Genome-wide association studies of survival in 1,520 cancer patients treated with bevacizumab-containing regimens

Abstract

Germline variants might predict cancer progression. Bevacizumab improves overall survival (OS) in patients with advanced cancers. No biomarkers are available to identify patients that benefit from bevacizumab. A meta-analysis of genome-wide association studies (GWAS) was conducted in 1,520 patients from phase III trials (CALGB 80303, 40503, 80405, and ICON7), where bevacizumab was randomized to treatment without bevacizumab. We aimed to identify genes and SNPs associated with survival independently of bevacizumab treatment or through interaction with bevacizumab. A cause-specific Cox model was used to test the SNP-OS association in both arms combined (prognostic), and the effect of SNPs-bevacizumab interaction on OS (predictive) in each study. The SNP effects across studies were combined using inverse variance. Findings were tested for replication in advanced colorectal and ovarian cancer patients from The Cancer Genome Atlas (TGCA). In the GWAS meta-analysis, patients with rs680949 in PRUNE2 experienced shorter OS compared to patients without it (p=1.02×10⁻⁷, HR=1.57, 95% CI:1.33–1.86), as well as in TCGA (p=0.0219, HR=1.58, 95% CI:1.07–2.35). In the GWAS meta-analysis, patients with rs16852804 in BARD1 experienced shorter OS compared to patients without it (p=1.40×10⁻⁵, HR=1.51, 95% CI:1.25–1.82) as well as in TCGA (p=1.39×10⁻⁴, HR=3.09, 95% CI:1.73–5.51). Patients with rs3795897 in AGAP1 experienced shorter OS in the bevacizumab arm compared to the non-bevacizumab arm (p=1.43×10⁻⁵). The largest GWAS meta-analysis of bevacizumab treated patients identified PRUNE2 and BARD1 (tumor suppressor genes) as prognostic genes of colorectal and ovarian cancer, respectively, and AGAP1 as a potentially predictive gene that interacts with bevacizumab with respect to patient survival.

Introduction

Germline variants can significantly improve predictions of the survival of cancer patients compared to clinical variables alone ¹. In addition to their prognostic potential, germline variants are also useful to individualize the selection of the most effective treatment ^2,3. This is clearly demonstrated by the example of deficient mismatch repair of germline DNA, which creates tumors responsive to checkpoint inhibition ⁴. Another example is tumors carrying BRCA1 germline variants, which are more responsive to poly(ADP-ribose) polymerase (PARP) inhibitors ⁵. When germline variants change the biology of the tumor phenotype or its responsiveness to therapy, new targets and/or changes in the tumor microenvironment increase the efficacy of novel therapies.

There are several anti-tumor therapies lacking validated biomarkers to select patients more responsive to treatment. One of them is bevacizumab, a humanized monoclonal antibody that binds to the vascular endothelial growth factor A (VEGF-A), a proangiogenic factor that plays an important role in tumor growth. Bevacizumab is approved by the U.S. Food and Drug Administration and the European Medicines Agency for the treatment of many tumor types, including metastatic colorectal cancer (CRC), metastatic breast cancer, advanced non-squamous lung cancer, metastatic renal cell carcinoma, recurrent glioblastoma, advanced cervical and ovarian cancers, and hepatocellular carcinoma ^6,7.

VEGF-A signaling is a key factor in angiogenesis, and the inhibition of the VEGF-A pathway impairs tumor angiogenesis, reducing tumor growth and distant metastatization. As angiogenesis is primarily a host-mediated process ^8,9, germline variation that affects the extent of angiogenesis might interfere with the efficacy of bevacizumab.

The identification of tumor-type agnostic markers to predict outcome in cancer patients is a challenge. However, pan-cancer studies have reported germline variants associated with outcome across different tumor types ^1,10. In addition, the identification of biomarkers that can indicate which patients would benefit or not from treatment with bevacizumab regardless of tumor type is even more challenging. Many studies have attempted to identify single nucleotide polymorphisms (SNPs) associated with overall survival (OS) in patients treated with bevacizumab, including a genome-wide association study (GWAS) ¹¹ and studies focused on VEGF-pathway genes (for a review ^12,13). However, for approved indications, there is no molecular marker that can identify the patients who are more or less likely to benefit from treatment with bevacizumab.

Using a GWAS meta-analysis of four phase III clinical trials of patients with different tumor types and randomized to treatment with bevacizumab, the objective of this study was to identify new genes, the genetic variation of which are associated with survival independently of treatment (prognostic). We also aimed at identifying new genes, the genetic variation of which interacts with bevacizumab with respect to patient outcome (predictive). We applied different approaches and methodologies to the newly discovered SNPs and genes, including, when available, external validation using data from The Cancer Genome Atlas (TCGA).

