Assessment of hepatocyte growth factor in ovarian cancer mortality

Abstract

Background

Invasive ovarian cancer is a significant cause of gynecologic cancer mortality.

Methods

We examined whether this mortality was associated with inherited variation in ~170 candidate genes/regions (993 SNPs) in a multi-stage analysis based initially on 312 Mayo Clinic cases (172 deaths). Additional analyses used The Cancer Genome Atlas (TCGA; 127 cases, 62 deaths). For the most compelling gene, we immunostained Mayo Clinic tissue micro-arrays (TMAs, 326 cases) and conducted consortium-based SNP replication analysis (2,560 cases, 1,046 deaths).

Results

The strongest initial mortality association was in HGF (hepatocyte growth factor) at rs1800793 (HR 1.7, 95% CI 1.3–2.2, p=2.0×10⁻⁵) and with overall variation in HGF (gene-level test, p=3.7×10⁻⁴). Analysis of TCGA data revealed consistent associations (e.g., rs5745709 [r²=0.96 with rs1800793]: TCGA 2.4, 1.4–4.1, p=2.2×10⁻³; Mayo Clinic+TCGA 1.6, 1.3–1.9, p=7.0×10⁻⁵) and suggested genotype correlation with reduced HGF mRNA levels (p=0.01). In Mayo Clinic TMAs, protein levels of HGF, its receptor MET, and phospho-MET were not associated with genotype and did not serve as an intermediate phenotype; however, phospho-MET was associated with reduced mortality (p=0.01) likely due to higher expression in early-stage disease. In eight additional ovarian cancer case series, HGF rs5745709 was not associated with mortality (1.0, 0.9–1.1, p=0.87).

Conclusions

We conclude that although HGF signaling is critical to migration, invasion, and apoptosis, it is unlikely that genetic variation plays a major role in ovarian cancer mortality; any minor role is not related to genetically-determined expression.

Impact

Our study demonstrates the utility of multiple data types and multiple datasets in observational studies.

INTRODUCTION

In the United States, ovarian cancer is the fifth leading cause of cancer death among women (1). Despite clinical responses to combination platinum/taxane-based chemotherapy in most women after surgical debulking, five-year overall survival lingers around 30% and, even with modern chemotherapy, most cases with advanced disease relapse and die of ovarian cancer (2, 3). Inherited variation may influence outcome. For example, cases with rare germline BRCA1 or BRCA2 mutations have improved chemoresponsiveness and survival (4). Common inherited variants may also be prognostic. Notably, we and others have reported evidence for a role of inherited variation in angiogenesis and inflammation genes in ovarian cancer survival (5–7). As initial ovarian cancer genome-wide association studies have not identified common mortality-associated alleles (8), in-depth analysis of additional candidate genes in key biological pathways holds promise for the identification of factors with functional relevance or prognostic utility.

Here, we examined key candidate genes encoding angiogenesis factors (9, 10) mitotic kinases (11), growth stimulatory mediators and stromal factors (12, 13), as well as genes and regions suggested by expression studies (14) and genome-wide association studies (15–17). We first evaluated the association between mortality and inherited single nucleotide polymorphism (SNP) variation among invasive epithelial ovarian cancer patients seen at the Mayo Clinic, and we pursued key findings via analysis of data from The Cancer Genome Atlas (TCGA). We then conducted expression analysis of tissue micro-arrays (TMAs) made from tumors of Mayo Clinic cases, and we examined genetic association in cases from eight additional ovarian cancer case series. In total, a multi-faceted approach integrating tumor and replication studies aimed to provide observational and functional insight into the role of SNPs in ovarian cancer mortality.

METHODS

Candidate Gene Analysis

Initial Study Participants

Recruitment of cases from Mayo Clinic’s gynecologic surgery and medical oncology departments (MAY1) used established protocols approved by the relevant Institutional Review Board (IRB) (5). All participants gave written informed consent. Eligible cases were women aged 20 years or older living in MN, IA, WI, IL, ND, or SD and ascertained within one year of a diagnosis of pathologically-confirmed primary invasive epithelial ovarian cancer. Between December 1999 and March 2006, 328 cases were enrolled; median time from diagnosis to recruitment was five days. Data on vital status through July 31, 2009 was obtained from the National Death Index, computerized medical records, and the Mayo Clinic Cancer Registry which annually follows cases diagnosed or receiving initial treatment at Mayo Clinic. Death certificates were available on 95 of 172 deceased cases, and dates of death were 94.7% concordant with dates obtained via registries (five certificates differed by a median of three days). Of 140 living cases, nine were lost to follow-up more than two years prior. Data on clinical features of disease including histology, surgical outcome, and chemotherapy were abstracted by experienced research nurses with review by gynecologic and medical oncologists. DNA was extracted from 10 to 15 mL fresh peripheral blood using the Gentra AutoPure LS Purgene salting out methodology (Gentra, Minneapolis, MN) and stored at −80°C; samples were bar-coded to ensure accurate processing and plated with duplicates and lab standards. We excluded 12 sequence-confirmed BRCA1 and BRCA2 mutation carriers and four cases with predicted non-European ancestry (Supplemental Figure 1) (18, 19), resulting in 312 analyzed cases (Supplemental Table 1).

