All patients included in this study had de novo GBM according to the 2007 World Health Organization Classification. A central pathological review was performed by DFB. In order to focus more specifically on treatment response, progression-free survival (PFS) and MacDonald's criteria of response were used as outcome measures rather than overall survival (OS), which may be influenced by the use of salvage treatment at relapse. Response to radiation therapy was defined in terms of PFS. Patients were considered as responders to radiotherapy if the PFS was >10 months and as non-responders if the PFS was <5 months. Patients treated with radiotherapy and concomitant temozolomide were excluded. In patients treated with first-line chemotherapy, response was evaluated before radiotherapy according to MacDonald's criteria, and all of these patients had an evaluable tumor 15 . These patients were treated with alkylating agents (BCNU or temozolomide). Radiotherapy was administered at progression or after six months of chemotherapy. Patients were considered as responders if they achieved either partial or complete response and as non-responders if they progressed during chemotherapy. Patients' clinical characteristics are available in Additional file 1 Table S1.

Samples were provided as snap-frozen sections of areas immediately adjacent to the region used for the histopathological diagnosis. Only samples representative of the tumor and from which high-quality DNA and/or RNA could be obtained were selected (n = 86). For the comparative genomic hybridization (CGH) array study, 67 samples were available: 21 responders to radiotherapy, 18 non-responders to radiotherapy, 11 responders to first-line chemotherapy and 17 non-responders to first-line chemotherapy. The gene expression array study was performed on 56 samples (including 37 samples common to the CGH study): 19 responders to radiotherapy, 15 non-responders to radiotherapy, 12 responders to first-line chemotherapy and 10 non-responders to first-line chemotherapy.

Approximately 50 mg of tissue from each tumor was used for total RNA extraction using the RNeasy Lipid Tissue mini kit (Qiagen, CA) according to the manufacturer's instructions. RNA quality was verified with the Bioanalyzer System (Agilent Technologies, Palo Alto, CA) using the RNA Nano Chip. RNA (1.5 μg) was processed and hybridized to the Genechip Human Genome U133 Plus 2.0 Expression array (Affymetrix, CA), which contains over 54,000 probe sets analyzing the expression levels of over 47,000 transcripts and variants. This roughly corresponds to 29,500 distinct Unigene identifiers. The processing was done according to the recommendations of the manufacturer.

Immunohistochemistry was performed on tissue microarrays (TMA) comprising 25 GBMs (15 responders and 10 non-responders to radiotherapy for whom enough material was available) that were constructed from routinely processed formalin-fixed paraffin-embedded tumor material. Areas of viable and representative tumor, as determined by a review of all blocks, were marked by a pathologist (DFB) prior to inclusion in the TMA (3 × 0.6-mm cores for each tumor).

After steam-heat-induced antigen retrieval, 5-μm sections of formalin-fixed paraffin-embedded samples were tested for the presence of CD3, CD20 and CD68 using a polyclonal rabbit antibody (1:2) (Dako, Trappes, France), a monoclonal mouse antibody (1:600, L26) (Dako) and a monoclonal mouse antibody (1:5000, KP1) (Dako), respectively. A Benchmark Ventana autostainer (Ventana Medical Systems SA, Illkirch, France) was used for detection, and TMA slides were simultaneously immunostained to avoid inter-manipulation variability. Immunostaining was scored by a pathologist (DFB) as follows: 0 = no positive cell; + = some positive cells; ++ = a clear CD3, CD20 or CD68 infiltration.