Materials and Methods

Patients and phase III clinical trials

The phase III clinical trials included in this study are from the Cancer and Leukemia Group B (CALGB, now part of the Alliance for Clinical Trials in Oncology, Alliance) and the Gynecologic Cancer InterGroup International Collaboration on Ovarian Neoplasms (ICON7). All studies were randomized, either bevacizumab versus placebo (CALGB 80303 and 40503), bevacizumab versus no bevacizumab (ICON7), or bevacizumab versus another targeted agent (cetuximab, CALGB 80405). All patients had either advanced or metastatic disease.

CALGB 80303 was conducted in advanced pancreatic cancer patients treated with gemcitabine and either 10 mg/kg bevacizumab or placebo. CALGB 40503 was conducted in hormone receptor-positive advanced-stage breast cancer patients treated with letrozole and either bevacizumab or placebo. CALGB 80405 was conducted in locally advanced or metastatic CRC patients treated with chemotherapy (FOLFIRI/FOLFOX) and either bevacizumab or cetuximab. ICON7 was conducted in advanced or high-risk early-stage ovarian cancer patients treated with carboplatin and paclitaxel with or without bevacizumab. Patient eligibility, characteristics, stratifications, and treatments have been previously described ^14–17.

The primary endpoint of this study is OS, defined as the time from trial registration (CALGB 80303), study entry (CALGB 80405, CALGB 40503), or randomization (ICON7) until death from any cause. Patients without reported deaths were censored at their last known follow-up.

Genotyping and quality control

Germline DNA was obtained from peripheral blood. The genotyping platforms used in each study are described in Table 1, and the number of SNPs included in the analysis in each study after quality control (QC) is described in Figure S1. Additional information on the QC procedures can be found in the publications of the GWAS of CALGB 80303 ¹⁸ and 80405 ¹⁹. The genotyping and QC for CALGB 40503 and ICON7 have not been reported before. In CALGB 40503, 964,058 SNPs were genotyped using the Illumina Human OmniExpressExome-8 BeadChip. Based on QC criteria, specifically call rates less than 95% (51,759) and minor allele frequencies (MAF) less than 0.05 (318,063), 369,822 SNPs were excluded from the analyses. The 594,236 SNPs passing QC in CALGB 40503 were included in the association analyses (Figure S1). In ICON7, 951,117 SNPs were genotyped using the Illumina Human OmniExpressExome-8 BeadChip. Based on QC criteria, specifically call rates less than 95% (4,634), MAF less than 0.05 (338,597), SNPs in chromosome 23 (15,748), and Hardy-Weinberg equilibrium with p-value <1.00×10⁻⁸ (323), 359,302 SNPs were excluded from the analyses. The 591,815 SNPs passing QC in ICON7 were included in the association analyses (Figure S1). Each SNP that passed QC and has genotype data for at least one study was tested for association.

Table 1. Clinical trials and GWAS in patients of European ancestry.

CALGB Cancer and Leukemia Group B; ICON7 International Collaboration on Ovarian Neoplasms; FOLFIRI 5-fluorouracil (400 mg/m² and 2400 mg/m²) + leucovorin (400 mg/m²) + irinotecan (180 mg/m²); mFOLFOX6 5-fluorouracil (400 mg/m² and 2400 mg/m²) + leucovorin (400 mg/m²) + oxaliplatin (85 mg/m²); SD standard deviation; OS overall survival; CI confidence interval; NR not reached; NE not estimated.

Clinical Trial		CALGB 80303		CALGB 40503		CALGB 80405		ICON7
Cancer type		Advanced pancreatic cancer		Hormone receptor-positive advanced-stage breast cancer		Locally advanced or metastatic colorectal cancer		Advanced or high-risk early-stage ovarian cancer
Genotype platform		Illumina HumanHap550-Quad		Illumina Human OmniExpressExome-8		Illumina HumanOmniExpress-12v1 and Illumina Human OmniExpressExome-8		Illumina Human OmniExpressExome-8
Treatment		Gemcitabine 1,000 mg/m2 on days 1, 8, and 15 plus either placebo or bevacizumab 10 mg/kg on days 1 and 15		Letrozole 2.5 mg orally/day plus either placebo or bevacizumab 15 mg/kg every 21 days		FOLFIRI or mFOLFOX6 every two weeks plus either cetuximab 400 mg/m2 on day 1, then 250 mg/m2 every week or bevacizumab 5 mg/kg every other week		Carboplatin AUC 5 or 6 and paclitaxel 286 mg/m2 or the same chemotherapy regimen plus bevacizumab 7.5 mg/kg every 3 weeks
Treatment arms		Bevacizumab (n=154)	Placebo (n=140)	Bevacizumab (n=105)	Placebo (n=100)	Bevacizumab (n=310)	Cetuximab (n=300)	Bevacizumab (n=212)	No Bevacizumab (n=199)
Age	Mean (SD)	64.1 (10.2)	63.6 (10.9)	56.9 (11.7)	58.5 (12.1)	58.5 (12.2)	59.3 (10.9)	54.6 (9.6)	57.7 (9.7)
Gender	Male (n, %)	90 (58.4)	69 (49.3)	0 (0.0)	0 (0.0)	119 (38.4)	112 (37.3)	0 (0.0)	0 (0.0)
	Female (n, %)	64 (41.6)	71 (50.7)	105 (100.0)	100 (100.0)	191 (61.6)	188 (62.7)	199 (100.0)	212 (100.0)
OS	Median (95% CI)	5.8 (4.9–7.01)	6.3 (5.2–7.9)	49.2 (40.0-NE)	46.0 (36.3–50.5)	29.4 (26.0–32.6)	30.1 (25.8–39.2)	NR (41.1-NE)	NR (38.8-NE)

