Skip to main content
NCBI home page

This is a preprint.

It has not yet been peer reviewed by a journal.

The National Library of Medicine is running a pilot to include preprints that result from research funded by NIH in PMC and PubMed.

medRxiv logoLink to medRxiv
[Preprint]. 2026 Jul 20:2026.07.17.26358051. [Version 1] doi: 10.64898/2026.07.17.26358051

Comprehensive molecular characterization of cutaneous squamous cell carcinoma reveals determinants of metastatic progression

Barbara Rentroia-Pacheco 1,2,3, Harsh Sharma 2,3, Lara Pozza 4, Joleen JH Traets 4, Bishal Tandukar 2,3, Olivia FM Steijlen 1, Romy Ruiter 4, Noel Cruz-Pacheco 2,3, Daphne Huigh 4, Arne Van Hoeck 1, Yan-Ting Chen 4, Beatriz Infante 4, Defne Baskurt 2,3, Vignesh Arunachalam 2,3, Celeste J Eggermont 1, Amanda Bas-Cristóbal Menéndez 4, Tamar Nijsten 1, Harmen J G van de Werken 5, Antien L Mooyaart 6, Domenico Bellomo 4,#, Marlies Wakkee 1,#, A Hunter Shain 2,3,#, Loes M Hollestein 1,#
PMCID: PMC13419656  PMID: 42539055

Abstract

Cutaneous squamous cell carcinoma (cSCC) is the second most common form of cancer worldwide. While most cSCCs are not life-threatening, 2–5% of patients develop metastases. To better understand what causes some cSCCs to progress to metastatic disease, we assembled a nationwide cohort of 19,120 patients with clinico-pathologically annotated tumors linked to metastatic outcome. RNA-sequencing was performed on 378 tumors, and whole-exome sequencing on 147, with balanced numbers of tumors that progressed to metastatic disease (cases) and did not (controls). UV radiation was the dominant mutational signature with additional contributions from aging, APOBEC activity, and, in immunosuppressed patients, azathioprine exposure. We identified 38 genes under selection across a core set of signaling pathways. Gene expression clusters were primarily associated with the differentiation state of tumor cells and secondarily with the composition of the tumor microenvironment. Several mutational and transcriptional programs were associated with metastasis, including a dedifferentiated gene expression signature, activating mutations in the RAS signaling pathway, loss-of-function alterations in the SWI/SNF chromatin remodeling complex, and specific arm-level copy number alterations. A 23-gene expression signature was built to predict metastasis from primary cSCC tissue. The signature was validated in two independent cohorts (N=102 and 52), where it predicted metastasis independently of staging systems. Together, these findings provide the most detailed molecular portrait of cSCC to date and establish an assay for risk stratification suitable for clinical implementation.

Introduction

Cutaneous Squamous Cell Carcinoma (cSCC) is a type of skin cancer that is the second most common form of cancer worldwide1,2. Given its high and rising incidence, cSCC poses a major public health burden3–5. However, it has been inaccurately perceived as a non-life-threatening cancer, and consequently understudied. Many national cancer registries exclude cSCC from registration6,7 and large-scale molecular atlases, such as The Cancer Genome Atlas (TCGA), have not profiled cSCC. Consequently, there is a lack of high-quality epidemiological studies and large-scale biomarker discovery studies describing cSCC, which has limited our ability to understand and effectively treat the disease.

The perception of cSCC as a non-life-threatening cancer stems from the relatively low risk of metastasis from primary cSCCs (2–5%)8,9 compared to other cancers. However, while an individual tumor may pose low risk to a patient, since cSCC is so common, the total morbidity and mortality across the entire population is high. Moreover, managing cSCC remains challenging. Some treatments, such as extensive surgeries and/or (adjuvant) immuno- or radio- therapy, require careful consideration because many patients are elderly, frail, or immunosuppressed, as these interventions may be disfiguring and carry substantial morbidity. When tallying the cumulative costs on our healthcare system, cSCC is arguably worse than melanoma. It kills a similar number of patients while affecting many more, thus burdening healthcare systems with billions of dollars in management costs on top of its death toll5, much of which could be avoided with better prognostic stratifications.

Most clinicians use staging systems developed by Brigham and Women’s Hospital (BWH)10 or the American Joint Committee on Cancer (AJCC)11, however, these are insufficient predictors of metastatic risk12. We recently developed a new system (the Erasmus MC model) based solely on clinical and histopathologic features that outperforms the BWH and AJCC systems13, but there remains limitations to clinical/pathologic staging systems. For instance, these systems require complete pathological assessment of the excised tumor, rendering them unsuitable for use with biopsy material and limiting their utility in pre-operative decision-making. Furthermore, observer variability in pathological assessment limits reproducibility across centers14.

As such, better biomarkers are needed to quantify metastatic risk in patients with primary cSCC. This knowledge could guide treatment strategies, indicating which patients may require margin-controlled surgery, additional imaging and/or adjuvant therapies, versus those who only need limited follow-up screening. To this end, gene expression signatures to predict cSCC metastasis have been developed15–17. Despite their potential, these signatures were developed in cohorts that do not reflect the actual distribution of cases across T-stages and have a low number of metastatic events. In addition, external validation in representative cohorts is still lacking.

To address these gaps in knowledge, we established the D-SQUAME study, one of the World’s largest and most deeply annotated datasets of cSCC18. Next, we performed a large-scale genomic and transcriptomic analysis of cSCCs drawn from this study (Fig. 1A). From these discovery analyses, we identified mutational and expression markers predictive of metastatic outcome. More generally, because cSCC was left out of large-scale sequencing studies, we envision the data here will serve as an important resource for the genomics community.

Figure 1. Mutational landscape of cutaneous squamous cell carcinoma.

Figure 1.

Open in a new tab

A. Overview of cohorts used in this study. NCC: Nested case-control B. Top: Mutation burden of each tumor (mutations per megabase). Bottom: Proportion of mutations assigned to each COSMIC mutational signature in each tumor. Immunocompetent indicates tumors from patients without any known immunosuppressive condition. Immunosuppressed indicates tumors from patients who received an organ transplant or had a hematological malignancy before cSCC diagnosis (see methods). C. Genome-wide chromosomal alterations (red = gain, blue = deletion).

Results

Cohort description

To facilitate molecular studies, we first assembled the D-SQUAME study18 – a nationwide cohort of cSCCs with detailed clinical/histopathologic information and long-term outcomes. We integrated data from the Netherlands Cancer Registry (NCR), the Dutch Nationwide Pathology Databank (Palga), and the Netherlands Organ Transplant Registry. The organ transplant registry was used to identify solid-organ transplant recipients, and the NCR to identify patients with hematological malignancies – these patients show an increased risk of cutaneous squamous cell carcinoma due to immunosuppression2,19. We identified all 19,120 patients diagnosed with a first cSCC between 2007 and 2009, selecting this interval to ensure adequate follow-up to distinguish lesions that progressed to metastatic disease from those that did not. From this cohort, we performed RNA-sequencing on 378 tumors and an additional round of whole-exome DNA sequencing on 147 tumors (Table S1, Fig. S1A). All patients were treatment-naïve and free of metastatic disease at the time their primary tumors were diagnosed.

We used a Nested Case-Control (NCC) study design to select balanced numbers of tumors that progressed to metastatic disease (cases) and did not (controls). A nested case-control study is a cost-effective way to sample a meaningful number of events (in this case, tumors that progress to metastatic disease) while retaining the ability to extrapolate absolute risk estimates to the whole population by weighting each sample according to how representative it is of the source cohort20.

As expected, >80% of cSCCs in the nationwide cohort of 19,120 tumors were early-stage, with an extremely low metastatic risk (Fig. S1B). Therefore, when selecting tumors for molecular analyses, we intentionally selected controls whose clinical and histopathologic features more closely resembled the cases (Fig. S1B, C). This was accomplished by pairing each case with a control from the same pathology lab that had a similar clinical-pathologic risk score13. Our strategy increased the likelihood that any molecular biomarker predictive of metastatic disease would provide prognostic information that complements, rather than simply recapitulates, established staging systems.

