Analyses of 7,635 Patients with Colorectal Cancer Using Independent Training and Validation Cohorts Show That rs9929218 in CDH1 Is a Prognostic Marker of Survival

Worldwide, over a million people are diagnosed with colorectal cancer each year. Several factors influence survival after diagnosis, but the only routinely used prognostic marker is clinical stage, which combines depth of tumor invasion, nodal status, and distant metastasis (1). Other factors thought to influence prognosis include lifestyle (2, 3), systemic inflammatory response to the tumor (4), the tumor immunologic microenvironment (5), and the tumor's somatic molecular profile (6–9).

The search for inherited factors that affect prognosis has primarily focused on candidate genes that either function within the pharmacologic pathways of the chemotherapeutic agents used in the treatment of colorectal cancer (10, 11) or that influence tumor progression (12). Recently, high-throughput SNP arrays have been used to search for colorectal cancer-susceptibility alleles by genome-wide association studies (GWAS) and, to date, identified 27 genome-wide significant low penetrance loci mapping to 8q24 (13, 14), 18q21 (15, 16), 15q13 (17, 18), 11q23 (16), 10p14 (19), 8q23 (19), 14q22 (20), 16q22 (20), 19q13 (20), 20p12 (20, 21), 1q41 (22), 3q26 (22), 12q13 (22), 20q13 (22), 6p21 (23), 11q13 (23), Xp22 (23), 2q32 (24), 12p13 (21, 25, 26), 5q31 (21), 1q25.3 (24, 25), 10q24 (25), 10q22 (26), 10q25 (26), 11q12 (26), 17p13 (26), and 19q13 (26). Studies have suggested that some of these risk alleles may also affect patient survival (27–32); however, none of the survival findings or prognostic biomarkers identified through the candidate gene analyses, have been validated in independent studies (33–35).

Here, we sought robust biomarkers of patient survival by analyzing 20 genome-wide significant colorectal cancer-susceptibility SNPs in a large training phase cohort, with subsequent validation of positive associations in an independent study group.

We prepared blood DNA samples from unrelated patients with advanced (stage IV) colorectal cancer from the MRC clinical trial COIN (NCT00182715; ref. 36). All patients had either previous or current histologically confirmed primary adenocarcinomas of the colon or rectum, together with clinical or radiologic evidence of advanced and/or metastatic disease, or had histologically/cytologically confirmed metastatic adenocarcinomas, together with clinical and/or radiologic evidence of a colorectal primary tumor. Patients were randomized 1:1:1 to receive continuous oxaliplatin and fluoropyrimidine chemotherapy (Arm A), continuous chemotherapy plus cetuximab (Arm B), or intermittent chemotherapy (Arm C). All patients gave informed consent for their samples to be used for bowel cancer research (approved by REC; 04/MRE06/60).

The validation phase consisted of samples from several different trials or prospective cohort studies. COINB is a MRC-funded phase II trial assessing cetuximab efficacy in intermittent oxaliplatin-fluoropyrimidine chemotherapy of advanced colorectal cancer (NCT00640081; ref. 37). FOCUS2 is a trial for patients with unpretreated advanced colorectal cancer judged unfit for full-dose combination chemotherapy (NCT00070213). FOCUS3 is a trial determining the feasibility of molecular selection of therapy using KRAS, BRAF, and topoisomerase-1 in advanced colorectal cancer (NCT00975897). PICCOLO is a trial of the treatment for fluorouracil-resistant advanced colorectal cancer (NCT00389870; patients from COIN or COINB that were subsequently recruited into PICCOLO were excluded). VICTOR is a trial of rofecoxib as postadjuvant therapy for colorectal cancer (NCT00031863). Six prospective cohort studies from the Genetics and Epidemiology of Colorectal Cancer Consortium (GECCO; refs. 24, 38) were also included: the Health Professionals Follow-up Study (HPFS), the Nurses' Health Study (NHS), the Physicians' Health Study (PHS), the VITamins And Lifestyle Study (VITAL), the Women's Health Initiative (WHI), and the Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial (PLCO; see Supplementary Information for references). All of these studies used a prospective design, with follow-up for incident cancer diagnoses and survival outcomes. Cases of incident colorectal cancer arising in these studies were identified from self-report and confirmed by their medical records (HPFS, NHS, PHS, PLCO, WHI) and/or linkage to cancer registries (VITAL). Two subsets of cases were genotyped in WHI: WHI1 included colon cancer patients diagnosed before September 2005 and WHI2 included unrelated colorectal cancer patients diagnosed before August 2009. Two subsets of cases were also genotyped in PLCO: PLCO1 included colon cancer patients and PLCO2 included unrelated colorectal cancer cases. All participants provided informed consent for genetic testing, and all studies were approved by their respective Institutional Review Boards. Protocols for assessing survival in the GECCO studies have been described previously (see Supplementary Information for references).

