Rare Variants in the DNA Repair Pathway and the Risk of Colorectal Cancer

Inherited genetic factors play an important role in the etiology of colorectal cancer (1). Genome-wide association studies (GWAS) in colorectal cancer have identified approximately 140 common risk loci predominantly in European and East Asian populations (2–4). However, risk variants, both high- and low-penetrance, collectively account for only a small proportion of the estimated familial risk (3). Part of the unexplained proportion of colorectal cancer heritability may be attributable to rare variants in genes/pathways with established or suspected roles in colorectal carcinogenesis, such as DNA repair, TGFβ signaling, vitamin D, and folate metabolism (5–8).

Rare variants by definition occur with a <1% frequency in the population and therefore large sample sizes are required to detect rare risk variants with modest effect sizes. Yet, next-generation sequencing approaches have enabled successful identification of population-specific rare variants and assessed their contributions to colorectal cancer risk. A recent exome sequencing study of early-onset colorectal cancer cases identified rare, highly penetrant pathogenic variants in 16% of cases, including novel variants in POT1, POLE2, and MRE11 (9). Whole-genome sequencing of sporadic colorectal cancer cases identified a rare variant in CHD1 with a strong protective effect (2). Studies using exome-wide genotyping arrays have also identified rare risk variants with large effect sizes [odds ratio (OR) ≥ 2.0] in genes previously unreported to be associated with colorectal cancer risk (i.e., TCF7L2, RAB11FIP5, COL27A1; refs. 10, 11). Recently, Bayesian model uncertainty techniques have been developed to maximize the power of rare variant association studies by incorporating uncertainty of both the inclusion of variants and the direction of association of each included variant in a model (12).

To further explore the contribution of rare coding variation to the risk of colorectal cancer, we examined the association between rare coding variants and colorectal cancer risk using participants of European ancestry from three large epidemiologic studies for a total of 5,293 individuals (2,327 cases and 2,966 controls). In addition to single variant testing, we investigated the aggregate effects of rare coding variation at the gene and pathway levels using the integrative Bayesian risk index (iBRI) method that incorporates external biological information to help refine the variant set selection procedure and ultimately localize the variant(s) most likely driving the association signal.

Study participants were individuals of European descent from North America, Europe, Australia, and Israel recruited from the Kentucky Case-Control Study (KY), the Colon Cancer Family Registry (CCFR), and the Molecular Epidemiology of Colorectal Cancer Study (MECC). Details on the study design and data collection for each study are provided in the Supplementary Materials and Methods. All cases had histologically confirmed adenocarcinoma of the colon or rectum. All participants provided written informed consent, and specimens and data were collected according to protocols approved by the Institutional Review Boards of the respective institutions.

Genomic DNA was extracted from blood samples and genotyped using the Infinium Human Exome BeadChip 12v1.0 (Illumina Inc.) plus custom content, which provides comprehensive coverage of >250,000 genetic markers across >20,000 genes in the human coding genome. Genotype calling was performed using standard methods.

Data from KY, CCFR, and MECC were cleaned on the basis of standard quality control (QC) metrics at the sample and variant levels as detailed in the Supplementary Materials and Methods. Among the 5,418 KY and CCFR samples genotyped, 2,691 KY samples (1,054 colorectal cancer cases and 1,637 controls) and 2,311 CCFR samples (1,132 colorectal cancer cases and 1,179 controls) passed QC filters and were retained for downstream analysis. Of the initial 1,392 MECC samples genotyped, 1,105 (559 colorectal cancer cases and 546 controls) passed QC filters, of which 303 were Arabs, 490 were Sephardi Jews, 291 were Ashkenazi Jews, and 21 were of unknown ethnicity. Only participants of Ashkenazi Jewish heritage (141 cases and 150 controls) were retained for downstream analyses due to their descent from Eastern Europe. Basic characteristics of the study subjects by case–control status are presented in Supplementary Table S1.

Overall, 142,390 polymorphic variants in 5,293 samples (2,327 cases and 2,966 controls) passed QC filters in at least one of the three studies and were assessed for single-variant association with colorectal cancer risk. A higher proportion of rare variants [minor allele frequency (MAF) <1%] was observed in KY (34.4%) and CCFR (33.2%) compared with MECC (10.1%; Supplementary Table S2).

