Integrative analysis of gene expression profiles reveals distinct molecular characteristics in oral tongue squamous cell carcinoma.

Oral tongue squamous cell carcinoma (OTSCC) is the most common type of oral cancer. Despite advances in knowledge regarding the genome-scale gene expression pattern of oral cancer, the molecular portrait of OTSCC biology has remained unclear over the last few decades. Furthermore, studies concerning OTSCC gene-expression profiles are limited or inconsistent owing to tissue heterogeneity in single-cohort studies. Consequently, the present study integrated the profile datasets of three cohorts in order to screen for differentially expressed genes (DEGs), and subsequently identified the potential candidate genes and pathways in OTSCC through gene enrichment analysis and protein-protein interaction (PPI) network construction. Using the selected Gene Expression Omnibus datasets GSE13601, GSE31056 and GSE78060, 206 DEGs (125 upregulated and 81 downregulated) were identified in OTSCC, principally associated with extracellular matrix (ECM) organization and the phosphoinositide 3-kinase/protein kinase B signaling pathway. Furthermore, 146/206 DEGs were filtered into the PPI network and 20 hub genes were sorted. Further results indicated that the two most significant modules filtered from the PPI network were associated with ECM organization and human papillomavirus infection, which are important factors affecting OTSCC pathology. Overall, a set of OTSCC-associated DEGs has been identified, including certain key candidate genes that may be of vital importance for diagnosis, therapy and prevention of this disease.

Introduction

Head and neck squamous cell carcinoma (HNSCC) ranked as the sixth leading incident cancer worldwide in 2012 (1). In contrast with the slightly decreased incidence rate of general HNSCC, the occurrence of oral HNSCC has increased over the last few decades, particularly oral tongue squamous cell carcinoma (OTSCC) (2–5). OTSCC represents malignancies of the oral cavity, with a significantly increasing incidence rate reported among younger individuals from 1975 to 2007 in the USA (6). Although OTSCC cases are considered as oral squamous cell carcinoma (OSCC) or HNSCC, their distinct histological and epidemiological characteristics have been verified (7,8). Owing to the complex lymphatic network and muscular structure of the tongue, patients with OTSCC present a more aggressive phenotype compared with those with tumors affecting other parts of the body, with a higher proportion of lymph node positivity, higher recurrence and metastasis rates post-therapy, and, therefore, poorer prognosis (9,10). However, the molecular mechanisms underlying these variations remain unknown.

Gene detection techniques based on gene expression and sequence variation, including gene microarrays and sequencing, facilitate the gathering of genetic information about numerous cancer types (11–14). A large amount of functional genomic data produced by these high-throughput techniques are archived in public repositories, including the Gene Expression Omnibus (GEO) database (www.ncbi.nlm.nih.gov/geo). Using these data, integrative analysis or re-analysis can provide valuable clues and new understanding regarding the underlying mechanism (15–17). To date, a considerable number of gene expression profiling studies on OSCC and HNSCC have been completed. However, only a few studies have focused on the transcriptome of OTSCC. The results from these independent studies are inconsistent partly due to sample heterogeneity. In the present study, two OSCC datasets and one OTSCC dataset were obtained from the GEO database and filtered prior to integrative analysis. Differentially expressed gene (DEG) screening, protein-protein interaction (PPI) network construction and gene functional annotation were performed, in order to investigate the distinct gene expression profile of patients with OTSCC.

Materials and methods

Acquisition, preprocessing and DEG screening of microarray data

The gene expression data and probe annotation files GSE13601 (18), GSE31056 (19) and GSE78060 (20) were downloaded from the GEO database for investigation. All of these datasets included microarray data of OTSCC samples. According to their anatomical definition, tongue samples were extracted from the three datasets. Raw microarray data in CEL format were processed with background correction, log2 transformation and quantile normalization using the Robust Multi-array Average (RMA) algorithm (21) in the Affy package (version 1.22.1; www.bioconductor.org/packages/2.4/bioc/html/affy.html) in R (version 3.4.3; www.r-project.org). Subsequently, the DEGs in OTSCC tissues compared with in normal tongue tissues were identified using linear models with the limma package (version 2.18.3; www.bioconductor.org/packages/2.4/bioc/html/limma.html) in R (22,23). |log2FC|≥1 (where FC is fold change) and adjusted P<0.05 were considered as the cut-off values for statistical significance. Furthermore, the intersection of the DEGs among the datasets was calculated, and the result was visualized as a Venn diagram using an online tool (bioinformatics.psb.ugent.be/webtools/Venn).

