Patient characteristics (age, sex and median survival) for the four datasets (AGS, the Mayo Clinic, GliomaSE and TCGA) are described in . The majority of the observed survival Cox regression P
values for 314,635 SNPs from the AGS discovery dataset conformed to the identity line in the Q-Q plot, whereas 90 SNPs showed significant deviation from expectation at P = 10−4
(Supplementary Figure 2
). We submitted these 90 SNPs for validation in Mayo Clinic patients of which 78 passed quality control. Ten of these SNPs had P < 10−5
in the combined analysis using a fixed-effect model. (25
) Examination of these 10 SNPs in two additional patient groups, GliomaSE and TCGA patients, yielded one SNP, rs7732320, that had discovery and validation combined P < 10−5
for survival and had proportionality of hazards in all four datasets ( and Supplementary Table 3
). The associations of this SNP with patient survival were in the same direction across the studies and had a combined validation P = 0.008 and a combined discovery-validation P = 1.3X10−6
. There was no evidence of heterogeneity of the HR estimates across the four studies (). Effect modification by age at diagnosis for rs7732320 was evaluated in the AGS discovery data by the significance of the interaction term between age at diagnosis and the SNP; no statistical significant interaction was detected. In the AGS discovery data, the median survival time for the three groups of patients with 0, 1, and 2 adverse alleles of rs7732320 were 17.8, 13.4 and 10.6 months respectively.
Patient characteristics of glioblastoma patients used in discovery (University of California, San Francisco, 1997–2008) and validation sets (Mayo Clinic, GliomaSE and TCGA).
Table 2 Association of rs7732320 genotype with survival for glioblastoma multiforme patients with initial standard of care (resection, radiation and temozolomide) treatment discovered in a genome-wide association study (University of California, San Francisco, (more ...)
Rs7732320 is located in the intronic region of SSBP2;
we therefore investigated whether patient survival was associated with the transcript levels of SSBP2
among 619 patients in three publically available glioblastoma gene expression datasets (Lee et al., (20
) Murat et al., (21
) and TCGA (22
);see Methods and Supplementary Figure 1
). We observed a strong and significant association of SSBP2
expression with poorer overall survival (HR = 1.22; 95% CI: 1.09 – 1.36; P = 5.3 X 10−4
) and the association was consistent across the three expression datasets (). No effect modification by age at diagnosis was found for the association of SSBP2
tumor expression with survival in any of the three expression datasets. Additionally, among TCGA glioblastoma patients, the HR for patient survival associated with tumor SSBP2
expression was highest and statistically significant only among patients with the previously described (22
) proneural signature (HR = 1.44; 95%CI: 1.10 – 1.89; P = 0.007) (). Consistent with this finding, we found that proneural glioblastoma patients expressed the lowest amount of SSBP2
compared to the other subtypes (Wilcoxon P = 2.16X10−12
; ). Intriguingly, even though the overall survival for patients of the proneural subtype was not significantly different from the other gene expression subtypes (log rank P = 0.21; ), significant survival differences were observed for the proneural SSBP2
-negative patients (), arbitrarily defined as the subset of patients with lower than 25 percentile of SSBP2
expression in the proneural group. We observed significantly better survival for proneural SSBP2
-negative patients (median survival time: 28.8 months) than proneural SSBP2
-positive patients (median survival of time: 12.4 months) and all other non-proneural glioblastoma patients (median survival time: 13.8 months). Proneural SSBP2
-negative status remained a significant prognostic factor for longer survival (Cox P = 9.7X10−3
) in Cox multivariate analysis after adjusting for patient age at diagnosis and sex.
Association of tumor gene expression and survival in glioblastoma multiforme cases using data from three different sources.
Figure 1 (A) Boxplots of SSBP2 tumor expression by previously assigned TCGA expression groups in 303 glioblastomas: C, classical; M, mesenchymal; N, neural; and P, proneural. (B) Kaplan-Meier survival curves for the four TCGA expression groups. (C) Kaplan-Meier (more ...)
