Introduction
Familial clustering of breast cancer is well recognized, having been described over 140 years ago [
1]; the familial relative risk of breast cancer is on average about twofold and is higher among relatives of patients with early-onset cases [
2,
3]. Three classes of breast cancer susceptibility sequence variants with different levels of risk and prevalence in the population are now well established [
4,
5]: rare high-risk variants, such as protein-truncating mutations in
BRCA1,
BRCA2,
PTEN and
TP53 (Mendelian Inheritance in Man numbers (MIMs) 113705, 600185, 601728 and 191170, respectively); rare intermediate-risk variants, such as protein-truncating mutations in
ATM [
6,
7],
BRIP1 [
8],
CHEK2 [
9] and
PALB2 [
10,
11] (MIMs 208900, 605882, 604373 and 610355 respectively); and, more recently, common modest penetrance variants such as the risk single-nucleotide polymorphisms (SNPs) detected by genome-wide association studies (GWASs) in
FGFR2,
TOX3 (
TNRC9),
MAP3K1 and
LSP1 [
12‐
14] (MIMs 176943, 611416, 600982 and 153432, respectively). High-risk variants in the known major breast cancer susceptibility genes
BRCA1,
BRCA2,
TP53 and
PTEN account for approximately 20% to 25% of the familial risk of breast cancer, and adding the known intermediate-risk genes increases the proportion by perhaps 1% for each gene [
15]. Moreover, the panoply of known modest-risk SNPs account for about 8% of the familial relative risk [
16]. Thus known genetic effects account for about one-third of the familial relative risk of breast cancer, leaving two-thirds unaccounted for, a phenomenon referred to as the "problem of missing heritability." Some of this so-called missing "heritability" is of course due to the familial component of environmental risk factors; the measured surrogates for these factors probably explain about 5% of the familial relative risk, but if measured more specifically and more precisely, they may explain considerably more familial aggregation [
17].
The gene
CHEK2 encodes a serine/threonine kinase, CHK2, that functions in the signaling pathways activated by DNA damage, particularly DNA double-stranded breaks [
18]. Inheritance of a
CHEK2 protein-truncating mutation such as the relatively well investigated Northern European founder mutation
c.1100delC confers a two- to threefold increased risk of breast cancer, an increased risk of a number of other cancer types and perhaps a decreased risk of some smoking-related cancers [
9,
19‐
21]. Some missense substitutions in
CHEK2 also alter cancer risk, as exemplified by the Ashkenazi
CHEK2 missense substitution p.S428F and the Slavic substitution p.I157T [
22‐
26]. Most large-scale genetic studies of
CHEK2 conducted to date have focused on genotyping known variants, such as founder mutations. Consequently, there has been little opportunity to assess the role of the potentially more numerous, rarer variants of this gene.
During the 1990s, linkage analysis proved to be an effective genome-wide approach for finding high-risk susceptibility genes for breast and colon cancer. Over the past few years, GWASs have proved to be an effective genome-wide approach to finding common, not necessarily causal, SNPs associated with modest risk. Case-control mutation screening, or its quantitative trait homolog of comparative mutation screening of individuals from the opposite ends of a trait spectrum, is emerging as a useful strategy for identifying and characterizing intermediate-risk susceptibility genes [
6‐
8,
10,
27‐
29]. While case-control mutation screening has been, to date, too technically demanding to examine a whole biochemical pathway, let alone the entire exome, one can imagine combining exon hybridization capture and massively parallel sequencing to accomplish such a study design. Beyond the laboratory challenge imposed by the implied scale of resequencing, a second challenge is to conduct a statistically powerful analysis of the large number of rare sequence variants that would be revealed if such a study design were applied to a common disease such as breast or colon cancer. Previously, we used data from mutation screening of
ATM in breast cancer patients and controls to demonstrate the ability to detect evidence of pathogenicity from both truncating and splice junction variants (T+SJV) and rare missense substitutions (rMS) [
7]. Here we apply the same analytic strategy to
CHEK2 and then extrapolate the results to determine the requirements for much larger-scale studies.
Materials and methods
Ethics statement
The CHEK2 mutation-screening studies and analyses described here were approved by the institutional review board (IRB) of the International Agency for Research on Cancer, the University of Utah IRB and the local IRBs of the Breast Cancer Family Registry (Breast CFR) centers from which we received samples. All participants gave written, informed consent.
Subjects
Patients were selected from among women gathered by population-based sampling by the Breast CFRs at three centers (Cancer Care Ontario, the Cancer Prevention Institute of California (formerly the Northern California Cancer Center) and the University of Melbourne) [
30]. Patients were recruited between 1995 and 2005.