For validation, the consistency between identified DEGs from the GEO datasets and the data from The Cancer Genome Atlas (TCGA; cancergenome.nih.gov) was assessed and visualized as a Venn diagram. HNSCC gene expression data were downloaded from TCGA and a filter was applied so as to retain only the data of patients with OTSCC. Subsequently, the edgeR package (version 1.2.4; www.bioconductor.org/packages/2.4/bioc/html/edgeR.html) was used to screen for DEGs with the cut-off values of |log2FC|≥1 and adjusted P<0.05 (24).

Functional enrichment analysis

The Database for Annotation, Visualization and Integrated Discovery (DAVID) online tool (version 6.8; david.ncifcrf.gov) was applied to map candidate DEGs onto their associated biological annotation (25,26), with Gene Ontology (GO; www.geneontology.org) (27,28) and pathway analyses using the Kyoto Encyclopedia of Genes and Genomes (KEGG; www.genome.jp/kegg) (29,30) and the Reactome (www.reactome.org) pathway databases (31,32). Adjusted P<0.05 was considered to indicate statistical significance. All significantly enriched terms were visualized in bubble chart using the ggplot2 package (version 3.1.0; docs.ggplot2.org) in R. The richness factor was calculated as the percentage of the enriched gene number relative to the background gene number for the same term.

PPI network construction

All candidate DEGs were searched in the Search Tool for the Retrieval of Interacting Genes/Proteins (STRING) database (version 10.5; string-db.org) and a combined score >0.4 was used as the criterion to establish the PPI network (33). All the isolated nodes were deleted from the network. The data of the PPI network were exported from the STRING website and imported into Cytoscape (version 3.5.1) software for visualization (34). Each protein in the network served as a node, and the degree and betweenness centrality were calculated using the CentiScaPe (version 2.2) plugin (35,36). The hub gene was defined as the node with a degree >10 within the top 30 betweenness centrality nodes in the present study.

Sub-network analysis

The MCODE plugin (version 1.4.2) (37) was used to identify highly interconnected regions, or clusters, in the PPI network. The degree cut-off was set to 2 and the κ-score was set to 2. The identified clusters with a score >10 were used to create a sub-network. The Cytoscape plugin ClueGO + CluePedia (version 2.5.0) (38,39), which facilitates GO and pathway enrichment analysis in a network, was applied to perform the enrichment analysis and subsequent visualization. The information from the GO and KEGG databases was combined, and the κ-coefficient threshold was set to 0.4. On the basis of the calculations, similar functional terms were marked with the same color.

Two-dimensional hierarchical clustering analysis

According to the anatomical site of samples in GSE31056, OTSCC samples and normal tongue samples were filtered into a subset of GSE31056. Following normalization, gene expression matrices of 206 DEGs from datasets GSE13601, GSE31056 and GSE78060 and the subset of GSE31056 were prepared. Unsupervised clustering was performed on the four matrices using the pheatmap package (CRAN.R-project.org/package=pheatmap) in R.

Univariate survival analysis

In order to distinguish prognostic factors for the outcome of patients with OTSCC, the 206 DEGs were subjected to overall survival (OS) analysis using the univariate Cox regression model. Owing to the unavailability of the clinical information of the samples in datasets GSE13601 and GSE78060, and the limitation in sample size of GSE31056, OTSCC gene expression data and clinical information from TCGA database were used in this analysis. Any causes of mortality were defined as events and survival was defined as a censored event. The OS analysis was performed with the R package Survival (version 2.43–3; CRAN.R-project.org/package=survival), and P<0.05 was considered to be statistically significant.