Statistical analysis

The objectives of the study were to test the association between SNPs and OS in both arms combined in each trial (prognostic effect) and the interaction between SNPs and bevacizumab treatment with respect to OS in each trial (predictive effect). A GWAS meta-analysis was performed including previously published data (CALGB 80405 ¹⁹) and unpublished data (CALGB 80303, CALGB 40503, and ICON7). For the principal component analysis, the genetic ancestral origin of patients was estimated based on clustering by Eigenstrat ²⁰. Following linkage disequilibrium (LD) pruning using an LD threshold of R² 0.2, pruned genotype data was compared to genotype data for the set of pruned variants from the 1,000 Genomes Project Phase 3 samples, including individuals from ASW, CEU, CHB, MEX, and YRI populations. Samples within each dataset that grouped with 1,000 Genomes Project CEU samples based upon the first and second principal components were categorized as of genetically determined European ancestry.

A cause-specific Cox model was fitted in each trial to test the association between SNPs and OS. The analysis was adjusted by various covariates in each study: prior radiation, extent of disease, and performance status (CALGB 80303); disease-free interval and presence of measurable disease (CALGB 40503); country group and FIGO score (ICON7); protocol chemotherapy (FOLFOX or FOLFIRI), prior radiotherapy, prior adjuvant therapy, and primary tumor location (CALGB 80405). For CALGB 80405, primary tumor location was added due to its prognostic value ¹⁹. An additive genetic model was used.

For the prognostic effect, the inverse variance formula was used to combine the SNP effect in each study to obtain the meta-analysis estimate (β) of the SNP-OS association and its standard error (SE). The exponential function was used to calculate the hazard ratio (HR) for each SNP. The SE was applied in function exp(β−1.96*(SE)) to calculate the lower bound of 95% confidence interval (CI) of HR and in the function exp(β+1.96*(SE)) to calculate the upper bound of 95% CI of HR.

For the predictive effect, the same methodology was applied, with a Cox model fitted in each study and the inverse variance formula used to combine the SNP-treatment interaction effect to obtain the meta-analysis estimate (β) of the SNP-treatment association and its SE.

The heterogeneity across studies was examined by the Cochran’s Q test, and the SNPs reported as associated with OS were those with Cochran’s Q p-value >0.20. The p-values were not corrected for multiple comparisons, as the p-values were used to prioritize associations for future analyses, including external replication in TCGA.

External replication of SNPs associated with OS (prognostic) in TCGA

We sought to replicate the prognostic associations observed in the meta-analysis using the data from cancer patients in TCGA ²¹ as we have done before ¹⁹. Stage IV patients with either colon adenocarcinoma (COAD) and rectum adenocarcinoma (READ) (n=80) or ovarian serous cystadenocarcinoma (OV) (n=86) were included. Patients with stage IV breast invasive carcinoma (n=20) or stage IV pancreatic adenocarcinoma (n=5) were not included due to their small sample sizes.

The ten most statistically significant prognostic SNPs associated with OS with the same direction of effect (either reduced or increased OS) in at least three out of the four trials were selected for replication. If the SNP for testing was not genotyped in TCGA, a SNP in high LD with it (R²>0.8 in Europeans, LDlink ²²) was used as a proxy. The SNPs were tested for association with OS using a Cox model as described above. The analysis was adjusted by primary tumor location in COAD+READ patients. The inverse variance formula was used to combine the SNP effect in COAD+READ and OV patients to obtain the meta-analysis β estimate and SE to calculate HR and 95% CI as described above (p-value <0.05 was used as the threshold for statistical significance).

Bioinformatics and downstream analyses

To test the hypothesis that germline variants associated with outcome might act as expression quantitative trait loci (eQTL) for tumor infiltrating immune cells, we investigated selected SNPs using the Database of Immune Cell eQTLs (DICE) ²³. Moreover, because germline variants might change gene expression in non-tumoral tissues that interact with cancer cells, for example aorta, blood cells, and fibroblasts ²⁴, we have also used the data from The Genotype-Tissue Expression project (GTEx v8) ²⁵. In order to postulate the mechanisms by which the SNPs may be exerting their effect, SNPs were referenced in the RegulomeDB ²⁶ for functional inference. For those SNPs with experimentally-determined binding site motifs according to RegulomeDB ²⁶, AtSNP was used to quantify the impact of SNPs on transcription factor binding ²⁷.