Selection criteria for cases (N = 1,313) were diagnosis at or before age 45 years and self-reported race or ethnicity plus grandparents' country of origin consistent with Caucasian, East Asian, Hispanic/Latino or African American racial or ethnic heritage.
The controls (N = 1,123) were frequency matched to cases within each center on racial or ethnic group, with age at selection not more than ± 10 years difference the age range at diagnosis of the patients gathered from the same center. Because of the shortage of available controls in some ethnic and age groups, the frequency matching was not one-to-one in all subgroups.
Mutation screening
Mutation screening started from whole-genome amplified (WGA) DNA for coding exons 1-9 and from genomic DNA for exons 10-14. A nested polymerase chain reaction (PCR) strategy was used, followed by high-resolution melting (HRM) curve analysis [
31,
32] and then dye terminator resequencing of samples that contained a melt curve aberration indicative of the presence of a sequence variant. For
CHEK2 amplicons harboring SNPs with a frequency ≥1% in either the Single Nucleotide Polymorphism Database (dbSNP) [
33] or initial amplicon testing, we applied a simultaneous mutation scanning and genotyping approach using HRM curve analysis to improve the sensitivity and efficiency of the mutation screening [
34]. The laboratory process used was the same as that described in detail for our recent case-control mutation screening for
ATM [
7], except that primary PCR assays for
CHEK2 exons 10-14 (which are involved in a subtelomeric repeat) relied on a long-range PCR assay as described by Sodha
et al. [
35].
All exonic sequence variants, plus splice junction consensus sequence variants that reduced splice junction sequence similarity to the standard consensus sequences AG^GTRRGT (donor) or Y16NYAG^ (acceptor) (where ^ indicates the position of the splice junction), were reamplified from genomic DNA for confirmation of the presence of the variant. Because of the presence of pseudogenes that partially matched the sequence of the CHEK2 long-range PCR exons (exons 10-14), sequence variants identified within these exons were subsequently tested using allele-specific PCR assays for the primary PCR to confirm that the sequence variants initially identified were true CHEK2 variants. To ensure amplification of the CHEK2 DNA sequence and not amplification of the potentially interfering CHEK2 pseudogenes, the positions of the specific primers were chosen so that the 3' extremity bases perfectly matched the CHEK2 wild-type sequence, while they mismatched the corresponding position of the pseudogenes.
All samples that failed at the primary PCR, secondary PCR or sequencing reaction stage were reamplified from WGA DNAs or genomic DNA. Samples that still did not provide satisfactory mutation-screening results for at least 80% of the
CHEK2 coding sequence were excluded from further analyses (
n = 24). Process and data management of the mutation screening were carried out as described by Voegele
et al. [
36]. Primer and probe sequences are available from FLCK upon request.
Alignments and scoring of missense substitutions
Previously, we used the T-Coffee (Tree-based Consistency Objective Function for alignment Evaluation) software suite of alignment tools [
37,
38] to prepare a CHK2 protein multiple sequence alignment in which the most diverged sequence was from sea urchin (
Strongylocentrotus purpuratus) to analyze a small number of
CHEK2 missense substitutions and in-frame deletions [
39]. We updated this alignment by replacing the partial pufferfish (
Tetraodon nigroviridis) sequence with a full-length zebrafish (
Danio rerio) sequence and including predicted CHK2 sequences from elephant (
Loxodonta africana), platypus (
Ornithorhynchus anatinus), tunicate (
Ciona intestinalis) and fruit fly (
Drosophila melanogaster). The alignment was characterized by (1) determining percentage sequence identity between each pair of sequences in the alignment, (2) using the Protpars routine of Phylogeny Inference Package version 3.2 software (PHYLIP; free software developed by Felsenstein [
40]) to make a maximum parsimony estimate of the number of substitutions that occurred along each clade of the underlying phylogeny and (3) recording the "median sequence conservation score" reported by the missense substitution analysis program Sorting Intolerant from Tolerant (SIFT) [
41,
42]. The sequence alignment, or updated versions thereof, is available at the Align-GVGD website [
43]. Missense substitutions observed during our mutation screening of
CHEK2 were scored using the Align-GVGD [
43‐
45] and SIFT [
41,
42] software programs with our curated alignments and with Polymorphism Phenotyping version 2 software, or PolyPhen-2, using its precompiled alignment [
46,
47].
