Genome-wide association study identifies the GLDC/IL33 locus associated with survival of osteosarcoma patients

Abstract

Survival rates for osteosarcoma, the most common primary bone cancer, have changed little over the past three decades and are particularly low for patients with metastatic disease. We conducted a multi-institutional genome-wide association study (GWAS) to identify germline genetic variants associated with overall survival in 632 patients with osteosarcoma including 523 patients of European ancestry and 109 from Brazil. We conducted a time-to-event analysis and estimated hazard ratios (HR) and 95% confidence intervals (CI) using Cox proportional hazards models, with and without adjustment for metastatic disease. The results were combined across the European and Brazilian case sets using a random-effects meta-analysis. The strongest association after meta-analysis, was for rs3765555 at 9p24.1, which was inversely associated with overall survival (HR=1.76; 95% CI 1.41-2.18, P = 4.84×10⁻⁷). After imputation across this region, the combined analysis identified two SNPs that reached genome-wide significance. The strongest single association was with rs55933544 (HR=1.9; 95% CI 1.5-2.4; P=1.3×10⁻⁸), which localizes to the GLDC gene, adjacent to the IL33 gene and was consistent across both the European and Brazilian case sets. Using publicly available data, the risk allele was associated with lower expression of IL33 and low expression of IL33 was associated with poor survival in an independent set of patients with osteosarcoma. In conclusion, we have identified the GLDC/IL33 locus on chromosome 9p24.1 as associated with overall survival in patients with osteosarcoma. Further studies are needed to confirm this association and shed light on the biological underpinnings of this susceptibility locus.

INTRODUCTION

Osteosarcoma is the most common primary bone cancer in children and adolescents.^1–4 The introduction of effective neo-adjuvant and adjuvant chemotherapy in the 1980s resulted in improved long-term survival for non-metastatic osteosarcoma patients, increasing survival from 20%–30% to more than 70%.^4–7 However, over the past three decades, there has been little improvement in survival rates for patients with metastatic disease at diagnosis⁸ with 5-year overall survival rates remaining at 25-30%.⁹, ¹⁰ Several factors have been suggested to influence survival of patients with osteosarcoma¹¹, including age at diagnosis, metastatic disease at presentation, tumor histology, blood alkaline phosphatase levels, tumor size and location, and response to chemotherapy (estimated by the percentage of tumor necrosis after chemotherapy).^12–14 Recently, we reported that germline genetic variants in the NFIB gene locus (9p23-9p22.3) are associated with metastatic disease at osteosarcoma diagnosis, suggesting that genetic susceptibility could contribute to survival.¹⁵

There is growing interest in whether germline genetic variants could influence outcomes of patients with cancer. A population-based family study showed that cancer-specific survival can be partly related to inherited factors within families, suggesting there are germline genetic determinants of survival.¹⁶ While earlier candidate gene studies have identified variants associated with prognosis that have not been confirmed, more recent large genome-wide association studies (GWAS) have identified associations between common SNPs and survival in adult cancers of the pancreas,¹⁷ breast,^18–23 ovary,²⁴ and in a rare pediatric cancer, neuroblastoma.^25–27 There have also been exploratory pharmacogenomic studies of pediatric Ewing sarcoma and osteosarcoma that have identified SNPs associated with response to treatment and survival,^28–31 although most await further validation. We performed a GWAS in order to explore whether germline genetics may contribute to survival in patients with osteosarcoma.

METHODS

Study populations

A summary of the participating studies is provided in Supplemental Table 1. Previously, we reported 689 histologically confirmed osteosarcoma patients of >80% European ancestry based on a STRUCTURE analysis³² employing principal components analyses of a set of 12,000 unlinked markers (pairwise r² < 0.004).¹⁵, ³³ A subset of 523 European ancestry osteosarcoma patients had available data on mortality (European set) for a survival analysis. After performing the GWAS for survival in this data set, we evaluated the most promising SNPs from the European set (P < 10⁻⁴) in 109 Brazilian osteosarcoma patients from the Instituto de Oncologia Pediátrica GRAACC/UNIFESP and Universidade Federal de Sao Paulo, Brazil (Brazilian set; Supplemental Table 1). Participating subjects provided informed consent under the auspices of local Institutional Review Boards.