Results

Identification of DEGs in OTSCC

Following data filtering, the data of 31 OTSCC samples and 26 normal tongue samples from dataset GSE13601 were termed dataset A, 12 OTSCC samples and 39 normal tongue samples from GSE31056 were dataset B, and 26 OTSCC samples and 4 normal tongue samples of GSE78060 were dataset C. Comparing the OTSCC tissues with the normal tongue tissues, a total of 1,562, 2,584 and 1,712 DEGs were identified in datasets A, B and C, respectively. Subsequently, when the DEGs were investigated for overlap, a total of 206 consistently aberrant genes were identified (Fig. 1), comprising 125 upregulated and 81 downregulated genes (Table I). Subsequently, these DEGs were subjected to survival analysis. The results revealed that four genes, NCLN, THBS2, SPARCL1 and YKT6, were associated with the outcome of patients with OTSCC in TCGA (Table II).

Figure 1.

Identification of 206 common DEGs. Cohort profile datasets A, B and C were selected from datasets GSE13601, GSE31056 and GSE78060, respectively, and are indicated in different colors. The DEGs were identified using the edgeR package in R with cut-off criteria of [log2fold change] >1 and adjusted P<0.05. The overlapping areas represent the common genes. DEGs, differentially expressed genes.

Table I.

List of 206 consistently aberrant genes identified from three Gene Expression Omnibus datasets.

Gene expression	Differentially expressed genes
Upregulated (n=125)	IFI27, CDH3, PYGL, MYO1B, MMP1, SCO2, TYMP, BNC1, COL4A1, MMP3, PTHLH, IRF6, F2RL1, COL4A2, IFI6, ACTN1, THBS2, RAB31, SLC16A1, ISG15, PRNP, KRT16, TPBG, MDFI, OSMR, PLAU, SERPINE1, PROCR, PXDN, DUSP7, ITGA6, COL1A2, SOX15, LAMB3, SHC1, NDRG1, LAMC2, ADORA2B, PDLIM4, COL5A2, GJA1, LGALS3BP, MMP13, DFNA5, IL1RAP, PDPN, RGS20, FSCN1, TPST1, STK3, SLC7A5, CTSC, ADAM10, COL7A1, UPP1, PTK7, CA2, ITGA3, GJB3, APOL1, SCG5, EIF6, PLAUR, SOX11, MMP10, COL3A1, TGFBI, MMP12, COL17A1, IRF9, ZWINT, STAT1, BPGM, PCDH7, NUP155, GNA15, POSTN, OAS1, IGFBP3, FAP, COL4A5, TUBB3, DUSP14, FST, TK1, SNAI2, FOXM1, GINS1, TRIP13, HIST1H2AE, IFIT3, PLOD2, DSG2, TGIF1, MYO10, IFI44, IFIT1, CXCL11, PRSS23, RBP1, SQLE, YKT6, KRT10, SNAPC1, BST2, HOMER3, SPP1, ENO1, DLGAP5, KIF23, OASL, COL4A6, RSAD2, CDC20, TNC, F3, FOLR3, EFNA1, PLSCR1, FN1, HIST1H2BD, GNLY, S100A3, LY6E, CCNB1
Downregulated (n=81)	ADH1B, GPRASP1, MEOX2, MYRIP, CBX7, ATP6V0E2, GPR64, C7, RNASE4, ITM2A, SLC25A20, CDO1, CLDN10, MAN2A2, GNG7, SATB1, TXNIP, SERPINA5, LPIN1, ABCA3, SELENBP1, LMO2, GYPC, CXCL12, KAT2B, ZNF529, RTN1, PRELP, ANG, CFD, SSBP2, CCDC69, ENPP4, BEX4, TSPYL5, MYOC, NCLN, SYNGR1, GDF10, P2RY14, CLU, PIP5K1B, ALDH1A1, CILP, MFAP4, FRZB, IGF1, TOX3, ZBTB20, RORC, NR3C2, PTGFR, CPEB3, LGI1, SUSD5, CLGN, GAS2, LCP1, SORBS2, HLF, DPT, CX3CR1, SERPINI1, ACOX2, ASPA, PCK1, MIA, LMOD1, NFIB, SLITRK5, CRISP3, DCLK1, ANGPT1, ABCA6, FAM149A, SPARCL1, NPY1R, PTGDS, AMPD1, FBLN5, STATH

Table II.

