Identification of proteins associated with development of metastasis from cutaneous squamous cell carcinomas (cSCCs) via proteomic analysis of primary cSCCs *

Cutaneous squamous cell carcinoma (cSCC) is one of the most common cancers capable of metastasizing. Proteomic analysis of cSCCs can provide insight into the biological processes responsible for metastasis, as well as future therapeutic targets and prognostic biomarkers.


Summary
Background Cutaneous squamous cell carcinoma (cSCC) is one of the most common cancers capable of metastasizing. Proteomic analysis of cSCCs can provide insight into the biological processes responsible for metastasis, as well as future therapeutic targets and prognostic biomarkers. Objectives To identify proteins associated with development of metastasis in cSCC. Methods A proteomic-based approach was employed on 105 completely excised, primary cSCCs, comprising 52 that had metastasized (P-M) and 53 that had not metastasized at 5 years post-surgery (P-NM). Formalin-fixed, paraffin-embedded cSCCs were microdissected and subjected to proteomic profiling after one-dimensional (1D), and separately two-dimensional (2D), liquid chromatography fractionation.
Results A discovery set of 24 P-Ms and 24 P-NMs showed 144 significantly differentially expressed proteins, including 33 proteins identified via both 1D and 2D separation, between P-Ms and P-NMs. Several differentially expressed proteins were also associated with survival in SCCs of other organs. The findings were verified by multiple reaction monitoring on six peptides from two proteins, annexin A5 (ANXA5) and dolichyl-diphosphooligosaccharide-protein glycosyltransferase noncatalytic subunit (DDOST), in the discovery group and validated on a separate cohort (n = 57). Increased expression of ANXA5 and DDOST was associated with reduced time to metastasis in cSCC and decreased survival in cervical and oropharyngeal cancer. A prediction model using ANXA5 and DDOST had an area under the curve of 0Á93 (confidence interval 0Á83-1Á00), an accuracy of 91Á2% and higher sensitivity and specificity than cSCC staging systems currently in clinical use. Conclusions This study highlights that increased expression of two proteins, ANXA5 and DDOST, is significantly associated with poorer clinical outcomes in cSCC.
What is already known about this topic?
• Keratinocyte cancer is the most common cancer in the UK, and the capacity for cutaneous squamous cell carcinomas (cSCCs) to metastasize presents a clinical problem.
• Although there are known clinical risk factors for cSCC metastasis, current staging systems are inaccurate at predicting the development of metastasis in patients with cSCC.
• It has been shown that mass spectrometry-based proteomic analysis can quantify and uncover potential key proteins in cancer development and metastasis.

What does this study add?
• This study has identified a number of proteins that are differentially expressed between primary cSCCs that metastasize and primary cSCCs that do not metastasize.
• Expression of the genes encoding several of these proteins influences the outcome in SCCs of other organs (lung, oropharynx, cervix and oesophagus).
• Higher abundances of two key proteins, annexin A5 (ANXA5) and dolichyldiphosphooligosaccharide-protein glycosyltransferase noncatalytic subunit (DDOST), are associated with the development of, and reduced time to, cSCC metastasis.
What is the translational message?
• This is the first study to undertake proteomic profiling using mass spectrometry to investigate proteins that are differentially expressed between human primary cSCCs that metastasize and those that do not metastasize.
• The results of this proteomic analysis of cSCCs will be useful for identifying potential therapeutic targets in this cancer.
• A prediction model incorporating ANXA5 and DDOST showed higher sensitivity and specificity than cSCC clinical staging systems for estimating the likelihood of cSCC metastases.
The number of keratinocyte cancers in the UK is > 211 120 annually, with cutaneous squamous cell carcinoma (cSCC) accounting for > 44 672, constituting one of the most common types of cancer capable of metastasizing. 1,2 The risk of metastasis for cSCC depends on clinical and histological parameters, including the site, depth of invasion, diameter, differentiation of the tumour, presence of lymphovascular or perineural invasion, and host immunosuppression. 3 Following surgical excision, cSCC metastasizes in 16% of cases with tumour depth > 6 mm, 4 and in 30% of tumours > 2 cm in diameter. 5 Whereas the 3-year disease-specific survival rate for patients with cSCC is 85%, 6 for patients with distant metastasis the median survival is < 2 years. 7 Staging systems assist identification of patients at greater risk of metastases after excision of primary cSCC. 8,9 However, current staging systems distinguish 'poorly to moderately' between patients who do and do not develop cSCC metastases, 8 and one-third of patients are classified incorrectly using these staging systems. 10 There is a need to undertake research into factors that contribute to more aggressive tumours, 11 to understand the mechanisms responsible for development of metastases in cSCC and to identify more accurately those patients at risk of metastases.
Proteomic analysis can aid in understanding the aetiology of cancer progression and provide information of prognostic relevance. 12 In this study we used a mass spectrometry-based proteomic approach on cSCCs to identify proteins involved in development of metastases. The results highlight a number of differentially expressed proteins that associate with occurrence of metastases from cSCC, and reduced survival in lung, cervical, oropharyngeal and oesophageal SCC.