Landscape of genetic alterations in cSCC

Studying the mutational landscape of cSCC is difficult because these tumors tend to have low neoplastic cellularity, owing to dense immune infiltrates and reactive hyperplasia mixed with neoplastic cells21. An additional challenge is that most tumors were derived from formalin-fixed, paraffin-embedded (FFPE) blocks that were more than 10 years old. To overcome these obstacles, we applied rigorous quality control standards to ensure that we were statistically powered to detect clonal mutations in each tumor (see methods). To our knowledge, no prior genomic study of cSCC has analyzed a cohort of this size while meeting comparable quality control criteria.

cSCCs had high mutation burdens (median of 44 mutations/megabase) (Fig. 1B). UV radiation was the dominant mutational signature with additional contributions from aging, APOBEC activity, and, in immunosuppressed patients, azathioprine exposure22 (Fig. 1B). Recurrent copy number alterations were observed, including gains of 3q, 5p, 8q and 9q as well as losses of 3p, 8p, 9p, and 18q (Fig. 1C, S2A–D). There were no statistically significant differences in the total mutation burden or fraction of the genome altered of cases versus controls or between tumors from immunocompetent versus immunosuppressed patients (Fig. S2E–F).

Due to its high mutation burden, it is challenging to discriminate driver mutations from passenger mutations in cSCC23. To maximize statistical power, we combined data from the 147 tumors sequenced here and 83 tumors from a previous meta-analysis of cutaneous squamous cell carcinoma21 and used IntOGen24 to identify genes under selection (Table S2). Mutations were also checked for overlap with the OncoKB25 database to detect drivers that were too rare in this dataset, alone, to show patterns of selection but are known to be under selection in other cancers (Table S3). Furthermore, we annotated genes with focal amplifications and deep deletions.

In total, we identified 38 driver genes, converging on a core set of signaling pathways (Fig. 2A, S3). Most tumors harbored alterations inactivating the Notch, p53, Hippo, and Rb pathways. Tumors also frequently carried loss-of-function mutations in KMT2D (53.7%), a key chromatin regulator of epidermal differentiation26, and gain-of-function mutations in the TERT promoter (58.5%). To a lesser extent, they had alterations disrupting the SWI/SNF chromatin remodeling complex (45.6%), activating the RTK–RAS–MAPK/PI3K pathway (32.0%), or inactivating TGF-β signaling (22.4%). TGF-β pathway mutations were now significant when compared to the previous meta-analysis of cSCC21, though Cammareri and colleagues had previously noted mutations in this pathway and provided functional evidence that it was tumor suppressive27. We also refined the specific genes disrupting other pathways already known to drive cSCC. Additional low-frequency driver genes are likely to exist, but this study represents, to our knowledge, the largest statistically powered cancer gene discovery effort to date in this tumor subtype.

Figure 2. Landscape of driver mutations in cutaneous squamous cell carcinoma.

Figure 2.

Open in a new tab

A. Tiling plot of genetic alterations (rows) in each tumor (columns). Genes with at least one driver mutation are shown. Genes are grouped into pathways and ordered by pathway mutation frequency (highest to lowest). White circles indicate the gene has multiple alterations within the same tumor - when this occurs, the tile is colored to indicate the most consequential alteration. For the computation of the percentage of sample with driver alterations in each gene or pathway, only driver mutations are counted. B. Number of genes with driver hits and number of total driver hits per tumor, stratified by metastatic outcome and immunosuppression status. Differences between groups were assessed using linear regression adjusted for callable coverage.

Tumor suppressor genes classically acquire two “hits” to eliminate the activity of both parental alleles28,29. We used copy number data and mutant allele frequencies to infer genes with bi-allelic alterations (see methods, Fig. S4). While IntOGen did not consider the number of hits when identifying genes under selection, most tumor suppressor genes had alterations affecting both alleles in the majority of the samples (Fig. S4A), further supporting their case as driver genes. A major exception was SWI/SNF genes (ARID2 and ARID1A), for which selection is more complex. SWI/SNF mutations alter, but do not fully ablate, chromatin remodeling activity, and hence bi-allelic mutations are sometimes not tolerated30. Several oncogenes, notably PIK3CA, also showed an increase in dosage of the oncogenic allele (Fig. S4B). While acknowledging the caveat that we are missing low-frequency pathogenic alterations, the median cutaneous squamous cell carcinoma is driven by 9 hits across 6 genes (Fig. 2B). We also tested whether mutations in driver gene pairs co-occurred or were mutually exclusive; however, no gene pair showed a significant association after multiple hypotheses testing (Table S4).

A major strength of this study is the ability to link genetic alterations with metastatic outcomes. Hazard ratios were calculated to compare metastatic event rates between tumors with and without mutations in individual genes, signaling pathways, or arm-level copy number alterations (Fig. 3A–C). We then applied a regularized Cox regression model to identify combinations of features that best predicted metastatic risk across bootstrap resamples (Fig. 3D, S5A). Bootstrapping was used to evaluate the stability of feature selection and to mitigate overfitting. Across bootstrap iterations, model performance was consistent in out-of-bag samples and effectively stratified tumors into risk groups (Fig. S5B–C). Poor prognosis was associated with mutations in the SWI/SNF, RTK–RAS–MAPK/PI3K, and TGF-β pathways; MYC gain (chr8q) or mutation; losses of 11q, 8p, or 3p; and mutations in CASP8. In contrast, a favorable prognosis was associated with KMT2D mutations, gain of 5p or loss of 4q or 10q.

Figure 3. Genomic alterations associated with metastatic outcome.

Figure 3.

Open in a new tab

A. Association between driver mutations and metastatic outcome for all genes with driver mutations present in at least 3 samples. Association was assessed using a Cox proportional hazards regression model, with gene hit count as the independent variable and sample callable coverage included as an adjustment covariate. Hazard ratios, associated 95% confidence intervals and Wald p-values for gene hit count are shown. p-values < 0.1 are colored in yellow, and p-values <0.05 are colored in red. B. Genes with driver mutations were grouped into pathways and association between the sum of all hits in a pathway and metastatic outcome was computed, with adjustment for sample callable coverage. C. Association between chromosomal alterations and metastatic outcome, adjusted for sample callable coverage. Only chromosomal alterations present in at least 10 samples are shown. D. The most predictive combination of pathway and copy number alterations was selected using a multivariable regularized Cox model. The selected variables (in bold) were frequently selected by regularized Cox models across bootstrap samples (N=200).

We also computed odds ratios to determine whether specific genetic alterations were enriched in tumors from immunocompetent versus immunosuppressed patients. Mutations in the RTK-RAS-MAPK/PI3K pathway were more common in tumors from immunosuppressed patients, though the association just missed statistical significance (Fig. S6).

Gene expression landscape in cSCC

Next, we clustered the RNA-sequencing data. The tumors were a mixture of surgical excisions and biopsies, and whole-slide RNA-sequencing, without macrodissection, was performed on each sample. To avoid batch effects due to the different composition of tissues from excisions versus biopsies, we performed unsupervised hierarchical clustering on the most variable genes within excisions and used the biopsies to validate the sample/gene clusters.

Consensus clustering identified three sample clusters and six gene expression clusters (Fig. 4A, Table S5). These clusters were reproduced in the biopsies (Fig. S7A–C). Samples clustered into groups with favorable, intermediate, and poor metastatic outcomes (Fig. 4B). Pathway overrepresentation analysis showed that gene expression clusters were largely driven by epithelial differentiation and were strongly associated with histopathologic differentiation (p<0.001). Differentiated tumors had (terminal) squamous cell differentiation programs with the longest metastasis-free survival. Basal-like tumors showed a higher expression of basal squamous cell programs with an intermediate metastatic outcome. Finally, Mesenchymal-like tumors exhibited poorly differentiated programs with the shortest metastasis-free survival. Further, there was a shift in metabolic signatures and proliferation. Differentiated tumors were enriched for lipid metabolic programs, and Mesenchymal-like tumors preferentially expressed the glycolytic pathway, consistent with the Warburg effect. They also showed higher expression of genes involved in the cell cycle. Our results support the model that tumor differentiation, metabolic and proliferation status are associated with time to metastasis.

