INTRODUCTIONOne of the most important risk factors for colorectal cancer is having a family
history of the disease. First-degree relatives of persons diagnosed with colorectal cancer
are, on average, at an approximately two-fold increased risk of colorectal cancer compared
with those without a family history (familial relative risk) (1). An estimated 3% to 5% of colorectal cancers are caused by
high-risk mutations in the identified major colorectal cancer susceptibility genes(2): DNA mismatch repair (MMR) genes(3) and constitutional 3′ end deletions of EPCAM(4, 5) implicated in
Lynch syndrome; the adenomatous polyposis coli (APC) gene implicated in
familial adenomatous polyposis(6–8); and the MUTYH gene implicated in
colorectal polyps and subsequently cancer (MUTYH-associated
polyposis)(9). Current estimates of MMR gene
mutation carriers in the general population, inferred from the prevalence of mutations in
cases and the risk of colorectal cancer for carriers, range widely from approximately 1 in
300 to 1 in 3,000 depending on differing assumptions and genes (10–16). With the
availability of cost-effective sequencing technologies, improved precision in estimates of
mutation prevalence would be useful for devising cost-effective genetic testing
protocols.
Less than half of the excess risk of colorectal cancer associated with family
history (familial aggregation) is explained by mutations in the above identified genes, and
only two studies have attempted to explain the remainder of the familial aggregation (17, 18). Aaltonen
et al could not confidently distinguish between different modes of
inheritance for the hypothetical unidentified major genes (17). Jenkins et al estimated that 1 in 588 of the population
carry major gene mutations associated with a recessively inherited risk, and these mutations
would explain 15% of all colorectal cancers diagnosed before age 45 years (18). Both these studies relied on relatively small
numbers of families and did not consider the existence of both polygenic and major
genes.
While much research has been conducted on the search for other major colorectal
cancer susceptibility genes in addition to those described above, only a few have been
confirmed (19). Genome-wide association studies have
identified at least 45 independent genetic susceptibility markers (single-nucleotide
polymorphisms, SNPs) that are reliably associated with small increments in the risk of
developing colorectal cancer (20).
The aim of this paper was to use population-based family data to estimate: the
prevalence of mutations in the identified major colorectal cancer susceptibility genes (MMR
genes and MUTYH); the prevalence, average penetrance, and likely mode of
inheritance for the unidentified major gene mutations; and the variance of the residual
polygenic component before and after allowing for different major gene scenarios.
MATERIALS AND METHODSSampleThe sample consists of nuclear families from the Colon Cancer Family Registry
which has been described in detail previously(21,22). The present study used data for
the first-degree relatives of the incident colorectal cancer cases (probands) who had been
recruited irrespective of family history from state or regional population cancer
registries in the USA (Washington, California, Arizona, Minnesota, Colorado, New
Hampshire, North Carolina), Australia (Victoria) and Canada (Ontario) between 1997 and
2012. Families were excluded if the proband was known to have an APC
mutation. Informed consent was obtained from all study participants, and the study
protocol was approved by the institutional research ethics review board at each recruiting
site of the Colon Cancer Family Registry.
Data CollectionInformation on demographics, personal characteristics, personal and family
history of cancer, cancer-screening history, history of polyps, polypectomy, and other
surgeries was obtained by questionnaires from all probands at baseline recruitment, which
was about 1–2 years after diagnosis of their colorectal cancer, and from all
participating relatives. The questionnaires are available from the Colon Cancer Family
Registry website(23). We sought confirmation of all
reported cancer diagnoses and ages at diagnosis for relatives using pathology reports,
medical records, cancer registry reports, and death certificates, where possible. We
attempted to obtain blood or buccal samples from all participants and tumor tissue from
all affected participants.