Tissue samples
Formalin-fixed paraffin-embedded (FFPE) human primary cSCCs were acquired from Histopathology, University Hospital Southampton NHS Foundation Trust (UHS-NHSFT) under ethics committee approval (South Central Hampshire B National Research Ethics Service Committee; LREC number 07/H0504/ 187). Samples were categorized as primary cSCCs that metastasized (P-M) or primary cSCCs that had not metastasized at 5 years post-surgery (P-NM), with the latter based on no evidence of metastasis during 5 years of follow-up and/or patient review for another reason after 5 years in Dermatology, UHS-NHSFT.
Sample preparation for mass spectrometry FFPE tissue sections were mounted onto glass slides, and tumour and surrounding immune infiltrate were microdissected and transferred into protein extraction buffer (Appendix S1; see Supporting Information). Samples were heated to 105°C for 30 min, cooled, then heated to 80°C for 2 h before reduction using dithiothreitol and alkylation with iodoacetamide. Samples were digested with sequencinggrade trypsin overnight and the resulting peptides were desalted using C18 reverse-phase clean-up plates.

Discovery liquid chromatography mass spectrometry
Samples were fractionated using a nanoACQUITY UPLC System (Waters, Milford, MA, USA) and electrosprayed into a Waters SYNAPT-G2-Si mass spectrometer operating in MS E mode with ion mobility activated (Appendix S1; see Supporting Information). Estimates of absolute quantification using the Top3 approach 13 were obtained using one-dimensional (1D) and two-dimensional (2D) liquid chromatography (LC) separation strategies. Data from the 1D and 2D LC procedures were analysed separately. Three blank runs were conducted between samples to ensure avoidance of carry-over into subsequent samples.

Multiple reaction monitoring
A spectral library from the discovery proteomic data was generated using Skyline software 14 to identify unique peptides for proteins of interest. Heavy stable-isotope-labelled (SIL) peptides were synthesized by Cambridge Research Biochemicals (Billingham, UK). Calibration curves were created using 1 µg cSCC 'proteomic-ready' sample as background. The High Definition MRM (multiple reaction monitoring) acquisition mode was used for targeted acquisition. Transitions for each peptide were identified using Skyline and imported into MassLynx (Waters) for targeted acquisition. Samples were analysed containing 100 fmoL of each heavy SIL peptide. Raw data were imported into Skyline for interpretation and calculation of native peptide quantity.

Gene expression in other squamous cell carcinomas
Expression levels of relevant genes were analysed in publicly available RNA-sequencing data from The Cancer Genome Atlas (TCGA) research network (http://cancergenome.nih.gov). 15 Computational analysis and statistical testing of next-generation sequencing data were conducted using R statistical programming language. 16 Filtered and log 2 -normalized RNA expression data, alongside available clinical data, were downloaded from the GDAC firehose database (run: std-data__2015_06_01; https://gdac.broadinstitute.org). Plotting of TCGA data was performed using the ggplot2 package in R. 17 Survival analysis was performed using the R packages survminer and survival. 18 Kaplan-Meier survival curves were constructed using TCGA clinical data. Statistical testing of differences between survival curves used the G-rho family of tests, as implemented in the survdiff function of the survival package.

Data analysis
A 1% false discovery rate was applied to searching for peptide identification. Each protein was inferred from identification of at least one unique peptide. Only proteins detected in ≥ 50% of samples were subsequently analysed. Data were normalized to the median protein concentration for each sample and Pvalues were obtained by Mann-Whitney U-test. Topological data analysis, using Ayasdi (Palo Alto, CA, USA), was performed on complete, normalized proteomic data with a hamming metric and two neighbourhood lenses. For Kaplan-Meier survival analysis, P-values were obtained by log-rank test. Machine learning was performed using the statistical programming language, R, with the packages caret and caretensemble.