Figure 4. Landscape of gene expression patterns in cutaneous squamous cell carcinoma.

Figure 4.

Open in a new tab

A. Clustering of cSCC excision samples (N=197). Within each cluster, samples are ordered according to hierarchical clustering using Spearman’s correlation and average linkage. Rows represent the 1000 genes with the highest association to the sample clusters. The tile plot above the heatmap shows the distribution of pathological characteristics and genomic alterations that were found to be significantly associated with the clusters (see Table S6), tiles with missing values are colored in white. B. Kaplan-Meier curve of the three cluster groups. Follow-up truncated at 5 years. C. Volcano plot showing differentially expressed genes between metastasizing and non-metastasizing CSCC, adjusted for sample type (excision vs biopsy). Significant genes were defined as those with an adjusted p-value < 0.05, indicated by a dashed horizontal line. D. Gene Set Enrichment Analysis (GSEA) showing the top 15 upregulated and top 15 downregulated Hallmark gene sets in metastasizing versus non-metastasizing CSCC, adjusted for sample type.

Tumor microenvironment signatures were also apparent. Favorable outcome tumors preferentially expressed stromal/fibroblast-rich gene expression programs, whereas intermediate outcome tumors were enriched for hypoxia-related signatures. Curiously, immune-related programs did not drive the unsupervised clustering. However, immune signatures were evident when the data were analyzed in a targeted manner. For example, tumors from immunocompetent patients showed enrichment of interferon-α and γ signaling and other immune-related pathways A-B). However, these immune signals were outweighed by differentiation- and metabolism-associated programs in unsupervised analyses. Prior studies have reported immune composition during progression from actinic keratoses to primary cutaneous squamous cell carcinoma31,32. If immune evasion is, indeed, an early event in skin carcinogenesis, it would already be present across most tumors in this cohort and likely overshadowed by other biological processes, such as differentiation.

We additionally tested for associations between sample clusters and clinical/genomic features (Table S6 and Fig. 4A, top tiles). Several pathological risk factors, including BWH and AJCC8 stages, increased from the differentiated to the mesenchymal tumors (p<0.001) as well as the proportion of genome altered (p=0.0015). Curiously, Basal-like tumors were associated with a higher percentage of Notch pathway mutations (p=0.029).

We also performed supervised differential gene expression analysis comparing cases and controls, identifying 2,300 genes that differed between the two groups (Fig. 4C). The transcriptional programs evident in this supervised analysis mirrored those identified by unsupervised clustering (Fig. 4D). A previous study profiled the gene expression landscape of earlier phases of skin carcinogenesis through comparison of normal skin, actinic keratoses, and primary cutaneous squamous cell carcinomas31. They defined a differentiation–progenitor–like (DvP) signature that also overlapped with our genes (Fig. S7D), confirming that differentiation is a dominant signal distinguishing cases and controls.

To distill these gene expression programs into a prognostic signature, amenable to clinical deployment, we implemented a regularized Cox regression model to select the combinations of genes whose expression is best able to discriminate cases from controls. As model input features, we tested the previously defined DvP genes (biology-inspired genes), genes defined exclusively in this study (data-driven genes), and integrations of the two gene sets (SFig. 9A). The entire procedure, including feature filtering and selection, was internally validated through bootstrapping (SFig. 9B) to get a reliable model performance estimate. The models based on DvP genes or exclusively on our genes performed well (SFig. 10A), but the model that combined features from both sets, with a late feature integration approach, achieved a high performance with fewer input genes, showing an internally validated weighted C-index of 0.75 (95%CI 0.66–0.80) (SFig. 10B). The final model, obtained by running the same pipeline on the entire dataset, yielded a 23-gene signature (Table S7), SCCore-GEP (Squamous Cell Carcinoma Outcome Risk Estimation Gene Expression Profile).

Out of the 23 SCCore-GEP genes, 10 were upregulated in cases and were associated with cell migration and stemness. The remaining 13 genes were upregulated in controls and were mostly associated with epidermal differentiation (Fig. 5A, left). We performed spatial transcriptomics on a subset of tumors to determine where the gene expression of the SCCore-GEP genes derived from (Fig. 5A, right). Most genes were primarily expressed in tumor keratinocytes and secondarily in normal epithelial cells or immune cells. PLA2G2A gene was an exception, as it was expressed mostly in fibroblasts.

Figure 5. SCCore-GEP: characterization of gene expression and validation performance.

Figure 5.

Open in a new tab

A. Scaled gene expression values of the 23 SCCore-GEP genes in the D-SQUAME discovery dataset. Samples are ordered according to the ranked SCCore-GEP risk output. Metastasis status, BWH stages, and sample type are indicated by colour annotations above the columns. Spatial transcriptomics on 4 case-control pairs (N=8 samples) from the discovery dataset reveals the read distribution of each of the 23 genes across cellular compartments. B. Assessment of the discriminative performance for metastasis prediction of the SCCore-GEP, BWH, and AJCC8 staging systems in T1-T2a tumors of the D-SQUAME validation and the Nassir et al datasets. C. Assessment of the discriminative performance for metastasis prediction of the SCCore-GEP, and EMC model in T1-T2a tumors of the D-SQUAME validation dataset with and without sample type stratification. Performances are reported in terms of weighted C-index and 95% confidence intervals (see methods). *Whenever an excision sample was available, EMC model predictions were computed using pathological variables scored on the excision sample. D. SCCore-GEP post-test risk probability across different risk thresholds. Varying the risk threshold on the predicted SCCore-GEP risk results in different stratifications of the patients into a high-risk and a low-risk groups. We show the post-test risk (i.e., Positive Predictive Value/precision) versus the percentage of metastases captured (i.e., sensitivity/recall) in the high-risk group, for different risk thresholds, within BWH T1 patients (N samples in NCC dataset=62). Color bands correspond to the 95% confidence intervals of the nodal metastasis prevalence in each BWH stage, as reported in Zakhem et al. Horizontal dotted lines indicate the pre-test risk in BWH T1 patients in the D-SQUAME validation dataset (0.70%). The lower panel shows the percentage of post-test low-risk and high-risk patients versus the percentage of metastases captured (sensitivity/recall). The interval from 0 to the dashed vertical line corresponds to risk thresholds for a potential upstaging of the patients.

Validation of a molecular signature to predict metastatic outcome in patients with primary cSCC

Mutational and gene expression features were both prognostic in the discovery cohort, but they were identified from datasets with incomplete overlap, making it difficult to directly compare their performance. To determine whether to pursue a prognostic biomarker based on mutations, gene expression markers, or a combination of both, we directly compared their performance on the subset of tumors with available DNA- and RNA-sequencing data (N=126 tumors). We developed (regularized) Cox regression models for each feature set and evaluated their performance through bootstrapping (SFig. 11). Within this sample subset, the combined model and gene expression-only model performed similarly, and both outperformed the mutation-only model. Considering this, for simplicity, we prioritized validation of the gene expression marker set, SCCore-GEP.

SCCore-GEP was validated in two independent cohorts: a more recent set of tumors from the D-SQUAME study, and an external RNA-sequencing cohort made available by Nassir and colleagues15 (Fig. 1A, SFig. 12A). Due to the scarcity of T2b samples (N=2 in the D-SQUAME study), validation was focused on BWH T1-T2a patients. SCCore-GEP showed high discriminative performance and outperformed BWH and AJCC8 staging systems in prediction of metastasis in both cohorts (Fig. 5B). In the D-SQUAME validation cohort, controls were sampled randomly to facilitate a population-based comparison with staging systems and the EMC model. SCCore-GEP achieved a weighted C-index of 0.83 (95%CI: 0.72–0.90), which corresponds to an increase of over 0.15 compared to the BWH and AJCC8 staging systems. Similarly, in the Nassir cohort, SCCore-GEP achieved an AUC of 0.79 (95%CI: 0.56–0.97), 0.34 higher than the BWH staging performance. The signature also showed robust performance across sample types (biopsy or excision, Fig. 5C) and discriminated equally well within BWH stages (SFig. 12B).