For the gene- and pathway-level analyses, common variants (MAF ≥ 1%, n = 24,425) and variants in downstream (n = 251), intergenic (n = 10,402) and upstream (n = 309) regions based on the GRCh37 assembly were excluded. We considered a variant to be rare or common based on the 1% MAF cut-off in the study-specific samples that passed QC. ANNOVAR (2019Oct24 release) was used to annotate variants according to location and predicted functional effect (13). Potential pathogenic effects of nonsynonymous variants were predicted using ClinVar and six in silico algorithms [Combined Annotation-Dependent Depletion (CADD), Polyphen2-HumDiv (HDIV), PolyPhen2-HumVar (HVAR), Likelihood Ratio Test (LRT), Mutation Taster (MT), and Sorting Tolerant From Intolerant (SIFT)] from dbNSFP (14). Top associated variants were analyzed for potential functional or regulatory effects using established tools and resources such as HaploReg v4.1 (https://pubs.broadinstitute.org/mammals/haploreg), GTEx (https://www.gtexportal.org), RegulomeDB 2.0 (https://regulomedb.org), ENCODE (https://www.encodeproject.org/), and FuncPred (http://www.funcpred.com/).

The associations between allele dosages for all common (MAF ≥ 1%) and rare (<1%) variants and the risk of colorectal cancer were estimated through a 1-degree of freedom likelihood ratio test (LRT) assuming a log-additive model. Per-allele ORs and 95% confidence intervals (CI) were also estimated using unconditional logistic regression adjusting for appropriate study-specific covariates and principal components (PC) that capture sub-European ancestry. For KY and MECC, models were adjusted for age, sex, and the top four PCs. For CCFR, models were adjusted for age, sex, the top four PCs, and recruitment site (Australia, FHCRC, Mayo, Ontario, and USC). To summarize evidence of association across the three studies, we performed a meta-analysis from the sample size-weighted average of the study-specific test statistics using METAL (15). Exome-wide statistical significance was set at Bonferroni adjusted P < 3.51 × 10^–7 in two-sided test based on the 142,390 unique polymorphic variants tested across the three studies. P < 0.05 and P < 0.001 were used to identify variants with nominal evidence of association.

We created quantile–quantile (Q–Q) plots of LRT P values and computed the genomic control factor (λ) to find evidence of residual population stratification after adjustment for PCs. Manhattan plots were generated to provide a visual illustration of the most significantly associated variants throughout the exome. Study-specific Q–Q plots and Manhattan plots were created using all variants and common variants only. The Q–Q plots including both common and rare variants showed evidence of inflation in both the study-specific P values (λ = 1.54–1.80; Supplementary Table S2) and meta P values (λ = 1.67; Supplementary Fig. S1A), which is likely due to the high sensitivity of the LRT to rare variants (16). When only common variants were plotted, a more uniform distribution between expected and observed P values was observed, with the genomic inflation factor indicating no appreciable evidence of population stratification after adjustment for PCs (λ = 0.99–1.07 for study-specific P values and λ = 1.02 for meta P values; Supplementary Fig. S1B). The Manhattan plots of meta-analysis estimates depict the magnitude of association of all variants (Supplementary Fig. S2A) and common variants only (Supplementary Fig. S2B) with colorectal cancer risk across the three studies (Supplementary Fig. S2).

The aggregate association of rare variants (MAF < 1%) with colorectal cancer risk was evaluated across gene-based regions, with variants assigned to genes based on chromosome position. Five methods were used for gene-level testing of rare variants: (i) weighted sum statistic by Madsen and Browning (17); (ii) sequence kernel association test (SKAT; ref. 18); (iii) computational step-up (19); and (iv) Bayesian Risk Index (BRI), which formally incorporates uncertainty of both the inclusion of variants in a region as well as the direction of effect of each included variant via Bayesian model uncertainty techniques (12). The BRI method is used to calculate a Bayes factor (BF) that quantifies the evidence that at least one variant within the specified region is associated with colorectal cancer. Details of each statistical method are described in the Supplementary Materials and Methods. Study-specific association models were adjusted for sex, age, the top four PCs, and for CCFR only, recruitment site. Evidence of gene-level associations across the three studies was found using a sample size-weighted meta-analysis of study-specific test statistics using METAL (15). The significance threshold for gene-level association analysis was P < 3.61 × 10^–6 in two-sided test based on the 13,866 unique genes with ≥2 rare variants in any of the three studies. P < 0.05 was used to identify variants with nominal evidence of association. For the BRI approach, we calculated the product of the three study-specific BFs to obtain a summary BF under the assumption that studies are independent, each with its own study-specific prior probability of the null hypothesis, consistent with the BRI model (20). Gene-level associations with colorectal cancer risk based on the BRI method were defined as extreme (summary BF ≥ 100), strong (summary BF ≥ 10), and moderate (summary BF ≥ 1) according to a priori categorization (21).