Discovery proteomics
This study investigated proteomic differences between P-M and P-NM cSCCs to identify proteins associated with metastasis in cSCC. As expected, more patients in the P-M than in the P-NM group had poorly differentiated tumours or perineural invasion or were immunosuppressed (Table 1). A discovery group of 24 P-M and 24 P-NM samples was subjected to proteomic profiling using 1D, and independently 2D, separation to identify and quantify differences in protein abundance between P-Ms and P-NMs. Microdissected cSCC samples included tumour keratinocytes and stromal regions containing the immune cell infiltrate (Figure 1a). Volcano plots demonstrated higher numbers of upregulated than downregulated proteins in P-M compared with P-NM cSCCs (Figure 1b, c). Overall, 4018 unique proteins were identified in the cSCCs (Figure 1d), of which 144 were significantly differentially expressed between P-Ms and P-NMs (P < 0Á05; Tables S1 and S2; see Supporting Information). This included 33 proteins identified via both 1D and 2D proteomics ( Figure 1e and Table 2). Topological data analysis of the 48 proteomes from the discovery set of 24 P-M and 24 P-NM cSCCs, performed without including input information on metastases or any other clinical data, demonstrated separation of samples in both 1D and 2D analyses according to development of metastases (Figure 1f, g). This provides support for distinct proteomic profiles of P-M and P-NM cSCCs.

Pathway analysis
Weighted gene coexpression network analysis (WGCNA) of the proteomics data was conducted and, following construction of a signed topological overlap matrix (TOM) of corresponding dissimilarity, hierarchical clustering was used on the dissimilarity TOM to produce modules of genes (Figure S1a, b; see Supporting Information). Modules were examined for correlation with clinical and immunohistochemical characteristics ( Figure S1c) in addition to analysis for pathway enrichment ( Figure S1d).
Immunohistochemical characterization showed significantly fewer CD8 + T cells and CD1a + Langerhans cells in P-M than in P-NM samples (Figure 2a Table S3; see Supporting Information). Lower numbers of CD8 + , and separately CD1a + , cells were significantly associated with reduced time to metastasis (P = 0Á0041 and P = 0Á0057, respectively; Figure 2c, d).
To identify cell-signalling pathways associated with development of cSCC metastasis, Search Tool for the Retrieval of Interacting Genes/Proteins (STRING) analysis with Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway mapping of significantly differentially expressed proteins in the 1D and 2D data was conducted. STRING analysis demonstrated highly connected structures with clusters ( Figure S2a, b; see Supporting Information). KEGG pathway enrichment highlighted ribosomal proteins, protein processing in the endoplasmic reticulum, focal adhesion, extracellular matrix-receptor interactions, phosphoinositide 3-kinase (PI3K)-Akt signalling, and antigen processing and presentation as key differences between P-Ms and P-NMs ( Figure S2c; see Supporting Information).

Comparison with The Cancer Genome Atlas
To determine whether proteins involved in development of cSCC metastases influence development of metastases in other SCC types, the 33 significantly differentially expressed proteins in the 1D and 2D proteomic data were compared against gene expression in cervical, oropharyngeal, oesophageal and lung SCC using TCGA. Expression levels of genes encoding for several proteins differentially expressed between P-M and P-NM cSCCs were identified as having significant effects on survival in SCCs arising at these other sites, with reduced survival associated, separately, with high expression of POSTN, DDOST, HNRNPK, COL6A3, ANXA5 and LCP1, and with low expression of CALML5 (Figure 3a-n; and Table S3; see Supporting Information). Furthermore, as immune dysfunction is important for cSCC development, and as L-plastin (LCP1) can stimulate the T-cell receptor and activate T cells, 19 immunohistochemistry for LCP1 was conducted on the discovery group of cSCCs. This demonstrated more LCP1 + cells in P-Ms than in P-NMs (Figure 3o, p; and Table S3; see Supporting Information).