The cBioPortal ²⁸ tool was used to obtain mRNA expression data of patients from TCGA. In TCGA, mRNA gene expression data was available for 51 stage IV COAD+READ patients and for 38 stage IV OV patients. The association between prognostic SNPs and mRNA expression (log2 RNA Seq V2) in patients with stage IV COAD+READ and OV from TCGA was tested by the Jonckheere-Terpstra test (patients having homozygous reference versus one or more alternate alleles, p-value < 0.05 was used for statistical significance).

Results

A total of 1,520 cancer patients of genetically determined European ancestry were genotyped and included in this study. The characteristics and demographics of patients are in Table 1. The Q-Q plots for the association between SNPs and OS show no evident deviation in distribution (Figure S2).

SNPs associated with OS in both arms (prognostic)

The Manhattan and Q-Q plots are shown in Figure S2. The ten most statistically significant SNPs that had the same direction of effect (either reduced or increased OS) in at least three out of the four studies are shown in Table 2. Variant rs680949 (A>G) in the intron 9 of PRUNE2 had the strongest evidence for association (p=1.02×10⁻⁷, HR=1.57, 95% CI 1.33–1.86) with a p-value <0.05 in each of the four studies (results in each study are shown in Table S1). Patients with the AG genotype of rs680949 (A>G) in PRUNE2 experienced significantly shorter OS than those with AA genotype (Figure 1A). The median (95% CI) OS (months) for the AG and AA genotypes in each CALGB study, respectively, was 4.9 (3.7–6.9) and 6.4 (5.7–7.2) in CALGB 80303, 36.3 (24.5-not estimated, NE) and 48.7 (42.2–51.3) in CALGB 40503, and 18.8 (12.9–27.6) and 31.1 (28.1–34.5) in CALGB 80405. Because the majority of genotypes did not reach median OS in ICON7, we report the 75^th percentile (95% CI) OS (months) for the AG and AA genotypes, which was 23.9 (17.1–29.2) and 33.1 (29.9–41.1), respectively. Due to the very small number of GG patients, their median or 75^th percentile (95% CI) OS (months) in each study is reported in Table 3.

Table 2. Results of the association between SNPs and OS as a prognostic effect and as a predictive effect. Ten most statistically significant SNPs associated with OS with same direction of effect (either reduced or increased OS), in at least three out of four trials.

Rows in bold are those with p-value<0.05 in at least three out of four trials. Ch chromosome, NA Intergenic SNP, MAF minor allele frequencies in CALGB 80303, 40503, 80405, and ICON7, respectively, HR hazard ratio, CI confidence interval. Direction of effect increasing (−) or decreasing (+) OS in CALGB 80303, 40503, 80405, and ICON7, respectively. “_” SNP not present in the genotyping platform of the trial.

Prognostic effect
SNP	Ch	Gene	Feature	5’ flanking	3’ flanking	Base change	MAF	Direction of effect	Effect size (β)	HR (95% CI)	p-value
rs680949	9	PRUNE2	intron	LOC392352	LOC392352	A>G	0.07/0.06/0.07/0.06	++++	0.45	1.57 (1.33– 1.86)	1.02x10 −7
rs218527	18	NA	NA	LOC100129774	LOC100128360	G>A	0.09/0.10/0.10/0.11	++++	0.39	1.48 (1.28–1.71)	1.83x10 −7
rs2795492	9	CORO2A	intron	TRIM14	TBC1D2	G>A	0.41/0.41/0.42/0.42	----	−0.22	0.81 (0.73–0.88)	5.41x10−6
rs476612	9	LOC100129762	intron	LOC392352	PRUNE2	A>C	_/0.06/0.07/0.06	_+++	0.52	1.68 (1.34–2.12)	8.45x10 −6
rs7615734	3	NA	NA	TPRG1	TP63	A>G	0.16/0.18/0.18/0.21	++++	0.26	1.29 (1.15–1.45)	9.58x10−6
rs6712389	2	NA	NA	KIAA1604	LOC729009	G>A	_/0.31/0.29/0.27	_---	−0.28	0.76 (0.67–0.86)	1.16x10−5
rs16852804	2	BARD1	intron	LOC100128203	ABCA12	G>A	0.06/0.10/0.07/0.06	++++	0.41	1.51 (1.25–1.82)	1.40x10−5
rs10950639	7	NA	NA	AGR3	LOC100131425	A>G	_/0.17/0.21/0.20	_+++	0.29	1.33 (1.17–1.52)	1.40x10−5
rs4807493	19	PIP5K1C	UTR-3’	C19orf29	TJP3	G>A	0.21/0.17/0.18/0.20	-+--	−0.25	0.78 (0.70–0.88)	2.26x10−5
rs2103445	8	NA	NA	CALB1	TMEM64	G>A	_/0.30/0.29/0.31	_+++	0.25	1.29 (1.15–1.45)	2.29x10−5
Predictive effect
SNP	Ch	Gene	Feature	5’ flanking	3’ flanking	Base change	MAF	Direction of effect	Effect size (β)		p-value
rs4969481	17	DCXR	near 3’	RAC3	DCXR	G>A	0.13/0.16/0.14/0.14	----	−0.70		1.18x10−7
rs11895736	2	LTBP1	intron	TTC27	RASGRP3	A>G	0.12/0.14/0.14/0.14	----	−0.68		6.85x10−7
rs448960	20	LOC100129869	near 3’	PMEPA1	LOC100129869	A>G	_/0.49/0.46/0.47	_---	−0.58		7.50x10−6
rs10763269	10	NA	NA	LOC389970	ZWINT	G>A	0.27/0.24/0.22/0.21	++++	0.48		1.09x10−5
rs13392750	2	NA	NA	ANKRD44	LOC729342	A>G	_/0.46/0.46/0.43	_+++	0.58		1.23x10−5
rs3795897	2	AGAP1	intron	LOC642692	LOC100130154	G>A	0.13/0.11/0.17/0.13	++++	0.68		1.43x10 −5
rs1150743	6	NA	NA	HLA-J	ETF1P1	G>A	0.11/0.11/0.10/0.10	++++	0.66		1.83x10−5
rs10915428	1	NA	NA	hCG_2036596	LOC644357	A>C	0.46/0.44/0.51/0.48	----	−0.39		1.86x10−5
rs11860804	16	NA	NA	LOC729847	DYNLRB2	G>A	_/0.49/0.50/0.49	_---	−0.49		1.87x10−5
rs6453031	5	RGNEF	intron	UTP15	ENC1	A>C	_/0.15/0.16/0.16	_---	−0.63		1.88x10−5

