A 13-gene expression-based radioresistance score highlights the heterogeneity in the response to radiation therapy across HPV-negative HNSCC molecular subtypes

We identified genome-wide expression profiles from four sets of established cancer cell lines with available radiation or drug sensitivity data. The first set included 59 cell lines established from 9 different cancer tissue types (breast, central nervous system, colon, leukemia, melanoma, non-small cell lung, ovarian, prostate, and renal) and included in the NCI-60 panel. Raw data was downloaded from the CellMiner database [27, 28] and information on the surviving cell fraction at 2 Gy (SF2) was retrieved [29]. Radiosensitive and radioresistant cancer cell lines were defined by a SF2 < 0.2 and > 0.8, respectively, as previously described [30]. Two or three replicates were available for each cell line. The second and third sets included 32 and 5 HNSCC cell lines. Raw data were downloaded from GSE79368 [23] and GSE21644 [9], respectively. In these two sets, intrinsic radioresistance was assessed by a 96-well plate clonogenic assay and the cutoff was set at a median area under the curve of 2.0 in order to identify radioresistant versus radiosensitive cell lines.

Finally, a set of 659 established cancer cell lines from 13 different cancer tissue types (nervous system, soft tissue, digestive system, bone, lung, urogenital system, aero-digestive tract, skin, kidney, breast, pancreas, blood and thyroid) was included in the analysis. Raw data was downloaded from ArrayExpress (E-MTAB-3610) [31]. The corresponding half maximal inhibitory concentration (IC₅₀) for 140 drugs was retrieved from the Genomics of Drug Sensitivity in Cancer data portal [32].

Detailed information (type and number of samples, platform, and normalization) is provided in Additional file 1: Table S1.

For the ‘The Cancer Genome Atlas’ (TCGA) dataset, RNAseqV2 normalized data (level 3, reads per kilobase per million mapped reads) of 518 primary HNSCC (421 HPV-negative and 97 HPV-positive) and 43 paired normal mucosa surrounding tumors was downloaded using the ‘TCGA2STAT’ R package [33]. From a previous publication of the comprehensive characterization of HNSCC [26], as well as from the Broad Institute TCGA GDAC Firehose using a q-value < 0.05 [34, 35], 73 genes recurrently altered at the genomic level in HNSCC were selected. Mutation status and copy number alterations from GISTIC [36] of these genes were downloaded using cBioPortal [37, 38]. Expression of 237 proteins and phosphoproteins generated by Reverse-Phase Protein Array, corresponding to 295 of the 421 primary HPV-negative HNSCC [39], was downloaded from The Cancer Proteome Atlas. For these 518 primary HNSCC, we queried the TCGA data portal as well as the cBioportal [37, 38] to retrieve clinical data, HPV status, the percentage of copy number alteration and the mutation count per sample. The molecular subtype of the 518 HNSCC was kindly provided by Vonn Andrew Walter (Lineberger Comprehensive Cancer Center, University of North Carolina at Chapel Hill, Chapel Hill, NC, USA). The percentage of tumor cells (tumor purity) was given for each tumor sample by the consensus purity estimation method and was retrieved from a previous publication [13].

Other datasets included 44 paired HNSCC and normal mucosa (GSE6631) [40], and two cohorts of 252 and 138 primary HNSCC retrieved from GSE65858 [41] and GSE39366 [25], respectively. Molecular subtypes, HPV status, and clinical information were available for tumors included in GSE39366 and GSE65858.

A detailed description of these datasets is available in Additional file 1: Table S1.

The following sequential approach was used to select genes associated with radioresistance in HPV-negative HNSCC. (1) In order to overcome the confounding effect of tumor purity [11‐13], genes had to be differentially expressed in tumor versus normal mucosa; (2) among them, genes had to be differentially expressed in radioresistant versus radiosensitive HNSCC cell lines; and (3) finally, genes had to be associated with DFS in patients with HPV-negative HNSCC and treated by surgery and adjuvant radiation. Data from TCGA and GSE79368 were used as training sets for gene selection.