Genotyping of 15 colorectal cancer risk alleles (rs6691170 and rs6687758 at 1q41, rs10936599 at 3q26, rs4444235 and rs1957636 at 14q22, rs9929218 at 16q22, rs10411210 at 19q13, rs961253 at 20p12, rs10795668 at 10p14, rs3802842 at 11q23, rs4925386 at 20q13, rs4939827 at 18q21, rs16892766 at 8q23, rs4779584 at 15q13, and rs6983267 at 8q24) was performed by Illumina's Fast-Track Genotyping Service using their high-throughput BeadArray technology. rs4925386 failed genotyping. For the remaining 14 SNPs, the genotyping concordance rate for duplicate samples (n = 110) was 100% (1,540/1,540 genotypes), GenTrain scores ranged from 0.6814 to 0.9500 and the overall genotype success rate was 99.44% (28,868/29,032 genotypes were called successfully). Genotyping of rs4925386 at 20q13, rs4813802 at 20p12, and rs16969681 and rs11632715 at 15q13 was carried out by LGC genomics using their KASPar technology with a genotype success rate of 99.17% (8,253/8,322 genotypes called successfully) and concordance rate for duplicate samples (n = 94) of 100% (376/376). Genotyping of rs11169552 and rs7136702 at 12q13 was carried out by Geneservice using TaqMan assays (Applied Biosystems) with a genotype success rate of 95.66% (3,966/4,146 genotypes called successfully) and concordance rate for duplicate samples (n = 94) of 100% (188/188).

rs16892766, rs9929218, and rs10795668 were genotyped in patients from COINB, FOCUS2, FOCUS3, and PICCOLO by LGC genomics (KASPar technology). In VICTOR, genotyping was carried out on Illumina HumanHap300 arrays and rs9929218 was directly genotyped, rs16892766 was imputed, and rs706771 was genotyped as a proxy for rs10795668 (R² = 0.965, D′ = 1). All three SNPs were genotyped in cases from HPFS, NHS, and PHS using the TaqMan Open Array SNP genotyping platform. For the other GECCO studies, genotyping was performed on Illumina 300/240S (PLCO1), 550K (WHI1), 610K (WHI1, PLCO1), and HumanCytoSNP (VITAL, WHI2, PLCO2) arrays; rs9929218 was directly genotyped on these platforms in all studies, and, rs16892766 and rs10795668 were directly genotyped on the platform used in WHI1 and PLCO1, and imputed (using MACH and HapMap2 Release 24) in WHI2, VITAL, and PLCO2. Note – different genotyping platforms were often used because susceptibility SNPs were identified and assayed at different times by different investigators.

All SNPs were tested for their genotypes being consistent with the Hardy Weinberg Equilibrium (HWE) using a Pearson χ² test. Linkage disequilibrium (LD) was examined using Haploview version 4.2. For survival analyses of the training phase, we used a Cox survival model with overall survival (time from trial randomization to death) as the primary measure. A codominant model was applied, analyses were stratified for treatment arm and type of fluoropyrimidine used, and P values were corrected for multiple testing by Bonferroni correction. Significant SNPs were tested for independence to known prognostic factors using a closed-test procedure multiple fractional polynomial model with P < 0.05 and the best-fitting genotype model (dominant or recessive) was identified. For survival analyses in the validation phase, overall survival was used for COINB, FOCUS2, FOCUS3, PICCOLO, and VICTOR, and time from diagnosis to death for HPFS, NHS, PHS, VITAL, WHI, and PLCO. A Cox survival model was fitted to the data from each trial or study separately, and an overall pooled result was calculated using a fixed-effects inverse-variance meta-analysis approach. Heterogeneity was assessed using the Q² and I² statistics. If the pooled validation data generated a significant result, additional analyses were conducted: (i) a further meta-analysis including the training and validation data together, (ii) a sensitivity analysis replacing time from randomization to death (considered left-truncated at randomization to account for the fact that randomization is conditional upon survival from diagnosis) with time from diagnosis to death—for those trials for which this information was available (COIN, COINB, and FOCUS3; n = 2,446 patients genotyped with survival data), and, (iii) the effect on 12-week response to chemotherapy in COIN Arms A and C (those arms not confounded by treatment with cetuximab; n = 1,369 patients genotyped with this data). Response was defined as complete response or partial response at 12 weeks and nonresponse was defined as stable disease or progressive disease.