Fig. 1. Kaplan-Meier estimates of OS for (A) rs680949 in PRUNE2 (A>G, MAF 0.06–0.07) and (B) rs16852804 in BARD1 (G>A, MAF 0.06–0.10), prognostic effect.

CALGB Cancer and Leukemia Group B, ICON7 International Collaboration on Ovarian Neoplasms.

Table 3.

Median (95% CI) OS (months) – CALGB 80303, 40503, 80405 – and 75^th percentile (95% CI) OS (months) – ICON7 – for rs680949 in PRUNE2 (A>G, MAF 0.06–0.07) and rs16852804 in BARD1 (G>A, MAF 0.06–0.10), prognostic effect, and for rs3795897 in AGAP1 (G>A, MAF 0.11–0.17), predictive effect.

Prognostic effect: rs680949 in PRUNE2 (A>G, MAF 0.06–0.07)
Genotype	CALGB 80303 Median OS (95% CI) - n		CALGB 40503 Median OS (95% CI) - n		CALGB 80405 Median OS (95% CI) - n		ICON7 75th percentile OS (95% CI) - n
AA	6.4 (5.7–7.2) - 253		48.7 (42.2–51.3) - 183		31.1 (28.1–34.5) −527		33.1 (29.9–41.1) - 360
AG	4.9 (3.7–6.9) - 40		36.3 (24.5- NE) – 21		18.8 (12.9–27.6) - 81		23.9 (17.1–29.2) - 50
GG	12.0 (NE-NE) - 1		16.3 (NE-NE) - 1		32.2 (19.8-NE) – 2		NR (NE-NE) - 1
Prognostic effect: rs16852804 in BARD1 (G>A, MAF 0.06–0.10)
Genotype	CALGB 80303 Median OS (95% CI) - n		CALGB 40503 Median OS (95% CI) - n		CALGB 80405 Median OS (95% CI) - n		ICON7 75th percentile OS (95% CI) - n
GG	6.5 (5.8–7.2) - 263		48.7 (40.4–51.3) - 168		29.6 (26.7–33.0) - 537		32.3 (27.6–38.9) - 362
GA	3.8 (2.7–6.8) - 29		42.2 (35.0-NE) - 35		29.0 (19.3–36.4) - 70		38.8 (22.9-NE) - 45
AA	13.9 (0.4-NE) - 2		39.9 (17.9-NE) - 2		15.9 (6.11-NE) - 2		NR (11.1-NE) - 4
Predictive effect: rs3795897 in AGAP1 (G>A, MAF 0.11–0.17)
	CALGB 80303 Median OS (95% CI) - n		CALGB 40503 Median OS (95% CI) - n		CALGB 80405* Median OS (95% CI) - n		ICON7 75th percentile OS (95% CI) - n
Genotype	Placebo	Bevacizumab	Placebo	Bevacizumab	Cetuximab	Bevacizumab	No bevacizumab	Bevacizumab
GG	6.31 (5.22–7.88) −107	6.54 (4.90–7.33) - 118	46.6 (36.4-NE) −75	49.2 (40.4-NE)- 87	23.2 (18.7–28.8) - 66	31.7 (25.2–41.0) - 71	27.6 (24.0-NE) - 160	37.0 (33.1-NE) −157
GA	5.45 (3.78–9.66) - 30	5.06 (3.71–6.64) - 33	40.1 (35.8-NE) - 22	33.4 (22.2-NE) −16	41.1 (22.5–60.9) - 35	30.1 (11.2–39.7) −20	NR (1.91-NE) - 36	26.1 (21.1-NE) - 49
AA	6.08 (4.99–7.88) - 3	7.23 (2.83-NE) - 3	NR (35.5-NE) - 2	NR (22.7-NE) −2	45.9 (45.9-NE) - 4	22.0 (NE-NE) - 3	NR (NE-NE) - 3	17.4 (16.3-NE) - 6