The single sample gene set enrichment analysis tool (ssGSEA) [42, 43] was used to compute the enrichment score (ES) using the selected genes, in the training sets (TCGA and GSE79368) and in the validation sets (GSE6631, GSE21644, NCI-60, and GSE39366). The ES of these selected genes was referred to as the RadR score. The RadR score was compared in normal versus tumor samples (GSE6631) and in radiosensitive versus radioresistant cancer cell lines (GSE21644 and NCI-60 panel), and its association with DFS was tested in GSE39366, which included patients who underwent surgical resection alone or followed by adjuvant radiotherapy alone or in combination with chemotherapy.

In the TCGA dataset, we performed an integrated analysis of the RadR score with known recurrent genomic alterations in HNSCC, genomic instability assessed by the percentage of copy number alterations (CNAs) and the mutation count provided by the cBioportal [37, 38]. Using the ssGSEA tool, we then calculated the Pearson’s coefficient of correlation of the RadR score with the ES of a collection of 48 biological hallmark gene sets previously reported [44] in three independent datasets, namely TCGA, GSE39366, and GSE65858. We then evaluated the correlation of the RadR score with the expression pattern of a panel of 237 proteins and phosphoproteins in 295 primary HPV-negative HNSCC from TCGA, as well as with expression of 188 proteins and phosphoproteins in 59 cancer cell lines from NCI-60. Finally, the RadR score was computed in 659 established cancer cell lines from various tumor types (E-MTAB-3610) and its correlation with the IC₅₀ for 140 drugs was tested.

Bioinformatics and statistical analysis were performed using the Array Studio software (Omicsoft Corporation), Bioconductor packages in the R language (http://​www.​bioconductor.​org) [45] and GraphPad Prism version 6.00 (San Diego, SA). Detailed information on processing and normalization of raw data are available in Additional file 1: Table S1. The ssGSEA tool from the GenePattern public server was used to compute the RadR score and the hallmarks’ ESs in various datasets [46]. The unpaired non-parametrical Mann–Whitney or Wilcoxon test was used to compare continuous variables between two groups and the Kruskal–Wallis test was used when more than two groups were compared. The Pearson correlation coefficient (r) was estimated to measure the strength of a linear association between two continuous variables. In the GSE39366 study, recurrence-free survival time was defined by the time in months from tumor biopsy to death, recurrence, or loss to follow-up. In the TCGA cohort, DFS was defined by the time in months from tumor biopsy to a new tumor event. Survival distributions were estimated using the Kaplan–Meier method and compared with the log‐rank test between groups of patients defined by the median RadR score (low versus high RadR score). Multivariate cox proportional hazard model, including age, node, and tumor stage, were built to select genes associated with DFS in patients treated by surgery and radiation. All statistical tests were two-sided, and P values of more than 0.05 were considered to be statistically significant. For multiple testing, a false-discovery rate was computed using the Bonferroni–Holm method in order to adjust P value.

A sequential approach was used to select genes associated with radioresistance, by selecting (1) genes differentially expressed in tumor versus normal head and neck samples; (2) genes associated with in vitro radioresistance, and (3) genes associated with a poor DFS in patients treated by surgery and radiotherapy (Fig. 1).