Polymorphisms and Genotyping

Key genes within angiogenesis, mitosis, growth and stromal factors, as well as expression-based genes, and genes in key chromosomal regions were identified using published literature (9–17) and the Kyoto Encyclopedia of Genes and Genomes (Supplemental Table 2) (20). For angiogenesis genes, all SNPs with minor allele frequency (MAF) ≥ 0.05 were selected; in 8q24 and 9p24, a combination of region-tagging, gene-tagging, and replication-based SNPs were included (21), and for remaining genes, we selected tagSNPs within 5 kb with MAF ≥ 0.05 based on European linkage disequilibrium (LD) in HapMap v. 22 (22, 23) (r² ≥ 0.8; expression-based genes r² ≥ 0.9). Genotyping of 897 samples (including analyzed cases, borderline cases, population controls, and 26 duplicates and laboratory standards including CEPH trios) was performed at the Mayo Clinic using the Illumina GoldenGate BeadArray Assay (24). Of 871 unique MAY1 participants, only one sample failed (call rate < 90%) which was from a control; thus, genotype data were available on all cases. We assessed departures from Hardy Weinberg equilibrium (HWE) using Pearson goodness-of-fit and Fisher exact tests using data from self-reported European-American controls. Manual review of the plots was performed in GenomeStudio (Illumina, San Diego CA) to verify optimal SNP clustering, using both replicate and inheritance information from the CEPH family trio. When data failed to reveal one to three distinct clusters (representing AA, AB and BB genotypes), a SNP was failed. Of 1,152 SNPs attempted (Supplemental Table 3), 25 failed including 15 with call rate < 90%, nine with poor clustering, and one with an unresolved replicate or Mendelian error. We excluded SNPs with MAF < 0.05 (N=123) or HWE p-value < 0.001 (N=11), leaving 993 SNPs in 168 genes. LD plots were created using Haploview v. 4.2 (25) (Supplemental Figure 2).

Statistical Methods

We used Cox proportional hazards regression (26) to estimate hazard ratios (HRs) and 95% confidence intervals (CIs) for association with mortality, adjusted for clinical covariates. We accounted for left truncation using the start-stop counting process style of input within the Cox regression framework; thus, cases did not contribute follow-up time until date of enrollment (27). For each SNP, HRs with 95% CIs were estimated per-allele (i.e., 0, 1, or 2 copies of minor allele), analogous to the Armitage test for trend for binary endpoints (28). As our aim was to evaluate the role of SNPs beyond known prognostic factors, we included as covariates the following clinical variables which were univariately associated with mortality in a stepwise Cox regression model (p < 0.05): age at diagnosis, pre-surgical log₁₀(CA125), tumor stage (I, II, III, IV, unknown), tumor grade (grade 1 or 2, grade 3, grade 4, unknown), volume of residual tumor following debulking surgery (≤ 1 cm, > 1 cm, unknown), and laterality of tumor (right, left, bilateral, unknown). Unadjusted HRs with 95% CIs were also estimated (Supplemental Table 4). The proportional hazards assumption was evaluated using scaled Schoenfeld residuals (29) for covariates and SNPs with p-values < 0.05. We conducted gene-level analyses by testing principal components (30) that explained 90% of SNP variance using multiple degree-of-freedom likelihood ratio tests. Analyses were performed on all MAY1 cases as well as on a subset of 192 cases with serous histology. Due to the exploratory nature of all analyses, no correction for multiple testing was performed.

Analysis of Top Hits using Public Data

For SNPs with p < 0.001, we analyzed publicly-available germline genotype and mortality data on white non-Hispanic invasive serous ovarian cancer cases (TCGA1, N=127, 62 deaths) from TCGA (31). These cases were enrolled at seven sites and genotyped with the Illumina 1MDuo panel; Mayo Clinic TCGA cases were excluded. Call rates ranged from 95.7% to 98.3%. As above, Cox proportional hazards regression assuming an ordinal model was used to assess risk of death associated with genotype at each SNP. Analyses were adjusted for study site, age at diagnosis, stage (II, III, IV, unknown), and grade (grade 2, grade 3 or 4, unknown); combined analyses of MAY1 and TCGA1 were adjusted for study site, age, stage (I, II, III, IV), and grade (grade 1 or 2, grade 3 or 4, unknown). Where results were consistent, we obtained TCGA data on additional SNPs within each gene as well as data on tumor mRNA levels acquired via the Affymetrix Human Genome U133 GeneChip Array (using the probe for most the highly-expressed transcript based on prior reports in normal ovarian tissue (32)). Genotype was correlated with tumor mRNA levels using an ANOVA comparison of means test in the R software program (33), and tumor mRNA levels were examined in relation to survival using Cox proportional hazards models adjusted for study site, age at diagnosis, stage (II, III, IV, unknown), and grade (grade 2, grade 3 or 4, unknown).