Compared to the clinicopathological EMC model, SCCore-GEP showed similar performance in all sample groups, except for samples from biopsy material, where SCCore-GEP was 0.12 higher (Fig. 5C). While the performance of SCCore-GEP, as measured by a C-index, was comparable to the EMC model for many subgroups, multivariable analysis showed that SCCore-GEP provides independent predictive value from staging systems and the EMC model (p-value <0.001, SFig. 12C). For completeness, signature performances for all available T-stages and validation cohorts are respectively reported in SFig. 13A–C and 14.

There is no consensus among clinicians on a post-test risk cutoff for patient upstaging13; however, we investigated the signature’s potential to identify patients at high-risk of metastasis within traditionally low-risk groups (BWH T1 and T2a) in the D-SQUAME validation cohort (Fig. 5D, SFig. 12D). Within T1 patients, the SCCore-GEP can identify candidates for upstaging – i.e. patients whose true metastatic risk is higher than the metastatic risk in T1 patients reported in the literature33. Depending on the desired post-test risk for upstaging, this high-risk group can represent up to 25% of T1 patients, while accounting for up to 75% of metastatic events in this population (dashed line in Figure 5D). SCCore-GEP can also identify up to a sixth of T2a patients with post-test risk compatible with upstaging: depending on the exact cutoff, this group can account for up to two-thirds of the metastatic events (SFig. 12D). Similar high-risk groups can be found within AJCC8 low-risk patients (T1-T2) (SFig. 14).

Conclusions

Given the enormous burden on public health systems, cSCC is an ideal example of a cancer subtype that would benefit from biomarkers to predict outcomes, thus reducing the burden of over- and under- treatment. However, it is difficult to identify biomarkers without first establishing large-scale clinical and molecular databases from which discovery analyses can be performed. Here, we generated, to our knowledge, the world’s largest and most comprehensive resource of cSCC by integrating national clinical registries and multi-omics profiling of rigorously selected tumors with known metastatic outcomes.

Our analysis of whole-exome sequencing and RNA-sequencing data encompasses more than twice as many tumors as the next largest studies21,34. Prior cancer gene discovery analyses were underpowered to identify driver genes in cSCC23, but this study provides the clearest portrait of genes under selection in cSCC to date. We did not observe genetic subtypes of cSCC, but some cSCCs are genetically more advanced, with mutations in the SWI/SNF, Ras, and TGF-beta pathways, and these tumors are more likely to progress to metastatic disease. The Pan-Cancer Analysis Working Group (PCAWG) estimates that most cancers are driven by 4–5 driver genes35, but cSCCs are driven by 9 hits spread across 6 genes, mostly affecting tumor suppressors. As the first line of defense against the environment, skin epithelial cells likely evolved mechanisms to avoid simple routes to transformation from one or a few powerful mutations.

Given the evidence of high immunogenicity in cSCC, it was somewhat surprising that immune gene expression programs did not drive unsupervised clustering. The tumors in the discovery cohort were mostly collected between 2007 and 2009, capturing salient metastatic biology without selective pressures introduced by drug treatments such as immunotherapy. A future case–control study stratified by response to immunotherapy would be expected to reveal stronger immune signals, but our findings here suggest that metastatic competence is mainly dictated by tumor differentiation state.

A 23-gene expression signature, SCCore-GEP, captures much of the prognostic information encoded in underlying molecular features. Notably, SCCore-GEP both outperforms and complements existing clinical staging systems, suggesting that it reflects biological characteristics not captured by standard clinicopathologic measures. Based on the known functions and cellular sources of these genes, the signature primarily reflects tumor differentiation state and, to a lesser extent, the contribution of surrounding normal epithelial and immune cells. The AJCC has acknowledged that poor differentiation is associated with unfavorable outcomes, but it was considered too subjective to be included as a risk factor in a staging system36. By contrast, gene expression can objectively measure differentiation to complement a purely clinical/pathologic predictor.

A major strength of SCCore-GEP is that it can predict metastatic risk from biopsy material, enabling a timely risk evaluation without waiting for a full pathological assessment based on the excision. Accurately assessing metastatic risk before any therapeutic intervention may also enable risk-informed decisions regarding the optimal surgical approach (e.g., whether to perform Mohs micrographic surgery) or (neo-) adjuvant treatment. This is particularly relevant in challenging cases, such as cSCCs in frail or immunosuppressed patients and cosmetically sensitive areas. While staging systems and nomograms remain valuable tools, their performance is contingent on the availability of complete pathological data.

While our nested case-control datasets enabled the estimation of the SCCore-GEP performance at the population level, further studies are needed to identify specific patient groups that would benefit most from SCCore-GEP risk stratification. Effective cSCC patient management must also consider the risks of local recurrence, local progression, and development of multiple primaries.

In summary, we generated the world’s largest population-based genomic and transcriptomic resources of cSCC, from which we developed and validated a gene expression-based signature (SCCore-GEP) with the potential to improve care of patients with cSCC.

Methods

Selection of tumors for molecular analyses

The tumors, selected for molecular analyses, came from a nested case-control (NCC) dataset, which was derived from a larger nationwide cohort of 19,120 patients (the D-SQUAME discovery cohort). Details of the epidemiological design, sample collection process, and tumor block slicing have been described elsewhere18, and the nested case-control sampling is discussed in Rentroia-Pacheco et al20. Briefly, through linkage of the Netherlands Cancer Registry (NCR)37 and the Dutch Nationwide Pathology Databank (Palga38), we identified all patients diagnosed with a first cSCC in the Netherlands between January 1, 2007 and December 31, 2009. Next, we retrieved their pathology reports documenting histologically confirmed primary cSCCs and metastases until 2022. Data was further linked to the Netherlands Organ Transplant Registry (NOTR39) to identify solid organ transplant recipients. Patients who developed metastatic disease (cases) were identified using a rule-based algorithm applied to the pathology reports40, followed by a manual review by two clinicians. A total of 305 cases were identified and matched to patients without metastasis (controls) using incidence density sampling, with two matching variables: pathology laboratory and metastatic risk, as assessed by the Erasmus MC clinicopathological model (EMC model13). The corresponding formalin-fixed paraffin-embedded (FFPE) tumor blocks were requested via the Dutch National Tissue Portal41. To maximize dataset completeness, pathological characteristics of all tumors were re-assessed on hematoxylin and eosin (H&E) slides by dermatopathologists blinded to both outcomes and the original pathology reports. Of the requested 305 case-control sets (N=610 primary cSCC samples), tumor blocks for 227 sets were received on time for RNA- and whole-exome sequencing.

Whole-exome sequencing and pre-processing

A subset of the tumors selected for molecular analyses (N=174) was sequenced with whole-exome sequencing. H&E slides were macrodissected by a dermatopathologist to enrich for areas of squamous cell carcinoma. Normal skin, adjacent to the cSCC, was macrodissected and used as a source of reference, germline DNA. Genomic DNA was isolated using the QIAamp DNA FFPE Tissue Kit from 20 unstained FFPE sections of 4 and 8 μm. The DNA was sheared to an average fragment size of 300 base pairs using the Covaris E220 focused ultrasonicator. Libraries were constructed with the KAPA HyperPrep Kit (Roche, cat. no 07962363001) and enriched for exomic fragments through hybridization with KAPA HyperExome V2 probes (Roche, cat. no 9718648001). After capture, they were further amplified with 13 PCR cycles using the KAPA HyperCapture Reagent Kit (Roche, cat. no 09075828001). Samples were then paired-end sequenced (2 × 100 bp) on an Illumina NovaSeq 6000 instrument, with a target of 100M reads per sample. All the nucleic acid inputs and outputs involving library preparation, hybridization, and sequencing were quantified using Qubit (dsDNA High Sensitivity quantification), Agilent Bioanalyzer 2100 (High Sensitivity DNA run), and/or QuantStudio 5 real-time PCR system (qPCR with the KAPA Quantification Kit, Roche, cat. no 07960298001).