In order to overcome the confounding effect of tumor purity on intrinsic radioresistance [13], we first filtered out genes that were not differentially expressed in cancer versus normal tissue using gene expression profiles of 86 paired HNSCC and normal mucosa from TCGA (FDR < 0.05, |Fold Change (FC)| > 1, paired Wilcoxon test) (Fig. 1). Among the genes differentially expressed in tumor versus normal tissue, we selected those associated with radioresistance in vitro in 32 HNSCC cell lines (18 and 14 radioresistant and radiosensitive cell lines, respectively) analyzed at baseline, 2 and 6 h after 4 Gy radiation (GSE79368) [23]. Using a two way-ANOVA independently of the time point (with a FDR < 0.05 and a |FC| > 1) (Fig. 1), all samples were included to increase the statistical power, allowing to identify 75 genes over-expressed and 146 genes under-expressed in radioresistant cell lines compared to radiosensitive ones (Fig. 1). These genes were also over- and under-expressed, respectively, in tumor compared to normal mucosa of the head and neck (Fig. 1). Among these 221 genes, we identified a set of clinically relevant genes by their association with DFS in 128 of the 421 patients with HPV-negative HNSCC from TCGA who were treated by surgery and adjuvant radiotherapy with or without concurrent chemotherapy, using a multicovariate Cox proportional hazard model, including age, node (pN0-N1 vs. pN2-N3) and tumor stage (pT1-T2 vs. pT3-T4) (Fig. 1). Clinical and pathological parameters of these 128 patients from TCGA are detailed in Additional file 2: Table S2. Median follow-up was 20.5 months. This led to the identification of 3 and 10 genes associated with increased (hazard ratio (HR) > 1, P < 0.05) and decreased risk (HR < 1, P < 0.05) of recurrence, respectively. The gene list is detailed in Table 1. Detailed information of the FC at each time points and P values using a Wilcoxon test for the 13 selected genes in radiosensitive and radioresistant HNSCC cell lines from GSE79368 are provided in Additional file 3: Table S3.

Using the ssGSEA tool, the 13-gene expression-based RadR score was computed in multiple independent datasets available from public repositories, namely GSE6631, GSE21644, GSE3966, and the NCI-60 panel. The RadR score was increased in 19/22 HNSCC tumors when compared to paired normal mucosa in GSE6631. Overall, the score was significantly increased when tumor and paired normal mucosa were compared (Wilcoxon paired test P < 0.0001) (Fig. 2a). Moreover, the RadR score was higher in radioresistant cell lines (UTSCC-77, UTSCC-24A) compared to the radiosensitive UTSCC-9 cell line (GSE21644) (Fig. 2b). Of note, among the five HNSCC cell lines that were profiled in duplicate by independent investigators (GSE21644), four were overlapping with the cell lines used in the training set (GSE79368). Thus, the RadR score was also computed in a larger panel of 59 cancer cell lines from various tumor types (NCI-60), which were profiled in two or three replicates (174 samples). Interestingly, the RadR score computed in each replicate was highly correlated to other replicates for the same cell line (r ≥ 0.9, P < 0.0001, see Additional file 4: Table S4). Using the first replicate, the RadR score was significantly different between the three groups of NCI-60 cell lines as defined by different ranges of survival fraction at 2 Gy (SF2) as shown in Fig. 2c (P = 0.0236, Kruskal–Wallis test). Radioresistant lines (SF2 > 0.8) had higher RadR scores compared to radiosensitive lines (SF2 < 0.2) (P = 0.0162, Mann–Whitney test).

The clinical relevance of the RadR score was assessed in a set of 63 primary HNSCC using GSE39366. In this study, HPV-positive tumors were excluded and patients were divided into two groups according to treatment. A high RadR score (i.e., greater than the median score) was significantly associated with a shorter recurrence-free survival (log-rank P = 0.0384) in 34 patients treated with surgery and adjuvant radiotherapy with or without concurrent chemotherapy (Fig. 2d). Interestingly, no difference was observed in 29 patients treated with surgery alone or with chemotherapy (P = 0.5218). This finding underlines the specificity of the RadR score and validates our strategy to include the selection of genes associated with in vitro radioresistance and poor DFS in patients surgically resected and treated with postoperative radiation therapy. When validated in independent cohorts, the RadR score may therefore help the clinician to tailor radiation dose according to the radioresistance level of a tumor.

In summary, these results validate the association of the RadR score with in vitro radioresistance and with the benefit from radiotherapy in patients treated with surgery and adjuvant radiotherapy but not in patients treated with surgery alone.

In large genomic profiling studies of HNSCC, four distinct molecular subtypes have been consistently reported, namely atypical, basal, classical, and mesenchymal [25, 26]. HPV-positive HNSCC, which arise from the oropharynx and often fall into the “atypical” class, are recognized to be more radiosensitive compared to HPV-negative tumors [14, 15]. In line with these reports, HPV-positive HNSCC showed lower RadR scores compared to HPV-negative tumors (P < 0.01) in three independent sets of HNSCC (TCGA, GSE65858 and GSE39366) (Additional file 5: Figure S1).

