Gene-expression signature regulated by the KEAP1-NRF2-CUL3 axis is associated with a poor prognosis in head and neck squamous cell cancer

We obtained RNA-Seq gene expression version2 (RNA-SeqV2) level 3 data (Illumina Hiseq platform) from HNSCC patients along with adjacent normal tissues from the Broad GDAC Firehose website (http://​gdac.​broadinstitute.​org/). We carried out the analysis of RNA-Seq data of 279 tumor samples and 37 adjacent normal samples listed in the TCGA network study [14]. All the alteration data for KEAP1-NRF2-CUL3 (KEAP1-mutation/deletion, NRF2-mutation/amplification, and CUL3-muatation/deletion) used in the present study was obtained from cBioportal [20, 21]. In addition to the TCGA-HNSCC RNA-Seq data, three independent HNSCC cohorts microarray data– Saintigny et al. (GSE26549) [22], Jung et al. (E-MTAB-1328) [23], and Cohen et al. (GSE10300) [24] – were also used for overall survival analysis. Our study meets the publication guidelines listed by the TCGA network.

The conventional method of differentially-expressed gene (DEG) analysis involves the comparison of tumor transcriptomic data with normal cell data. However, in recent studies, due to the availability of large sets of tumor samples and fewer adjacent normal datasets, researchers have performed DEG analysis of TCGA data by applying a new method in which the DEGs are identified by comparing altered or mutated tumor samples (including a particular gene/set of genes) with wild-type tumors (caused by factors other than alterations or mutations) [15, 25, 26].

Despite the fact that these two methods have been used separately for DEG analysis, in this study, we applied a combinatorial approach to obtain DEGs from HNSCC patients by using both conventional and new methods. We then integrated the resulting upregulated genes from both datasets to obtain overlapping genes. This approach led to the robust identification of more markedly upregulated genes specific to the samples with altered KEAP1-NRF2-CUL3 than in both normal and wild-type samples. Moreover, our method not only identified specific genes targeted by the KEAP1-NRF2-CUL3 axis but also minimized false-positive results.

We segregated the 279 HNSCC tumor samples into two groups: 54 altered KEAP1-NRF2-CUL3 samples (referred to below as ‘altered’) and 225 wild-type samples. Before performing transcriptomic data analysis, the TCGA barcodes of patient data were cross-checked to avoid technical errors. First, we carried out DEG analysis in the 54 altered versus 37 normal samples followed by 54 altered versus 225 wild-type samples using the R/Bioconductor package [27] – edgeR [28]. To crosscheck how our combinatorial approach effectively found specific genes targeted by the KEAP1-NRF2-CUL3 axis, we also subjected the 225 wild-type and 37 normal samples to DEG analysis. Briefly, the raw counts of RNA-SeqV2 level 3 data were filtered by removing the genes containing zero values. We then considered the genes with >100 counts per million in at least two samples for normalization using the trimmed mean of M-values method, followed by the estimation of dispersions using generalized linear models. Up- and down-regulated genes for altered versus normal and altered versus wild-type samples were identified separately by applying a Benjamini-Hochberg (BH) false-discovery rate (FDR) p < 0.01 with a log-fold change (logFC) > 1.5 and <−1.5. Finally, we used the overlapping upregulated genes obtained from both datasets using ‘Venny 2.1’ (http://​bioinfogp.​cnb.​csic.​es/​tools/​venny/​index.​html) for further analysis. Hierarchical clustering of overlapping upregulated genes was performed using the ‘Heatmapper’ web tool [29]. Box plots of the overlapping upregulated genes that represent the log (counts per million) expression values were generated using R-package ‘ggplot2’ [30]. The overall workflow of the study design is presented in Fig. 1.

Functional annotation (Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis) of overlapping upregulated genes was performed using the updated version of the Database for Annotation, Visualization and Integrated Discovery (DAVID) v6.8 web tool [31]. PPI network analysis was performed using the STRING v10 database [32].

To identify the NRF2 binding sites within the promoter regions of the putative KEAP1-NRF2-CUL3-regulated genes, we used the transcription factor-binding site finding tool LASAGNA-Search 2.0 [33] with cutoff p-values ≤ 0.001. The search was limited to the -5 kb upstream promoter region relative to the transcription start site.