We analyzed blood DNA samples from 2,083 unrelated patients with advanced colorectal cancer from the UK national trial COIN (36). In total, 34% of patients were female with a mean age at diagnosis of 62 years (range 18–84 years, Table 1). We assayed 20 independent, genome-wide significant, colorectal cancer-risk alleles (13, 15–17, 19, 20, 22) representing 14 loci; with a single SNP at nine loci, two SNPs at four loci, and three SNPs at one locus (loci with ≥2 SNPs contain multiple independent risk alleles; refs. 20, 22). Fifteen SNPs were genotyped using the Illumina GoldenGate platform (one failed), four (including a repeat of the failed SNP) were successfully genotyped using KASPar technology and two were successfully genotyped using TaqMan assays. All 20 SNPs, apart from rs7136702 (P = 0.027), had genotype distributions consistent with the HWE with no imbalances between the treatment arms or according to the somatic mutation status of the colorectal cancers (42.27%, 9.01%, and 3.56% of colorectal cancers were KRAS, BRAF, and NRAS mutant, respectively; ref. 39).

Fourteen SNPs did not influence survival under a codominant model (Table 2). Six SNPs were significant in the univariate analyses, of which three (rs16892766 at 8q23, rs9929218 at 16q22, and rs10795668 at 10p14) remained significant after correction for multiple testing (Table 2). We have previously shown that the WHO performance status, number of metastatic sites, white blood cell count, alkaline phosphatase levels, and KRAS and BRAF mutation status are independent prognostic factors affecting survival in patients from COIN (36). We therefore applied a multivariate model with these factors, together with the best genetic models that fitted the data, and showed that all three SNPs independently influenced survival (Supplementary Table S1).

We used samples from numerous independent trials and cohort studies to provide sufficient power to carry out our validation analyses. In total, we assayed rs16892766, rs9929218, and rs10795668 in 5,552 patients with colorectal cancer (196 from COINB, 337 from FOCUS2, 172 from FOCUS3, 334 from PICCOLO, 918 from VICTOR, 259 from HPFS, 355 from NHS, 278 from PHS, 531 from PLCO1, 478 from PLCO2, 281 from VITAL, 450 from WHI1, and 963 from WHI2; Table 1). No significant heterogeneity was detected in any of the meta-analyses (I² = 0%). Only rs9929218 was found to be significantly associated with survival (P = 2.5 × 10⁻²; Supplementary Table S2).

Patients homozygous for the minor allele of rs9929218 (AA genotype), equating to approximately 8% of patients, showed significantly poorer survival as compared with patients with the AG or GG genotypes (training phase HR = 1.47; 95% CI, 1.24–1.75, P = 1.4 × 10⁻⁵ unadjusted, HR, 1.43; 95% CI, 1.20–1.71, P = 5.8 × 10⁻⁵ after adjustment for age, sex, and time from diagnosis to randomization; validation phase HR, 1.19; 95% CI, 1.02–1.38, P = 2.5 × 10⁻² unadjusted, HR, 1.18; 95% CI, 1.01–1.37, P = 3.2 × 10⁻² adjusted; combined HR, 1.30; 95% CI, 1.16–1.46, P = 6.1 × 10⁻⁶ unadjusted, HR, 1.28; 95% CI, 1.14–1.43, P = 2.2 × 10⁻⁵ adjusted; Fig. 1 and Table 3). This equated to a median decrease in life expectancy of 4.3 months (based on training phase data). Patients with a single variant allele (AG genotype) had similar survival outcomes to those with a wild-type (GG) genotype (Supplementary Table S3).

We combined the training and validation phase data and analyzed by disease stage. rs9929218 genotype did not deviate from the HWE according to stage (Supplementary Table S4). rs9929218 was not significantly associated with survival amongst patients with stage I–III (premetastatic) disease (HR, 1.19; 95% CI, 0.93–1.52, P = 0.18), with little statistical evidence of heterogeneity amongst the individual studies (P = 0.39; Fig. 2). In contrast, rs9929218 was highly associated with survival in patients with stage IV (metastatic) colorectal cancer (HR, 1.34; 95% CI, 1.17–1.53, P = 2.7 × 10⁻⁵), with no heterogeneity amongst the individual trials and cohorts (P = 0.91; Fig. 2). There was, however, no significant difference between the associations of rs9929218 genotype and survival in patients with stage I–III and stage IV disease (P_interaction = 0.48).

