cDNA microarray analysis was performed using a panel of ten HNSCC cell line series from the University of Michigan (UM-SCC), derived from eight patients with aggressive HNSCC (survival <2 years), and representing a distribution of different anatomic sites (Table 1). Many of the molecular alterations and biologic characteristics of these UM-SCC cell lines have been confirmed to reflect those identified in HNSCC tumors from patients in laboratory and clinical studies. These include the roles of activation of epidermal growth factor receptor, IL-1, and IL-6 signal transduction pathways; altered activation of transcription factors p53, NF-κB, AP-1, and STAT3; expression of cytokines and other genes; and variation in radiation and chemosensitivity 521222324252630313233414243. The p53 mutation and expression status of UM-SCC cells lines were evaluated using bidirectional genomic sequencing of exons 4 to 9 (Figure 1a), and confirmed with immunocytochemistry using monoclonal antibody to p53 (DO-1 clone) (Figure 1b). No mutation was detected in those exons in four cell lines, namely UM-SCC 1, 6, 9, and 11A. Mutation of p53 was detected in five cell lines, namely UM-SCC 5, 22A, 22B, 38, and 46 (Figure 1a). A mutation was also detected in UM-SCC 11B cells, but immunocytochemistry of p53 protein suggested there might be a mixed population of UM-SCC 11B cells with a heterogeneous expression pattern for nuclear p53 protein (Figure 1b). The findings regarding p53 mutation in UM-SCC 1, 5, 6, 11B, and 46 cells are consistent with a previous report by Bradford and coworkers 41.

Gene expression profiles were determined using a 24,000 element cDNA microarray by comparing 10 UM-SCC cell lines with four cultured primary human keratinocyte (HKC) lines as normal controls. The expression of 9,273 of 12,270 evaluable known genes was submitted for principal components analysis and hierarchical clustering 44. Both methods grouped UM-SCC 11B together with its parental cell line UM-SCC 11A, as well as the other UM-SCC cells with wild-type p53, and these findings were statistically significant (P < 0.001, class prediction analysis; BRB-Array Tools 45). Based on mixed p53 protein staining and wild-type p53 associated gene expression pattern, we classified UM-SCC 11B cells with the wild-type p53 group as having a 'wild-type p53-like' expression pattern.

Next, we studied a total of 1,011 genes that exhibited twofold or greater differences in gene expression when comparing HKCs with all UM-SCC, or comparing UM-SCC cells with either wild-type p53-like (UM-SCC cell lines 1, 6, 9, 11A, and 11B) or mutant p53 expression patterns (UM-SCC cell lines 5, 22A, 22B, 38, and 46; Figure 2a). The 1,011 genes, including 371 over-expressed and 640 under-expressed genes, were subjected to hierarchical clustering, as shown in Figure 2a. The expression profile of 1,011 genes clustered all samples into three groups, namely HKCs, UM-SCC wild-type p53-like, and UM-SCC with mutant p53 (Figure 2a). Six major clusters (A to F) of differentially expressed genes were identified, with most over-expressed genes included in two distinct clusters, A and B, and three subclusters within cluster C (subclusters C1, C2, and C3) on the top portion of the expression tree (Figure 2a).

The unique gene signatures of clusters A and B consisted of 34 and 37 genes (Figure 2b,c and Table 2), respectively. We used the mixed model based F-test to examine the statistical difference of gene expression within clusters among HKCs and UM-SCC cell lines with different expression patterns. Within both cluster A and cluster B genes, a significant difference in gene expression (probability, Pr [F] < 0.001) was observed when comparing the two groups of UM-SCC cells. A significant difference was also observed when comparing HKCs with UM-SCC cells with mutant p53 in cluster A, and comparing UM-SCC with the wild-type p53-like expression pattern in cluster B. Thus, cluster A genes were over-expressed in UM-SCC cells with mutant p53 (Figure 2b), whereas cluster B genes were over-expressed in UM-SCC cells with wild-type p53-like expression pattern (Figure 2c).

In addition to genes of clusters A and B, we defined another group of over-expressed genes, namely cluster C, including three subclusters C1, C2, and C3 (Figure 2a and Additional data file 1). Overall, cluster C contained 240 genes that were over-expressed by 10 cancer cell lines when compared with HKCs (Pr [F] < 0.001). However, two of the subclusters (C1 and C2) identified exhibited a degree of differential expression in UM-SCC cells similar to the various p53-associated expression patterns (Additional data file 1).

To determine the functional classification of the various gene clusters, we conducted Gene Ontology (GO) annotation using Onto-Express, which constructs statistically significant functional profiles from a set of genes 46. Additional data file 2 shows functional categories that are significantly enriched in the six clusters. The top categories of GO biologic processes in cluster A were nucleosome assembly, chromosome organization, and biogenesis. These included genes involved in regulation of chromosome structure or function (such as H2B histone family B, C, D, R, L and Q, and H2A histone family L and N), transport (such as MYST3, ABCC5, ATP1B3, and HBE1), and DNA repair (such as XPA). The main GO molecular function was DNA binding, including eight genes in histones H2A and H2B, and MYST3, THAP11 and ARID1A (Table 2 and Additional data file 2).