Multiple reaction monitoring
MRM was used to validate the discovery proteomics. MRM is a highly sensitive and specific mass spectrometry method that involves filtering the mass spectrometer on specific peptides of interest and quantifying these against known concentrations of isotopically labelled peptides spiked into the samples, enabling greater sensitivity and more accurate quantification of protein The data are presented as n (%) unless stated otherwise. Some samples were used for both proteomic and immunohistochemical analysis, whereas other samples were used for proteomic or immunohistochemical analysis according to the amount of tissue available. Two samples from each group were removed during multiple reaction monitoring analysis due to the limited amount of tissue available.
concentrations. Firstly, machine learning (using a generalized linear model) was conducted on significantly differentially expressed proteins between P-M and P-NM cSCCs, in which a model predicting cSCC metastases was produced for every combination of two proteins on a training set and tested on a holdout cohort (two-thirds and one-third split, respectively). From > 300 models, the combination of ANXA5 and DDOST gave one of the best area under the curve results, and because expression of both of these genes had been identified via TCGA as being important in reducing survival in SCCs of other organs, ANXA5 and DDOST were selected for targeted verification and validation using MRM.
Three unique peptides per protein were identified using Skyline software and synthesized as SIL peptides ( Figure S3; see Supporting Information). MRM of the discovery cSCC group (22 P-Ms and 22 P-NMs) verified that there was more DDOST and ANXA5 in P-M than P-NM cSCCs (DDOST P = 0Á0036, ANXA5 P = 0Á0046; Figure 4a-d; and Table S3; see Supporting Information). MRM for DDOST and ANXA5 was then conducted in a different (i.e. validation) group of cSCCs, comprising 28 P-Ms and 29 P-NMs. Again, DDOST and ANXA5 levels were significantly higher in P-M than P-NM cSCCs (DDOST P < 0Á001, ANXA5 P < 0Á001; Figure 4e-h; and Table S3; see Supporting Information). Fold changes for individual proteins were calculated by dividing the mean of P-M by the mean of P-NM; blue P > 0Á05, green P < 0Á05, red P < 0Á01. (d, e) Venn diagrams of (d) the total number of unique proteins identified in 1D and 2D proteomes and (e) the number of significantly differentially expressed proteins between P-M and P-NM cSCCs. (f, g) Topological data analysis (which analyses datasets using systems derived from topology) of (f) the whole 1D proteome and (g) the whole 2D proteome demonstrates separation of samples according to metastasis status; nodes represent a cluster of samples (two or more) with highly similar proteomes; edges (lines between nodes) indicate similarity between the clusters. Survival analyses were conducted to investigate the relationship between ANXA5 and DDOST expression and clinical outcome. High expression of ANXA5 and DDOST was associated with reduced time to cSCC metastasis (P < 0Á001, Figure 5a). P-M cSCCs were associated with a reduced time to death compared with P-NM cSCCs (P < 0Á001, Figure 5b), and high expression of ANXA5 and DDOST was also associated with reduced 5-year overall survival (P = 0Á024, Figure 5c). Moreover, TCGA analysis demonstrated that high coexpression of ANXA5 and DDOST significantly reduces survival in cervical and oropharyngeal SCC (P = 0Á046 and P = 0Á0072, respectively; Figure 5d, e).
A stacked ensemble prediction model with the ANXA5 and DDOST MRM data was created using R software and base-level algorithms comprising k nearest neighbours, naive Bayes, glmnet, AdaBoost, xgbDART and the stochastic gradient boosting GBM. The predictions of these individual algorithms were then subjected to a top-layer algorithm, xgbTree, to form final predictions for each sample. Data were split into two-thirds (n = 67) for training and one-third (n = 34) for testing, and models were trained using 10-fold cross-validation repeated three times. The resulting prediction-model receiver operating characteristic curve gave an area under the curve of 0Á93 (Figure 5f). This ANXA5-DDOST prediction model was compared on the same cSCC samples with cSCC clinical staging systems, including American Joint Committee on Cancer 7th and 8th editions, 20,21 Brigham and Women's Hospital, 9 British Association of Dermatologists, 22 Breuninger et al., 23 European Dermatology Forum 7 and Union for International Cancer Control, 24 and with results of the validation study of some of these staging systems by Roscher et al. on their patient cohort. 8 This comparison showed that the ANXA5-DDOST prediction model has higher sensitivity and specificity than each of these staging systems.