^*genotyped only in patients from the GWAS-Batch 1 (n=199, Fig. S1). CALGB Cancer and Leukemia Group B, ICON7 International Collaboration on Ovarian Neoplasms; OS overall survival; CI confidence interval; NR not reached; NE not estimated.

Replication in TCGA of the SNPs associated with OS in both arms (prognostic)

Seven out of the ten SNPs in Table 2 had genotype data (same SNP or a proxy) in TCGA and were tested for replication in COAD+READ and OV patients (Table S2). Out of the seven SNPs, rs680949 in PRUNE2 and rs16852802 in BARD1 showed a p-value <0.05 in the TCGA meta-analysis (Table S2). Patients with the CC and TC genotypes in rs680949 in PRUNE2, genotyped in the opposite strand in TCGA (T>C in TCGA corresponding to A>G in the GWAS), experienced shorter OS than patients with the TT genotype (p=0.0219, HR=1.58, 95% CI 1.07–2.35) (Figure 2A). The median (95% CI) OS (months) for the CC, TC, and TT genotypes, respectively, was 8.07 (1.63-NE), 17.0 (14.1-NE) and 36.5 (27.4-NE) in COAD+READ and 4.67 (0.8-NE), 38.53 (22.7-NE) and 31.73 (27.4–56.2) in OV.

Fig. 2. Kaplan-Meier estimates of OS for (A) rs680949 in PRUNE2 (T>C, MAF 0.11–0.12) and (B) rs16852802 in BARD1 (A>G, MAF 0.02–0.08, complete LD, R²=1.0, with rs16852804) in TCGA patients.

It should be noted that the frequency of the AG genotype of rs16852802 in BARD1 (A>G) in COAD+READ patients is low (n=2). TCGA The Cancer Genome Atlas, COAD colon adenocarcinoma, READ rectum adenocarcinoma, OV ovarian serous cystadenocarcinoma.

Patients with the AG genotype of rs16852802 (A>G) in BARD1 experienced shorter OS than patients with the AA genotype (p=1.39×10⁻⁴, HR=3.09, 95% CI 1.73–5.51) (Figure 2B). The median (95% CI) OS (months) for the AG and AA genotypes, respectively, was 5.1 (1.63-NE) and 36.5 (27.4-NE) in COAD+READ and 16.0 (10.6-NE) and 38.5 (31.4–60.0) in OV. No patients with the GG genotype were found. Variant rs16852802 (A>G) in BARD1 is in complete LD (R²=1.0) with rs16852804 (G>A) genotyped in the GWAS, and A>G of rs16852802 corresponds to G>A of rs16852804. In the GWAS, patients with the AA and GA genotypes of rs16852804 (G>A) in BARD1 experienced significantly shorter OS than those with GG genotype (p=1.40×10⁻⁵, HR=1.51, CI 1.25–1.82) (Figure 1B, Table 3).

The association between prognostic SNPs in PRUNE2 and BARD1 and their tumor mRNA expression in TCGA was tested, and only the AG genotype of rs16852802 (A>G) in BARD1 was associated with lower BARD1 mRNA expression in OV patients compared to the AA genotype (p=0.0169, Figure S3).

Bioinformatic functional analyses of the rs680949 in PRUNE2 and rs16852804 in BARD1

These analyses show that rs680949 (T>C corresponding to A>G in the GWAS) in PRUNE2 is located within a binding motif for Myc and Max transcription factors in human embryonic stem cell lines, where T>C creates a binding site for the Myc transcription factor (p=0.006) ²⁷. The Myc transcription factor binds preferentially to the C than the T allele (p=0.0038, log-likelihood=−23.7 for the C allele, and p=0.094, log-likelihood=−46.7 for the T allele) (Figure S4) ²⁷.

The A allele of rs16852804 (G>A) decreases BARD1 expression in whole blood in GTEx (p=1.90×10⁻⁵, normalized effect size, NES=−0.25) ²⁵. The A allele of rs368962482 in BARD1 (G>A, complete LD, R²=1.0, with rs16852804) decreases BARD1 expression in non-classical monocytes in DICE (p=0.013, effect size=−1.3) (Figure S5) ²³.