Genotyping and quality control

All subjects were previously genotyped as part of our osteosarcoma susceptibility GWAS (dbGaP Study Accession: phs000734.v1.p1).³³ In brief, genotyping was conducted on the Illumina OmniExpress BeadChip, SNPs included in the analyses were autosomal with a minor allele frequency (MAF) of 5% or more; had a 90% or more completion rate; and no evidence of deviation from Hardy-Weinberg equilibrium (P > 10⁻⁷). SNPs were also excluded if they had abnormal heterozygosity values. After quality control assessment, 510,856 SNPs were advanced in our survival analysis.

Statistical analyses

We conducted a time-to-event analysis to investigate the effect of genetic variation on overall survival. The outcome variable of interest was time until the event of death. The overall survival time was calculated as the time from the date of osteosarcoma diagnosis until the date of death for those deceased or the last date known to be alive; patients were censored at the last date known to be alive or when lost to follow-up. All events were identified and verified through medical record review and/or death certificates at each participating study center. We modeled each genome-wide association between one or more SNPs and survival using Cox proportional hazards regression and estimated hazard ratios (HR) and 95% confidence intervals (CI) per copy of the minor allele (log-additive genetic model).³⁴ We tested the hazards proportionality assumption (i.e., the hazard ratio is constant over time)³⁵ of the Cox model using Schoenfeld’s residuals;^{36, 37} we did not detect nominally significant violations of the proportional hazards assumption.

Cox models were adjusted for age at diagnosis, gender, significant principal components, and study/center (except for the Brazilian study, since all samples were from the same hospital). We did not adjust for metastatic disease agnostically at the genome-wide level, because some SNPs may be associated with metastatic disease and survival, as we have shown previously,¹⁵ and thus adjusting for potential intermediate factors that lie on the causal pathway could introduce bias.^38–40 However, since metastatic disease is a prognostic factor for overall survival,¹² we performed sensitivity analyses for the top signals by also adjusting for metastatic disease .⁴¹ We constructed Kaplan-Meier survival curves⁴² for the top SNPs under dominant models and compared their statistical equivalence for each model with the log-rank test.

SNPs that reached P < 10⁻⁴ in the European set were followed-up in the Brazilian set. We used a random-effects meta-analysis with inverse-variance weighting to estimate the summary effect across the sets.⁴³ We evaluated between-sets heterogeneity using Cochran’s Q chi-squared statistic and quantified heterogeneity with the I² metric.⁴⁴ SNPs that demonstrated significant between-set heterogeneity (P < 0.05) were excluded.

Based on our GWAS results, we conducted region-specific imputation analysis of flanking SNPS, namely, 1 Mb region on either side of the strongest SNP using the IMPUTE2 software and the reference data from the 1000 Genomes project (Phase 3 genotype data).⁴⁵

To investigate whether signals in the same genomic region represent independent associations, we conducted conditional analyses by adjusting the Cox models for the top SNP in each region as applicable.

Statistical analyses were performed in Stata version 13 and R 3.0.2. All P-values are two-sided.

eQTL and survival-expression analyses

We performed expression quantitative trait locus (eQTL) based analyses using publicly available genotype and expression data from 29 osteosarcoma tumors (GSE33383)⁴⁶ and separately in osteosarcoma tumors from 89 patients included in the Therapeutically Applicable Research to Generate Effective Treatments (TARGET) osteosarcoma dataset (http://ocg.cancer.gov/programs/target).⁴⁷ The data used for TARGET are available at dbGaP accession phs000218, accession phs000468, in which we conducted survival-expression analysis.⁴⁷ For the 127 patients from the combined dataset (Kuijjer et al⁴⁶ and Buddingh et al⁴⁸) survival-expression was analyzed using the R2: Genomics analysis and visualization platform (http://r2.amc.nl/).⁴⁹

Bioinformatic analyses

Linkage disequilibrium (LD) was evaluated with r² based on the 1000 Genomes Project (Phase 3 genotype data)⁴⁵ using LDlink (http://analysistools.nci.nih.gov/LDlink/).²³ Chromosome location for human genome assembly hg19 was retrieved from the National Center for Biotechnology Information’s (NCBI) Gene database (http://www.ncbi.nlm.nih.gov/gene/). Genomic annotation of SNP markers was conducted using the Encyclopedia of DNA Elements (ENCODE)⁵⁰ tool HaploReg⁵¹ and RegulomeDB⁵² for all cell lines. Surrogate SNPs were identified using the 1000 Genomes data for individuals of European ancestry with an r² > 0.4 and within ±500kb. LocusZoom⁵³ was used to plot regional associations.