Mismatch Repair (MMR) gene mutation screeningAll probands had their colorectal cancers tested for MMR deficiency, defined by
either tumor microsatellite instability (MSI) and/or lack of MMR protein expression by
immunohistochemistry (IHC). Probands with a MMR-deficient tumor were screened for germline
mutations in MMR genes. MLH1, MSH2 and MSH6 mutations
were identified using Sanger sequencing or denaturing high performance liquid
chromatography (dHPLC), followed by confirmatory DNA sequencing. Large duplication and
deletion mutations including those involving EPCAM, which lead to
MSH2 methylation, were detected by Multiplex Ligation Dependent Probe
Amplification (MLPA) according to the manufacturer’s instructions (MRC Holland,
Amsterdam, The Netherlands) (21,24,25).
PMS2 mutations were identified using a modified protocol from Senter
et al(26) where exons
1–5, 9 and 11–15 were amplified in three long range PCRs followed by
nested exon specific PCR/sequencing. The remaining exons (6, 7, 8 and 10) were amplified
and sequenced directly from genomic DNA. Large-scale deletions in PMS2
were detected using the P008-A1 MLPA kit according to manufacturers specifications (MRC
Holland, Amsterdam, The Netherlands). Germline variants were classified for pathogenicity
based on 5 class system for quantitative assessment of variant pathogenicity(27) and the application of a multifactorial likelihood
model developed for MMR gene variants(28) as
applied to variants catalogued within the InSiGHT database (29) where classes 4 and 5 were considered pathogenic (30). For variants not yet classified by InSiGHT, we
considered a variant as pathogenic if it resulted in a stop codon, frameshift, large
deletion, or if it removed a canonical splice site. The relatives of probands with a
pathogenic MMR germline mutation, who provided a blood sample, underwent testing for the
specific mutation identified in the proband.
<italic>MUTYH</italic> mutation testingPopulation-based probands were tested for 12 previously identified
MUTYH variants: c.536A>G p.(Tyr179Cys), c.1187G>A
p.(Gly396Asp), c.312C>A p.(Tyr104Ter), c.821G>A p.(Arg274Gln),
c.1438G>T p.(Glu480Ter), c.1171C>T p.(Gln391Ter), c.1147delC
p.(Ala385ProfsTer23), c.933+3A>C p.(Gly264TrpfsX7), c.1437_1439delGGA
p.(Glu480del), c.721C>T, p.(Arg241Trp), c.1227_1228dup p.(Glu410GlyfsX43), and
c.1187−2A>G p.(Leu397CysfsX89) using the MassArray MALDI-TOF Mass
Spectrometry (MS) system (Sequenom, San Diego, CA) (31). To confirm the MUTYH mutation and identify additional
mutations, screening of the entire MUTYH coding region, promoter, and
splice site regions was performed on all samples exhibiting MS mobility shifts using
denaturing high-performance liquid chromatography (Transgenomic Wave 3500HT System;
Transgenomic, Omaha, NE). All MS-detected variants and WAVE mobility shifts were submitted
for sequencing for mutation confirmation (ABI PRISM 3130XL Genetic Analyser). That is, if
a heterozygous MUTYH mutation was identified, then the
MUTYH gene was screened for any additional mutations not captured by
the Sequenom genotyping screen to ensure all potential compound heterozygous carriers were
identified. The relatives of probands with a pathogenic MUTYH germline
mutation, who provided a blood sample, underwent testing for the specific variant
identified in the proband. For the present study, MUTYH gene mutation
status was recorded as monoallelic or biallelic mutation-positive or negative, with no
distinction between different variants.
Statistical MethodsWe used modified segregation analysis to fit a range of genetic models to the
observed colorectal cancer family histories for the proband and their first-degree
relatives. Individuals were assumed to be at risk of colorectal cancer from birth until
the earliest of the following: diagnosis of colorectal cancer or any other cancer (except
skin cancer); first polypectomy; death; and the earlier of last known age at baseline
interview or age 80 years.