Genes significantly associated with overall survival in oral tongue squamous cell carcinoma.

Gene	HR	95% CI	P-value
NCLN	38.678	1.117–6.192	0.0047
THBS2	2.050	0.153–1.283	0.0127
SPARCL1	3.333	0.097–2.310	0.0330
YKT6	13.765	0.179–5.065	0.0354

HR, hazard ratio; CI, confidence interval.

Gene enrichment and functional annotation analysis of DEGs in OTSCC

With adjusted P<0.05 as the cut-off criterion, the GO analysis was performed. A total of 206 DEGs were enriched significantly into 27 diverse GO terms, being categorized into three functional groups: Biological process (BP), cellular component (CC) and molecular function (MF) (Fig. 2). Among these terms, extracellular matrix (ECM) organization, extracellular space and ECM structural constituent were the most significant in the BP, CC and MF groups, respectively. Furthermore, the candidate genes were enriched in terms of cell adhesion, response to virus and angiogenesis. Subsequently, the pathway enrichment analysis was performed to assess the aberrant gene-associated pathways. A total of 25 significantly enriched pathways were observed (Fig. 3), a number of which are associated with ECM organization. Overall, the greatest number of genes were involved in the phosphoinositide 3-kinase (PI3K)/protein kinase B (Akt) signaling pathway.

Figure 2.

Visualization of the GO enrichment analysis for 206 differentially expressed genes in oral tongue squamous cell carcinoma. The GO enrichment analysis was performed with the Database for Annotation, Visualization and Integrated Discovery online tool, and the detailed information is presented as a bubble chart. The y-axis represents the GO terms, the x-axis represents the BP, CC and MF functional group categorization, the size of bubbles represents the number of assigned genes, and the color of the bubbles indicates the -log10 (Q-value). The larger the number of genes associated with the term, the larger the bubble. The more significant the GO category, the higher on the color bar the bubble is. GO, Gene Ontology; BP, biological process; CC, cellular component; MF, molecular function; Q-value, Bonferroni-adjusted P-value.

Figure 3.

Visualization of the pathway enrichment analysis for 206 differentially expressed genes in oral tongue squamous cell carcinoma. The pathway enrichment analysis was performed with Database for Annotation, Visualization and Integrated Discovery online tool, and the detailed information is presented as a bubble chart. The y-axis represents the significantly enriched pathways, the x-axis represents the richness factor, the size of the bubbles represents the number of assigned genes, and the color of bubbles represents the -log10 (Q-value). The larger number of genes classified into the pathway, the larger the node size is. The more significant the pathway, the higher on the color bar the bubble is. KEGG, Kyoto Encyclopedia of Genes and Genomes; PI3K, phosphoinositide 3-kinase; Akt, protein kinase B; ECM, extracellular matrix; PDGF, platelet-derived growth factor; Q-value, Bonferroni-adjusted P-value.

PPI network analysis in OTSCC

The 206 DEG-encoded proteins were searched in the STRING database, and 206 proteins in Homo sapiens matched the input. Among these, 142 proteins were filtered into the PPI network with 523 edges, whereas the remaining 64 disconnected nodes were hidden (Fig. 4). For the 142 connected nodes, the 20 central nodes were selected with the filtering criterion of degree >10 within the top 30 betweenness centrality nodes. These were FN1, IGF1, TIMP1, ISG15, STAT1, SPP1, COL17A1, SERPINE1, CXCL12, PLAU, MMP1, COL7A1, ITGA6, PLAUR, CCNB1, ACTN1, PLSCR1, CLU, CXCL11 and FOXM1, among which STAT1 and FOXM1 were identified as transcription factors. Subsequently, two modules with score >10 were identified using the MCODE plugin and marked in different colors. Module 1 consisted of 103 edges and 20 nodes in light salmon, and module 2 consisted of 85 edges and 14 nodes in yellow-green. For a further analysis, functional enrichment of these two modules was conducted using ClueGO + CluePedia plugin. The function annotation results demonstrated that module 1 was principally associated with ECM organization (Fig. 5A) and module 2 was principally associated with human papillomavirus (HPV) infection (Fig. 5B).