The association between rare variants (MAF < 1%) and colorectal cancer risk was evaluated across gene sets grouped by candidate pathways chosen based on a priori evidence for their involvement in colorectal carcinogenesis: DNA repair (n_gene = 137), TGFβ signaling (n_gene = 88), vitamin D metabolism (n_gene = 10), and folate metabolism (n_gene = 42; refs. 2, 5, 8). We further integrated information about specific DNA repair subpathways such as ataxia telangiectasia mutated (ATM), base excision repair (BER), Fanconi anemia/homologous recombination (HR/FA), mismatch repair (MMR), nucleotide excision repair (NER), nonhomologous end joining (NHEJ), recQ helicases (RECQ), translesion synthesis (TLS), cross-link repair (XLR), and other minor pathways (OTHER). DNA repair genes were annotated according to previous curations (22, 23). For TGFβ signaling, vitamin D metabolism and folate metabolism, genes were downloaded from KEGG version 91 (https://www.genome.jp/kegg/, release 2019/10) and REACTOME version 70 (https://reactome.org/, release 2019/09). A full list of genes assigned to each pathway as well as the number of common and rare variants identified in each study are summarized in Supplementary Table S3.

The pathway-level analysis was performed using the iBRI approach, which extends over the BRI method by integrating multiple region-specific risk indices within a given model and including external predictor-level covariates to help guide the region and variant selection procedure (24). Details of the iBRI method are described in the Supplementary Materials and Methods. Briefly, we used pathway information to estimate the probability that at least one variant within the specified pathway is associated with colorectal cancer risk through estimation of the BF via set-specific posterior probability, which is the sum of the posterior model probabilities for every model that includes at least one variant within the set (12). The iBRI was computed separately for each gene in the pathway, and multiple gene-specific risk indices were included in a single model adjusted for sex, age, the top four PCs, and for CCFR only, recruitment site. Given evidence of an association with the DNA repair pathway, we further investigated the likely genes and variants that are driving the pathway-level association through estimation of BFs under the iBRI method. For each study, the top 10 DNA repair genes within the top 25 models found using the iBRI method were plotted. Summary BFs were calculated as the product of the three study-specific BFs, using a weighted average of the posterior probabilities of inclusion for the three studies (20). Similarly to the BRI method, strength of association with colorectal cancer risk based on iBRI was defined on a priori categorization of the summary BF (21).

All statistical analyses were conducted using the R statistical computing platform and PLINK v. 1.07 (25).

A flow diagram representing the analysis pipeline and key results from the study is provided in Fig. 1.

The exome-wide single variant analysis was carried out on 142,390 polymorphic variants that passed QC in at least one of the three studies investigated. Although none of the variants reached exome-wide statistical significance, 7,099 variants were nominally associated with colorectal cancer risk (meta P < 0.05), including 84 variants at the 0.001 significance level. These 84 variants with meta P < 0.001 are summarized in Supplementary Table S4. Of these, 56 (66.7%) were rare or monomorphic among controls in the overall study population. The strongest association was found for three low-frequency variants located on chromosome 17: rs139401613 in ST6GALNAC2 (meta P = 1.01 × 10^–6; 0.034% in overall controls), rs35467001 (meta OR = 2.05; 95% CI, 1.5–2.81; meta P = 7.72 × 10^–6; 3.9% in overall controls) and rs34322745 (meta OR = 2.1; 95% CI, 1.47–3.02; meta P = 5.16 × 10^–5; 3.6% in overall controls) in SDK2. A noteworthy association was also observed for rs61736607 in ADAMTS14 (meta OR = 1.67; 95% CI, 1.3–2.14; meta P = 7.1 × 10^–5; 7.3% in overall controls). While coding variants rs35467001 and rs34322745 in SDK2 were in substantial LD across the three studies (r² = 0.74–0.79), none of these were in LD with rs139401613 (r² ≤ 0.00052). The two top associated variants (rs139401613 and rs35467001) were each predicted to be deleterious based on three out of the six algorithms examined.