Discussion
This proteomics-based study identified multiple proteins associated with development of cSCC metastases and ascertained that high levels of expression of several respective genes   encoding for these proteins associate with reduced survival in SCCs of the cervix, oropharynx, oesophagus and lung. Although mass spectrometry for proteomic analysis of cSCCs has been employed previously, 25 to our knowledge, the current study is the first to investigate differential expression of proteins in primary cSCC with respect to metastasis and clinical outcome. Our topological data analysis was largely able to separate cSCCs according to development of metastases, providing strong support for involvement of the detected proteins in the metastatic process, although it is not possible to  conclude from this study what proportion of these are drivers or passengers in this process.
Some differences in protein expression between P-M and P-NM cSCCs may be due to variation in tumour parameters (e.g. cell proliferation, differentiation status) or composition of the immune infiltrate between the two tumour groups. However, bioinformatic analysis highlighted several pathways and processes likely to be causally involved in permitting cSCC metastases. STRING/KEGG identified differences between P-Ms and P-NMs in PI3K-Akt signalling, which influences development of cancer metastasis 26 and can affect cSCC growth. 27 Indeed, PI3K-Akt signalling pathways differ between well-differentiated and moderately or poorly differentiated cSCCs, 28    STRING/KEGG also identified extracellular matrix-receptor interaction and enrichment of focal adhesion, which are important for cancer invasion and metastases, 30,31 in P-M compared with P-NM samples. Additionally, STRING/KEGG identified 'antigen processing and presentation' differences between P-M and P-NM, consistent with our observations that lower numbers of CD1a + Langerhans cells and CD8 + T cells in cSCCs associate with metastasis, and our previous work demonstrating that cSCC Tregs suppress effector T cells in this tumour. 32 Furthermore, the current study shows that P-Ms have higher levels of transforming growth factor-b1, which exerts immunosuppressive effects via Tregs 33 and induction of programmed death (PD)-1 on CD8 + T cells. 34 More proteins were upregulated than downregulated in the comparison of P-M with P-NM cSCCs, which may relate to limitations with mass spectrometry in detecting reduced protein expression below the sensitivity threshold. There were also substantial variations between samples, confirming our previous observations that cSCCs and their immune infiltrates are highly heterogeneous. 32 In addition, although many proteins that were differentially expressed between P-Ms and P-NMs were identified using both 1D and 2D separation, the 1D and 2D separation methodologies yielded differences in the overall numbers of unique proteins. Moreover, correction for multiple parameters was not feasible given the large number of variables, including varying levels of infiltration of different immunocyte populations. However, we processed cSCC samples including the tumour and surrounding stroma/immune infiltrate, instead of microdissecting the tumour without the stroma, because there is evidence that immune, as well as tumour, parameters are determinants of clinical outcome in cSCC. 3,4,32,35 We acknowledge there is likely to have been a loss of resolution with this approach, and that future studies undertaking proteomic profiling of cSCCs following purification of separate tumour regions, and deconvolution of data based on heterogeneous cell populations, would allow identification of additional pathways relevant to development of metastases and clinical outcome.
MRM verified differential expression of ANXA5 and DDOST in the discovery group of P-M and P-NM cSCCs and validated this in a separate cohort of tumours, highlighting the relevance of ANXA5 and DDOST in development of cSCC metastasis. However, as both proteins were expressed in tumour and immune cells ( Figure S4; see Supporting Information), it is unclear whether the mechanism underlying this association is due to expression of the proteins in the tumour or immune infiltrate, or both these sites. High ANXA5 expression is associated with metastases from colorectal cancer, 36 and reduced survival in renal cell carcinoma. 37 Additionally, the Human Protein Atlas indicates that, using TCGA data, ANXA5 is an unfavourable prognostic marker in renal, liver, urothelial, and head and neck cancers, but a favourable marker in endometrial and stomach cancers. 38 ANXA5 has also been identified as a potential biomarker in a DNA microarray study of cSCC cell lines and tissue, 39 and in a proteomic analysis of head and neck SCC. 40 The mode of action of ANXA5 in relation to development of metastases is not fully understood, but it has been shown to promote migration and invasion of keratinocyte, 41 oral SCC, 41 renal cell carcinoma 37 and hepatocarcinoma 42 cell lines in ANXA5 knockdown experiments. Potential mechanisms for this include effects of ANXA5 on regulation of genes implicated in cell motility (including S100A4, TIMP3 and RHOC), 41 activation of PI3K-Akt-mammalian target of rapamycin signalling leading to tumour cell proliferation, 37 promotion of migration and invasion via upregulation of matrix metalloproteinases 2 and 9, 37 and effects on integrin signalling and mitogen-activated protein kinase kinase-extracellular signalregulated kinase pathways. 