Fastq files were preprocessed with fastp42 (v.0.23.4) and aligned to the hg19 reference genome with the BWA-MEM algorithm43 (v.2.0.5). Duplicated reads were removed with Picard (v1.97) and base quality was recalibrated with GATK BaseRecalibrator44 (v.3.3.0). A total of 27 samples were excluded due to cross-contamination (N=2), mismatches between RNA-seq/WES (N=3) and insufficient power to detect somatic mutations (N=22). The last criterion was based on the percentage of callable coverage across the sequenced exome, where callable coverage corresponds to the genomic positions with sufficient power to detect a somatic mutation (STable 1, “Callable coverage (%)” column). A minimum callable coverage of 80% was required. Further details are provided in the supplementary text.

RNA-sequencing and pre-processing

For each tumor, RNA was collected from 13 unstained sections of 4 μm FFPE tumor tissue. In contrast to the DNA isolation, the RNA was not dissected to enrich for tumor tissue – whole slide sequencing was performed. The RNA was extracted from the tissue curls using the Qiagen RNeasy FFPE kit (Qiagen, cat. no 73504). Libraries were prepared using the Illumina RNA Prep with Enrichment kit (Illumina, cat. no 20040537) combined with the Illumina Exome Panel (Cat.no 20020183), and they were sequenced on an Illumina NextSeq™ system (Illumina). Resulting RNA-sequencing data were pre-processed, aligned to the hg38 genome, and evaluated for quality control (which led to the exclusion of 12 samples). Further details pertaining to quality control exclusion criteria are provided in the Supplementary text. No clear batch effects, linked to purely technical variables, were observed, except for differences due to surgical procedure, namely, biopsy vs excision. Out of 60,295 profiled genes, 14,958 were retained for downstream analyses after filtering on transcription type, variance, and expression level.

Somatic mutation calling

Somatic mutations (single to triple nucleotide substitutions) were identified in each sample using Mutect245 (v4.2.0.0), using tumor and matched normal samples as well as the hg19 GATK panel of normals. To remove artifacts, variant calls were filtered using an orientation bias model created with the GATK LearnReadOrientationModel function, and variants with fewer than four supporting altered reads or an allele fraction below 4% were excluded.

Since the normal tissue was macrodissected from skin adjacent to each tumor, we explored whether tumor contamination into the normal was occurring. In approximately 5% of samples, we detected tumor contamination in the reference DNA (up to 16% of total reads). To rescue mutations potentially missed by MuTect2 due to tumor contamination, we ran Mutect2 in tumor-only mode, using a panel of normals, to rescue somatic mutations that were incorrectly filtered out. Mutations identified in this analysis were further filtered to identify variants that: (i) were not germline in 1000genomes/dbSNP, (ii) had a mutant allele fraction (MAF) in the normal <0.25, (iii) had at least 6 fold coverage in the normal sample, (iv) normal MAF was less than half of tumor MAF, and (v) tumor MAF > 0.2 ×tumor cellularity. Mutations were annotated with Funcotator (v4.2.0.0, v1.6.20190124s data source). Additionally, the TERT promoter region (chr5: 1,295,120–1,295,350) was inspected in the Integrative Genomics Viewer (IGV) in all samples, and somatic mutations showing (i) at least two reads in the normal sample, (ii) at least 2 mutated reads in the tumor sample, and (iii) tumor cellularity adjusted MAF higher than 10% were included even if they were not detected as somatic mutations by Mutect2.

Indels were identified using Pindel46 (v0.2.5a7). Indels were inspected in IGV and removed if the reference also had reads that misaligned in the region. This typically occurred for Indels that resided in repetitive tracks of DNA or were adjacent to germline SNPs– common causes artifacts noted in the PCAWG study35.

Mutations present in at least 40% of tumor cells were considered clonal. Unless specified, only these mutations were included in the analyses. Tumor mutation burden was also calculated based on clonal mutations as described by Chang et. al.21. Table S8 contains all somatic mutations called.

Copy number alterations

Copy number alterations were identified using cnvkit47 (v0.9.6), using normal samples of each sequencing batch as panel of normals. For some sequencing batches, separate references were generated for samples with coverage above and below 100x. Chromosomal arm-level amplifications and deletions were identified based on a 0.1 cutoff applied to absolute log2 copy number ratios. Samples with tumor cellularity lower than 13.5% were excluded from copy number analyses because in these samples the cutoff of 0.1 is not sensitive enough to detect copy number deletions, as the expected absolute log2 ratio of single-copy gains would be lower than 0.1. To identify focal copy-number alterations, candidate amplifications and deep deletions were first called using CNVkit. Candidate events were then evaluated through a structured expert review process, modeled after the approach used by the Pan-Cancer Analysis of Whole Genomes working group, to assign each focal event to the most likely target gene when supported by the genomic context. Although focal copy-number alterations were uncommon in cSCC, this analysis identified additional tumors with deep deletions of CDKN2A, amplifications of CCND1, among other established cSCC driver genes.

Tumor cellularity

We inferred tumor cellularity as previously described21,48. Briefly, we used the following methods, whose assumptions and formulas are detailed in the supplementary text: “Modal Somatic MAF × 2”, “Median Somatic MAF in sex chromosomes” for male patients, “Copy number deletion log ratios”, “Allelic Imbalance of SNPs over copy number neutral LOH” and “Allelic Imbalance of SNPs over copy number deletions”, as described by Chang et. al.21. We used the average of at least two of these methods when possible. The specific methods used for each sample are indicated in Table S1; further details are in the supplementary text. Tumor cellularity was used to compute adjusted MAFs, by dividing the unadjusted MAF by the sample tumor cellularity fraction.

Mutational signatures

COSMIC reference mutational signatures were assigned to individual samples using SigProfilerAssignment49(v.0.2.3), using exome=True option and cosmic version 3.4. Samples where the signature linked to azathioprine exposure (SBS32) was detected were labeled as immunosuppressed in the whole-exome sequencing analyses.

Nomination of driver genes

To increase statistical power, we combined the mutation calls from the 147 tumors sequenced here with those from a meta-analysis of 83 cSCCs previously performed by our group21.

We used IntOGen24 (v2024.11) to nominate cancer genes, using the default settings, but without filtering out drivers with more than 2 mutations in a sample. Genes that were identified by at least one individual tool within IntOGen and with q-value <0.01 were considered as significant. Revision of all IntOGen hits (N=27) revealed that out of 6 genes with an oncogenic role, as determined by IntOGen, 2 (TPM3, SDHC) did not have any OncoKB25 oncogenic mutation in the entire dataset. Additionally, a known false positive21, which is driven by annotation errors, affecting KNSTRN was called. As such, these genes were not considered as drivers. There were two genes, LAMC1 and IGF2BP2, that were considered significant by IntOGen, but we did not include them as bona fide drivers because their mutations tended to be mono-allelic, and they do not operate in pathways known to be important in cancer or keratinocyte biology (Fig. S4C, Table S2).

Additionally, we queried OncoKB (accessed 06 October 2025), to find mutations in driver genes that are infrequently altered yet still pathogenic. Outside of IntOGen tumor suppressor driver genes, a total of 79 nonsynonymous mutations were annotated as likely gain/switch of function according to OncoKB. These were considered probably pathogenic in CSCC, except when the mutations were not in a known hotspot according to COSMIC (N=16); or have been shown to not be under selection in melanoma (N=3). TERT mutations outside C228T, C250T and CC242TT (N=18) were also not considered probably pathogenic, based on the work of Eliott et al50. For Loss-of-function mutations, OncoKB annotates any truncating mutations in known tumor suppressor genes as likely oncogenic. Given the high mutation burden in cSCC, we did not consider any mutation annotated as loss-of-function/likely loss-of-function as probably pathogenic in cSCC, unless they were missense mutations on a hotspot and present in COSMIC (N=2). Mutations with unknown or inconclusive effect according to OncoKB were still considered as pathogenic if they were in a clear hotspot according to cosmic (N=1). Table S3 shows all OncoKB annotated mutations in our cohort.