A total of 12,967 genes in KY, 12,729 genes in CCFR, and 5,582 genes in MECC harbored ≥2 rare variants and were assessed for association with colorectal cancer risk. On the basis of SKAT, none of the genes reached exome-wide statistical significance (P < 3.61 × 10^–6), with the strongest associations observed for SPTBN5 (4.66 × 10^–5), CAVIN1 (9.88 × 10^–5), and ST6GALNAC2 (1.1 × 10^–4). Using the BRI approach, 17 genes were extremely associated (summary BF_BRI ≥ 100), 300 genes were strongly associated (summary BF_BRI ≥ 10), and 3,744 genes were moderately associated (summary BF_BRI ≥ 1) with colorectal cancer risk. The top associated genes with summary BF_BRI ≥ 100 are presented in Supplementary Table S5. The strongest association was found for ST6GALNAC2 on chromosome 17 (summary BF_BRI = 2604.23), followed by OSTM1, COL22A1, EPHA7, TTC28, SPTBN5, FSIP1, AKR1D1, NOTCH3, C6orf120, OR11H4, NAT1, EVI2B, CENPQ, SMPDL3A, CEP43, and GPC3 with summary BF_BRI ranging from 104.43 to 948.28. All these genes were nominally associated with colorectal cancer risk (meta P < 0.05) in at least one of the alternative gene-level tests (weighted sum, SKAT, and computational step-up), except for FSIP1 that showed a borderline association in the weighted sum test (meta P = 0.054).

Table 1 summarizes the number of genes and rare variants included in each of the pathways investigated as well as the pathway-level association results based on the iBRI method. The study-specific BF_iBRI ranges were 21.15 to 34.13 for DNA repair pathway, 0.50 to 0.96 for TGFβ signaling, 0.51 to 1.00 for vitamin D metabolism, and 0.40 to 2.16 for folate metabolism. The summary BF from the iBRI method highlighted an extreme association between DNA repair pathway and colorectal cancer risk (summary BF_iBRI = 17852.4). Figure 2 plots the study-specific inclusion of the top 10 DNA repair genes in the top 25 models found using iBRI. Given evidence of association with the DNA repair pathway, we further estimated the risk associated with individual DNA repair subpathways under the iBRI method (Table 2). We found an extreme association with NHEJ (summary BF_iBRI = 437.85) and a strong association with NER (summary BF_iBRI = 36.96). The other DNA repair subpathways were reported with BF_iBRI<10 in either the summary or study-specific results.

The iBRI method also identified the most likely genes and variants driving the association with DNA repair pathway. Table 3 reports genes with summary BF_iBRI ≥ 1 and, for each gene, the variant with the highest summary BF_iBRI (defined as the most likely associated variant). The top associated gene was RPA2 (summary BF_iBRI = 164.28) in the group of genes representing minor DNA repair pathways (OTHER), and the most likely associated variant in this gene was rs28988897 (summary BF_iBRI = 657.56). An extreme association was also found for PRKDC (summary BF_iBRI = 109.77) in NHEJ, with the most likely associated variant being rs8178232 (summary BF_iBRI = 17.68). Strong associations were observed for ERCC5 (summary BF_iBRI = 65.52) and ERCC8 (summary BF_iBRI = 21.9) in NER. The most likely associated variants in these genes were rs141369732 (summary BF_iBRI = 6.39) and rs201642761 (summary BF_iBRI = 0.92), respectively. Additional information on the DNA repair genes and variants associated with colorectal cancer risk based on the iBRI method are provided in Supplementary Tables S6 and S7. RPA2, ERCC5, ERCC8, XPC, LIG4, and MSH3 were nominally associated with colorectal cancer risk (meta P < 0.05) in at least one of the alternative gene-level tests (weighted sum, SKAT, and computational step-up), with XPC being the only gene consistently associated across all tests. In addition, there was a substantial increase in summary BF estimates from the basic BRI approach to the iBRI method for the top associated genes (RPA2, PRKDC, ERCC5, and ERCC8; summary BF_iBRI = 21.9–164.28 vs. summary BF_BRI = 2.038–6.53; Supplementary Table S6).