The colorectal cancer incidence
λi(t,k) for
individual i at age t in sex group k
(k = 1 for males or 2 for females) was assumed to depend on
genotype according to a parametric survival analysis model
λi(t,k)=λ0(t,k)
exp(Gi+Pi(t)), where
λ0(t,k) is the
sex-specific baseline incidence at age t. Gi
is the natural logarithm of the relative risk associated with the major genotype and
Pi(t) is the polygenic component for age
t.
The major genotype was defined by six components representing each of the genes
MLH1, MSH2, MSH6, PMS2, MUTYH and one representing the hypothetical
unidentified major genes. We fitted models in which the unidentified major genes were
autosomal with a normal and a mutant allele unlinked to mutations in the MMR genes or
MUTYH. We also fitted models in which the average relative risk for the
unidentified major genes was assumed to be age dependent. We used the published age-, sex-
and country-specific incidences for MLH1 and MSH2
mutation carriers (32), and published age- and
sex-specific incidences for MSH6, PMS2 and MUTYH
mutation carriers (26, 33, 34).
The polygenic component for age t,
Pi(t), was assumed to be normally
distributed with zero mean and variance
σ2p(t).
P was approximated by the hypergeometric polygenic model (35, 36). We also
fitted models where the variance of the polygenic ‘modifying’ component
was allowed to take a different value σ2m
for MMR gene and MUTYH carriers.
To compute the baseline colorectal cancer incidence
λ0(t), we constrained the overall
incidence of colorectal cancer to agree with the national age- and sex-specific incidences
(1998–2002) separately for Australia, Canada and USA (37). Other cancers were ignored in this model.
We assumed that the sensitivity of the mutation testing of probands for MMR
genes and MUTYH was 80%,(38) and we examined the effect of varying this sensitivity. For relatives, we
assumed the mutation screening for the proband’s mutation (i.e. predictive
testing) was 100% sensitive and specific.
The genetic models were specified in terms of colorectal cancer incidence for
MMR gene and MUTYH mutation carriers, the frequency
(qA) of the putative high risk allele “A” of
the unidentified major genes component, the average relative risk of colorectal cancer for
carriers of mutations in the unidentified major genes, and the variances of the polygenic
and modifying components (σ2p and
σ2m). Maximum likelihood estimation was
used to estimate parameters. The estimates we present are the values that were the most
likely (i.e. most consistent) with the data. Maximum likelihood is the optimal method for
making such estimates, and provides confidence intervals (CIs). We adjusted for
ascertainment by maximizing the likelihood of each pedigree conditioned on the colorectal
cancer status of the proband and his or her age of diagnosis (but not the mutation carrier
status as this information was not known at the time of recruitment).
The relative goodness of fit for nested models was tested by the likelihood
ratio test. The Akaike’s Information Criterion(39) [AIC=−2×log-likelihood +
2×(no. of parameters)] was used to assess goodness of fit between non-
nested models (40).
The expected versus observed number of affected relatives under each fitted
model was assessed using the Pearson χ2 goodness of fit statistic. The
expected number of probands with MMR and MUTYH mutation carriers for
families that had undergone mutation testing based on their cancer family history was
computed using Bayes theorem (41). Statistical
methods are described further in the Appendix.
RESULTSA total of 5,744 families was eligible for inclusion, including 37,634
first-degree relatives of probands of whom 50% were female and 806 (2%) had
been diagnosed with colorectal cancer (Table 1).
Nearly two-thirds of the families were recruited from the USA (63%), with
16% and 21% of families recruited from Australia and Canada, respectively.
Seventy-three percent of the probands were Caucasian whereas the rest were African American
(17%), Asian (6%), Latino (1%), Native American (1%) and
unknown (2%).
Approximately 7% of all probands (N=386) had been found to have a
MMR-deficient colorectal tumour and therefore had been screened for germline mutations in
the MMR genes, while two-thirds of all probands (N=3,796) had been tested for
germline mutations in MUTYH. Of the probands who were screened, 136 had a
MMR gene mutation (49 in MLH1, 39 in MSH2, 24 in
MSH6 and 24 in PMS2) and 81 had a MUTYH
mutation (63 monoallelic and 18 biallelic) (Table
2).