HGF Tissue Microarray Expression Analysis

Study Participants and Tissue Blocks

TMAs were created from formalin-fixed paraffin-embedded tumors of 326 Mayo Clinic cases enrolled through January 2009. All participants provided written informed consent for an IRB-approved protocol. We used an automated Beecher Instruments ATA-27 arrayer (Reutlingen, Germany) following gynecologic pathologist review indicating tumor location. Three 0.6 mm cores were removed from each case paraffin block and placed in a recipient paraffin block according to a randomized electronic TMA map. Recipient blocks were sliced into 5 µm sections and mounted on charged slides. Characteristics of cases with and without protein expression and HGF genotype data are shown in Supplemental Table 5.

Antibodies, Immunohistochemistry, and Scoring

TMA slides were immunostained with primary antibodies recognizing HGF (polyclonal H-145, Santa Cruz; Santa Cruz, CA), MET (monoclonal clone 3D4 Zymed-Invitrogen; Carlsbad, CA), and phospho-MET (polyclonal Y1003) Upstate/Millipore; Charlottesville, VA) after optimizing staining conditions on positive control tissues colorectal cancer and normal liver (34). Negative controls included a non-specific isotype match (for MET) or rabbit IgG (HGF and pMet); conditions were appropriate in that no background staining was observed without the specific primary antibodies. Slides were prepared according to manufacturer’s instructions (Dako; Carpinteria, CA), and digital images were created as described previously (35). Imaging instrument: Bacaus Laboratories Inc. Slide Scanner (BLISS) system, utilizing a Zeiss Axioplan 2 microscope; 20× objective magnification was used during scanning; however, provided as 10× via Figure 1; Tracer Version 0.91 acquisition software was used and no additional techniques done; temperature during scanning was approximately 23C; no oil immersion was used, therefore imaging medium was dry; and no fluorochromes were use. The viewing software used: WebSlide Enterprise, utilizing ActiveX in conjunction with the Slide Tray v5.0 program. The saved format for the scanned pictures is SlideScan.ini, which means that, for these particular core scans, each original image actually is composed of twelve 752×480 pixel jpg scans. Extent was defined as the percentage of epithelial tumor cells staining positive for each antigen (negative, 0%–10% of cells expressing; positive, >10% of cells expressing), and intensity was defined as the strength of antibody staining in epithelial tumor cells (absent, weak, moderate, strong). Slides were scored by two reviewers, and discrepancies were resolved by a gynecologic pathologist. Extent and intensity measures for each core were combined as negative (extent negative), weak (intensity weak; Figure 1), moderate (intensity moderate), and strong (intensity strong; Figure 1), and the strongest protein expression value over the multiple cores for each case was used.

Figure 1. HGF, MET, and phospho-MET staining in epithelial ovarian cancer TMAs.

Immunohistochemical staining of a tissue microarray of epithelial ovarian cancer samples. Representative images are shown of weak (panels A, C, and E) and strong (B, D, and F) staining. Panels A and B show cores stained with antibodies recognizing HGF; Panels C and D represent MET staining, and Panels E and F represent phospho-MET staining.

Statistical Methods

We assessed associations of HGF genotype with TMA-based protein expression values using linear regression analyses (genotype as exposure and expression value as outcome). We examined associations of protein expression values with mortality using Cox proportional hazards regression analyses, accounting for left truncation as described above (27). HRs and 95% CIs were calculated for each antibody, modeling negative or weak protein expression as the referent group, and, as with genotype analyses, models were fit both unadjusted and adjusted for potential confounding variables. We evaluated whether protein expression values played a role in the association between genotype and mortality by fitting five sets of Cox regression models with genotype as exposure: unadjusted, adjusted for HGF, adjusted for MET, adjusted for phospho-MET, and adjusted for all three protein expression values. Attenuation of the genotype HR estimate due to adjustment for protein expression value(s) was considered evidence that genotype may associate with mortality via expression. All statistical tests were two-sided, and all analyses were carried out using SAS (SAS Institute, Cary, NC).

HGF Replication Analysis

Study Participants and SNP Genotyping

White non-Hispanic invasive epithelial ovarian, fallopian tube, and peritoneal cancer cases from additional independent cases collections (SRO, MAL, BEL, LAX, PVD, BAV, and MAY2) were genotyped at HGF rs5745709 and rs2074725 (a highly-correlated SNP) using germline DNA obtained from peripheral blood lymphocytes. As above, protocols were approved by each IRB, and all participants provided written informed consent. Two studies, BAV and BEL, included prevalent cases thus absolute survival was relatively high and median survival could not be estimated; nonetheless, appropriately accounting for the left-truncated nature of the data enabled analysis (Supplemental Figure 3). Genotyping used a custom Sequenom iPLEX MassArray multiplex assay at the Queensland Institute for Medical Research and included at least eight study duplicates per 384-well plate. Quality control was reviewed for each study requiring ≥ 98% concordance, ≥ 90% call rate for each 384-well plate, ≥ 95% call rate overall, and HWE p-value > 0.05. For each questionable quality control metric, cluster plots were reviewed to ensure appropriate genotype calling; SNPs with appropriate genotype calling showed one to three distinct clusters. In addition, new genotype data on additional invasive serous cases enrolled in the TCGA (TCGA2), as described above, became available and was used. Additional detail on these participants (2,560 cases, 1,046 deaths) is provided in Supplemental Table 1 and Supplemental Table 6.