Cox proportional hazard regression was performed using the online survival analysis and biomarker validation tool SurvExpress [34]. We considered the data from a total of 502 patients in 4 independent HNSCC cohorts available in the SurvExpress database: the TCGA-HNSCC cohort (n = 283) with other three HNSCC cohorts – Saintigny et al. (GSE26549) (n = 86) [22], Jung et al. (E-MTAB-1328) (n = 89) [23], and Cohen et al. (GSE10300) (n = 44) [24] – for survival analysis. In the case of microarray-based survival data, we considered the average values for genes whose expression was associated with multiple probe sets such as duplicates or alternatives. SurvExpress separated the patient samples into two groups, high - and low-risk, based on average expression of the 17 genes signature values, and performed statistical analysis of survival probability of the two groups using the log-rank method. SurvExpress used the log-rank test to generate Kaplan-Meir plots based on the ‘Survival’ package of the R platform, which is integrated into its website. Log-rank test p-values < 0.05 were considered to be statistically significant.

In HNSCC, changes in the KEAP1-NRF2-CUL3 axis occurred in ~20% of patients; of these, KEAP1 alterations accounted for 4.6%, NRF2 for 11.8%, and CUL3 for 5.7%. However, few samples overlapped (Fig. 2a). In order to better understand the KEAP1-NRF2-CUL3 mutational landscape in HNSCC, we used the cBioportal cancer genomics website [20, 21] to examine the types of mutation and their positions in the domain structure of proteins. All 13 KEAP1 and 18 NRF2 mutations were missense mutations, while 70% of the CUL3 mutations (7/10) were missense, 20% (2/10) were nonsense, and 10% (1/10) were splice mutations (Fig. 2b).

KEAP1 consists of 605 amino-acids with 3 domains in which 6 mutations were reported in the BTB (broad-complex, tramtrack, and bric-a-brac) domain, 1 in the IVR (intervening region), 1 in the C-terminal, 1 in the N-terminal region, and another 4 were in the Kelch domain, which is essential for the binding of NRF2. In the case of NRF2 structure, the majority of mutations (16) occurred in the crucial KEAP1-binding domain Neh2, and another 2 were found in each of the Neh7 and Neh3 domains. CUL3 contained 4 mutations in the N-terminal domain, 5 in the C-terminal domain, and 1 in the cullin repeat 3 domain (Fig. 2c). Overall, two samples contained both KEAP1 and NRF2 mutations, while one sample contained both NRF2 and CUL3 mutations. KEAP1 and CUL3 mutations were mutually exclusive.

In order to identify the genes regulated by the KEAP1-NRF2-CUL3 axis in HNSCC, we focused on the identification of differentially expressed genes by analyzing the RNA-Seq expression profiles in 54 altered versus 37 normal, and 54 altered versus 225 wild-type samples. A total of 215 upregulated genes and 9 downregulated genes were found in the altered versus normal analysis (Additional file 1: Table S1), and 25 upregulated genes and 13 downregulated genes in the altered versus wild-type analysis (Additional file 2: Table S2) with logFC >1.5 (p < 0.01 with BH-FDR adjustment). Since the ultimate effect of KEAP1-NRF2-CUL3 axis gene alterations leads to overexpression of NRF2 and its downstream genes, we focused on the upregulated genes for further analysis. By integrating both datasets using Venny web tool (http://​bioinfogp.​cnb.​csic.​es/​tools/​venny/​index.​html), we obtained 17 overlapping upregulated genes (Fig. 3a). We carried out literature survey to verify whether the downregulated genes obtained from both methods contains previously reported NRF2 regulated genes or not. Notably, we didn’t observe any previously reported NRF2 target genes among all downregulated genes.

We also carried out DEG analysis in 225 wild-type versus 37 normal samples to assess the specificity of the 17 genes regulated by the KEAP1-NRF2-CUL3 axis. Strikingly, none of the 17 genes were found in the list of upregulated genes in the wild-type versus normal samples with logFC > 1.5 (p < 0.01 with BH-FDR adjustment; Additional file 3: Table S3). Thus, our analysis clearly showed that these 17 genes were significantly overexpressed in altered KEAP1-NRF2-CUL3 samples compared with their normal and wild-type counterparts (Fig. 3b, c). We then designated these 17 genes as the signature of gene expression regulated by the KEAP1-NRF2-CUL3 axis based on their specificity and higher expression (Table 1). Among these 17 genes, 13 – AKR1B10, AKR1C1, AKR1C2, AKR1C3, G6PD, GCLC, GCLM, GSTM3, OSGIN1, SRXN1, TXNRD1, SLC7A11 [11, 35, 36], and SPP1 [37]– are well-known NRF2-regulated genes, listed and reviewed in a wide variety of studies.