The MGMT promoter's (MGMTP) methylation status was assessed in patients treated with first-line chemotherapy. The DNA methylation status of the MGMT promoter was determined by bisulfite modification and subsequent Nested Methylation Specific PCR as previously described 16 . Sodium bisulfite specifically modifies non-methylated cytosines, but not methylated cytosines, to uracil. The sodium bisulfite treatment was carried out using the EZ DNA Methylation Kit (Zymo Research). The stage-1 PCR amplifies a 289-bp fragment of the MGMT gene using primers that do not discriminate between methylated and unmethylated alleles. The primer sequences are as follows: Forward 5′ GGATATGTTGGGATAGTT 3′, Reverse 5′ CCAAAAACCCCAAACCC 3′. PCR conditions were as follows: 95°C for 15 min, then 30 cycles of 95°C for 30 s, 52°C for 30 s and 72°C for 30 s, and finally 10 min at 72°C. The stage-1 PCR products were then diluted 100-fold, and 5 μl was subjected to a stage-2 PCR in which primers were specific to methylated or unmethylated alleles. Primer sequences are as follows: for the methylated reactions, Forward 5′ TTTCGACGTTCGTAGGTTTTCG 3′, Reverse 5′ GCACTCTTCCGAAAACGAAACG 3′; for the unmethylated reactions, Forward 5′ TTTGTGTTTTGATGTTTGTAGGTTTTTGT 3′, Reverse 5′ AACTCCACACTCTTCCAAAAACAAAACA 3′. PCR conditions were as follows: 95°C for 15 min, then 30 cycles of 95°C for 30 s, 62°C for 30 s and 72°C for 30 s, and finally 10 min at 72°C. PCR products were separated on 2% agarose gels stained with ethidium bromide and visualized under UV illumination. As a positive control for methylated alleles, we used DNA from lymphocytes treated with SssI methyltransferase (New England Biolabs: Ozyme, St-Quentin-Yvelines, France) and modified by bisulfite treatment.

All raw and normalized data files for the microarray analysis have been deposited under accessions E-TABM-897 and E-TABM-898 at the European Bioinformatics Institute http://www.ebi.ac.uk/microarray-as/ae. All genomic and transcriptomic analysis was carried out using R software http://www.R-project.org. For details, please refer to Additional file 2.

The influence of the p16 homozygous genomic deletions was studied in an independent data set of 222 GBMs from the Pitié-Salpêtrière Hospital's neuro-oncology department; the p16 deletion and MGMTP methylation statuses of these data set samples were available. This data set consists largely of patients whose data have been previously published 20 . These patients were treated with either radiation therapy alone or with radiation therapy followed by adjuvant chemotherapy with alkylating agents (BCNU or temozolomide).

First, in order to assess if responders and non-responders to either radiation therapy or first-line chemotherapy corresponded to different transcriptomic subgroups of GBMs, we performed an unsupervised hierarchical clustering analysis of the 56 GBMs. As shown in Figure 1, three main transcriptomic subgroups were identified. This clustering was robust and conserved across different gene lists and clustering methods. The centroids generating this classification are provided in Additional file 1 Table S2. However, none of the three clusters was enriched in responders or non-responders, and neither the PFS nor the OS differed between the three clusters. As shown in Figure 1, some responders and non-responders had very similar gene expression profiles.

Next, in order to assess if transcriptomic subgroups of GBMs previously identified in larger series of patients were associated with a specific pattern of response to radiotherapy or to chemotherapy, we classified our 56 samples according to Phillips et al.'s and Verhaak et al.'s transcriptomic classifications and estimated the response rate in each subgroup 8 10 . As shown in Figure 1, there was a significant but not complete overlap between our three subgroups of GBMs and the classes of GBMs identified in these studies. As shown in Table 1, Verhaak et al.'s classes (but not Phillips et al.'s classes) were significantly associated with response to treatment. Indeed, we found that GBMs classified as mesenchymal were more likely to respond to radiotherapy (8 out of 10) than to chemotherapy (1 out of 7) (Fisher's exact test, p = 0.01), whereas GBMs classified as classical were more likely to respond to chemotherapy (7 out of 8) than to radiotherapy (3 out of 11) (Fisher's exact test, p = 0.02). Accordingly, as shown in Figure 2, patients with a mesenchymal GBM had a better outcome when treated with radiotherapy, whereas patients with a classical GBM had a better outcome when treated with first-line chemotherapy. In the neural GBMs, the overall response rate to either radiotherapy or chemotherapy was higher than in the other subgroups (8 responders out of 9 neural GBMs versus 23 responders out of 47 non-neural GBMs, Fisher's exact test, p = 0.03).