Figure 4.

PPI network for DEGs in oral tongue squamous cell carcinoma. Using the STRING online database, a total of 142 DEG-encoded proteins were filtered into the PPI network, with the remaining 64 disconnected nodes hidden. Topology analysis was performed using the MCODE plugin and two significant sub-networks (modules), termed 1 and 2, were identified. Nodes in light salmon color represent the DEG-encoded proteins of module 1, and nodes in yellow-green color represent the DEG-encoded proteins of module 2. PPI, protein-protein interaction; DEGs, differentially expressed genes.

Figure 5.

Functional annotation of the two significant modules from the protein-protein interaction network analysis. The function annotation of the two sub-networks was performed using the ClueGO and CluePedia plugins. (A) Module 1 consists of 20 proteins that are principally associated with extracellular matrix organization. (B) Module 2 consists of 14 proteins that are principally associated with human papillomavirus infection. Solid rounded rectangles represent enriched Gene Ontology terms, solid circles represent enriched Kyoto Encyclopedia of Genes and Genomes pathways.

Validation of the distinctive expression profile in OTSCC

The 206 identified DEGs were subjected to a two-dimensional hierarchical clustering analysis. In dataset GSE31056, distinct clusters of tumor and normal tissues were formed for all tongue samples (Fig. 6A), whereas for all oral samples, two tumor tissues were grouped within the cluster of normal tissues (Fig. 6B). As expected, separate clusters between tongue tumors and normal tongue tissues were also observed for the samples of GSE13601 (Fig. 7A) and GSE78060 (Fig. 7B). These results revealed the differences in gene expression profiles between OSTCC and OSCC. Finally, to confirm the reliability of the identified DEGs, aberrant genes in OTSCC were screened from data of OTSCC and normal tongue tissue from TCGA database in order to investigate the overlap between the data of these two databases. In TCGA OTSCC data, 1,724 downregulated and 792 upregulated genes were identified. Although more DEGs were identified in TCGA data, a total of 119 genes (72 upregulated and 47 downregulated genes) were identified as concordant between the data of the two databases (Fig. 8) and are listed in Table III.

Figure 6.

Expression heat maps of DEGs in different samples of GSE31056. (A) In total, 206 DEGs were identified in oral tongue squamous cell carcinoma samples compared with normal tongue tissues. (B) A total of 206 DEGs in oral squamous cell carcinoma samples compared with normal oral tissues. The rows represent the genes, and the columns represent the samples. The red and blue colors indicate upregulated and downregulated genes, respectively. DEGs, differentially expressed genes.

Figure 7.

Expression heat maps of the 206 identified DEGs in oral tongue squamous cell carcinoma samples compared with normal tongue tissues in datasets (A) GSE13601 and (B) GSE78060. The rows represent the genes, and the columns represent the samples. The red and blue colors indicate upregulated and downregulated genes, respectively. DEGs, differentially expressed genes; OTSCC, oral tongue squamous cell carcinoma.

Figure 8.

Intersection of differentially expressed genes identified from the GEO microarray datasets and a TCGA RNA sequencing dataset of oral tongue squamous cell carcinoma samples. As indicated in the Venn diagram, the results from the two databases had a total of 119 genes (72 upregulated and 47 downregulated genes) in common. TCGA, The Cancer Genome Atlas; GEO, Gene Expression Omnibus.

Table III.

List of 119 aberrant genes in oral tongue squamous cell carcinoma identified from the Gene Expression Omnibus and the cancer Genome Atlas databases.