Counting driver events

For each gene, all alterations labeled as probably pathogenic above were considered as driver alterations. Additionally, focal amplifications in oncogenes were considered as driver events; as were all nonsynonymous alterations and focal deletions affecting tumor suppressor genes identified under selection by IntOGen. To identify tumor suppressor genes with bi-allelic alterations or oncogenes with increased gene dosage of the oncogenic allele, we identified genes with multiple hits by identifying genes harboring more than one driver mutation, as well as genes with driver mutations showing higher than expected allele frequencies (defined as tumor cellularity adjusted MAF > 0.75). For tumor suppressor genes, we also considered the gene to have 2 hits if a chromosomal arm deletion was detected in the same tumor with a point mutation. For USP9X, which is located in the X chromosome, a single hit in male samples was considered equivalent to a double hit.

Association of genomic alterations with metastatic outcome

We examined the association between each altered gene, pathway and copy number alteration with metastatic outcome, using univariate Cox proportional hazards regression models adjusted for callable coverage. For gene models, we used gene hit count as the independent variable, corresponding to the number of driver alterations in the gene, capped at 2 (i.e., 0,1 or 2). For pathway models, we grouped genes into pathways (as indicated in figure 2), summed all gene hits within the pathway and used the total hits within a pathway as the independent variable. For instance, if a tumor had bi-allelic loss-of-function mutations affecting NOTCH1 (2 hits) and an additional heterozygous mutation affecting NOTCH2 (1 hit), then the Notch pathway would have 3 total hits. For copy number alteration models, alteration presence (0 or 1) was used as the independent variable. Only genes/pathways with driver mutations present in at least 3 samples and chromosomal alterations present in at least 10 samples were assessed. The combination of pathway and chromosomal alterations that was most predictive of metastasis was identified using a multivariable regularized Cox regression model, with α=0.5 and λ chosen with cv.glmnet lambda.1se option. Only alterations present in at least 20% of samples were used as input features. The model was internally validated with bootstrap resampling (N=200).

Mutual exclusivity and co-occurrence analysis

Co-occurrence and mutual exclusivity were assessed for all driver mutation pairs with driver mutations in at least 10 samples (total of 19 genes), using two-sided Fisher’s exact tests. P-values were corrected for multiple testing with the Benjamini-Hochberg correction.

Clustering analysis

Sample clusters were found on the excision samples (N=197) using the cola51 R package (v.2.16.0) on the count data after DESeq2 variance stabilizing transformation, blinded to the metastatic outcome. Three feature selection methods were used: standard deviation, coefficient of variance and median absolute deviation, with the default clustering methods. Four solutions with K=2 showed perfect stability as measured by 1-Percentage of ambiguous clustering (PAC) (1-PAC=1), but there was intra-cluster heterogeneity in gene expression, implying that a two-cluster solution was missing some sample subtypes. As such, spherical k-means using median absolute deviation filtering with K=3 was selected as the final clustering solution because it showed high stability (1-PAC=0.953), the highest number of confident samples (i.e. samples with silhouette index higher than 0.5, N=194), and explained more outcome differences between tumors. Clusters were visualized using the 1000 genes with the highest association with the obtained sample clusters (determined using F-tests with the get_signatures function from the cola package). Overrepresentation analysis of each gene cluster was performed using clusterProfiler52 (v4.14.6).

Clusters found in the excision samples were validated in the biopsy samples (N=181) by projecting the biopsy samples onto the PCA space defined by the expression of the top 1000 genes in the excision samples. Clusters were then assigned to the biopsy samples using a nearest centroid classifier based on the first 3 principal components. Statistical tests used to assess the association between clinico-pathological variables/genomic alterations and genomic clusters are provided in the supplementary text.

Differential gene expression analysis

Differential gene expression analyses between cases and controls and between immunosuppressed and immunocompetent patients were performed using the DESeq253 R package (v1.46.0), with statistical adjustments for sample type (biopsy vs. excision). Gene set enrichment analysis was performed using fgsea54 (v1.32.4) with the fgseaMultilevel function. Genes were ranked by the DESeq2 Wald statistic, and enrichment was assessed against MSigDB Hallmark gene sets55. Pathways with an adjusted P < 0.05 were considered significant.

SCCore-GEP development

Candidate genes for modeling were selected using a structured workflow that combined expression-based filtering, biology-informed prioritization of differentiation/progenitor-like genes, and unbiased data-driven selection of remaining genes. To select highly expressed genes for the modelling, we additionally filtered the 14,958 genes used for differential gene expression analysis down to 11,144 genes based on the average log2(TPM+1). Then, these genes were divided into two sets: (i) genes belonging to the differentiation-progenitor-like signature31 (DvP genes) (n=403); and (ii) all other genes (n=10,741), referred to as non-DvP genes. Within each set, another pre-filtering step was performed using DESeq2 to identify differentially expressed genes between cases and controls. This filter was performed separately within the DvP genes (biology-inspired feature selection) and within the non-DvP genes (data-driven feature selection) to promote the selection of genes involved in a diverse set of pathways.

We evaluated different combinations of modeling approaches and feature integration strategies. Specifically, we compared regularized Cox’s proportional hazards (linear model) and Random Survival Forest (RSF, a nonlinear model) models, each paired with different feature integration approaches of gene sets (no integration, early integration, or late integration at the feature or prediction level). These models were implemented as described in the Supplementary text, and only samples from complete case-control sets were used in signature development (N=366 samples). The final approach - based on Cox regression - entails two steps: 1) Cox regressions with elastic net regularization were fitted separately to the two gene sets to identify a reduced set of features associated with metastasis; then, 2) the selected features were integrated (late integration) in a single Cox regression with L2 regularization. The model was internally validated using 200 bootstrapped samples. Discriminative ability was evaluated by weighted C-index, an adaptation of the C-index adjusted to the NCC design56.

SCCore-GEP genes assessment on spatial transcriptomics

Spatial transcriptomics was performed on a 10X Genomics FFPE Visium HD platform (following protocols CG000684 Rev A and CG000685 Rev B), and the resulting sequencing data were processed with the SpaceRanger pipeline (v.4.0). Supplementary Text contains further details on cluster annotation. To quantify cell compartment representations of the 23 genes included in SCCore-GEP, reads from cells assigned to each annotated cell type were aggregated per patient and expressed as a fraction of the total reads for that gene across all cells in that sample. Average proportion across patients (N=8) are reported, reflecting the relative contribution of each cell compartment to the overall expression of the signature genes.

Validation datasets

SCCore-GEP was validated in an independent RNA-sequencing NCC dataset (N=102) derived from a second nationwide Dutch cohort of 25,921 patients, which was also part of the D-SQUAME study18. This dataset is a subset of the complete NCC D-SQUAME validation dataset and encompasses those samples whose sequencing data were available at the time of the study. This validation dataset differs from the discovery dataset in that patients had their first cSCC more recently (2017–2018), and no risk-matching of the controls was performed. Additionally, the SCCore-GEP was validated on an external RNA-sequencing dataset of 52 primary cSCCs15, where tumors were matched based on BWH staging and immunosuppression status. Limited clinical information was available for this dataset: the metastatic status was available as a binary variable (i.e., no time-to-event data), and BWH stages were extracted from samples’ annotation. Further information on how these datasets were processed is available in the supplementary text.

Ethics statement

The study was approved by the Medical Ethics Review Committee of the Erasmus Medical Center for the discovery and validation datasets (MEC-2020–0147), as well as the scientific and privacy committees of the NCR, PALGA, and NOTR.

Patents, Royalties, Other Intellectual Property

SCCore-GEP was patented (EP26151074, Methods for predicting the risk of metastasis in cutaneous squamous cell carcinoma (cSCC)).