RESULTS

Patient characteristics

Table 1 shows the characteristics of the patients included in this study. There were 632 total osteosarcoma patients included in the overall survival analyses. In the European set, there were 170 (33%) mortality events, and 37 (34%) in the Brazilian set. Age (P < 0.001) and presence of metastases at diagnosis (P < 0.001) were associated with patient survival (Table 1, Supplemental Figure 1).

Table 1.

Patient characteristics in the European and Brazilian set

Endpoint	European*	Brazil	Combined Analysis
Overall Survival	N=523	N=109	N=632	75% ST† (years)	P-value††
Vital status, N (%)
Dead	170 (33)	37 (34)	207 (33)	2.8
Alive	353 (67)	72 (66)	425 (67)
Age (years)					<0.001
<25	432 (83)	93 (85)	525 (83)	3.3
≥25 to<60	69 (13)	12 (11)	81 (13)	3.4
≥60	21 (4)	0 (0)	21 (3)	4.3
Missing	1 (0)	4 (4)	5 (1)
Gender, N (%)					0.737
Males	295 (56)	58 (53)	353 (56)	2.9
Females	228 (44)	51 (47)	279 (44)	2.7
Metastasis at diagnosis, N (%)					<0.001
Yes	131 (25)	40 (37)	171 (27)	4.2
No	392 (75)	69 (63)	455 (72)	1.5

^*All subjects included in the European set were of >80% European ancestry.

^†Shows the time at which 75% of patients had not experienced the event of interest (i.e. death or progression).

^††P-values are from log-rank test.

ST: survival time; NA: non-applicable

SNPs associated with overall survival

In a case-case analysis, 81 SNPs were associated with overall survival at P < 10⁻⁴ in the European set (Supplemental Table 2) and were followed-up in the independent set of 109 osteosarcoma cases from Brazil. The strongest association with overall survival in the European set is at chromosome 5q11.1 tagged by rs1030228 (located 14kb 5' of NDUFS4) with a HR for mortality of 1.71 (95% CI 1.39-2.12, P = 7.10×10⁻⁷; Supplemental Figure 2, Supplemental Table 2). However, this SNP is not associated with survival in the Brazilian study (P=0.80). There is a large degree of between-set heterogeneity (P_het=0.052, I²=73.5%) for this variant (Supplemental Table 2).

In the combined analysis, the strongest association was SNP rs3765555, which is inversely associated with overall survival (HR=1.76 per copy of the A allele, 95% CI 1.41-2.18, P=4.84×10⁻⁷; I²=0%; Table 2, Figure 1A). This SNP is located in intron 23 of the glycine dehydrogenase (decarboxylating) gene (GLDC) on chromosome 9p24.1. The MAF for rs3765555 in the European (MAF=0.26) and Brazilian (MAF=0.21) populations are similar and we did not observe significant between-set heterogeneity (P_het=0.34; Table 2, Supplemental Table 2).

Table 2.

SNPs associated with overall survival.

SNP*	Method	Gene Locus**	Position†	Set	MAF	Unadjusted for metastatic disease		Adjusted for metastatic disease
						HR (95% CI)	P-value	HR (95% CI)	P-value
rs3765555-C|A	Genotyped	GLDC	Chr9: 6535956	European	0.259	1.67 (1.32–2.13)	2.70×10−5	1.71 (1.34–2.18)	1.60×10−5
				Brazil	0.205	2.23 (1.31–3.81)	3.29×10−3	2.12 (1.27–3.53)	3.87×10−3
				Combined		1.76 (1.41–2.18)	4.84×10−7	1.78 (1.43–2.21)	2.74×10−7
rs55933544-C|T	Imputed	GLDC	Chr9: 6534080	European	0.231	1.85 (1.44–2.37)	1.58×10−6	1.91 (1.49–2.45)	3.20×10−7
				Brazil	0.231	2.11 (1.21–3.69)	8.44×10−3	1.98 (1.16–3.38)	0.012
				Combined		1.89 (1.50–2.37)	4.81×10−8	1.92 (1.53–2.41)	1.34×10−8

^*Alleles are shown as major/minor.

