the same patient and sampling spatially contiguous N, D, and T regions, our approach examined morphology-anchored proteomic alterations while reducing inter-patient confounding. Furthermore, the use of FFPE material is also has practical relevance, given the limited availability of fresh OSCC specimens and the increasing compatibility of MS-based proteomics with archived tissue [33–35].

Stage-resolved pairwise comparisons indicated different biological programs along the N–D–T continuum. The N-to-D transition was marked by proteins related to proliferation and metabolism, with enrichment in cell-cycle, hypoxia-associated, and central carbon metabolic pathways. This suggests that detectable biosynthetic and hypoxia-adaptive alterations may already be present at dysplasia. The D-to-T transition showed a more focused DEP profile centered on cytoskeleton-related processes. Across the broader N-to-T comparison, tumor regions showed a wider increase in tumor-associated proteomic signals, with enrichment converging on glycolysis/pyruvate metabolism and motility-related features. These findings are consistent with previous studies describing increased proliferative biomarkers such as Ki-67 and cyclin D1 in dysplasia and early OSCC [36, 37], cytoskeletal and extracellular matrix remodeling during dysplasia-to-OSCC progression [38], and Warburg-like metabolic adaptation in tumor tissue [39]. Our data extend those observations by resolving these programs at the proteomics level within a continuous, patient-matched tissue context.

One feature of the dataset is that the N-to-D transition yielded more statistically defined DEPs than the D-to-T transition. This finding suggests that threshold-detectable proteomic alterations may emerge early, during the shift from histologically normal epithelium to dysplasia. Given that the N, D, and T regions in this study were spatially contiguous within the same block, the pattern may be compatible with the concept of field cancerization, in which adjacent normal, premalignant, and malignant tissues may share features of a local molecular field [4, 40–43]. Similar early-stage prominence has also been reported in extracellular vesicle proteomics related to oral carcinogenesis [44]. Nevertheless, the present data do not directly establish field cancerization as a mechanism. DEP counts are affected by statistical threshold, sample size, inter-sample heterogeneity, and effect-size distribution. The RMS analysis showed comparable and heterogeneous proteomic distances between the N-to-D and D-to-T transitions, with three patients showing larger N-D distances and three showing larger D-T distances. The PCA and correlation patterns also revealed that D was more dispersed and partially overlapped with both N and T. Therefore, dysplasia already exhibits threshold-detectable proteomic alterations, and behaves more like a heterogeneous transitional compartment. Similarly, the smaller number of D-to-T DEPs should not be equated with reduced biological importance.

Trajectory-based module analysis further indicated that OSCC evolution is not purely linear, but instead comprises early-elevated, progressive, and dysplasia-restricted transient proteomic patterns. EW showed an abrupt proteomic shift at the dysplasia stage, with abundance remaining elevated in tumor relative to normal tissue. This pattern supports the interpretation that dysplasia marks an early departure from epithelial homeostasis, rather than only an intermediate morphology [45, 46]. Pg increased stepwise at the Mfuzz cluster-center level from N to D to T, consistent with gradual consolidation of lesion-associated traits involving metabolism, genome organization, vesicle-mediated transport, host–microbe and innate immune responses, and cytoskeletal remodeling. This dynamic interpretation extends prior proteomic studies of OPMD and OSCC, which reported extracellular exosome-, extracellular matrix-, and adhesion-related processes but did not resolve their stage-dependent temporal architecture [13, 31, 44]. DsP and DsV should be interpreted more cautiously. Their protein numbers were limited, and formal enrichment analysis was therefore underpowered. Still, their cluster-center trajectories and protein-level annotations suggest that dysplasia may involve transient activation of growth-, RNA regulatory, stress-adaptive, and cytoskeleton-associated programs, alongside temporary attenuation of epithelial differentiation, polarity, and barrier-related functions. These transient patterns may be biologically important, but they remain hypotheses for future validation.

MYO1B and HSP90AB1 represent candidates that emerged from different parts of the analysis. MYO1B, a class I actin-associated motor protein, was assigned to the Pg module and showed increasing protein expression from morphologically N to D and T. IHC analysis was consistent with this pattern: staining was absent or weak in N, heterogeneous in D, and more broadly positive in T regions. Although some D areas showed relatively strong staining intensity, T regions generally showed a wider distribution of MYO1B-positive cells. As the Pg module was defined by the overall N–D–T proteomic trajectory and IHC-based IOD analysis incorporated both staining intensity and positive staining area, interpretation was based on regional expression patterns rather than focal intensity alone. Together with the functional enrichment of the Pg module, this pattern supports the interpretation that MYO1B is associated with cytoskeletal remodeling and adhesion-related programs. This is consistent with prior evidence showing that MYO1B promotes migration, invasion, and lymph node metastasis in HNSCC [47–50], as well as with work identifying MYO1B among genes activated during oral carcinogenesis in association with chromatin compartment reorganization [8]. However, the validation cohort was small. Additional studies will be needed before MYO1B can be evaluated as a diagnostic or risk-stratification marker.