In a second step, we focused on all patients treated with radiotherapy alone to identify the molecular characteristics associated with a response to radiation therapy. Comparison of the genomic profiles of responders (n = 21) and non-responders (n = 18) demonstrated that the two groups of patients had very similar genomic profiles (Figure 3). In both groups, there was a high frequency of chromosome 7 gain and chromosome 10 loss, consistent with the most frequently observed genomic abnormalities in GBMs. Only three Minimal Common Regions (MCR) were significantly different between the two groups (Fisher's exact test p-value <0.05), albeit at a low frequency (Table 2). Several genes located in these MCRs were also differentially expressed, but to our knowledge, none of them has been reported to play a role in response to radiation therapy (Table 2 and Additional file 1 Table S3).

Therefore, we focused on the comparison of the gene expression profiles of responders (n = 19) and non-responders (n = 15). As suggested by the observation that responders and non-responders to radiotherapy could have very similar expression profiles, the differences between responders and non-responders were modest; nevertheless, 417 genes were up-regulated in non-responders and 449 up-regulated in responders with p < 0.05 and fold change > 1.5 (Additional file 1 Table S4). To characterize the differences between the two groups, we performed a gene set analysis on the 1000 genes most differentially expressed (500 genes up-regulated in responders and 500 genes up-regulated in non-responders). This demonstrated that these gene lists were enriched in genes with very different ontologies (Table 3, Table 4 and Additional file 1 Table S5). The list of up-regulated genes in responders was most significantly enriched in genes involved in the immune response, namely in immune genes previously reported to be associated with an improved outcome after radiochemotherapy (Cluster G24, Table 4) 12 . This enrichment was seen in non-specific inflammatory response genes as well as in genes involved in the B cell-mediated response and T cell activation (Table 3, Table 4). In order to validate these findings at the protein level, an immunohistochemical study of CD3, CD20 and CD68 markers was performed in the tumors of responders and non-responders to radiotherapy. Neither GBM samples of responders nor those of non-responders were infiltrated by CD20 cells. However, infiltration by CD3 cells and by both CD3 and CD68 was much more frequent in responders than in non-responders to radiotherapy (Table 5, Fisher's exact test, p-value = 0.04). On the other hand, the list of up-regulated genes in non-responders was most significantly enriched in genes induced by hypoxia, suggesting a higher level of hypoxia in the non-responders (Table 3, Table 4). As hypoxia is a well-known mechanism of radiation resistance, these results suggest that even in GBMs, which are highly hypoxic tumors, a higher level of hypoxia raises the level of resistance to radiation therapy.

Comparison of the genomic profiles (gains, losses, homozygous deletions and amplifications) of responders (n = 11) versus non-responders (n = 17) to first-line chemotherapy (in contrast to radiation therapy) demonstrated substantial genomic differences (Figure 4). CDKN2A (p16) locus homozygous deletions on 9p21, EGFR amplification and 24 MCRs (3 loss and 21 gains) were significantly associated with response to chemotherapy (Fisher's exact test, p < 0.05) (Table 2 and Additional file 1 Table S6). CDKN2A (p16) locus homozygous deletion was the most significant event; it was observed in 82% of responders but in none of the non-responders (Fisher's exact test, p < 10^-4). It was associated with a significant down-regulation of the expression of CDKN2A in responders (p = 0.04). Concerning MTAP and CDKN2B, which are also located in the p16 locus, a significant down-regulation of the expression in responders was observed only for MTAP (p = 0.0003). Among the other genomic alterations that have been reported to alter the retinoblastoma (RB) signaling pathway in GBMs (i.e., CDK4 or CDK6 or CCND2 amplification and RB1 or CDKN2C homozygous deletion), we found a CDK4 amplification in three non-responders and a CCND2 amplification in one responder who had also a p16 deletion. Thus, in contrast to p16 deletion, other genomic alterations disrupting the RB pathway did not seem to be associated with response to chemotherapy.