Statistical Methods

Associations of genotypes with mortality were examined using Cox proportional hazards regression as above, accounting for left-truncation (27) using an ordinal (log-additive) genotypic effect. Analyses were conducted adjusted for study site, age at diagnosis, grade, and stage, as well as for study site and age only. Replication data were analyzed separately as well as combined with initial data (MAY1, TCGA1). We examined heterogeneity in genetic associations across study site by fitting and testing genotype*site interaction terms. As above, all analyses were carried out for all cases, and then subset to high-grade serous cases; all statistical tests were two-sided, and all analyses were carried out using SAS (SAS Institute, Cary, NC).

RESULTS

Characteristics of 312 Mayo Clinic invasive epithelial ovarian cancer cases genotyped in a large-scale candidate gene screen (MAY1) are shown in Supplemental Table 1; 172 cases (55%) died during the study period, including 129 cases with serous subtype. Examination of associations between SNP genotypes and mortality among these cases revealed minor alleles at 29 SNPs in ten genes with p < 0.01 among either case group (all cases or serous subtype only) including 21 SNPs with r² < 0.9 (Table 1). Gene-level analyses for four genes yielded p < 0.01. There was no violation of the Cox proportional hazards assumption for covariates or SNPs. At HGF (encoding hepatocyte growth factor) rs1800793, the minor allele was associated with increased risk of death (all cases: HR 1.69, 95% CI 1.33–2.16, p=2.0 × 10⁻⁵), representing the most statistically-significant single SNP result. Additional correlated SNPs (r² > 0.9, Supplemental Figure 2) showed this association, as did the less-correlated rs2214825 (r²=0.76; all cases: HR 1.44, 95% CI 1.13–1.84, p=3.3 × 10⁻³). Gene-level analyses using principal components also detected an association of HGF with mortality (all cases, p=3.7 × 10⁻⁴) representing the most statistically-significant gene-level test. Analysis of 12 common HGF haplotypes suggested that the individual SNP results were not a reflection of specific haplotypes (data not shown).

Table 1.

SNPs associated with mortality among Mayo Clinic ovarian cancer cases (MAY1)

				All Cases (N=312)		Serous Subtype (N=192)
Gene/Region	Chr	bp (distance)a	rsidb	HR (95% CI)c	p-trend	HR (95% CI)c	p-trend
Angiogenesis
RAF1	3	12,622,547	rs17771249	0.80 (0.63 – 1.00)	0.05	0.68 (0.52 – 0.89)	5.0 × 10−3
PLG	6	161,092,932	rs783173	1.24 (0.99 – 1.54)	0.06	1.57 (1.21 – 2.04)	8.1 × 10−4
HGF	7	81,184,621	rs1800793	1.69 (1.33 – 2.16)	2.0 × 10−5	1.63 (1.20 – 2.21)	1.8 × 10−3
		+24,934	rs2214825	1.44 (1.13 – 1.84)	3.3 × 10−3	1.39 (1.03 – 1.88)	0.03
PLAU	10	75,342,967	rs2227562	1.30 (0.98 – 1.72)	0.07	1.62 (1.17 – 2.26)	4.0 × 10−3
THBS1	15	37,665,559	rs753598	1.50 (1.07 – 2.10)	0.02	1.73 (1.16 – 2.56)	6.5 × 10−3
Mitosis
PRKACB	1	84,307,885	rs12031680	1.26 (1.00 – 1.59)	0.05 1.47	1.47 (1.11 – 1.94)	7.9 × 10−3
		+21,226	rs12405120	0.70 (0.55 – 0.89)	4.2 × 10−3 0.60	0.60 (0.45 – 0.81)	6.7 × 10−4
		+43,990	rs1402694	1.51 (1.20 – 1.92)	5.6 × 10−4	1.67 (1.26 – 2.21)	3.2 × 10−4
		+28,142	rs12129768	1.68 (1.18 – 2.38)	3.7 × 10−3	2.03 (1.32 – 3.13)	1.3 × 10−3
NEK2	1	209,909,699	rs697003	0.74 (0.59 – 0.93)	9.0 × 10−3	0.74 (0.57 – 0.96)	0.02
CUL1	7	148,098,380	rs3807446	0.42 (0.22 – 0.79)	6.8 × 10−3	0.55 (0.27 – 1.11)	0.10
WEE1	11	9,541,835	rs7929469	1.64 (1.15 – 2.34)	6.4 × 10−3	1.91 (1.26 – 2.92)	2.5 × 10−3
KIAA0999	11	116,227,036	rs10047459	1.50 (1.13 – 1.99)	5.2 × 10−3	1.27 (0.90 – 1.81)	0.18
		+41,208	rs499910	1.53 (1.18 – 1.99)	1.5 × 10−3	1.37 (1.00 – 1.87)	0.05
DCTN5	16	23,589,772	rs12447304	1.83 (1.29 – 2.60)	7.1 × 10−4	1.71 (1.13 – 2.58)	0.01
Growth Factors, Stromal Proteins
PIK3R1	5	67,605,502	rs171649	0.79 (0.62 – 1.00)	0.05	0.61 (0.45 – 0.83)	1.8 × 10−3
SHMT2	12	55,909,088	rs7301155	0.82 (0.63 – 1.05)	0.11	0.67 (0.49 – 0.90)	9.0 × 10−3
ERBB2	17	35,119,531	rs1810132	0.70 (0.55 – 0.90)	4.9 × 10−3	0.68 (0.51 – 0.92)	0.01
Expression-Based
CC2D1A	19	13,905,287	rs2305778	1.50 (0.94 – 2.39)	0.09	1.93 (1.17 – 3.18)	9.6 × 10−3
Regional
8q24	8	128,530,789	rs10094059	0.66 (0.50 – 0.87)	3.7 × 10−3	0.72 (0.52 – 1.00)	0.05