SNP-bevacizumab interactions associated with OS (predictive)

The Manhattan and Q-Q plots are shown in Figure S2. The ten SNPs with the smallest interaction p-values that had the same direction of effect (conferring either reduced or increased OS in the bevacizumab arm versus the non-bevacizumab arm) in at least three out of the four studies are shown in Table 2. The only SNP in Table 2 with p-value <0.05 in three out of the four studies was rs3795897 (G>A, MAF 0.11–0.17) in AGAP1 (p=1.43×10⁻⁵) (details on the results of each study are shown in Table S1). The genotype-bevacizumab interaction shows that patients with the GA genotype of rs3795897 in AGAP1 tended to experience shorter OS in the bevacizumab arm compared to the non-bevacizumab arm. In contrast, patients with the GG genotype of rs3795897 in AGAP1 tended to experience longer OS in the bevacizumab arm compared to the non-bevacizumab arm. This is more evident in ICON7, where the 3-year survival for the GA and GG genotypes, respectively, was 57.1% and 81.5% in the bevacizumab arm, and 96.7% and 64.3% in the non-bevacizumab arm. The effect of the AA genotype is more difficult to detect due to its small sample size (Figure S6, Table 3).

Bioinformatics functional analysis of rs3795897 in AGAP1

Variant rs3795897 (G>A) is an eQTL for AGAP1-IT1, with the A allele significantly increasing gene expression in many different cell types in GTEx, the strongest of these effects being in cultured fibroblasts (p=7.8×10⁻¹⁴, NES=0.59) ²⁵.

Discussion

This GWAS meta-analysis of 1,520 cancer patients treated with bevacizumab-containing regimens has selected PRUNE2 and BARD1 as genes with potential prognostic significance. These findings were supported by validation in stage IV cancer patients from TCGA. Moreover, AGAP1 was selected as a potentially novel gene with predictive significance, interacting with bevacizumab with respect to OS. To the best of our knowledge, this is the largest GWAS meta-analysis of patients treated with bevacizumab. Here, we discuss the main findings from our studies in the context of available knowledge of the biology and genetics of these genes in oncology.

PRUNE2, prune homolog 2 with BCH domain, acts as a tumor suppressor gene through the suppression of Ras homolog family member A activity ^29,30. Variant rs680949 (A>G) in PRUNE2 was identified in this study as being prognostic of OS, with a p-value close to the cut-off of genome-wide significance (p-value=1.02×10⁻⁷) and a p-value <0.05 in each of the four studies. The AG genotype of rs680949 (A>G) in PRUNE2 was associated with shorter OS and was replicated in the meta-analysis of COAD+READ and OV patients from TCGA. However, considering COAD+READ and OV independently, the effect is not present in OV. Bioinformatics functional analysis showed that the G allele of rs680949 (A>G) in PRUNE2 increases the likelihood of Myc binding to PRUNE2 compared to the A allele ²⁷. Myc amplifies the expression of oncogenes ^31,32 and decreases the expression of tumor suppressor genes ³³. We then hypothesize that the G allele of rs680949 (A>G) in PRUNE2 might decrease the expression of PRUNE2 by increasing Myc binding. The reduced expression of PRUNE2 and its tumor suppressor activity might lead to reduced survival in stage IV CRC patients with the G allele of rs680949 (A>G) in PRUNE2.

This study also proposes BARD1 as a prognostic gene in stage IV ovarian cancer. BARD1, BRCA1-associated RING domain 1, acts as a tumor suppressor gene and its interaction with BRCA1 is essential for the tumor suppression activity of BRCA1 ³⁴. In this study, the AA and AG genotypes of rs16852804 (G>A) in BARD1 were associated with shorter OS. Although the p-value did not reach the cut-off of genome-wide significance (p-value=1.40×10⁻⁵), the association was also observed in COAD+READ and OV patients from TCGA. The AG genotype of rs16852802 (A>G, G allele equivalent to the A allele of rs16852804, G>A) in BARD1 was associated with lower BARD1 mRNA expression in OV but not in COAD+READ tumors. This finding is consistent with the tumor suppressor activity of BARD1, with a more prominent effect in ovarian cancer, where higher mRNA expression leads to a higher tumor suppressor activity and hence longer OS. Consistently, the GA genotype of rs16852804 (G>A) in BARD1 might be decreasing BARD1 expression in the tumor of patients in our study, which leads to a reduced tumor suppressor activity, contributing to the poor prognosis. Our analysis also showed that the A allele of rs16852804 (G>A) decreases the expression of BARD1 in whole blood, and the A allele of rs368962482 (G>A, in LD with rs16852804, R²=1.0) decreases the expression of BARD1 in non-classical monocytes. These findings indicate that genetic variation in BARD1 might be exerting its effect also through indirect, non-tumoral mechanisms. Non-classical monocytes are essential to maintain vascular homeostasis in endothelial cells ³⁵. When BRCA1 was knocked out in mice, an impaired endothelial function was observed ³⁶, suggesting an important role of BRCA1 and its regulation by BARD1 in the normal function of vascular endothelial cells. A dysfunctional endothelium promotes inflammatory signaling and aggressive properties of cancer cells ³⁷. Thus, we postulate an additional mechanism where the A allele of rs16852804 decreases the expression of BARD1 in non-classical monocytes and leads to a higher endothelial dysfunction and higher tumor aggressiveness, resulting in poor prognosis.