^**Gene locus information is based on the GENCODE data from HaploReg v2.

^†Position is based on hg19.

Hazard ratios are shown per copy of the minor allele in the discovery stage.

MAF: minor allele frequency; HR: hazard ratio; CI: confidence interval

Figure 1.

Regional plots of the combined association results, recombination hotspots and linkage disequilibrium (LD) for the 9p24.1 region that harbors rs55933544 and rs74438701 that are associated with overall survival. Results are shown for unconditional (A, B) and conditional (C) analyses. Also shown is the Kaplan-Meier curve (D) for overall survival for the strongest SNP (rs55933544) under a dominant model in the combined European and Brazilian sets. In panels A-C, Y-axes represent the statistical significance (−log10 transformed P values) of SNP association results from a trend test (left) and the recombination rate (right). SNPs are color-coded based on pairwise linkage disequilibrium (r²) with the most statistically significant SNP. The most statistically significant SNP is labeled and shown in purple. Allelic P values are the P-values from Cox models. Physical locations of the SNPs are based on NCBI human genome build 36, and gene annotation was based on the NCBI RefSeq genes from the UCSC Genome Browser.

In a further exploration of the promising regions, we imputed SNPs across a 1 Mb region centered on rs3765555 to further evaluate this locus. After a random-effects meta-analysis for the imputed SNPs, we identified rs55933544 (Table 2, Figure 1A) as the SNP most strongly associated with overall survival (HR=1.89 per copy of the T allele, 95% CI 1.50-2.37; P=4.81×10⁻⁸; I²=0%), which is in strong LD with rs3765555 (r²=0.86 in Europeans and r²=0.94 in admixed Americans). The results remained the same after adjustment for metastatic disease at diagnosis (Table 2), suggesting that the 9p24.1 locus marked by rs55933544 affects overall survival independent of metastatic disease at genome-wide significance (HR=1.92 per copy of the T allele, 95% CI 1.53-2.41; P=1.34×10⁻⁸; I²=0%). rs55933544 (chr9:6534080) was not correlated with SNP rs7034162 at 9p23-9p22.3 (chr9:14190287) that we previously identified as associated with metastatic disease at osteosarcoma diagnosis (r²=0.0008, 1,000 Genomes Project CEU data).¹⁵

A second SNP at 9p24.1, rs74438701 located approximately 25kb downstream of the interleukin 33 (IL33) gene, is also inversely associated with overall survival in the combined analysis (HR=2.00 per copy of the C allele, 95% CI 1.56-2.57, P=4.90×10⁻⁸; I²=0%; Figure 1B). However, the conditional analysis showed that rs74438701 is not an independent signal (Figure 1C).

Kaplan-Meier curve analysis indicate a statistically significant difference between survival rates over time (log-rank test P < 0.001) for both the dominant (Figure 1D) and a multiplicative model (data not shown) for rs55933544. In addition, we confirmed this association in an independent dataset of 89 cases (TARGET⁴⁷; log-rank P=0.013; Figure 2A).

Figure 2.

IL33 expression levels associated with survival in osteosarcoma patients. (A) Kaplan-Meier curve for overall survival by rs55933544 genotype; (B) eQTL of rs55933544 and IL33; and, (C) patient survival by IL33 low and high expression levels (independent of genotype), all using TARGET data.

We examined the set of highly correlated surrogate SNPs (n=31) across the 9p24.1 region (based on an r²>0.4, 1,000 Genomes Project CEU data) to identify putative regulatory elements using the ENCODE data resource⁵⁰ tools HaploReg⁵¹ and RegulomeDB⁵² (Supplemental Table 3). A subset of the surrogate SNPs (n=29) are located in predicted promoter and/or enhancer histone marks, DNAse sensitivity regions, and/or transcription factor binding sites in a variety of different cell types and may have a functional impact (Supplemental Table 3).