^abp on genome build 36.3; within genes, distance is represented in bp from previously-listed SNP (+bp).

^b29 SNPs yielded per-allele p < 0.01 in either case group; only 21 SNPs with r² < 0.90 are shown.

^cHazard ratios (HR) represent the estimated increase (or decrease) in the risk of death for each one-copy increase in the number of minor alleles carried by an individual adjusted for age at diagnosis, pre-surgical log₁₀(CA125), stage (I, II, III, IV, unknown), grade (grade 1 or 2, grade 3, grade 4, unknown), volume of residual tumor following debulking surgery (≤ 1 cm, > 1 cm, unknown), and laterality of tumor (right, left, bilateral, unknown).

Three other genes had SNP p-values < 0.001 including PRKACB (encoding protein kinase, cAMP-dependent, catalytic, beta) with rs1402694 (all cases HR 1.51, 95% CI 1.20–1.92, p=5.6 × 10⁻⁴; serous subtype HR 1.67, 95% CI 1.26–2.21, 3.2 × 10⁻⁴). Genotypes at DCTN5 (dynactin 5 (p25)) rs12447304 were associated with mortality among all cases (HR 1.83, 95% CI 1.29–2.60, p=7.1 × 10⁻⁴), thus we also conducted regional analysis with neighboring gene PLK1 and found that in combination, DCTN5 and PLK1 SNPs were associated with differential survival (all cases p=1.6 × 10⁻³; serous subtype 4.4 × 10⁻³). PLG (encoding plasminogen) rs783173 was not associated with survival among all cases, but an association was observed with serous subtype (HR 1.57, 95% CI 1.21–2.04, p=8.1 × 10⁻⁴) (Table 1). Additional genes with suggestive SNP associations (p < 0.01) are listed in Table 1, and full gene-level results are provided in Supplemental Table 7.

Using publicly-available data on 127 genotyped serous invasive cases of the TCGA (TCGA1, 62 deaths), we analyzed SNPs with p < 0.001 in MAY1 analysis. Results were not consistent at PRKACB rs1402694 (HR=1.13, 95% CI 0.76–1.67, p=0.54) or at PLG rs783173 (HR=0.70, 95% CI 0.48–1.00, p=0.05); no genotypes were available at DCTN5 rs12447304. However, at several HGF SNPs with MAY1 p < 0.001, compelling, consistent associations with mortality were observed including rs5745709 (HR 2.36, 95% CI 1.36–4.09, p=2.2 × 10⁻³) and the modestly-correlated rs2214825 (r²=0.72; HR 2.27, 95% CI 1.38–3.72, p=1.2 × 10⁻³; Table 2). Combining these Mayo Clinic and TCGA data (MAY1+TCGA1, N=439), the strongest association with mortality was observed for rs5745709 (HR 1.56, 95% CI 1.25–1.94, p=7.0 × 10⁻⁵, Table 2, Figure 2); this SNP is highly correlated with the initially most-significant SNP in MAY1, rs1800793 (r²=0.96; Supplemental Figure 2), which was not genotyped in TCGA. Similar results were observed when analyses were performed only on cases with advanced-stage disease (data not shown). To evaluate a possible mechanism for this association, we then explored HGF ovarian tumor mRNA levels in TCGA1 cases. Carriers of minor alleles at HGF SNPs had reduced HGF mRNA levels (e.g., rs2214825 p=0.03; Supplemental Figure 4). Although sample size was small and HGF mRNA levels themselves were not directly associated with mortality (HR=0.39, 95% CI 0.02–3.79, p=0.52), these results suggested further tumor and germline analysis of HGF in ovarian cancer mortality.

Table 2.