All seven models that incorporated a polygenic component and the hypothetical
unidentified major genes provided significantly better fits than the model that included
only MMR gene and MUTYH mutation carriers (all P<0.001) (Supplementary Table 1). The mixed
dominant model was essentially identical to a mixed codominant model in terms of fit
(likelihood ratio test, P=0.94), but was more parsimonious given it used less
parameters. All other models were rejected when compared with the mixed codominant model
(likelihood ratio test, all P<0.001).
When we allowed the polygenic variance to vary by age, the mixed dominant model
for the unidentified major genes was the most parsimonious (i.e., had the lowest AIC)
compared with all other models fitted (Table 3).
Under this model, we estimated 0.19% (95% CI, 0.04% –
1.08%) of the population carry mutations in unidentified major genes, and these are
associated with on average a 31-fold (95% CI, 12 – 83) increased risk of
colorectal cancer. The estimated variance of the polygenic component was 3.28 for age
<40 years, 0.92 for age 40–49 years, 0.46 for age 50–59 years, 0.79
for age 60–69 years, and 0.52 for age ≥70 years. The proportion of polygenic
variance after adjusting for the identified major genes explained by the unidentified major
genes was 13%, 54%, 58%, 33% and 36% for ages
<40, 40–49, 50–59, 60–69 and ≥70 years, respectively
(Figure 1). The estimated population carrier
frequency for mutations in MLH1, MSH2,
MSH6, PMS2, and monoallelic and biallelic
MUTYH are shown in Table 4.
Table 5 (A) shows the expected versus
observed number of relatives of the probands, who developed colorectal cancer before age 80
years. Consistent with the AIC, the expected numbers from the mixed dominant model is
closest to the observed numbers.
Table 5 (B) shows the expected and observed
number of probands who are mutation carriers for each MMR gene and monoallelic and biallelic
MUTYH mutations. The expected numbers from the mixed dominant model with
an age-dependent polygenic variance were closest to the observed numbers and had the lowest
χ2 compared with other models. In general, all the models closely
predicted the number of mutation carriers.
In all the fitted models above, the sensitivity of mutation testing was fixed at
0.80. When we re-fitted the models assuming the sensitivity was 0.90, the impact was
negligible. Model estimates were virtually identical when the unidentified major genes were
fitted as a separate locus to the MMR mutations and MUTYH (not shown).
Results were not materially different when we restricted analyses to Caucasian
families (not shown). The relative risks for the unidentified major genes did not vary
appreciably by age in the major gene models (not shown). There was virtually no evidence of
a difference between the size of the polygenic variance for non-carriers
σ2p and the modifying variance
σ2m for any of the models (not shown).
DISCUSSIONWe have used a large population-based family data set from the Colon Cancer Family
Registry, and existing penetrance estimates, to produce new estimates of the population
prevalence of high-risk mutations in the identified major susceptibility genes for
colorectal cancer: the DNA mismatch repair genes and MUTYH. We estimated
that 1 in 279 (95% CI, 192 – 403) of the population carry mutations in
mismatch repair genes (MLH1 = 1 in 1946, MSH2
= 1 in 2841, MSH6 = 1 in 758, PMS2
= 1 in 714), and 1 in 45 carry mutations in MUTYH.
Previously, researchers have inferred these carrier frequencies from the carrier
frequency for cases, risk for the general population and risk for mutation carriers (Supplementary Table 2)(10–16). None, except
those estimated by Song et al(16),
were gene specific. Previous estimates of population carrier frequencies for the four MMR
mutations combined (or MLH1 and MSH2 mutations combined)
were similar to our estimates, except for those obtained by Dunlop et
al(11). This discrepancy might be
explained by different screening methods, and that knowledge about which mutations are truly
pathogenic has improved substantially over time (30).