Since the ultimate effect of KEAP1-NRF2-CUL3 gene alterations results in the overexpression of NRF2 and its target genes, it was not surprising that the majority of genes in our results were well-characterized NRF2-regulated genes. In addition, we found 4 putative KEAP1-NRF2-CUL3-regulated genes, NTRK2 (neurotrophic receptor tyrosine kinase 2), RAB6B, TRIM16L, and UCHL1 and investigated whether they were also regulated by NRF2. Interestingly, further in silico analysis using the ‘LASAGNA-Search 2.0’ [33] bioinformatics tool identified NRF2-ARE sequences within the -5 kb upstream promoter regions of the human RAB6B, UCHL1 and TRIM16L genes (Fig. 4a,b,c; Additional file 4: Table S4). However, we did not find an ARE sequence in the promoter region of the NTRK2 gene. Together, our results suggest that NRF2 directly binds with the promoter regions of 16 of the genes in the signature and triggers their overexpression; NTRK2 is the exception.

Functional annotation analysis from GO and KEGG pathway predictions using both DAVID and STRING v10 revealed that the 17 genes were significantly enriched (p < 0.001) in the biological processes daunorubicin metabolic process, doxorubicin metabolic process, oxidation-reduction process, cellular response to jasmonic acid stimulus, progesterone metabolic process, response to oxidative stress, and steroid metabolic process. In KEGG pathway analysis, we found significant enrichment (p < 0.005) in the three pathways glutathione metabolism, steroid hormone biosynthesis, and metabolism of xenobiotics by cytochrome P450 (Table 2).

To evaluate the prognostic value of the 17-gene signature in patient survival, we first analyzed overall survival in the TCGA-HNSCC cohort available in the SurvExpress web tool. A total of 283 patient samples were divided into high-risk (n = 141) and low-risk groups (n = 142) based on their expression pattern (Fig. 5a). The survival probability estimates in the two risk groups were visualized as Kaplan-Meier plots. Strikingly, overall survival analysis revealed that the patients in the high-risk group had poorer survival (HR = 2.28; CI = 1.56–3.32; p = 1.221e-05) than the low-risk group (Fig. 5b). Thus, our analysis strongly suggests that genes regulated by the KEAP1-NRF2-CUL3 axis are powerful predictors of a poor prognosis in HNSCC patients. In addition, we also carried out the multivariate analysis with the limited variables present in Survexpress database. Consistent with the above results, patients with high-risk scores for clinical variables such as tumor grades G2 and G3, pathological stages T1 and T2, and pathological disease stages II and III were significantly associated with poor survival whereas the results were insignificant in other variables (Additional file 5: Table S5). Kaplan-Meier survival plots with log-rank test results for the significant clinical variables are shown in Additional file 6: Figure S1.

After analyzing the prognostic value of the 17-gene signature in the TCGA cohort, we evaluated its prognostic value in another 3 HNSCC cohorts containing DFS, MFS, and recurrence data. Among these, Saintigny et al. (GSE26549) [22] contains DFS data, while Jung et al. (E-MTAB-1328) [23] contains MFS data. The third cohort, Cohen et al. (GSE10300) [24], contains recurrence data. Interestingly, our DFS analysis using the Saintigny et al. (GSE26549) [22] cohort showed that patients in the high-risk group with increased expression of the 17-gene signature had poorer survival (HR = 2.28; CI = 1.56–3.32; p = 1.221e-05) than the low-risk group (Fig. 6a). Likewise, we found a markedly shorter MFS (HR = 2.83, CI = 1.47–5.48; p = 0.001) in the high-risk group of the Jung et al. (E-MTAB-1328) [23] cohort (Fig. 6b). In the Cohen et al. (GSE10300) [24] cohort, we found lower recurrence-free survival (HR = 4.15; CI = 1.14–15.05; p < 0.01) in the high-risk group with the17-gene signature than in the low-risk group (Fig. 6c). Thus, log-rank analysis revealed that the 17-gene signature was associated with a significantly increased risk of recurrence in HNSCC. The multivariate analysis results for the above cohorts were listed in Additional file 5: Table S5.