A sensitivity analysis excluding MECC subjects showed robust associations only for the DNA repair pathway (summary BF_iBRI = 721.85) and the NER and NHEJ subpathways (summary BF_iBRI = 50.65 and 1129.16, respectively), indicating that a limited sample size is not driving the lack of association with other pathways. At the gene-level, there was variability in summary BF when the association was mainly driven by MECC, with RPA2 being no longer strongly associated with colorectal cancer risk (BF_iBRI = 2.2). This heterogeneity in results across cohorts reinforces the hypothesis that population-specific rare variants are driving the association signal within the same gene.

This study examined the association between rare coding variants and colorectal cancer risk using Bayesian approaches that maximize the power to detect rare variant associations at multiple levels (i.e., pathway, gene and variant levels). The more efficient model search algorithm and the inclusion of external information (i.e., biological pathway) to help guide the variant set selection procedure allowed the identification of novel rare risk variants and provided additional evidence for a role of DNA repair pathways and genes in colorectal cancer tumorigenesis.

Our analysis based on the iBRI method identified NHEJ and NER as the only DNA repair subpathways strongly associated with colorectal cancer risk. These findings are consistent with the well-recognized role of these pathways in carcinogenesis as well as with previous studies reporting associations between genetic polymorphisms in genes involved in these pathways and colorectal cancer risk (26, 27). Using DNA repair subpathways as predictor-level covariates to help estimate the likelihood that any variant in a gene is associated with colorectal cancer risk, we found extreme gene-level associations with RPA2 and PRKDC. RPA2 encodes for a subunit of the heterotrimeric replication protein A (RPA) complex that is essential for DNA replication, repair, recombination, and cell-cycle regulation (28). PRKDC in the NHEJ pathway encodes for a nuclear serine/threonine protein kinase that plays a pivotal role in DNA damage response and maintenance of genomic stability (29). Aberrant expression of RPA2 and PRKDC has been previously associated with colorectal cancer development and adverse clinical outcome (30, 31). We also found strong evidence of association for ERCC5 and ERCC8, which are essential components of the NER pathway. Genetic polymorphisms in ERCC5 have been previously associated with risk and prognosis of colorectal cancer (32). ERCC8 has been identified as a potential susceptibility factor for certain cancers (33); however, our study is the first to report an association specifically with colorectal cancer. Among the other DNA repair genes with moderate evidence of association with colorectal cancer risk, REV3L, XPC, MSH3, POLQ, and TDG have been previously associated with either familial (34) and sporadic colorectal cancer (35, 36). Although many of these genes are recognized tumor suppressor genes or oncogenes, further work is required to investigate their potential role as genetic biomarkers and therapeutic targets for colorectal cancer.

In addition to gene-level associations, the iBRI method allowed to identify the most likely associated variant in each of the top associated genes. There is no prior evidence of association between these variants and colorectal cancer risk in the literature. However, except for rs121965090 in XPC, all these variants were predicted to be deleterious according to at least one of the algorithms examined. On the basis of ENCODE data, the most likely associated variants in RPA2, rs28988897, overlaps with multiple histone marks associated with enhancer and promoter activity in colon and rectal samples. In addition, this variant is likely to affect binding of transcription factors according to RegulomeDB (probability score >0.8), supporting its role as regulatory variant. These findings demonstrate the usefulness of predictor-level covariates such as biological pathways in identifying the variants that are most likely driving the associations observed at the gene level and that would not be detected by other statistical methods.

In the exome-wide gene-level analysis, ST6GALNAC2 (17q25.1) emerged as the strongest association based on the BRI method. ST6GALNAC2 encodes for a type II transmembrane Golgi-localized enzyme involved in protein glycosylation, which is a key post-translational modification that plays a role in regulating multiple cellular processes including cell adhesion, migration, signaling and immune response (37). Aberrant protein glycosylation is a common phenotypic alteration occurring in many types of cancer, including colorectal cancer (38). A rare variant located in exon 9 of ST6GALNAC2, rs139401613, was the top associated variant in the single variant analysis. This nonsynonymous variant results in the Arg340Gln amino acid change that was predicted to be deleterious by SIFT, HDIV, and HVAR. On the basis of ENCODE data specific to colon and rectal samples, rs139401613 overlaps with histone marks that are indicative of enhancer and promoter activity. In addition, this variant might affect binding of transcription factors and influence gene expression through enhancer activity.