For MUTYH mutations, a systematic review and meta-analysis estimated the
population carrier frequency of monoallelic MUTYH mutations to be 1 in 60
and biallelic MUTYH mutations to be approximately 1 in 7,000, similar to
our estimates(42).
We then sought to explain the residual familial aggregation of this disease. We
considered a polygenic component that proposes there are multiple independent loci, and
across loci and at each locus, the alleles have a multiplicative effect on risk. We also
considered the existence of one or more unidentified major genes (genes for which there are
mutations associated with a high risk of colorectal cancer), and allowed for different modes
of disease inheritance (dominant, recessive and codominant).
We found evidence that there exist as yet unidentified major colorectal cancer
susceptibility genes, and their mode of inheritance was most likely dominant (thought this
does not necessarily mean that they were all dominant). It is important to note that the
apparent dominant component might also reflect missed mutations in MMR genes,
MUTYH or APC because the mutation screening techniques
used were not 100% sensitive and not all probands had been screened. We estimated
that the 1 in 504 (95% CI, 93 – 2778) of the population carry unidentified
mutations associated with an average 31-fold increased risk of colorectal cancer. The
estimated polygenic variance was reduced by 30–50% after allowing for these
unidentified major genes, after which it decreased from 3.3 for age <40 years to 0.5
for age ≥70 years (equivalent to sibling relative risks of 5.1 to 1.3,
respectively).
The term ‘missing heritability’ has been variously defined over
the last decade to refer to the fact that not all the causes of familial aggregation, or of
familial aggregation considered to be due to genetic factors, have been found (43). The latter has been addressed by assuming an
all-or-nothing unmeasured liability model that makes untestable assumptions (44). For the purposes of discussion here, we assume that
heritability encapsulates both genetic and non-genetic causes of familial aggregation. In
this regard, it is plausible for common cancers that non-trivial heritability is due to
non-genetic factors (45). In this paper, we have
fitted a polygenic component to capture familial aggregation not explained by the major
genes. It is based on an underlying genetic model of Fisher (1918)(46), but given are studying nuclear families it also represents
non-genetic familial factors. That is, although it is labelled polygenic, it could also
reflect the effect of environmental and lifestyle factors shared by first-degree relatives.
Given that the polygenic variance is proportional to the log of the familial relative risk
attributable to the polygenic component, the unidentified major genes might explain
one-third to one-half of the missing heritability of colorectal cancer across the ages of 40
to 70 years.
The polygenic component will also capture the currently identified, and as yet
unidentified, common SNPs associated with colorectal cancer risk. For example, the current
45 independent susceptibility SNPs explain 22% of familial aggregation (20). It is likely this proportion will increase as larger
studies are conducted, such as the OncoArray initiative, and as more informative statistical
strategies are used to devise risk-prediction SNP-based scores other than the current highly
conservative paradigm of considering each SNP individually and applying stringent penalties
for multiple testing. The common SNPs identified to date are not necessarily causal, and
they could also be tagging rare causal variants (as was the case for HOXB13
and prostate cancer (47)).
Our analyses suggest a role for rare variants in as yet undiscovered
susceptibility genes associated with high risk. Individually they could be very rare, and
difficult to discover. One recent attempt to resolve this issue was a whole exome sequencing
study that identified some high-risk mutations in candidate susceptibility genes such as
POT1, POLE2 and MRE11 (19). The authors concluded that the study “probably
discounts the existence of further major high-penetrance susceptibility genes, which
individually account for >1% of the familial risk”. Therefore, if
both their and our findings are correct, there is likely to be perhaps hundreds of major
genes each contributing little to the missing heritability. As well as sample size, the
authors recognized that restriction to exomes limited their ability to identify pathogenic
mutations outside of transcribed regions, and that targeted capture is insufficiently
sensitive to detected copy number variation. We, therefore, agree with the authors in their
conclusion that there is a need for very large-scale sequencing studies that would benefit
from including highly informative families.