EGFR amplification was more frequently observed in responders than in non-responders (81% vs. 35%, Fisher's exact test, p = 0.02). All lost MCRs were located on chromosome 9p22-24, and these loci were not independent. Among the genes located in these MCRs (Table 2), none has been reported to be involved in chemosensitivity. Gained MCRs were located on chromosomes 11p, 11q, 17p, 19p, 19q, 20p and 20q. Genes located in these MCRs and significantly overexpressed in responders are summarized in Table 2. To our knowledge, none of these genes has been associated with chemosensitivity.

Surprisingly, MGMTP methylation analysis demonstrated that most patients in this series were MGMTP unmethylated (8 out of 11 responders and 14 out of 17 non-responders); thus, these genomic abnormalities may actually represent alternative mechanisms of chemosensitivity in MGMTP unmethylated patients.

Next, we compared the gene expression profiles of responders (n = 12) and non-responders (n = 10). The differences between the two groups were less important than the analogous differences between responders and non-responders to radiotherapy, with 292 genes being up-regulated in non-responders and 203 genes being up-regulated in responders with a p-value <0.05 and a fold change >1.5 (Additional file 1 Table S7). Gene set analysis on the 1000 genes most differentially expressed between the two groups (500 genes up-regulated in responders and 500 genes up-regulated in non-responders) was performed. In agreement with the genomic analysis, the list of up-regulated genes in non-responders was enriched in genes located on 9p (Table 6). Interestingly, this gene list was also enriched in genes reported to be up-regulated in embryonic and neural stem cells, whereas the list of genes up-regulated in responders was enriched in genes up-regulated in the normal brain, suggesting a link between resistance to chemotherapy and a more undifferentiated phenotype of the tumor (Table 6, Table 7 and Additional file 1 Table S8). Consistently, the list of genes up-regulated in responders was enriched in a set of normal brain genes (Cluster 18, Table 7) associated with improved outcome after chemoradioherapy whereas the list of genes up-regulated in non-responders was enriched in a set of stem-cell genes (Cluster 28_98, Table 7) associated with a worse outcome after concomitant chemoradiotherapy 12 . Among the stem-cell genes, HOXA10 and HOXC6, were the most up-regulated genes in non-responders to chemotherapy.

As the p16 locus homozygous deletion was the most consistent finding in patients responding to first-line chemotherapy, we focused on the study of this genomic abnormality in an independent series of patients treated either by radiotherapy alone (n = 79) or by radiotherapy followed by adjuvant chemotherapy (n = 143) for whom the MGMTP methylation status was also available (Table 8). We hypothesized that, if this genomic abnormality was consistently associated with chemosensitivity, then, as has been demonstrated for MGMTP methylation, the benefit of adjuvant chemotherapy would be more pronounced in the group of patients with the p16 deletion.

In the patients with the p16 deletion, we found that, regardless of MGMTP methylation status, those treated with radiotherapy and adjuvant chemotherapy had a significantly longer PFS (10.7 vs. 7.9 months, p = 0.007, if unmethylated; 9.7 vs. 7.9 months, p = 0.01, if methylated) and OS (25.3 vs. 12.2 months, p = 0.002, if unmethylated; 18.6 vs. 12.2 months, p = 0.01, if methylated) than those treated with radiotherapy alone (Figure 5). This association was independent of age, Karnofsky performance status and type of surgery. In contrast, among patients without the p16 deletion who were treated with radiotherapy and adjuvant chemotherapy, only those patients with a methylated MGMTP had a longer PFS (11.8 vs. 6.3 months, p = 0.008) and a longer OS (18.4 vs. 15 months, p = 0.05) than those treated with radiotherapy alone (Figure 5). Patients without the p16 deletion treated with radiotherapy and adjuvant chemotherapy and with an unmethylated MGMTP did not fare better than patients without the p16 deletion who were treated with radiotherapy alone.

The finding that MGMTP unmethylated patients benefit from adjuvant chemotherapy when the p16 locus is deleted is consistent with the association between p16 deletion and chemosensitivity. However, we found no additive effect between MGMTP methylation and p16 deletion. Indeed, in the group of patients with a p16 deletion who were treated with radiotherapy and adjuvant chemotherapy, the outcomes were similar between MGMTP methylated and unmethylated patients (Figure 5).