Supplementary Material

Supplement 1
media-3.xlsx (21.3KB, xlsx)
Supplement 2
media-4.xlsx (131.9KB, xlsx)
Supplement 3
media-5.xlsx (16.2KB, xlsx)
Supplement 4
media-6.xlsx (2.1MB, xlsx)
Supplement 5
media-7.xlsx (14.3KB, xlsx)
Supplement 6
media-8.xlsx (12.5KB, xlsx)
Supplement 7
media-9.xlsx (46.2MB, xlsx)
1

Acknowledgements

We thank the data managers and registrars of the NCR, Palga and NOTR for compiling and linking patient and tumor data, thereby enabling tumor identification at the nationwide level. We thank the staff of the pathology laboratories for retrieving the requested tumor blocks, as well as Selin Tokez, Amber Khan, Rosa Almand, Andrya Hollatz, and Nikita Hulscher for assistance in reviewing pathology reports and processing the tumor samples. We thank Thierry van den Bosch and members of Erasmus MC Pathology Research and Trial Service (PARTS) for DNA/RNA isolation; David van Klaveren for statistical consulting and Erik Valent, Daniëlle van Keulen, Rowan Kuiper, Sheril Alex and Jvalini Dwarkasing for continuous support in the project.

Funding

The project was co-funded by the PPP Allowance made available by Health~Holland, Top Sector Life Sciences & Health, to stimulate public-private partnerships (EMCLSH21019); and it was also supported by grants from the National Institutes of Health (NIH) (U01CA294536, R01AR080626, and R01CA265786).

Competing interests

BR-P is funded by a PPP Allowance made available by Health ~ Holland, Top Sector Life Sciences & Health, to stimulate public–private partnerships, of which SkylineDx is a contributing member. LP, JJHT, RR, DH, Y-TC, BI, AB-CM, DoB are SkylineDx employees. MW has served as a paid speaker for Regeneron and an advisory board member for Regeneron and SunPharma. Employees of Erasmus MC (BR-P, MW and LMH) and SkylineDx (LP, JJHT, RR, DH and DoB) are inventors on the patent EP26167218.2, related to the development of SCCore-GEP. All other authors have no competing interests.

Funding Statement

The project was co-funded by the PPP Allowance made available by Health~Holland, Top Sector Life Sciences & Health, to stimulate public-private partnerships (EMCLSH21019); and it was also supported by grants from the National Institutes of Health (NIH) (U01CA294536, R01AR080626, and R01CA265786).

Code Availability

Code is available at https://github.com/barentroia/onco_cSCC_genomic_characterization/.

Data Availability

Raw exome sequencing data and RNA sequencing data were made available through the European Genome Archive (Accession number EGAS50000001852). Clinical data according to the Human Tumor Atlas Network (HTAN) phase 2 data model is available upon request at the Netherlands Cancer Registry (gegevensaanvraag@iknl.nl) upon the following description: “HTAN dataset from data request numbers: K19.054 (D-SQUAME discovery cohort) and K23.132 (D-SQUAME validation cohort)”.