HGF SNPs and mortality in initial analyses

			MAY1 (N=312)			TCGA1 (N=127)			MAY1+TCGA1 (N=439)
rsid	r2 to next	alleles	MAF	HR (95% CI)	p-trend	MAF	HR (95% CI)	p-trend	HR (95% CI)	p
rs2074725	0.96	C>A	0.20	1.59 (1.24 – 2.04)	2.1 × 10−4	0.17	1.63 (1.05 – 2.55)	0.03	1.47 (1.19 – 1.81)	2.9 × 10−4
rs1800793	0.97	G>A	0.20	1.69 (1.33 – 2.16)	2.0 × 10−5	---	---	---	---	---
rs5745720	0.99	A>G	0.20	1.63 (1.28 – 2.08)	9.0 × 10−5	0.17	1.87 (1.11 – 3.15)	0.02	1.52 (1.23 – 1.88)	1.3 × 10−4
rs5745709	0.93	G>A	0.19	1.66 (1.29 – 2.13)	9.0 × 10−5	0.17	2.36 (1.36 – 4.09)	2.2 × 10−3	1.56 (1.25 – 1.94)	7.0 × 10−5
rs2887069	0.01	A>G	0.19	1.59 (1.24 – 2.03)	2.1 × 10−4	---	---	---	---	---
rs5745687	0.01	G>A	0.06	0.78 (0.48 – 1.25)	0.3	---	---	---	---	---
rs2214825	---	G>A	0.22	1.44 (1.13 – 1.84)	3.3 × 10−3	0.20	2.27 (1.38 – 3.72)	1.2 × 10−3	1.44 (1.17 – 1.78)	6.3 × 10−4

r² based on LD in MAY1 cases, the first five SNPs listed were in strong LD (r² > 0.92) and in modest LD with rs2214825 (r² > 0.76), rs5745687 was independent (r²=0.01); alleles column indicates major>minor allele; MAY1 analyses adjusted for age at diagnosis, pre-surgical log₁₀(CA125), stage (I, II, III, IV, unknown), grade (grade 1 or 2, grade 3, grade 4, unknown), volume of residual tumor following debulking surgery (≤ 1 cm, > 1 cm, unknown), and laterality of tumor (right, left, bilateral, unknown); TCGA1 analyses adjusted for age at diagnosis, grade (grade2, grade 3, unknown), stage (II, III, IV, unknown), and study site; combined analysis adjusted for study site, age at diagnosis, grade (grade 1 or 2, grade 3 or 4, unknown ), and stage (I, II, III, IV, unknown); hazard ratios (HR) represent the estimated increase (or decrease) in the risk of death for each one-copy increase in the number of minor alleles carried by an individual; sorted by position on genome build 36.3.

Figure 2. Kaplan-Meier ovarian cancer survival curves by HGF rs5745709 genotype, initial Mayo Clinic and TCGA analysis.

Genotype-specific Kaplan-Meier survival curves based on (a) MAY1, (b) TCGA1, and (c) MAY1+TCGA1. Numbers superimposed on curves represent genotype-specific cumulative number of deaths/number remaining at risk

We therefore performed immunohistochemical analysis of HGF, MET, and phospho-MET on TMAs created from 326 Mayo Clinic cases. We examined these three antigens because HGF binds to the transmembrane tyrosine kinase receptor MET, which, in turn, results in conversion of MET to its activated form, phospho-MET. This signaling initiates the ERK1/2, PI3K/AKT, and p38 mitogen-activated protein kinase (MAPK) cascades resulting in regulation of cell proliferation, apoptosis, migration, and invasive growth in normal ovarian surface epithelium and in ovarian cancers (36). Representative immunostained cores are provided in Figure 1. The distribution of protein expression values is provided in Table 3; as expected, there were no cases with activated phospho-MET in the absence of MET (data not shown). We assessed whether protein expression values were associated with mortality and found a suggestion that stronger phospho-MET protein expression correlated with decreased mortality (p=0.01; moderate v negative/weak HR 0.77, 95% CI 0.39–1.52; strong v negative/weak HR 0.55, 95% CI 0.27–1.10; Table 3). As known biology predicts that stronger protein expression of phospho-MET would correlate with increased (rather than decreased) mortality, we examined additional clinical characteristics and observed stronger protein expression of phospho-MET among cases with early-stage disease (p=0.003); inclusion of stage in phospho-MET Cox models attenuated the HR. Thus, the association between phospho-MET and decreased mortality is likely driven by an inverse association between phospho-MET and stage. In 255 cases with genotypes (MAY1 or MAY2), we evaluated whether rs5745709 genotype was associated with HGF, MET, and phospho-MET protein expression values, and no significant associations were observed (p ≥ 0.41 Table 3), although trends were consistent with TCGA1 mRNA data. Results were similar for other HGF SNPs (data not shown). We then evaluated relationships among HGF rs5745709 genotype, HGF/MET/phospho-MET protein expression, and mortality; inclusion of protein expression values, alone or in combination, did not at all attenuate the association between genotype and mortality (data not shown); results were similar considering time to recurrence. These results, therefore, suggest that the observed genetic association in the Mayo Clinic dataset is not a result of modified gene expression.

Table 3.