BARD1 germline mutations have been linked to the risk of ovarian cancer ^38–40. Most of them are exonic, but some intronic germline variants have been reported in BRCA1/2 negative high-risk breast and /or ovarian cancer patients, including rs71579844 (T>-, in complete LD with rs16852804, R²=1.0) and rs5031010 (T>G, in very high LD with rs16852804, R²=0.9) ³⁸. Germline variation of risk of sporadic cancer can be associated with worse patient survival ⁴¹. Moreover, the BARD1 coding region is sequenced in several mutation panels, such as Foundation Medicine (FoundationOne^® CDX) and MSK-IMPACT. Because rs16852804 in BARD1 is germline and intronic, BARD1 could be fully sequenced by including intronic regions in these panels. Orthogonal confirmation of the rs16852804 genotype in each patient will aid the collection of outcome data and response to therapy in larger cohorts in the future.

The discovery of new SNPs and genes that interact with a drug with respect to patient survival is much more challenging than the discovery of SNPs and genes with prognostic significance. We have provided evidence of AGAP1 as a potentially novel predictive candidate gene interacting with bevacizumab with respect to OS. In this study, patients with the GA genotype of rs3795897 (G>A) in AGAP1 experienced shorter OS in the bevacizumab arm compared to the non-bevacizumab arm, and patients with the GG genotype experienced longer OS in the bevacizumab arm compared to the non-bevacizumab arm (Tables 3, Figure S6). AGAP1 encodes a member of an ADP-ribosylation factor GTPase-activating protein family, and its role in the bevacizumab interaction with respect to patient survival needs to be further explored.

This study has some limitations. None of the SNPs reached the cut-off of genome-wide significance; however, the test for replication of the effect in TCGA patients and the concordance of effects in three out of the four studies potentiate the validity of these associations. The independent relative contribution of the effect of each variant in each tumor type cannot be ascertained in the TCGA datasets due to the small sample sizes and the relatively low frequency of these variants. Our study included patients with different tumor types and treatments, and the heterogeneity tests and the selection of SNPs based on a Cochran’s Q p-value >0.20 might mitigate the contribution of the confounding effects of these differences across studies. Only two out of the four studies were randomized to placebo, limiting the detection of predictive effects of SNPs that are specific to bevacizumab. Also, tumor molecular data are not always available to confirm some of our hypotheses. Still, we have applied different approaches using mRNA data from TCGA to circumvent this limitation. We did not perform imputation to these data, and this approach will be further applied to larger future studies aimed at discovering additional variants. Lastly, the mechanistic hypotheses proposed in this study should be evaluated in experimental models.

In conclusion, we have identified genes and new germline variants associated with prognosis in cancer patients. Two germline variants were also associated with prognosis in cancer patients from TCGA, rs680949 (A>G) in PRUNE2 in CRC cancer patients and rs16852804 (G>A) in BARD1 in ovarian cancer patients. Furthermore, we have also identified rs3795897 (G>A) in AGAP1 as a potentially new predictive germline variant that should be further evaluated as a marker affecting the response of patients to bevacizumab. Prospective testing of the association of these germline variants with survival should be performed to translate these findings to clinical practice.

Novelty and Impact.

We have provided evidence of potentially prognostic germline variants in PRUNE2 and BARD1 (tumor suppressor genes) in advanced colorectal and ovarian cancer, respectively. We have also identified a potentially predictive germline variant in AGAP1 that should be further evaluated as a marker affecting the response of patients to bevacizumab.

Abbreviations

CALGB

Cancer and Leukemia Group B

CI

confidence interval

COAD

colon adenocarcinoma

CRC

colorectal cancer

DICE

Database of Immune Cell eQTLs

eQTL

expression quantitative trait loci

GTEx

Genotype-Tissue Expression project

GWAS

genome-wide association study

HR

hazard ratio

ICON7

International Collaboration on Ovarian Neoplasms

LD

linkage disequilibrium

NE

not estimated

NES

normalized effect size

OS

overall survival

OV

ovarian serous cystadenocarcinoma

PARP

poly(ADP-ribose) polymerase

QC

quality control

READ

rectum adenocarcinoma

SE

standard error

SNPs

single nucleotide polymorphisms

TCGA

The Cancer Genome Atlas

VEGF-A

vascular endothelial growth factor A

Data availability statement

The GWAS summary statistics are available through the NHGRI-EBI GWAS Catalog under study accession numbers GCST90020034 and GCST90020035. Further details and other data that support the findings of this study are available from the corresponding author upon request.