Statistical analysis and power calculations
To assess risk associations using the case-control frequency distribution of T+SJVs and rMSs, we constructed a single table with one entry per participant; zero or one rare sequence variant per participant; and annotations for type of sequence variants, study center, case-control status, race or ethnicity, and age. For the two participants who carried more than one rare variant of interest (one participant carried p.I448S (C15) plus p.E394D (C35), and one participant carried p.E239K (C15) plus p.R346H (C25)), only the variant belonging to the more likely evolutionarily deleterious grade (that is, higher C-number as scored by Align-GVGD) was considered.
Most analyses were performed using multivariable unconditional logistic regression using Stata version 11 software (StataCorp, College Station, TX, USA). Differences in the case-control ratio between ethnic groups and age categories were accounted for by including categorical variables for each age category and ethnic group. Adjustment was also made for study center. We explored the possibility of interactions between ethnic group and study center, checking both improvement of model fit by the likelihood ratio statistic and comparing the estimates of the parameter of interest (log odds ratio (OR) per Align-GVGD grade) in different models. Adjustment for ethnic group should also capture confounding of genetic and social factors with interaction terms, allowing that this confounding effect may be different for the broadly labeled ethnic groups in different centers. Because the Breast CFR matched cases and controls for age in 5-year categories, and because the maximum age of Breast CFR patients included in this study was 45 years, all participants ages 41 years and older (at diagnosis for patients and at ascertainment for controls) were combined into a single age category.
Logistic regression trend tests were formatted such that participants who did not carry any T+SJV or any rMS, as well as carriers of the seven grades of rMSs (C0, C15, C25, C35, C45, C55 and C65) defined by Align-GVGD [
45], were assigned the default row labels 0, 1, 2, 3, 4, 5, 6 and 7, respectively. These row labels were then used as a continuous variable in the logistic regression analyses. Regression coefficients and trend test
P values (
Ptrend) were estimated from the resulting lognormal ORs using the logit function of Stata software. Carriers of T+SJVs were analyzed against the same noncarrier group defined above. Two strategies were used to combine evidence of association with T+SJV and rMS variants: (1) carriers of T+SJVs were combined with carriers of C65 rMSs in category 7, and (2) T+SJV carriers were assigned row label 8. We used the Fisher's exact test to obtain the lower bound of the 95% confidence interval (95% CI) for associations with categories that contained one or more patients but zero controls.
Post hoc power calculations were performed by specifying a hypothetical OR and population prevalence for each class of variant, together with the cumulative probability of breast cancer prior to age 70 years. The ORs and control carrier frequencies that we specified for the individual grades of sequence variants, relative to the noncarriers, were based on data from the population-based Breast CFR sample series. For the grades for which there were a reasonable number of observations, that is, C0, C15, C25, C65 and T+SJV, we used the adjusted ORs and observed carrier frequencies. Because of the very low numbers of observations in grades C35-C55, ORs for these categories were estimated from the logistic regression OR coefficient and population carrier frequencies defined to obtain the specified OR, given the number of observations in patients. On the basis of these OR and frequency estimates, we calculated expected values and variances of the test statistics for the types of test considered: Pearson's χ2 test for the two-category tests and the Wald statistic from a logistic regression for the trend test. We then calculated the probability of these statistics exceeding a series of desired P value thresholds using a normal approximation.
Attributable fractions were estimated according to the method described by Greenland [
48], and familial relative risks were estimated according to the methods described by Goldgar [
49]. Both calculations used the same frequency and risk association estimates as those used for the
post hoc power calculations.
Discussion
That protein-truncating variants in
CHEK2 confer a moderately increased risk of breast cancer is well established. The OR that we observed for T+SJVs is numerically somewhat higher than that reported in the 2004 CHEK2 Breast Cancer Case-Control Consortium study of
c.1100delC [
50], but not significantly, as our 95% CIs do include the point estimate from that study. Moreover, as previous studies have observed higher ORs for
c.1100delC in familial versus sporadic cases and in early-onset versus later-onset cases [
9,
50], we should expect that this study's focus on early-onset breast cancer cases with oversampling of familial cases would result in relatively high OR estimates.