Among the other genes with strong evidence of association from the BRI method, EPHA7, TTC28, EVI2B, GPC3, CENPQ, NOTCH3, and SMPDL3A were previously reported to play a role in the development and prognosis of colorectal cancer (39–45). Our study is the first to report an association of C6orf120, COL22A1, FSIP1, AKR1D1, and CEP43 specifically with colorectal cancer risk. Consistency in the associations between these genes and colorectal cancer risk in at least one of the alternative gene-level tests (weighted sum, SKAT, and computational step-up) increased our confidence in the validity of the results. These genes encode for proteins involved in signal transduction and cell differentiation (GPC3, EPHA7, NOTCH3; refs. 46–48), immunoregulation (C6orf120; ref. 49), structural and physiological integrity (COL22A1; ref. 50), cell-cycle progression and growth regulation (TTC28, CEP43, EVI2B, FSIP1, CENPQ; refs. 51–55), bile acid synthesis and steroid hormone metabolism (AKR1D1; ref. 56), and oxysterol or lipid metabolism SMPDL3A (57). Thus, our findings not only confirm the importance of established colorectal cancer-related genes, but also highlight novel genes previously unreported to be involved in colorectal cancer tumorigenesis.

In the exome-wide single variant analysis, 84 variants were associated with colorectal cancer risk at the meta P < 0.001 significance level. Of the 7,099 variants nominally associated with colorectal cancer risk (meta P < 0.05), 13 common variants were previously reported to be associated with colorectal cancer risk in previous exome- or genome-wide association studies (Supplementary Table S8; refs. 2, 3, 58). On the other hand, we failed to replicate associations with rare colorectal cancer risk variants previously identified in populations of European descent (2, 9, 11, 58), except for two variants (rs2427284 and rs13042941) that were previously associated with colorectal cancer risk in conditional analyses (Supplementary Table S8; ref. 58). There are several factors that may have hampered the replication of previous findings. First, the majority of prior studies were conducted using familial colorectal cancer cases that are enriched for rare high-risk alleles as compared with sporadic cases. Second, rare risk variants are usually population-specific and associations with these variants may not be replicated across different ancestral groups (59). Third, it is possible that previous colorectal cancer studies were underpowered for detecting associations with low-penetrance, rare risk variants. Finally, risk variants may interplay with other genetic and/or environmental factors that could influence their expressivity and pathogenic effect (60). Further studies using admixed populations, leveraging large sample sizes, and evaluating gene–gene and gene–environment interactions are warranted to gain new insights into the pathogenesis of colorectal cancer.

Substantial heterogeneity across study-specific BFs from the BRI and iBRI methods was observed in our study. As rare variants are often population-specific and the number of recruited participants varies widely across the study populations investigated (i.e., MECC has a far smaller sample size than KY or CCFR), it is reasonable to observe variability in rare variation patterns across these different European-descent populations that may lead to divergent risk estimates. It is also possible that environmental or lifestyle factors that vary across populations could modify the associations between genetic variation and risk of colorectal cancer. Caution should be taken when generalizing the predicted risk obtained from one study to the general European-descent population. Nonetheless, the summary BF calculated as the product of the three study-specific BFs highlights an overall magnitude of association that warrants further investigation.

Taken together, the observed gene-level associations with colorectal cancer risk based on the Bayesian approaches are biologically plausible. As shown in Fig. 2, many of the top 25 models from the iBRI method were comprised of risk indices that included at least one variant in multiple DNA repair genes. Thus, our results support the polygenic inheritance model where the risk of colorectal cancer is influenced by multiple low-penetrance genes rather than a few high-penetrance cancer-predisposition genes. This study is the largest exome-focused association analysis of colorectal cancer using array-based technology and takes advantage of a large sample size derived from the combination of three large epidemiologic studies. However, there are several limitations such as the use of exome array that does not provide whole exome coverage and could miss a substantial proportion of very rare genetic variation. We did not investigate large indels that could result in loss-of-function of the entire protein. Also, our study was limited to the investigation of the coding genome and we cannot exclude potentially important effects from rare regulatory variation in noncoding regions. As many of the controls had a family history, the meta OR for colorectal cancer could be biased (presumably downwards); however, the P value for significance is expected to be robust, thus preserving the ranking of variants. Finally, we were unable to assess the generalizability of our results through an appropriate external validation dataset, although the combined analysis of three independent studies was important to identify and replicate potential candidate risk variants and genes for future studies.