Gene expression	Differentially expressed genes
Upregulated (n=72)	COL4A5, CCNB1, SHC1, ITGA3, GINS1, FOXM1, PXDN, TPBG, FN1, IFI27, FST, COL5A2, SPP1, ITGA6, PLOD2, MMP1, MMP12, BNC1, KIF23, GNLY, CDH3, COL4A2, MMP3, POSTN, FSCN1, PLSCR1, DLGAP5, COL4A6, COL4A1, LAMC2, TPST1, ACTN1, COL1A2, PROCR, SLC16A1, FOLR3, IFIT3, MYO1B, PLAU, MMP13, HOMER3, PTHLH, CXCL11, MYO10, PTK7, ADAM10, CDC20, RAB31, OASL, PRNP, TRIP13, DFNA5, ISG15, PDPN, TK1, TNC, FAP, BST2, IFI6, PYGL, IFIT1, THBS2, PRSS23, SERPINE1, RSAD2, SOX11, RBP1, TGFBI, SNAI2, SCG5, IFI44, CTSC
Downregulated (n=47)	GDF10, CPEB3, TOX3, HLF, SORBS2, NPY1R, CLDN10, MIA, SSBP2, NR3C2, CBX7, MYOC, SLC25A20, GAS2, GNG7, RORC, PIP5K1B, LPIN1, CX3CR1, ATP6V0E2, SERPINA5, SYNGR1, CFD, RNASE4, SATB1, KAT2B, ENPP4, FAM149A, LMOD1, ASPA, AMPD1, ANG, BEX4, CRISP3, STATH, DPT, PTGDS, NFIB, SLITRK5, ALDH1A1, ITM2A, GPRASP1, ADH1B, MYRIP, FRZB, ACOX2, SELENBP1

Discussion

Microarrays have been extensively applied to gene expression studies of human cancer, describing the genetic profiles of the disease. In the present study, gene expression data of multiple cohorts were obtained from the GEO database for the screening of OTSCC-associated genes. Consistent with the results of previous studies on gene expression in OTSCC (40–42) and other carcinomas, including hepatocellular carcinoma, ovarian cancer and nasopharyngeal carcinoma (43–46), the present study revealed numbers of DEGs in the order of 103 in each cohort. However, the majority of previous studies were performed on a single cohort and focused on a single genetic event (40–42). Patient and sample heterogeneity in independent studies is inevitable, and consequently inconsistencies exist among these single cohorts. Furthermore, OTSCC has been classified as OSCC for investigation, thus the distinct gene expression profile underlying OTSCC remains undefined. By integrating multiple cohorts, the combination of integrative bioinformatics methods and expression datasets is an innovative way to solve these problems. Therefore, a multiple-cohort integrative analysis with a relatively stringent sample filtering was applied in the present study. The term OTSCC was used as a query to screen the candidate microarray datasets in the GEO database. First, datasets with ambiguous anatomical information were removed. Secondly, those with <30 tongue samples were removed. Thirdly, certain datasets concerning formalin-fixed paraffin-embedded samples in long-term archives were removed, owing to the poor quality of RNA. Finally, three datasets with a total of 69 tumors and 69 normal tissues of the tongue were included in the present study and an overlap of 206 DEGs was identified. Numerous studies have demonstrated that the anatomical site is one of the factors that influences progression and prognosis of OSCC (10,47). Owing to a rich blood supply and lymphatic drainage, OTSCCs are more likely to metastasize compared with other types of OSCC (48). A number of studies suggested that OTSCC possesses distinct epidemiological characteristics (7,8), and a corresponding distinct gene profile is therefore expected. From the expression data of 206 DEGs in datasets A, B and C, the distribution of the samples in two separate clusters was in agreement with their classification as tumor or normal tissue. Notably, taking all samples of GSE31056 into account, one tongue tumor sample (GSM771224) and one buccal carcinoma sample were classified within the cluster of normal tissues, which might suggest these two samples have distinct gene expression compared with the other OTSCC samples.