Previous studies have shown that some CHEK2 missense substitutions are pathogenic, but the scale of their contribution to breast cancer susceptibility relative to that of T+SJVs is not known. Although we hesitate to extrapolate our current data to true population-attributable risks (within the age groups that we sampled) or familial relative risks, the data do provide a basis on which to compare the relative contributions of these two classes of variants. Working from the control carrier frequencies and the OR point estimates (adjusted for race or ethnicity, study center, and age) observed from the population-based Breast CFR sample series, we calculate attributable fractions of 0.014 for T+SJVs as compared with 0.015 for the sum of C15-C65 rMSs. In addition, we calculate a familial relative risk among first-degree relatives of 1.036 for T+SJVs as compared with 1.033 for a product across the C15-C65 rMSs. Thus, as a first approximation, the attributable fractions and familial relative risks of truncating variants and rare missense substitutions are virtually identical. It is important to remember that these attributable fraction and familial relative risk point estimates are inflated compared with those that would be obtained from a population-based study that included patients diagnosed in their 70s or older. In addition, as more than 25% of the T+SJVs observed in this study were nonsense and frame shift mutations other than c.1100delC, these data also speak to the importance of full open reading frame mutation screening to observe the majority of genetically relevant sequence variants in this cancer susceptibility gene.
Several of the missense substitutions observed in this study have been subjected to functional assays in one or more published works. For the 14 missense substitutions that Align-GVGD scored C0 and which we would consequently predict to be neutral or nearly so, assay results have been reported for 4 (p.P85L, p.R137Q, p.R180H and p.T323P). Using a
Saccharomyces cerevisiae Rad53 complementation assay, Shaag
et al. [
22] found that
p.P85L is equivalent to wild-type
CHEK2. While Bell
et al. [
55] found this allele to have modestly reduced activity in an
in vitro kinase function assay, both Bell
et al. and Shaag
et al. concluded that the allele is effectively neutral. Sodha
et al. [
39] assayed the
p.R137Q allele and found that it encodes a protein with normal stability and normal response to DNA damage. Bell
et al. [
55] also assayed the
p.R137Q allele and found that it has normal kinase activity. In addition, Sodha
et al. [
39] assayed the
p.R180H allele and found that it encodes a protein with slightly reduced stability but normal response to DNA damage. Thus existing functional assay results for these three variants are consistent with their being either neutral or at most weakly pathogenic. Wu
et al. [
56] found the fourth C0 substitution, p.T323P, to have moderately reduced autophosphorylation and Cdc25C kinase activity. Classification of this substation as C0 is probably a true Align-GVGD error, because the crystal structure of the protein reveals that T323 is located in an α-helix, which would not typically be permissive of substitution to proline. The algorithmic problem is that the atomic composition and polarity of proline (the amino acid side chain characteristics considered by the original Grantham difference [
57] and Align-GVGD are atomic composition, polarity and volume) are intermediate between those of threonine and isoleucine, which are the two amino acids observed at position 323 in our alignment. The consequence is that proline is only slightly outside the range of variation represented by these two wild-type residues and is consequently predicted to be neutral or nearly so. Although unpublished, misclassification of substitutions to proline that map within an α-helix is a problem that we have observed before and is an obvious issue to bear in mind when considering missense substitution analyses made using Align-GVGD. p.I157T is perhaps the most interesting of the substitutions observed in our study that have been subjected to functional assays. Align-GVGD scores the variant as C15, indicative of modest evidence in favor of pathogenicity. Initially, Lee
et al. [
58] found that kinase activity of the
p.I157T allele was comparable to the wild type. More recent studies have reported that the allele is at least partially defective in dimerization and autophosphorylation, binding and phosphorylating Cdc25, and binding
BRCA1 [
59‐
62]. In populations in which
p.I157T and
c.1100delC are both present at appreciable frequencies and have been subject to independent risk estimates,
p.I157T does appear to confer increased risk of breast cancer, but the OR or penetrance associated with the missense substitution appears to be more modest than that associated with the frame shift
c.1100delC [
63]. At the other end of the spectrum, of the five C65 substitutions that we observed, only one, p.R117G, has been subjected to functional assays. Summing across several studies, the protein encoded by this allele is phosphorylated by ATM in response to DNA damage, shows slightly to markedly reduced autophosphorylation, probably fails to oligomerize and has severely compromised kinase activity toward Cdc25C [
39,
56,
62]. Therefore, the
p.R117G allele encodes a functionally defective protein and is in all likelihood pathogenic. Thus, for the missense substitutions that were observed in our mutation-screening study and subjected to functional assays, there is a qualitative trend toward agreement between the Align-GVGD classification and the functional assay result, consistent with the trend in ORs that we observed across the Align-GVGD-defined ordered series of missense substitution grades. However, since concordant results between
in silico assessments and functional assays are not yet considered sufficient for formal clinical classification of missense substitutions observed in
BRCA1 and
BRCA2 [
64‐
66], it does not appear that the state-of-the-art of
CHK2 functional assays has reached the point at which concordant results from an
in silico assessment and a functional assay would be sufficient for clinically relevant classification of a
CHEK2 missense substitution.