Strengths of our study include a large number of families ascertained regardless
of a family history, standardized questionnaires and protocols used by the Colon Cancer
Family Registry, and sophisticated statistical techniques that properly adjust for
ascertainment and account for residual familial aggregation of disease (thereby avoiding
bias). We also used a systematic approach for screening and testing of germline mutations in
both MMR genes and MUTYH.
When predicting the number of relatives with colorectal cancer, we did not
differentiate family history of colorectal cancer in terms of tumor location within the
bowel. This approach was supported by findings from a large study in Utah, which reported
similarly elevated risks of colorectal cancer associated with a family history of colorectal
cancer regardless of tumor location (proximal colon, distal colon, and rectum) (48).
The response of the population-based probands approached to participate was
72% (49). MMR gene and MUTYH
mutation carriers have both been associated with better colorectal cancer survival than
non-carriers (50–52). Therefore, if probands with better prognosis are more likely
to participate in the study, survivor bias could potentially lead to an overestimation of
the mutation frequency. Data on participation differences by prognostic characteristics were
not available to assess this.
A potential limitation of our study is inaccurate reporting of family colorectal
cancer history. Of the 806 colorectal cancer diagnoses reported by first-degree relatives,
26% were confirmed by pathology reports, clinic records or cancer registries.
Previous studies have found reported colorectal cancer history in first-degree relatives to
be reasonably accurate (85–90% agreement)(53) so even though the colorectal cancer diagnoses in relatives were not
confirmed, it is unlikely to have a great impact on our results.
Another potential limitation of our study is the reliance on external estimates of
colorectal cancer relative risks for carriers of MMR gene and MUTYH
mutations. To help mitigate this weakness, we used estimates based on the largest studies
available, and all used data from the same source, the Colon Cancer Family Registry (26, 32–34). Future studies should focus
on incorporating the explicit effects of other colorectal cancer susceptibility genes such
as STK11(54)
BMPR1A(55), SMAD4,
PTEN(56), POLE and
POLD1(57) as well as the explicit
effects of identified common low risk alleles(20). In
addition to colorectal cancer risk, it is known that MMR gene mutations increase the risks
of other cancers such as endometrial and ovarian cancer (58). Our analyses can be extended to incorporate such information.
The polygenic variance describes the range of familial risk across a population at
a given age. For example, given the estimated variances by age for the mixed dominant model,
the familial relative risk was 5.1, 1.6, 1.3, 1.5 and 1.3 for ages <40,
40–49, 50–59, 60–69 and ≥70 years, respectively. Although we
found no evidence that the polygenic effects differed for carriers of a MMR gene mutation
compared with non-carriers, this does not imply that they are due to the same variants. Some
studies have shown that the common genetic variants identified through GWAS to be associated
with the risk for the general population are not relevant for MMR gene mutation carriers
(59). If future studies identify specific genetic
modifiers of colorectal cancer risk for MMR gene or MUTYH mutation
carriers, it should be possible to extend the current analyses to allow for this level of
complexity.
In conclusion, we have used a large population-based family study to estimate the
prevalence of mutations in the identified major colorectal cancer-susceptibility genes, as
well as the prevalence and relative risk of yet-to-be-discovered, high-risk susceptibility
genes. This is an essential step in the development of a high quality-risk prediction model
for colorectal cancer and is a major clinical and public health goal. Subsequently,
screening programs can be optimized at an individual level to attain maximum benefit,
however that may be defined. This study also provides a guidepost for future new gene
discovery research and will justify, and guide, the use of next-generation sequencing to
find these genes. The results show that our current understanding of hereditary
predisposition to colorectal cancer is incomplete and supports the existence of yet
undiscovered rare but highly penetrant mutations, while also underscoring that the polygenic
component is still largely unresolved.