As a sensitivity analysis, we investigated whether overall survival accurately reflected survival from the time of diagnosis to death. For 2,444 trial patients (from COIN, COINB, and FOCUS3), we had relevant clinical information available and we found little difference in the effect of rs9929218 between the two survival measures (overall survival HR, 1.50; 95% CI, 1.27–1.76, P = 1.5 × 10⁻⁶; survival time from diagnosis HR, 1.46; 95% CI, 1.24–1.73, P = 6.3 × 10⁻⁶, Supplementary Figure).

We also investigated whether the type and duration of treatment influenced survival, by evaluating rs9929218 according to trial arm in COIN (the largest trial for which we had high-quality clinical data). We did not find significant heterogeneity between the treatment arms (P = 0.38) suggesting that treatment did not influence the association between rs9929218 genotype and survival (Supplementary Table S5).

We also sought whether rs9929218 was associated with response to treatment (likely to be correlated with survival). In COIN Arms A and C, treatment was identical for the first 12 weeks apart from the choice of fluoropyrimidine. At 12 weeks, patients from these arms that were homozygous for the rs9929218 minor allele had significantly worse response (36/112 responded, 32%), as compared with patients that were heterozygous or homozygous wild-type (626/1,257 responded, 50%; OR = 0.47; 95% CI, 0.31–0.72, P = 3.9 × 10⁻⁴, adjusted for choice of fluoropyrimidine; Table 4).

The literature contains many reports of potential common inherited biomarkers of survival for colorectal cancer; however, most of these have been derived from poorly designed studies, with small numbers of samples and/or no validation of their results. As a consequence, very few of these prognostic biomarkers have been validated by independent groups. To address the critical shortcomings of previous studies, we have carried out an analysis using large independent training and validation phase cohorts as recommended by the REMARK guidelines (40) and produced robust evidence for the first common inherited genetic variant affecting survival in patients with colorectal cancer. As such, this finding represents an important clinical milestone.

Our data suggest that patients homozygous for the minor allele of rs9929218, equating to approximately 8% of patients, have worse survival, with a median decrease in life expectancy of approximately 4 months (in the advanced disease setting). Another study recently reported that the major allele of rs9929218 was associated with improved prognosis (30), providing further support for this variant having a genuine prognostic effect. Although the effect size of rs9929218 identified herein is modest (HR, 1.28; 95% CI, 1.14–1.43), the identification of further prognostic alleles by well-powered GWAS-based approaches may help clinicians model the combined effects of common germline variants together with their somatic mutation profiles to help inform patient outcome. Our study therefore represents a critical first step in this endeavor.

We have shown a clear effect of rs9929218 on survival amongst patients with stage IV disease. However, many of these patients would have received similar therapies raising the possibility that rs9929218 influences survival based upon an interaction with treatment, and we noted that patients carrying both minor alleles had poor response to chemotherapy. However, survival and response are likely to be related and we found similar prognostic effects across all arms of the COIN trial (including in those patients receiving intermittent therapy) and amongst many of the other trials and cohorts used in this study. These data suggest that the prognostic effect may therefore reflect an underlying influence on a biologic process or pathway. rs9929218 lies within intron 2 of CDH1 encoding E-cadherin, in strong LD with rs16260 (41) in the CDH1 promoter, which downregulates CDH1 expression (42). Patients homozygous for the minor allele of rs9929218 would be expected to have reduced E-cadherin expression. E-cadherin functions as a transmembrane glycoprotein that is critical in the establishment and maintenance of intercellular adhesion, cell polarity, and tissue morphology and regeneration (43) and its loss represents a defining feature of the epithelial to mesenchymal transition during metastasis. A clear mechanism therefore exists for the potential prognostic effect of rs9929218 by influencing this process.

D. Kerr is an employee of and has ownership interest (including patents) in Oxford Cancer Biomarkers, reports receiving commercial research grants for Roche, and is a consultant/advisory board member for Amgen and Bayer. No potential conflicts of interest were disclosed by the other authors.