Immunostaining of Mayo Clinic cases

		Antibody
		HGF	MET	Phospho-MET
Protein expression values, N (%)	Negative (0)	1 (0.3%)	1 (0.3%)	2 (0.6%)
	Weak (1)	52 (16.1%)	6 (1.9%)	13 (4.0%)
	Moderate (2)	209 (64.7%)	124 (38.9%)	165 (51.4%)
	Strong (3)	61 (18.6%)	188 (58.9%)	141 (43.9%)
Association of protein expression with mortality, HR (95%CI)	Negative/Weak	Ref.	Ref.	Ref.
	Moderate	1.26 (0.83– 1.91)	1.44 (0.45 – 4.56)	0.77 (0.39 – 1.52)
	Strong	1.21 (0.73 – 2.01)	1.35 (0.43 – 4.25)	0.55 (0.27 – 1.10)
	p-value	0.48	0.90	0.01
Protein expression value by HGF rs5745709 genotype, mean (S.D.)	GG	2.01 (0.60)	2.57 (0.53)	2.35 (0.58)
	GA	2.01 (0.60)	2.56 (0.55)	2.42 (0.60)
	AA	1.92 (0.49)	2.55 (0.52)	2.42 (0.51)
	p-value	0.80	0.89	0.41

Due to core drop-off, data were unavailable for three cases on HGF, seven on MET, and five on phospho-MET.

Finally, we characterized HGF SNP associations with mortality in additional ovarian cancer cases populations, including six new studies as well as new cases enrolled at the Mayo Clinic and in the TCGA (Supplemental Table 1, Supplemental Table 6). rs5745709 was successfully genotyped in each population, and MAFs were similar to that expected (range, 0.17–0.25); 284 MAY1 cases were genotyped in both laboratories and were 99.6% concordant (duplicates subsequently excluded). Analyses of the replication set adjusted for site, age, grade, and stage (as done in MAY1+TCGA1 analysis) revealed no association with ovarian cancer survival (HR 0.99, 95% CI 0.89–1.10, p=0.87; Table 4). Figure 3 displays study-specific covariate-adjusted HRs for all cases and reveals the borderline-significant heterogeneity between initial and replication studies (p=0.06). A large number of sensitivity analyses were conducted to examine the root of such heterogeneity. Results were similar when adjusted only for site and age and when restricted to high-grade serous cases; results were also similar at rs2074725 due to high LD (r²= 0.97). Data on time to recurrence were available on some studies (MAY1, SRO, MAL, MAY2), and results were similar. When cases with peritoneal or fallopian tube cancer were excluded and when cases were right-censored at five years, results were also similar. Although we note that other SNP effects appear to exhibit themselves only in the context of optimal debulking (37) and that both Mayo Clinic series included a relatively large proportion of optimally-debulked patients (78% and 87% in MAY1 and MAY2, respectively), analysis of 1,139 optimally-debulked replication cases (BAV, BEL, LAX, MAL, MAY2, SRO) also yielded null results (HR 0.95, 95% CI 0.78–1.15, p=0.60). Finally, no compelling common feature of study populations with risk estimates greater than 1.0 (MAY1, TCGA1, MAL, BAV, and MAY2) could be identified. Thus, we conclude that the initial Mayo Clinic/TCGA HGF association a due to the winner’s curse phenomenon or unexplained heterogeneity.

Table 4.

HGF rs5745709 and ovarian cancer mortality

			Adjusted for site, age, grade, stage			Adjusted for site and age only
Histology	Group	N Cases (N Deaths)	HR (95% CI)	p-trend	p-het	HR (95% CI)	p-trend	p-het
All Histologies	Initial	435 (234)	1.56 (1.25–1.94)	6.9 × 10−5	0.83	1.45 (1.16–1.81)	1.1 × 10−3	0.82
	Replication	2,560 (1,046)	0.99 (0.89–1.10)	0.87	0.63	1.01 (0.91–1.12)	0.92	0.97
	Combined	2,995 (1,280)	1.06 (0.97–1.17)	0.21	0.06	1.07 (0.97–1.17)	0.17	0.38
High-Grade Serous	Initial	296 (179)	1.60 (1.23–2.09)	5.4 × 10−4	0.90	1.62 (1.25–2.11)	3.0 × 10−4	0.95
	Replication	1,061 (422)	0.99 (0.84–1.17)	0.91	0.70	1.01 (0.86–1.20)	0.89	0.62
	Combined	1,357 (601)	1.13 (0.98–1.30)	0.10	0.25	1.14 (0.99–1.31)	0.07	0.22

Initial includes MAY1, TCGA1; Replication includes SRO, MAL, BEL, LAX, TCGA2, PVD, BAV, and MAY2.

Figure 3. rs5745709 and ovarian cancer mortality, all studies.

Hazard ratios and 95% confidence intervals by study site, study phase (initial vs. replication) and overall, from Cox proportional hazards regression analysis. Site-specific analyses adjust for age at diagnosis, tumor stage, and tumor grade. Combined initial, replication and overall analyses adjust additionally for study site.

DISCUSSION

Although less productive than hoped for, the study of candidate genes in cancer epidemiology has yielded a handful of replicated consistent associations (e.g., TERT, CASP8, and NAT1 in ovarian (38), breast (39), and bladder cancers respectively (40)), even in the GWAS era. In ovarian cancer, candidate gene survival analyses have suggested angiogenesis, inflammation, and other pathways as drivers of genetically-determined variation in outcome (5–7). Here, we examined approximately 170 genes and regions and found evidence for association between variants in HGF and mortality among cases enrolled in Mayo Clinic and TCGA studies. With consistency of results and evidence suggesting expression as a mechanism, we then stained ovarian tumors for the primary signaling molecules, and we expanded our association analysis. In these important follow-up studies, we did not observe a clear relationship between SNP genotype, expression of the relevant proteins, and outcome; we also did not replicate the genetic association in a broader collection of samples.