42 Conversely, ANXA5 may have a protective role in some cancers because ANXA5 overexpression can inhibit proliferation and metastasis, including in uterine and cervical carcinoma cell lines. 43 In addition, administration of ANXA5 in a murine model of human papillomavirus 16-associated cancer augmented antitumour immunity by binding to phosphatidylserine externalized by apoptotic tumour cells, which enhanced the immunogenicity of tumour antigens. 44 While there is limited published research on DDOST in cancer, the Human Protein Atlas documents DDOST as an unfavourable prognostic marker in renal, liver, and head and neck cancers but as a favourable marker in endometrial cancer. 45 Gene expression profiling interactive analysis of TCGA, and genome-scale CRISPR-Cas9 knockout screening data have demonstrated DDOST as an essential gene across many cancer cell lines, with DDOST upregulated in colon adenocarcinoma and overlapping with expression of genes required for cell growth and viability (although in that study, higher DDOST expression was associated with increased survival in colon adenocarcinoma). 46 Furthermore, another study investigating susceptibility variants for oesophageal SCC reported missense variants in DDOST in two cases. 47 The mechanism whereby DDOST permits metastasis is unclear, but it may involve protein glycosylation and the impact of this via various biological processes relevant to cancer. 48 For example, DDOST functions as a subunit for an accessory protein required for stabilization of the STT3 protein subunits of oligosaccharyltransferase (OST), 49,50 which promotes tumour immune evasion via PD ligand 1 (PD-L1). 51,52 Moreover, STT3, which is induced by epithelial-to-mesenchymal transition, is required for PD-L1 Nglycosylation, which stabilizes and upregulates PD-L1 in breast cancer stem cells. 53 OST is also required for epidermal growth factor receptor (EGFR) cell-surface localization and signalling in non-small lung cancer cells. Furthermore, in EGFR-driven tumour cells, OST inhibition induces senescence. 54 Likewise, OST inhibition reduces tumour growth in EGFR-mutant nonsmall lung cancer 55 and glioma 56 xenografts.
The absolute quantification of ANXA5 and DDOST via MRM in primary cSCCs in this study, and confirmation of higher levels of these proteins in P-M tumours in the discovery and validation groups, suggest that they may have potential for use as biomarkers for development of metastasis in cSCC following surgical excision of the tumour. This is supported by our findings that high expression levels of ANXA5 and DDOST are associated with shorter time to metastasis and reduced 5year overall survival in patients with cSCCs, and, similarly, reduced survival in cervical and oropharyngeal SCC. Indeed, the incorporation of our ANXA5 and DDOST MRM data into a prediction model demonstrated higher sensitivity and specificity than commonly used clinical staging systems for cSCC, indicating that ANXA5 and DDOST offer the potential to provide additional useful information on the likelihood of metastatic spread in this cancer.
As MRM was conducted on FFPE cSCC samples in the current study, future evaluation of ANXA5 and DDOST will be possible in larger cohorts of FFPE samples, and in their subsequent study and use in clinical practice as an adjunct to current staging systems that use FFPE samples. 10 Although conjectural, based on evaluation of ANXA5 and DDOST in larger cohorts of patients, the future incorporation of these markers with other relevant clinicopathological risk factors into a prediction model may offer clinical benefits through improved staging and consequently more personalized treatment and/or follow-up of patients with cSCC.
In conclusion, this proteomics study has identified multiple proteins associated with cSCC metastasis, with several of our findings relevant to other types of SCC. Importantly, high expression of ANXA5 and DDOST in primary cSCCs is associated with subsequent metastatic spread. These results highlight that proteomic analysis has the potential to offer useful insights into biological factors that influence development of metastases from primary cSCCs, and can be a useful adjunct to other 'omics' approaches aimed at identifying potential biomarkers in this cancer.

Supporting Information
Additional Supporting Information may be found in the online version of this article at the publisher's website: Appendix S1 Supplementary materials and methods. Figure S1 Weighted gene coexpression network analysis reveals clusters of proteins that can be related to clinical and histological characteristics. Figure S2 STRING analysis with KEGG pathway mapping identified several pathways significantly enriched in both onedimensional and two-dimensional data. Figure S3 Multiple reaction monitoring of cutaneous squamous cell carcinoma samples. Figure S4 Immunostaining of DDOST and ANXA5 shows presence of these proteins in tumour cells and in cells within the surrounding immune infiltrate.
Table S1 List of significantly differentially expressed proteins in one-dimensional discovery proteomics between primary metastatic and nonmetastatic cutaneous squamous cell carcinoma. Table S2 List of significantly differentially expressed proteins in two-dimensional discovery proteomic data between primary metastatic and nonmetastatic cutaneous squamous cell carcinomas.
Table S3 Medians, interquartile ranges and P-values for comparison of primary metastatic and nonmetastatic cutaneous squamous cell carcinoma groups.