Conception and design: C.G. Smith, I. Tomlinson, J.P. Cheadle

Development of methodology: C.G. Smith

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): C.G. Smith, R. Harris, T.S. Maughan, S. Richman, M. Seymour, I. Tomlinson, D. Kerr, A.T. Chan, U. Peters, P.A. Newcomb, H. West, A. Meade, R. Kaplan, J.P. Cheadle

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): C.G. Smith, D. Fisher, A.I. Phipps, D. Rosmarin, A.T. Chan, U. Peters, H. West, J.P. Cheadle

Writing, review, and/or revision of the manuscript: C.G. Smith, D. Fisher, T.S. Maughan, A.I. Phipps, S. Richman, M. Seymour, I. Tomlinson, D. Kerr, A.T. Chan, U. Peters, P.A. Newcomb, A. Meade, R. Kaplan, J.P. Cheadle

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): C.G. Smith, R. Harris, P.A. Newcomb, S. Idziaszczyk, R. Kaplan

Study supervision: C.G. Smith, I. Tomlinson, J.P. Cheadle

Other (chief investigator of the COIN trial from which over 2000 patients contributed samples for this study, was also the lead clinical collaborator with Cheadle and Smith during the time this work was conceived and undertaken): T.S. Maughan

Other (directed study): J.P. Cheadle

The authors thank Ayman Madi, Richard Adams, Sarah Kenny, and the COIN trial management group for their advice or support; the patients and their families who participated in COIN and gave their consent for this research; and the investigators and pathologists throughout the UK who submitted samples for assessment. COIN, COINB, FOCUS3, and PICCOLO were conducted with the support of the National Institute of Health Research Cancer Research Network. The authors also thank all staff at the GECCO Coordinating Center. For HPFS, NHS, and PHS, they thank Patrice Soule and Hardeep Ranu of the Dana Farber Harvard Cancer Center High-Throughput Polymorphism Core who assisted in the genotyping for NHS, HPFS, and PHS under the supervision of Immaculata Devivo and David Hunter, Qin Guo and Lixue Zhu who assisted in programming for NHS and HPFS, and Haiyan Zhang who assisted in programming for PHS. They also thank the participants and staff of the Nurses' Health Study and the Health Professionals Follow-Up Study, for their valuable contributions as well as the following state cancer registries for their help: AL, AZ, AR, CA, CO, CT, DE, FL, GA, ID, IL, IN, IA, KY, LA, ME, MD, MA, MI, NE, NH, NJ, NY, NC, ND, OH, OK, OR, PA, RI, SC, TN, TX, VA, WA, WY. For WHI, the authors thank the WHI investigators and staff for their dedication, and the study participants for making the program possible. For PLCO, the authors thank Christine Berg and Philip Prorok, Division of Cancer Prevention, NCI, the Screening Center investigators and staff of the PLCO Cancer Screening Trial, Tom Riley and staff, Information Management Services, Inc., Barbara O'Brien and staff, Westat, Inc., and Bill Kopp, Wen Shao, and staff, SAIC-Frederick. The authors also thank the study participants for their contributions to making this study possible.

COIN and COINB were funded by Cancer Research UK and the Medical Research Council, and the associated translational studies were supported by the Bobby Moore Fund from Cancer Research UK, Tenovus, the Kidani Trust, Cancer Research Wales and the National Institute for Social Care and Health Research Cancer Genetics Biomedical Research Unit (2011–2015). FOCUS2 (ISRCTN21221452) was funded jointly by the Medical Research Council and Cancer Research UK. PICCOLO (ISRCTN93248876) was funded by Cancer Research UK, with support from Amgen. FOCUS3 was funded by the Medical Research Council Efficacy and Mechanism Evaluation Programme. Core funding to the Wellcome Trust Centre for Human Genetics was provided by the Wellcome Trust (grant 090532/Z/09/Z). GECCO was supported by the National Cancer Institute (NCI), National Institutes of Health (NIH), and U.S. Department of Health and Human Services (DHHS; U01 CA137088 and R01 CA059045). NIH also supported HPFS (P01 CA055075, UM1 CA167552, R01 CA137178 and P50 CA127003), NHS (R01 CA137178, P01 CA087969 and P50 CA127003), PHS (R01 CA042182), and VITAL (K05 CA154337). PLCO was supported by the Intramural Research Program of the Division of Cancer Epidemiology and Genetics, and supported by contracts from the Division of Cancer Prevention, NCI, NIH, and DHHS. WHI was supported by the National Heart, Lung and Blood Institute, NIH, and DHHS (HHSN268201100046C, HHSN268201100001C, HHSN268201100002C, HHSN268201100003C, HHSN268201100004C, and HHSN271201100004C).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.