In many ways, this work exemplifies the challenges of modern molecular epidemiologic investigation. A biologically-based hypothesis was comprehensively examined with regard to both the number of genes and the coverage of inherited variants within them. We capitalized initially upon a homogeneous patient population with detailed clinical data with 80% power to detect a HR as small as 1.46, assuming MAF=0.20, α= 0.05, and dominance. Prior to committing additional expenditure, we conducted in silico analysis of the top hits using publicly-available data including exploration of a possible intermediate phenotype (41). In addition to the statistical significance of the combined association, the biology of the most-significant gene was compelling (as is always the case in candidate gene studies). Here, HGF signaling plays a key role in ovarian cancer cell growth, migration, and invasion (36), and tumor expression has been shown to associate with outcome (36, 42). Some studies have reported improved chemoresponsiveness (43–45), consistent with the associations we observed between minor alleles, reduced expression, and increased mortality.

To follow-up in terms of depth, we created and immunostained TMAs from a large group Mayo Clinic cases, most of whom had genotype and outcome data, for protein expression of HGF, MET, and phospho-MET. Cores were randomized with respect to both genotype and survival, removing a source of potential bias in TMA analysis. In contrast to the initial microarray-based TCGA mRNA analyses, no genotype-expression association was seen. We must note that there are complexities in measuring HGF with immunohistochemistry, particularly because it is a soluble ligand; positive staining for the protein may reflect HGF generated at a distant source which bound to the tumor cells only in situ. Random measurement error of HGF protein levels could have attenuated the HRs. In contrast, with mRNA expression studies, mRNA is clearly isolated from the relevant tissue source (tumor or tumor stroma) used in the study. Our inclusion of MET and phospho-MET in some ways addresses these concerns about cellular specificity and also revealed null results. We also note that while genetic variation in HGF does not appear to account for dramatic changes in protein expression, other functional roles for HGF SNPs can not be ruled out as alterations in binding affinity, efficacy of signal transduction, or turnover rate are most likely not detectable using this methodology. We also analyzed protein expression among Mayo Clinic cases restricted to high-grade serous (similar to TCGA eligibility), and results remained null. In construction of TMAs, preferential sampling of cores from the periphery of ovarian tumors has been recommended due to loss of expression centrally (46). Because we did not do this, it is possible that our TMA-based mean protein expression levels (and variance estimates) are lower than they are in truth, perhaps resulting in some loss of statistical power. However, our distribution of protein expression values were similar to those reported in other studies (47, 48).

In terms of breadth, we expanded our initial study of the HGF SNP association with ovarian cancer mortality via an international consortium. Use of consortia has become standard in SNP-oriented molecular epidemiology in order to minimize the false discoveries of smaller sample sizes, to increase precision of true risk allele effect estimates, and to evaluate generalizability (49, 50). Indeed, the large sample size of our replication set provided 80% power to detect relatively small effect sizes of 1.21 in the replication set and 1.19 in the combined set of participants (assuming MAF=0.20, α= 0.05, and dominance), minimizing the possibility of a false negative result. Genotyping was centrally done and included a subset of the cases initially studied; the quality of the data appeared good suggesting laboratory issues were not a concern. Further challenges lie in aligning clinical data across studies. We used a consistent data dictionary with thorough examination of outlying observations and differences in eligibility or follow-up across studies; in addition, sensitivity analyses considering clinical and study design factors yielded consistent results. Factors such as covariate adjustment, restriction to certain clinical subsets, and consideration of time to enrollment or follow-up did not explain the discrepancy of results between initial and replication analyses. Unlike the two Mayo Clinic case collections, results differed between TCGA datasets; however, no systematic differences in clinical or follow-up features appear to exist across TCGA sets. We note that if data on all TCGA cases had been available at the time of our initial investigation, our follow-up strategy would likely have differed. Thus, this work serves as a cautionary reminder of the fluidity of public data. In addition, because the majority of cases were treated with standard chemotherapies, the likelihood is low that a true association exists only in a certain treatment group. Although an unanalyzed factor may account for differences in association across studies, we observed that the majority of data are null and thus variation in risk estimates around unity is likely random.

In summary, we report on a comprehensive analysis of ovarian cancer mortality and SNPs in several genes and regions of interest to cancer biology, and we described a two-pronged approach to follow-up the most promising result with TMA and collaborative studies. Although, in Mayo Clinic cases, HGF SNPs appear associated with increased mortality, and phospho-MET protein expression was associated with early stage disease and reduced mortality, common genetic variation in HGF is unlikely to account for a significant proportion of deaths in ovarian cancer. This body of work is characteristic of the state-of-the-art in molecular epidemiology and demonstrates the importance of incorporating multiple data types and study populations in interpretation of promising genetic associations.