HSP90AB1 was prioritized in the invasion-focused ND–T comparison. In the present dataset, it was differentially expressed in the N–T and ND–T comparisons, included among the one-way ANOVA-selected proteins across the N-D-T sequence, and assigned to the Pg module. PPI analyses also identified HSP90AB1 among the highly connected candidate nodes. External transcriptomic datasets and survival analysis provided additional support. Given the established role of HSP90AB1 in promoting proliferation, migration, and glycolysis in HNSCC [51], its prioritization here is biologically plausible and aligns with carbon metabolism-related enrichment observed in the invasion-focused analysis. However, its role as an invasion-associated hub should be further validated by additional protein-level spatial validation and functional experiments.

Additionally, the ND–T comparison was not intended to merge N and D into a single biological state. Indeed, a central finding of this study is that the N–D and D–T transitions are characterized by distinct biological programs. The rationale for combining N and D was based on histological invasion status: both are non-invasive epithelial regions, whereas T represents invasive carcinoma. The invasion status is a biological and relevant threshold, important in pathological diagnosis, risk stratification, and treatment decision-making [4, 52]. Thus, the ND–T comparison was used as an exploratory binary invasion-status analysis to complement the stage-wise N–D–T analyses. Considering the complexity and non-linearity of the N–D–T trajectory, directional concordance analysis was performed. The observation that most ND–T DEPs changed in the same direction in the D–T comparison supports the view that ND–T captures invasion-associated alterations. However, larger cohorts, together with sampling strategies more specifically focused on the invasive front, will be required to validate these invasion-associated proteomic changes.

Overall, this study provides a spatially resolved proteomic framework for OSCC evolution across histologically normal, dysplasia, and tumor regions. The data suggest that the N–D–T continuum is not a simple linear accumulation of change, but instead comprises early-shift, progressive, and dysplasia-restricted transient programs protein programs. This framework extends current OSCC proteomics by placing protein-level changes within a local histopathological trajectory in FFPE tissues. However, several limitations should be acknowledged. First, the discovery cohort was small and derived from a single center, limiting statistical power and generalizability. The differential-expression, trajectory, and subgroup analyses should therefore be viewed as exploratory and hypothesis-generating. Second, although microdissection improved spatial resolution, finer intraregional heterogeneity and protein features such as post-translational modifications or higher-order complexes were not comprehensively captured. Finally, Functional validation also remained limited. Larger independent cohorts, longer follow-up, and complementary spatial multi-omics approaches will be required to clarify the mechanistic and translational relevance of these proteomic programs in OSCC progression.

Conclusion

In conclusion, this spatially resolved proteomic analysis suggests that OSCC evolution from normal epithelium to dysplasia and carcinoma is not purely linear. Instead, it involves early-shift, progressive, dysplasia-specific transient protein programs within a spatially contiguous histopathological continuum. By defining a morphology-anchored proteomic workflow to FFPE tissues, this study provides insight into protein-level changes during oral malignant progression and identifies candidate proteins for further validation in larger OSCC cohorts.

Abbreviations

OSCC

Oral squamous cell carcinoma

N

Normal

D

Dysplasia

T

Tumor

HNSCC

Head and neck squamous cell carcinoma

OPMDs

Oral potentially malignant disorders

OED

Oral epithelial dysplasia

MS

Mass spectrometry

FFPE

Formalin-fixed paraffin-embedded

NED

No evidence of disease

LCM

Laser capture microdissection

LC–MS/MS

Liquid Chromatography-Tandem mass spectrometry

FA

Formic acid

DIA

Data-independent acquisition

GO

Gene Ontology

KEGG

Kyoto Encyclopedia of Genes and Genomes

DEPs

Differentially expressed proteins

RMS

Root mean square

PCA

Principal component analysis

LN

Lymph node

PPI

Protein–protein interaction

IHC

Immunohistochemistry

H&E

Hematoxylin and eosin

IOD

Integrated optical density

EW

Early Warning Module

Pg

Progression Module

DsP

Dysplasia-Specific Peak Module

DsV

Dysplasia-Specific Valley Module

Authors’ contributions

CP contributed to conceptualization, data curation, formal analysis, software, visualization, and writing of the original draft. XC contributed to data curation and funding acquisition. MH contributed to data curation and funding acquisition. YZ acquired funding and developed the software. JZ contributed to conceptualization, data curation, funding acquisition, supervision, and revised the manuscript. TL contributed to funding acquisition, supervision, and revised the manuscript. All authors have read and approved the final manuscript.

Funding

This work was supported by Beijing Science & Technology Development of Traditional Chinese Medicine (BJZYQN-2025-11), China Postdoctoral Science Foundation (2025M781685, GZB20240038), China Oral Health Foundation (COHF-KQZS2026023), Talent development plan for the future in Medical-Engineering Integration by BRA-CDCHE and ZTA (MBRC0012025013), Beijing Natural Science Foundation (Z240003), Education Research Project of Peking University School and Hospital of Stomatology (2024-ZD-05), Natural Science Foundation of Education Department of Henan Province (25A320004), Foundation for Young Scholar in School of Stomatology of Henan University (HUSSYS2024005), and CAMS Innovation Fund for Medical Sciences (2019-I2M-5-038). The funder played no role in study design, data collection, analysis and interpretation of data, or the writing of this manuscript.

Data availability

The datasets generated and analysed during the current study are available in the iProX repository (accession: PXD078064).

Declarations

Ethics approval and consent to participate

This study was approved by the Ethics Committee of Peking University School and Hospital of Stomatology. All samples and clinical data were obtained with informed consent, and all procedures were conducted in accordance with the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.