We then focused on the RadR score distribution according to the molecular subtypes and the anatomical subsites in HPV-negative HNSCCs from these three independent datasets (TCGA, n = 421; GSE65858, n = 179; GSE39366, n = 82). The RadR score was found to be statistically different across molecular subtypes, and consistently lower in atypical compared to classical, basal, and mesenchymal HPV-negative HNSCC (P < 0.0001 in all three datasets) (Fig. 3a). Furthermore, the RadR score was different according to the anatomical subsite with a trend to be higher in the oral cavity in the TCGA (P < 0.0001) and GSE39366 (P = 0.0002) datasets, although no difference was observed in GSE65858 (P = 0.1243) (Fig. 3b).

Overall, our results provide evidence that, among HPV-negative HNSCC, “atypical” HNSCCs may be more sensitive to radiotherapy when compared to other subtypes, whereas “mesenchymal” HNSCCs may be more radioresistant. In contrast, the association of the anatomical subsite with the RadR score was not consistent across the three sets of data analyzed.

Since gene expression-based molecular subtypes have been shown to be enriched with some genomic alterations, such as EGFR amplification and HRAS-CASP8 co-mutations in the classical and basal subtypes, respectively [25, 26], we then tested the association of the RadR score with genomic instability and known recurrent genomics alterations in HNSCC.

No clear association was observed between the RadR score and genomic instability assessed by the percent of CNAs (r = 0.012; P = 0.803) and the mutation count (r = –0.204, P = 0.002). We then further investigated the association between the RadR score with somatic mutations and CNAs of 73 genes recurrently altered in HNSCC as reported previously [26], and defined by the Broad Institute TCGA GDAC Firehose [34, 35]. Somatic mutations as well as putative CNAs (GISTIC) of these genes were retrieved from cBioPortal and used to perform an integrated analysis of RadR score with overall genomic instability, mutation status and CNAs in 220 HPV-negative HNSCC (Fig. 4a). A decreased RadR score was significantly associated with NSD1 (P < 0.0001) and PIK3CA (P = 0.0042) mutations, and to a lesser degree with TP63 amplification (P = 0.0108); while HNSCC harboring CCND1 amplification had a higher RadR score (P = 0.0354) (Fig. 4b). Noticeably, percentages of NSD1 and PIK3CA mutations as well as TP63 amplification were significantly higher in the atypical and classical subtypes between HPV-negative HNSCC (P = 0.0013, P = 0.0003, and P = 0.0002, respectively), whereas the percentage of CCND1 amplification was not different between the four subtypes (P = 0.7130) (Fig. 4c).

The influence of tumor purity on genomic analysis has been well established [11‐13]. We thus evaluated the association between the RadR score and the tumor purity given by the consensus purity estimation method [13]. No significant association was found in the 421 HPV-negative HNSCC from TCGA (r = –0.088, P = 0.0707) (Additional file 6: Figure S2), suggesting that the RadR score is independent of the tumor purity.

In summary, the RadR score was associated with some specific known recurrent genomic alterations that are known to be present at different rates in the molecular subtypes of HPV-negative HNSCC, suggesting that genomic-driven radioresistance may be different among different molecular subtypes. The lack of correlation between the RadR score and the tumor purity greatly enhances our confidence in this result.

Since few actionable associations were observed with known recurrent genomic alterations in HNSCC, we seek to gain more insights into the biological context associated with radioresistance, using an in silico functional approach based on the integrative analysis of biological pathways and proteomics data provided by The Cancer Proteome Atlas [39].

We computed the Pearson’s correlation of the RadR score with the ESs of 48 gene sets characterizing biological hallmarks in three independent datasets of primary HPV-negative HNSCC (TCGA, n = 421; GSE65858, n = 179; GSE39366, n = 82). In this analysis, seven hallmarks were significantly correlated with the RadR score in the three datasets (P < 0.01, Fig. 5a): “MYC_targets v2” (average r = 0.30), “E2F targets” (average r = 0.29), “unfolded protein response” (UPR gene set; average r = 0.30), “TGFb signaling” (average r = 0.38), “angiogenesis” (average r = 0.31), “DNA repair” (average r = 0.31), and EMT (average r = 0.38) (Fig. 5a). The ESs of these seven radioresistance-associated hallmarks were significantly different between the molecular subtypes (Kruskal–Wallis test, FDR < 0.0001). Interestingly, the mesenchymal subtype had higher ES for the hallmarks “EMT” and “angiogenesis”, whereas the classical subtype was associated with a higher ES for the hallmark “E2F targets” and “MYC targets”. Notably, the ES of the “UPR” hallmark were lower in the atypical tumors compared to others.