References

  • 1.Zhang W. et al. Global, regional and national incidence, mortality and disability-adjusted life-years of skin cancers and trend analysis from 1990 to 2019: An analysis of the Global Burden of Disease Study 2019. Cancer Med 10, 4905–4922 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2.Nehal K. S. & Bichakjian C. K. Update on Keratinocyte Carcinomas. N. Engl. J. Med. 379, 363–374 (2018). [DOI] [PubMed] [Google Scholar]
  • 3.Tokez S., Hollestein L., Louwman M., Nijsten T. & Wakkee M. Incidence of Multiple vs First Cutaneous Squamous Cell Carcinoma on a Nationwide Scale and Estimation of Future Incidences of Cutaneous Squamous Cell Carcinoma. JAMA Dermatol 156, 1300–1306 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4.Keim U. et al. Incidence, mortality and trends of cutaneous squamous cell carcinoma in Germany, the Netherlands, and Scotland. Eur J Cancer 183, 60–68 (2023). [DOI] [PubMed] [Google Scholar]
  • 5.Mudigonda T., Pearce D. J., Yentzer B. A., Williford P. & Feldman S. R. The economic impact of non-melanoma skin cancer: a review. J Natl Compr Canc Netw 8, 888–896 (2010). [DOI] [PubMed] [Google Scholar]
  • 6.Venables Z. C. et al. Epidemiology of basal and cutaneous squamous cell carcinoma in the U.K. 2013–15: a cohort study. Br J Dermatol 181, 474–482 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7.Olsen C. M., Pandeya N., Ragaini B. S., Neale R. E. & Whiteman D. C. International patterns and trends in the incidence of melanoma and cutaneous squamous cell carcinoma, 1989–2020. Br J Dermatol 190, 492–500 (2024). [DOI] [PubMed] [Google Scholar]
  • 8.Tokez S. et al. Cumulative incidence and disease-specific survival of metastatic cutaneous squamous cell carcinoma: A nationwide cancer registry study. J Am Acad Dermatol 86, 331–338 (2022). [DOI] [PubMed] [Google Scholar]
  • 9.Urban K., Mehrmal S., Uppal P., Giesey R. L. & Delost G. R. The global burden of skin cancer: A longitudinal analysis from the Global Burden of Disease Study, 1990–2017. JAAD Int 2, 98–108 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10.Jambusaria-Pahlajani A. et al. Evaluation of AJCC tumor staging for cutaneous squamous cell carcinoma and a proposed alternative tumor staging system. JAMA Dermatol 149, 402–410 (2013). [DOI] [PubMed] [Google Scholar]
  • 11.Lydiatt W. M. et al. Head and Neck cancers-major changes in the American Joint Committee on cancer eighth edition cancer staging manual. CA Cancer J Clin 67, 122–137 (2017). [DOI] [PubMed] [Google Scholar]
  • 12.Venables Z. C. et al. Validation of four cutaneous squamous cell carcinoma staging systems using nationwide data. Br J Dermatol 186, 835–842 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13.Rentroia-Pacheco B. et al. Personalised decision making to predict absolute metastatic risk in cutaneous squamous cell carcinoma: development and validation of a clinico-pathological model. EClinicalMedicine 63, 102150 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14.Diede C., Walker T., Carr D. R. & Shahwan K. T. Grading differentiation in cutaneous squamous cell carcinoma: a review of the literature. Arch Dermatol Res 316, 434 (2024). [DOI] [PubMed] [Google Scholar]
  • 15.Nassir S. et al. Multiomic Sequencing of Intermediate- to High-Risk Cutaneous Squamous Cell Carcinoma Identifies Critical Genes and Expression Patterns Associated with Disease and Poor Outcomes. J Invest Dermatol 145, 2060–2070.e5 (2025). [DOI] [PubMed] [Google Scholar]
  • 16.Wysong A. et al. Validation of a 40-gene expression profile test to predict metastatic risk in localized high-risk cutaneous squamous cell carcinoma. J Am Acad Dermatol 84, 361–369 (2021). [DOI] [PubMed] [Google Scholar]
  • 17.Wang J. et al. Transcriptomic analysis of cutaneous squamous cell carcinoma reveals a multigene prognostic signature associated with metastasis. J Am Acad Dermatol 89, 1159–1166 (2023). [DOI] [PubMed] [Google Scholar]
  • 18.Steijlen O. F. M. et al. The Dutch cutaneous squamous cell carcinoma and metastasis (D-SQUAME) study: a nationwide discovery cohort and nationwide validation cohort with nested case-control designs for risk prediction modeling. Eur J Epidemiol https://doi.org/10.1007/s10654-026-01384-4 (2026) doi: 10.1007/s10654-026-01384-4. [DOI] [Google Scholar]
  • 19.Eggermont C. J. et al. Skin cancer risk in more than 200 000 patients with haematological malignancies over 30 years: a nationwide population-based study in the Netherlands. Br J Dermatol 192, 1029–1037 (2025). [DOI] [PubMed] [Google Scholar]
  • 20.Rentroia-Pacheco B. et al. Nested Case-Control and Case-Cohort: Efficient Study Designs to Develop Biomarker-Based Prediction Models for Rare Outcomes. J Invest Dermatol 146, 591–601 (2026). [DOI] [PubMed] [Google Scholar]
  • 21.Chang D. & Shain A. H. The landscape of driver mutations in cutaneous squamous cell carcinoma. NPJ Genom Med 6, 61 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22.Inman G. J. et al. The genomic landscape of cutaneous SCC reveals drivers and a novel azathioprine associated mutational signature. Nat Commun 9, 3667 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 23.Lawrence M. S. et al. Discovery and saturation analysis of cancer genes across 21 tumour types. Nature 505, 495–501 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 24.Martínez-Jiménez F. et al. A compendium of mutational cancer driver genes. Nat. Rev. Cancer https://doi.org/10.1038/s41568-020-0290-x (2020) doi: 10.1038/s41568-020-0290-x. [DOI] [Google Scholar]
  • 25.Chakravarty D. et al. OncoKB: A Precision Oncology Knowledge Base. JCO Precis Oncol 2017, (2017). [Google Scholar]
  • 26.Egolf S. et al. MLL4 mediates differentiation and tumor suppression through ferroptosis. Sci Adv 7, eabj9141 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 27.Cammareri P. et al. Inactivation of TGFβ receptors in stem cells drives cutaneous squamous cell carcinoma. Nat Commun 7, 12493 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 28.Knudson A. G. Two genetic hits (more or less) to cancer. Nat. Rev. Cancer 1, 157–162 (2001). [DOI] [PubMed] [Google Scholar]
  • 29.Knudson A. G. Mutation and Cancer: Statistical Study of Retinoblastoma. Proc Natl Acad Sci U S A 68, 820–823 (1971). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 30.Helming K. C., Wang X. & Roberts C. W. M. Vulnerabilities of mutant SWI/SNF complexes in cancer. Cancer Cell 26, 309–317 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31.Bailey P. et al. Driver gene combinations dictate cutaneous squamous cell carcinoma disease continuum progression. Nat Commun 14, 5211 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32.Tandukar B. et al. Genetic evolution of keratinocytes to cutaneous squamous cell carcinoma. Nat Commun 16, 10663 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 33.Zakhem G. A., Qiblawi S., Shelton E. & Xu Y. G. Prevalence of poor outcomes in cutaneous squamous cell carcinoma by AJCC and BWH tumor stages: A systematic review and meta-analysis. Journal of the American Academy of Dermatology 92, 1064–1071 (2025). [DOI] [PubMed] [Google Scholar]
  • 34.Bencomo T. & Lee C. S. Gene expression landscape of cutaneous squamous cell carcinoma progression. Br J Dermatol ljae249 (2024) doi: 10.1093/bjd/ljae249. [DOI] [Google Scholar]
  • 35.ICGC/TCGA Pan-Cancer Analysis of Whole Genomes Consortium. Pan-cancer analysis of whole genomes. Nature 578, 82–93 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 36.Karia P. S., Morgan F. C., Califano J. A. & Schmults C. D. Comparison of Tumor Classifications for Cutaneous Squamous Cell Carcinoma of the Head and Neck in the 7th vs 8th Edition of the AJCC Cancer Staging Manual. JAMA Dermatol 154, 175–181 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 37.(NCR) NCR. The Netherlands Comprehensive Cancer Organisation. 2023. https://iknl.nl/en/ncr.
  • 38.Casparie M. et al. Pathology Databanking and Biobanking in The Netherlands, a Central Role for PALGA, the Nationwide Histopathology and Cytopathology Data Network and Archive. Analytical Cellular Pathology 29, 19–24 (2007). [Google Scholar]
  • 39.Nederlandse Transplantatie Stichting (NTS). Nederlandse Transplantatie Stichting, Leiden, The Netherlands. https://transplantatiestichting.nl/samenwerken/data.
  • 40.Eggermont C. et al. Development and Validation of an Algorithm to Identify Patients with Advanced Cutaneous Squamous Cell Carcinoma from Pathology Reports. Journal of Investigative Dermatology 143, 98–104.e5 (2023). [DOI] [PubMed] [Google Scholar]
  • 41.Verjans R. et al. The Dutch National TissueArchive Portal enables efficient, consistent, and transparent procurement of diagnostic tissue samples for scientific use. Cell Tissue Bank 22, 727–736 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 42.Chen S., Zhou Y., Chen Y. & Gu J. fastp: an ultra-fast all-in-one FASTQ preprocessor. Bioinformatics 34, i884–i890 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 43.Li H. Aligning sequence reads, clone sequences and assembly contigs with BWA-MEM. Preprint at 10.48550/ARXIV.1303.3997 (2013). [DOI]
  • 44.Van Der Auwera G. A. et al. From FastQ Data to High-Confidence Variant Calls: The Genome Analysis Toolkit Best Practices Pipeline. CP in Bioinformatics 43, (2013). [Google Scholar]
  • 45.Benjamin D. et al. Calling Somatic SNVs and Indels with Mutect2. Preprint at 10.1101/861054 (2019). [DOI]
  • 46.Ye K., Schulz M. H., Long Q., Apweiler R. & Ning Z. Pindel: a pattern growth approach to detect break points of large deletions and medium sized insertions from paired-end short reads. Bioinformatics 25, 2865–2871 (2009). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 47.Talevich E., Shain A. H., Botton T. & Bastian B. C. CNVkit: Genome-Wide Copy Number Detection and Visualization from Targeted DNA Sequencing. PLoS Comput Biol 12, e1004873 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 48.Shain A. H. et al. The Genetic Evolution of Melanoma from Precursor Lesions. N Engl J Med 373, 1926–1936 (2015). [DOI] [PubMed] [Google Scholar]
  • 49.Díaz-Gay M. et al. Assigning mutational signatures to individual samples and individual somatic mutations with SigProfilerAssignment. Bioinformatics 39, btad756 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 50.Elliott K. et al. Mechanistic basis of atypical TERT promoter mutations. Nat Commun 15, 9965 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 51.Gu Z., Schlesner M. & Hübschmann D. cola : an R/Bioconductor package for consensus partitioning through a general framework. Nucleic Acids Research 49, e15–e15 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 52.Xu S. et al. Using clusterProfiler to characterize multiomics data. Nat Protoc 19, 3292–3320 (2024). [DOI] [PubMed] [Google Scholar]
  • 53.Love M. I., Huber W. & Anders S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol 15, 550 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 54.Korotkevich G. et al. Fast gene set enrichment analysis. Preprint at 10.1101/060012 (2016). [DOI] [Google Scholar]
  • 55.Liberzon A. et al. The Molecular Signatures Database Hallmark Gene Set Collection. Cell Systems 1, 417–425 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 56.Rentroia-Pacheco B. et al. Weighted metrics are required when evaluating the performance of prediction models in nested case-control studies. BMC Med Res Methodol 24, 115 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplement 1
media-3.xlsx (21.3KB, xlsx)
Supplement 2
media-4.xlsx (131.9KB, xlsx)
Supplement 3
media-5.xlsx (16.2KB, xlsx)
Supplement 4
media-6.xlsx (2.1MB, xlsx)
Supplement 5
media-7.xlsx (14.3KB, xlsx)
Supplement 6
media-8.xlsx (12.5KB, xlsx)
Supplement 7
media-9.xlsx (46.2MB, xlsx)
1
NIHPP2026.07.17.26358051V1-supplement-1.pdf (8.5MB, pdf)

Data Availability Statement

Code is available at https://github.com/barentroia/onco_cSCC_genomic_characterization/.

Raw exome sequencing data and RNA sequencing data were made available through the European Genome Archive (Accession number EGAS50000001852). Clinical data according to the Human Tumor Atlas Network (HTAN) phase 2 data model is available upon request at the Netherlands Cancer Registry (gegevensaanvraag@iknl.nl) upon the following description: “HTAN dataset from data request numbers: K19.054 (D-SQUAME discovery cohort) and K23.132 (D-SQUAME validation cohort)”.

Articles from medRxiv are provided here courtesy of Cold Spring Harbor Laboratory Preprints

RESOURCES