The proneural expression subtype has recently been linked to a subset of tumors exhibiting a glioma-CpG island methylator phenotype (G-CIMP)(26
). To understand the relationship between SSBP2
and the G-CIMP signature, we compared the SSBP2
genotype and tumor expression in the set of TCGA glioblastoma samples with available G-CIMP status. Of the 241 TCGA samples with concomitant tumor expression and G-CIMP information, 24 were G-CIMP positive and they expressed a much lower level of SSBP2
than the 217 G-CIMP negative tumors (Wilcoxon P = 3.54X10−4
). Of the 151 TCGA samples with attendant SSBP2
genotype and G-CIMP information, 2 out of the 16 (12.5%) G-CIMP positive glioblastoma patients belonged to the group with at least one copy of the adverse allele T, in contrast to a much higher proportion (28.4%, 38 out of 135) in the G-CIMP negative glioblastomas. Because of small sample sizes, validating the relationship between SSBP2 genotype, expression and G-CIMP status will require further studies.
Interestingly, IDH1 mutation status was not found to be associated with the SSBP2 genotype in either of the AGS and TCGA datasets, nor was it linked to SSBP2 tumor expression in the TCGA dataset. For TP53 mutation, we detected an increased frequency of the risk allele T of SSBP2 in TP53 mutated glioblastoma patients (OR = 2.35; 95% CI = 1.06–5.19; P = 0.03) in the TCGA dataset. However, this association was not found in the AGS dataset. Next, in order to perform a multivariate analysis incorporating both patient genotypes and tumor markers that are related to survival in glioblastoma patients, we used the AGS dataset, for whom 143 of the 315 patients with standard-of-care treatment had data on TP53 and IDH1 mutation status, and EFGR amplification. Unfortunately only 35 of the 115 TCGA patients with standard-of-care treatment had both IDH1 and TP53 mutation data. In a Cox multivariate regression including age, sex, IDH1 mutation status, EGFR copy number, TP53 mutation and SSBP2 rs7732320 genotype, SSBP2 genotype remained an independent predictor of poorer survival (HR=1.99; 95%CI: 1.32–3.00, P=0.001, n=143)
Taken together, the findings above present a consistent connection by showing that both the adverse SSBP2 inherited variant and increased SSBP2 expression in tumors are associated with shorter survival time in glioblastoma patients and that the relationship is most evident among patients with the proneural expression signature. A test for the statistical interaction between the SSBP2 SNP rs773232 and its tumor expression was performed in the TCGA dataset by inclusion of the cross-product term in the Cox model and assessed by use of the likelihood ratio test. No statistical significant interaction effect (P=0.66) was observed.
To further localize the association with survival in the 5q14.1 region around rs7732320, we imputed non-genotyped SNPs in the entire genomic locus of SSBP2
with a 100kb extension at its 3′ end from 80,680,000 – 80,980,000 on chromosome 5. The Hapmap II CEU dataset (27
) contained 217 SNPs in this region (the AGS dataset had 31 SNPs). Out of the 186 (217 minus 31) imputed SNPs, 159 had good imputation quality for AGS, Mayo, and TCGA. Meta-analysis using a fixed effect model to combine study-specific HR estimates from age-gender adjusted Cox models shows a genome-wide statistically significant association of patient survival with the SNP rs17296479 (P = 1.0 X 10−7
; see and Supplementary Table 4
), which is located ~8kb centromeric of rs7732320. Two SNPs, rs12187089 and rs11738172, located between these two markers, also displayed strong associations with patient survival, with P = 1.2 X 10−7
and 2.3 X 10−7
respectively. These four SNPs are highly linked with each other (r2
> 0.8). The smallest combined nominal P value from multivariate Cox models of patient survival with the remaining SNPs adjusting for rs17296479 genotype was 0.061, suggesting that there were no residual independent survival signals remaining.
Figure 2 Association of genetic variants near SSBP2 with survival using data from uniformly treated glioblastoma patients. We used data from the San Francisco Adult Glioma Study, the Mayo Clinic and The Cancer Genome Atlas studies for imputation. Evidence for (more ...)