In contrast to cluster A, the top ranked GO biologic processes in cluster B belonged to signal transduction (such as cell-cell signaling, cell surface receptor linked signal transduction), including AKAP12, CAP2, IL6, IL8, RAB17, SHANK2, STC1, PTPRJ, TXNRD1, and YAP1 (Table 2 and Additional data file 2). Other enriched functional categories included cell cycle (AIF1, BCAT1, RAD54L, and STK6), regulation of transcription (ARID3A, DMAP1, and ZNF239), cell proliferation and apoptosis (CROC4, BIRC2, PLK1, and PORMIN), adhesion (ICAM1), and structural proteins related to tumor progression (KRT8 and KRT18; Table 2 and Additional data file 2). Interestingly, several genes in this cluster or their homologs involved in angiogenesis and inhibition of apoptosis have previously been associated with metastatic tumor progression in murine SCC or human HNSCC (IL6, IL8, YAP1, and BIRC2) 922232430333738474849, and shown to be regulated by NF-κB 222324253749.

Genes in cluster C exhibited annotations for DNA replication, ubiquitin cycle, cell division, and oxidoreductase and catalytic activities. The gene list and ontology of the subclusters in C are presented in the Additional data files 1 and 2, respectively. Several genes in subclusters C1 and C2 exhibited weaker clustering but similar functions as those in cluster A or B. In subcluster C1, in which over-expressed genes were mainly found in UM-SCC cells with mutant p53 as in cluster A, there are two additional genes identified that encode proteins involved in chromosome structure and functions (HIST1H2AL and HDAC5). Other genes previously associated with cancer included a member of epidermal growth factor receptor family (ERBB3); a target gene of p53/p63 (IGFBP3); and a gene whose product is involved in calcium storage and signaling (CALR). In subcluster C2, in which over-expressed genes were found in UM-SCC cells with wild-type p53-like expression pattern as in cluster B, another apoptosis related gene (BAG2), genes encoding signal-related molecules (MYBL2 and UBE2C) and a cell cycle related molecule (CCNB2) were identified (Additional data file 1). The rest of the cluster C genes were over-expressed by more than half of ten UM-SCC cells when compared with HKCs. Several genes encoding protein products that are important in cancer and have functions related to cell cycle, growth, DNA replication and protein translation (such as CCND1, TOP2A, TOPBP1, TFRC; three members of H4 histone family [HIST1H4C, HIST1H4B, and HIST1H4E]; and EIF4G1). Some genes encode proteins with functions related to signal transduction (PIK3R3, MAPK8IP1, and GATA2), and one gene encodes a protein that regulates tumor invasion and metastasis (TIMP2).

Genes downregulated in UM-SCC cells were included in cluster D, which represents functional categories that are involved in epidermis development, cell adhesion, and cell-cell signaling (Additional data file 2). Downregulated cluster E genes included those encoding molecules with functions in other signal pathways, cell cycle, calcium ion regulation, and actin binding activities (Additional data file 2). The categories over-represented among the downregulated genes in cluster F included cell adhesion, differentiation, and morphogenesis (Additional data file 2). A more detailed analysis of the downregulated genes will be presented elsewhere (Yan, unpublished data).

Based on the gene expression profiling data, we hypothesized that transcriptional regulation by multiple transcription factors may be key elements that contribute to the expression of unique gene clusters. To test this hypothesis, in silico computational analyses were performed to determine whether dominant cis-regulatory elements are present in the proximal promoter region of over-expressed genes. We evaluated five transcription factors that were previously found to be altered and functionally important in HNSCCs and other cancers, including p53, NF-κB, AP-1, STAT3, and EGR1. We compared the frequencies of their binding sites with those from vertebrate promoters from the Genomatix promoter database (GPD), which consists of information from human, mouse, and rat. The five transcription factors examined have been shown by our laboratory and others to contribute to regulation of individual gene expression with functional importance in cancer, such as cell proliferation, cell cycle, apoptosis, DNA repair, and angiogenesis 21222436373850515253.

Table 2 shows a list of genes included in clusters A and B and corresponding binding sites for the five transcription factors in proximal promoter regions that are predicted with high probability. The detailed location and sequences of the putative TFBSs are shown in Additional data file 3. Significant differences in the prevalence of predicted TFBSs were observed for genes from different clusters when compared with vertebrate promoters (Figure 3). In cluster A, putative p53 binding sites were detected in 50% of the 34 gene promoters, which is significantly higher than observed in vertebrate promoters (P < 0.05; Figure 3). Conversely, predicted NF-κB binding sites were observed in about 66% to 70% of the promoters in clusters B and C, which was significantly more than in vertebrate promoters. There was no significant difference in the prevalence of NF-κB binding sites between the promoters of cluster A and vertebrates. There were also differences in the prevalence of TFBSs predicted between different clusters. For example, the p53 binding motif was significantly greater in cluster A than cluster B (χ² analysis; P < 0.05), and the greatest frequency of NF-κB binding sites was observed in cluster B (26/37 [70%]; Figure 3). There were significantly fewer genes with AP-1 binding sites in cluster A and subcluster C1 (12% and 13%, respectively) compared with vertebrate promoters (Figure 3). A relatively higher frequency of AP-1 binding sites was observed in cluster B genes when compared with frequencies in cluster A and subcluster C1 genes (χ² analysis; P < 0.01). In contrast, a relative increase in prevalence of STAT3 and EGR1 binding sites was observed and distributed among all of the upregulated clusters relative to vertebrate promoters (Figure 3), with increasingly higher frequencies of EGR1 motifs detected in clusters B and C (60% to 76%; using Genomatix matrix EGR1.02).