We then analyzed the expression of 237 proteins and phosphoproteins in 295 HPV-negative HNSCC from TCGA and tested their correlation with the RadR score. Interestingly, the most positively correlated proteins were fibronectin and PAI1 (r = 0.3486, P < 0.0001; r = 0.3507, P < 0.0001, respectively), while E-cadherin was negatively correlated (r = –0.3016, P < 0.0001), consistent with the association of the RadR score with EMT. Using a coefficient threshold of 0.3, we found no relevant correlation of phosphoproteins with the RadR score.

In order to assess the association of proteins and phospho-proteins with in vitro radioresistance, we looked for proteins that were correlated to the RadR score as well as to SF2 in 59 cancer cell lines from NCI60 for consistency, using a Pearson’s correlation. A significant correlation was found for Akt (vs. RadR score: r = 0.5351, P < 0.0001; vs. SF2: r = 0.2836, P = 0.0295), fibronectin (vs. RadR score: r = 0.3687, P = 0.0041; vs. SF2: r = 0.259, P = 0.0476), cyclin D1 (vs. RadR score: r = 0.3587, P = 0.0053; vs. SF2: r = 0.4027, P = 0.0016), Irs1 (vs. RadR score: r = 0.2921, P = 0.0248; vs. SF2: r = 0.2885, P = 0.0267), Rictor pT1135 (vs. RadR score: r = 0.2742, P = 0.0356; vs. SF2: r = 0.3047, P = 0.0189), and Paxillin (vs. RadR score: r = 0.2641, P = 0.0433; vs. SF2: r = 0.452, P = 0.0003). Of note, cyclin D1 and fibronectin protein levels were increased according to the range of SF2 in cancer cell lines from NCI-60 (P = 0.0015 and P = 0.0325, respectively) (Fig. 6a). Moreover, fibronectin was significantly higher in the mesenchymal subtype of HPV-negative HNSCC, whereas no difference was observed for cyclin D1 protein expression between the four subtypes.

The association of specific biological hallmarks and proteomic alterations with the RadR score and molecular subtypes of HPV-negative HNSCC suggests that drugs targeting these hallmarks may be relevant to treat tumors with a high RadR score, according to their molecular classification. We computed the RadR score in a large panel of 659 cancer cell lines with available IC₅₀ (half maximal inhibitory concentration) for 140 drugs [32]. By correlating the RadR score with the IC₅₀, four drugs were positively correlated to the RadR score (r > 0.2; P < 0.01) (Additional file 4: Figure S3). They included cisplatin, a potent radiosensitizer routinely used in patients concurrently with radiation therapy, the pro-apoptotic Bcl-2 inhibitor TW37, the BRAF inhibitor PLX4720, and elesclomol, an inhibitor of HSP90 that modulates the UPR pathway [47, 48]. Of note, the RadR score was significantly higher in cell lines sensitive to elesclomol compared to resistant cell lines as previously defined [31] (Additional file 7: Figure S3), suggesting that targeting the radioresistance-associated UPR hallmark may be relevant in patients with a high RadR score.

Our results evidence an important role of seven specific hallmarks into the radioresistance of HPV-negative HNSCC, particularly the EMT hallmark, which is in line with an increased radioresistance observed in the mesenchymal subtype, and the UPR hallmark, which may be modulated by elesclomol in HPV-negative HNSCC with a high RadR score. The association of these hallmarks with molecular subtypes of HPV-negative HNSCC may lead to the generation of hypotheses to refine combination therapies of radiation with modulators of EMT-driven cancer cell plasticity or of the UPR pathway.