GO analysis is a method used to annotate genes and provide evidence-based statements associating them with specific ontology terms (27,28), whereas pathway databases, including KEGG (29,30) and Reactome (31–32), are web resources for understanding high-level functions and interpreting pathway knowledge to support basic and clinical research. In the present study, the samples were analyzed using the DAVID online tool, which integrates a comprehensive set of functional annotation tools including the three aforementioned databases (25,26), in order to decipher the biological functions of the identified DEGs. Regarding GO annotation, the most significantly enriched terms were all associated with the ECM, in agreement with the results of a previous study (41). Certain biological ECM molecules, including fragments of glycosaminoglycan and hyaluronan, are key regulators of injury and inflammatory response during carcinogenesis (49). Certain other ECM proteins, including MMP, regulate cell motility, which may account for the high probability of distant metastasis of OTSCC. Furthermore, the appearance of cell adhesion and angiogenesis terms in OTSCC is reasonable, as these processes are associated with cancer development and metastasis. Regarding the pathway enrichment, the majority of significantly enriched terms were associated with ECM organization, in agreement with the results of the GO enrichment. The majority of detected DEGs were involved in the PI3K/Akt signaling pathway, which is a critical pathway regulating diverse cellular functions, including metabolism, growth, proliferation, survival, transcription and protein synthesis (50). Aberrant Akt signaling is the underlying defect in a number of diseases, including cancer (51). Numerous studies have demonstrated that the PI3K/Akt signaling pathway serves an essential function in the origin and progression of OTSCC (52,53). In addition, Yu et al (54) suggested that the pathway may be a key regulator of radiosensitization in patients with OSCC. Therefore, the result that the PI3K/Akt signaling pathway is affected in OTSCC is reasonable.

In order to delineate complex biological processes, including cancer initiation and progression, it is helpful to consider DEGs in the context of a complex molecular network. The STRING database is an online resource curating known and predicted PPIs for constructing functional protein association networks. Although human PPI maps represent only a fraction of the complete interaction network, their utility in interpreting complex cancer signatures has led to them being more widely used and a valuable aid in research. In the present study, a PPI network consisting of 142 nodes and 523 edges was established. Topology analysis suggested that FN1, IGF1, TIMP1, ISG15, STAT1, SPP1, COL17A1, SERPINE1, CXCL12, PLAU, MMP1, COL7A1, ITGA6, PLAUR, CCNB1, ACTN1, PLSCR1, CLU, CXCL11 and FOXM1 were the hub molecules. Among them, IGF1, FN1, SERPINE1, SPP1, COL17A1, COL7A1, MMP1, TIMP1, ITGA6 and ACTN1 are involved in the regulation of cancer cell adhesion and motility (55–58). Additionally, CCNB1, SERPINE1 and IGF1 are involved in the cellular tumor antigen p53 signaling pathway (59), and COL17A1 has been identified as a novel target of p53 with an inhibitory effect on breast cancer migration and invasion (60). Furthermore, STAT1 and FOXM1 are transcription factors. A previous study suggested that BCL10 promotes OSCC progression through activating STAT1 and ATF4 (61). Yang et al (62) provided evidence that FOXM1 is a mediator of epithelial-mesenchymal transition, facilitating OTSCC migration and invasion. Another study noted the importance of PLAU and PLAUR in complement and coagulation cascades that are linked to immune responses to tumors (63). Therefore, these molecules may represent promising candidates for molecular diagnosis and therapeutic intervention for patients with OTSCC. In further exploration, a sub-network analysis was performed and three representative modules were identified. As expected, an ECM organization-associated module was represented. Notably, the other module was associated with HPV infection, which has been identified as an emerging risk factor for OTSCC (64). Together, the results support the reliability of functional analysis of DEGs, and propose these hub genes as promising candidates for further functional experimentation.

Finally, an analysis of overlap further verified the reliability of the results of the present study. In total, 119/206 were also differentially expressed in TCGA OTSCC samples. Although the DEG lists derived from the two datasets were not identical, the disparity is explicable. First, different detecting platforms may partly account for the differences between the results from the two datasets, as neither RNA sequencing data nor microarray data cover the complete genome. Secondly, the absence of probes in certain datasets may result in fewer identified DEGs in the present analysis. Specifically, HOXD11, CDK1 and CCL15 were identified as DEGs in TCGA data. However, HOXD11, CDK1 and CCL15 are not present on the GSE13601, GSE31056 or GSE78060 arrays, respectively. Furthermore, the different genetic background of individuals and tumor heterogeneity may also be part of the reason.

In conclusion, using multiple cohort profile datasets and integrative bioinformatics analysis, the present study has identified a set of DEGs that may help in better distinguishing OTSCC from normal tongue tissue. The identified gene set may contain candidate molecular targets for disease-specific diagnosis and therapy.