The genetic results described in this work, combined with the above functional assay summary, have implications for potential clinical genetic susceptibility tests that might include
CHEK2 and other genes with similar mutation profiles. In the 2003 American Society of Clinical Oncology Policy Statement Update on Genetic Testing for Cancer Susceptibility, the second and third "indications for genetic testing for cancer susceptibility" were that "2) the genetic test can be adequately interpreted, and 3) the test results will aid in diagnosis or influence the medical or surgical management of the patient or family members at hereditary risk of cancer" (pp. 2398) [
67]. With regard to the third criterion, some investigators have argued that in the context of a high-risk family, the difference in risk between carriers and noncarriers of clearly pathogenic
CHEK2 sequence variants is sufficient to justify a difference in cancer surveillance strategies [
68‐
70]. However, our results in addition to similar work regarding
ATM [
7,
71] point toward an issue under the second criterion. If roughly one-half of the genetically relevant risk that the test can pick up actually resides in rare missense substitutions that will be considered unclassified variants at their initial detection, it may not currently be possible to adequately interpret the test results. Therefore, while it is now technically feasible to design a massively parallel sequencing-based test that can accurately and relatively inexpensively identify mutations in a panel of breast cancer susceptibility genes that includes
ATM and
CHEK2 [
72], it may be inappropriate to introduce such a test into widespread use before a clinically validated method of assessing unclassified missense substitutions in these genes has been developed.
The rare missense substitution analysis model combining Align-GVGD with the logistic regression test for trends grew out of the
in silico analysis of missense substitutions that has now become a standard component in the integrated evaluation of unclassified variants in
BRCA1 and
BRCA2 [
65,
73]. We proposed the model on the basis of clinical
BRCA1 and
BRCA2 mutation-screening data and then demonstrated its effectiveness by an analysis of
ATM case-control mutation-screening data [
7,
45]. Thus the
CHEK2 analysis presented here stands as a methodological confirmation of our approach to the inclusion of rare missense substitution data in case-control mutation-screening studies. The logistic regression test for trends that we used also provides a simple approach to combining evidence from rare missense substitutions with evidence from protein-truncating sequence variants to build a more complete and statistically powerful approach to assessing case-control mutation-screening data than would be afforded by either method alone. From a technological perspective, we can envision combining exon capture and massively parallel sequencing to extend case-control mutation screening to entire biochemical pathways and beyond. On the basis of our
post hoc power calculations, at least 2,000 patients and 2,000 controls would be required for a whole pathway (such as DNA double-stranded break repair and allied cell cycle checkpoints) study, and 3,300 patients and 3,300 controls would be required to undertake a whole-exome study. On the one hand, these numbers could be an underestimate because
CHEK2 might be among the most important (in terms of familial relative risk) of the intermediate-risk class of breast cancer susceptibility genes. On the other hand, it could turn out that a test based on observations of evolutionarily unlikely sequence variants has an intrinsically lower false-positive rate than anonymous marker GWASs and consequently would not require a full Bonferroni multiple testing correction to reasonably constrain the rate of false-positive results.
Competing interests
The authors declare that they have no competing interests.
Authors' contributions
FLCK contributed to study design, led the laboratory team and helped to draft the manuscript. FL contributed to study design, led the data analysis and helped to draft the manuscript. FD contributed to the mutation screening and data analysis and helped to refine the laboratory platform. MV contributed to the sequence alignment and data analysis. CV was responsible for data management throughput for the project and helped to refine the laboratory platform. DB contributed to the sequence alignment and method for analysis of rare missense substitutions. GD contributed to the mutation screening and data analysis and helped to refine the laboratory platform. NF contributed to the mutation screening and data analysis and helped to refine the laboratory platform. SMC contributed to the mutation screening and data analysis and helped to refine the laboratory platform. NR contributed to the mutation screening and data analysis and helped to refine the laboratory platform. TND contributed to the sequence alignment and data analysis. AT contributed to statistical analyses and helped to draft the manuscript. GBB contributed to statistical analyses and helped to draft the manuscript. JLH was responsible for patients gathered through the University of Melbourne and helped to draft the manuscript. MCS contributed to study design and contributed to the management of samples obtained through the University of Melbourne. ILA was responsible for patients gathered through Cancer Care Ontario. EMJ was responsible for patients gathered through the Northern California Cancer Center (now the Cancer Prevention Institute of California). SVT was responsible for overall study design, contributed to data analysis and helped to draft the manuscript. All authors read and approved the final manuscript.