IL33 expression levels associated with survival

We performed expression quantitative trait locus (eQTL) analyses to evaluate whether rs55933544 was associated with expression of GLDC, IL33 or other neighboring protein-encoding genes, using publicly available expression and genotyping data on osteosarcoma tumors. Interestingly, a previous eQTL was observed between rs55933544 and IL33 expression in human skin⁵⁴ and human brain tissue⁵⁵ (Supplemental Table 3). We found that the risk allele (T) of rs55933544 was significantly associated with a decrease in IL33 expression in both osteosarcoma tumor data sets from TARGET⁴⁷ (N=83, P=0.041) and Kuijjer et al.⁴⁶ (N=29, P=0.020) (Figure 2B and Supplemental Figure 3). In addition, lower expression of IL33 in osteosarcoma tissue was independently associated with worse osteosarcoma patient survival in TARGET⁴⁷ (log-rank test P=0.023; Figure 2C) and Kuijjer et al.⁴⁶ (raw P=7.9x10⁻³; Supplemental Figure 3). There was no association between rs55933544 genotypes and expression of GLDC or other nearby protein-encoding genes (TPD52L3, UHRF2 and KDM4C; data not shown).

DISCUSSION

We conducted a genome-wide association study for overall survival in osteosarcoma cases using data from a multi-stage, international collaborative effort.³³ One locus, GLDC/IL33 at 9p24.1, was associated with overall survival of patients with osteosarcoma, which suggests that germline genetics can influence osteosarcoma outcomes, independent of metastatic disease. Here we observed moderate to large effect sizes for a SNP associated with overall survival, a finding similar to that observed in our GWAS of metastatic disease at osteosarcoma diagnosis.¹⁵ The observed effect sizes are also comparable to GWAS of other pediatric and young adulthood cancers,^{25, 56–58} and higher than those observed in adult GWAS of common cancer susceptibility.^59–61

The most promising signal for overall survival in patients with osteosarcoma localizes to the 9p24.1 region, downstream and independent of the NFIB gene locus (9p23-9p22.3) previously reported for metastatic disease.¹⁵ The SNP marker, rs55933544, in the GLDC gene region is associated with decreased survival. High expression of GDLC has been associated with poor survival in other cancers,^{62, 63} however, we did not detect an eQTL for rs55933544 and GDLC in osteosarcoma cells. Interestingly, rs55933544 alleles have also been associated with expression of the nearby gene, IL33^{54, 55} and we detected an eQTL with IL33 in osteosarcoma cells. In addition, lower expression of IL33 was associated with poor survival in patients with osteosarcoma. IL-33 is an inhibitor of bone reabsorption, blocking osteoclastic activity,⁶⁴ which may be important in osteosarcoma. Lower levels of IL-33 have also been associated with worse prognosis or more advanced disease in several other tumor types,^65–67 consistent with our data.

This exploratory study requires further follow up and has limitations. Treatment likely varied among patients in our studies, but we could not control for this because treatment information was not available. However, it is unlikely that individual treatment modalities varied systematically by germline genotype. We also performed a combined analysis with osteosarcoma patients from Brazil, with a relatively small sample size. LD patterns differ in many parts of the genome between admixed Brazil populations and Europeans, and this could lead to false negative results in our analysis. This could explain the lack of replication in Brazilian patients of some of the top SNPs in the European osteosarcoma patients. However, an important strength, is that this two-stage design also reduces the likelihood of false positive results.

In conclusion, we provide evidence that germline genetic variants are associated with overall survival in osteosarcoma patients. These findings warrant follow-up in additional populations and functional characterization to investigate the biologic mechanisms by which polymorphisms at this locus impact survival.

Novelty and Impact.

To date, few prognostic factors have been identified associated with survival in patients with osteosarcoma. The authors conducted a genome-wide association study (GWAS) of overall survival in two sets of patients with osteosarcoma. They identified a common single nucleotide polymorphism (SNP), rs55933544, located in the GLDC gene on chromosome 9, associated with poor survival. The rs55933544 risk allele was associated with lower expression of the nearby gene, IL33. These findings, if replicated in additional populations, form the foundation for future studies of the molecular basis of the association of the GLDC/IL33 (rs55933544) variant with survival in osteosarcoma.

Abbreviations used

GWAS

genome-wide association study

HR

hazard ratio

CI

confidence intervals

MAF

minor allele frequency

LD

linkage disequilibrium

ENCODE

Encyclopedia of DNA Elements

NCBI

National Center for Biotechnology Information

eQTL

expression quantitative trait loci

SNP

single nucleotide polymorphisms

GLDC

glycine dehydrogenase (decarboxylating) gene

IL33

Interleukin 33 gene

NFIB

Nuclear Factor I B gene

TPD52L3

Tumor Protein D52 Like 3 gene;

KDM4C

Lysine Demethylase 4C gene