In conclusion, the increased power and more efficient model search algorithms to identify rare risk variants provided new insights into the genetic and biological landscape of colorectal cancer by identifying novel colorectal cancer genes involved in DNA repair mechanisms and other key cellular processes such as cell division, signaling and immune regulation. Future work should focus on replicating our findings in independent cohorts and using more comprehensive methodologies such as targeted sequencing or fine-mapping to identify potential novel genetic targets for improved risk prediction.

No disclosures were reported.

M. Matejcic: Data curation, formal analysis, investigation, writing–original draft, writing–review and editing. H.A. Shaban: Data curation, formal analysis. M.W. Quintana: Methodology. F.R. Schumacher: Data curation, investigation. C.K. Edlund: Resources, data curation. L. Naghi: Resources, writing–review and editing. R.K. Pai: Data curation, writing–review and editing. R.W. Haile: Resources, data curation, writing–review and editing. A.J. Levine: Writing–review and editing. D.D. Buchanan: Resources, data curation. M.A. Jenkins: Resources, data curation, funding acquisition, writing–review and editing. J.C. Figueiredo: Writing–review and editing. G. Rennert: Resources, funding acquisition. S.B. Gruber: Writing–review and editing. L. Li: Resources, data curation, funding acquisition. G. Casey: Data curation, methodology, writing–review and editing. D.V. Conti: Conceptualization, supervision, validation, investigation, methodology, writing–review and editing. S.L. Schmit: Conceptualization, formal analysis, supervision, investigation, project administration, writing–review and editing.

The Colon Cancer Family Registry (CCFR) was supported by the NCI of the NIH under award number U01 CA167551. Additional support for case ascertainment was provided from the Surveillance, Epidemiology, and End Results (SEER) Program of the NCI and the following U.S. state cancer registries: AZ, CO, MN, NC, NH; and by the Victoria Cancer Registry (Australia) and Ontario Cancer Registry (Canada). The content of this manuscript does not necessarily reflect the views or policies of the National Cancer Institute or any of the collaborating centers in the Colon CFR, nor does mention of trade names, commercial products, or organizations imply endorsement by the U.S. Government, any cancer registry, or the Colon CFR. The Kentucky Case-Control Study (KY) was supported by Damon Runyon Cancer Research Foundation Clinical Investigator Award, CI-8 (to L. Li); the Case Center for Transdisciplinary Research on Energetics and Cancer, U54 CA-116867-01 (to L. Li); NCI K22 Award, K22 CA120545-01 (to L. Li); State of Ohio Biomedical Research and Technology Transfer Commission; and the NCI Award R25 CA094186-06.

The Molecular Epidemiology of Colorectal Cancer Study (MECC) was supported by the NCI at the NIH: R01 CA197350, R01 CA81488, P30 CA014089, U19 CA148107 (to S.B. Gruber), and P01 CA196569 and a generous gift from Daniel and Maryann Fong. This work was also supported by the National Human Genome Research Institute at the NIH [T32 HG000040] and the National Institute of Environmental Health Sciences at the NIH [T32 ES013678].

The CCFR acknowledges the generous contributions of their study participants, dedication of study staff, and the financial support from the U.S. NCI, without which this important registry would not exist.

We would also like to acknowledge the recipients of the following grants: State of Ohio Biomedical Research and Technology Transfer Commission (to G. Casey), the NCI Award R25 CA094186-06 (to C.L. Thompson), P01 CA196569 (to D.C. Thomas), the National Human Genome Research Institute at the NIH (T32 HG000040, to M. Boehnke), and the National Institute of Environmental Health Sciences at the NIH (T32 ES013678, to W. Gauderman).

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.