Abstract

Background

Lynch syndrome is caused by inherited mutations in DNA mismatch repair genes (primarily MSH2 , MLH1 , MSH6 , and PMS2 ) and is one of the most prevalent inherited cancer syndromes. Several models have been developed to predict the occurrence of Lynch syndrome in high-risk patients and families, but it is not known how these models compare with one another or how they perform for colorectal cancer patients from the general population. We used data from such patients to test the ability of four models—Leiden, MMRpredict, PREMM 1,2 , and MMRpro—to distinguish between those who did and did not carry DNA mismatch repair gene mutations.

Methods

We studied a consecutive series of 725 patients who were younger than 75 years at colorectal cancer diagnosis and whose DNA mismatch repair gene mutation status was available; 18 of the 725 patients carried such a mutation. For each model, we calculated the risk score, compared the observed number of mutations with the expected number, and determined the receiver operating characteristics. All statistical tests were two-sided.

Results

Although all four models overestimated the probability of a mutation (range = 1.2- to 4.3-fold), especially in low-risk patients, they could discriminate between carriers and noncarriers of a mismatch repair mutation. The areas under the receiver operating characteristics curves from the four models ranged from 0.91 to 0.96. Carriers of mutations in the MSH6 or PMS2 genes had lower risk scores than carriers of MSH2 or MLH1 mutations. For example, the MMRpredict model gave median risk scores of 24% and 94% ( P < .015) for MSH6PMS2 and MSH2MLH1 mutation carriers, respectively. For the Leiden, MMRpredict, and PREMM 1,2 models, correcting the risk scores for bias introduced by family size improved their power to discriminate between carriers and noncarriers. After correcting for family size, the best model was MMRpredict, which achieved a sensitivity of 94% (95% confidence interval [CI] = 73% to 99%) and a specificity of 91% (95% CI = 88% to 93%) and identified a smaller proportion of patients than the revised Bethesda criteria as those who should undergo additional molecular or immunohistochemical testing (11% vs 50%).

Conclusion

MMRpredict was the best-performing model for identifying colorectal cancer patients who are at high risk of carrying a DNA mismatch repair gene mutation and thus should be screened for Lynch syndrome.

Context and Caveats
Prior knowledge

Inherited mutations in DNA mismatch repair genes (primarily MSH2 , MLH1 , MSH6 , and PMS2 ) cause Lynch syndrome, which is one of the most prevalent inherited cancer syndromes. Several models have been developed that predict the occurrence of Lynch syndrome in high-risk patients and families.

Study design

Data from 725 patients who were younger than 75 years at colorectal cancer diagnosis and whose DNA mismatch repair gene mutation status was available (18 of whom carried such a mutation) were used to test the ability of four models—Leiden, MMRpredict, PREMM 1,2 , and MMRpro—to distinguish between those who did and did not carry DNA mismatch repair gene mutations.

Contribution

All four models overestimated the probability of a mutation, especially in low-risk patients, but could discriminate between carriers and noncarriers of mismatch repair mutations. The best model was MMRpredict when corrected for family size. This model identified fewer patients who should be screened for Lynch syndrome than previous models or criteria.

Implications

MMRpredict should be tested further as a prediction model for Lynch syndrome in the general population of patients with colorectal cancer.

Limitations

The cohort contained only 18 carriers of a DNA mismatch repair gene mutation. Patients who were older than 74 years when diagnosed with colorectal cancer were not included in this study.

From the Editors

There are several forms of inherited colorectal cancer in which disease susceptibility is inherited as a Mendelian (ie, single gene) trait and the genetic cause has been well characterized [as reviewed by Rustgi ( 1 )]. The most common type of inherited colorectal cancer has been referred to as either Lynch syndrome or hereditary nonpolyposis colorectal cancer (HNPCC). The Amsterdam criteria have been used to define HNPCC ( 2 , 3 ). Mutations in genes involved in the DNA mismatch repair pathway appear to be responsible for colorectal cancer in 1%–3% of patients ( 4 , 5 ), although there are few good population-based studies of the incidence. It has been proposed that the term “Lynch syndrome” be restricted to HNPCC patients who carry pathogenic mutations in DNA mismatch repair genes (ie, mutations that interfere with the DNA mismatch repair system) and that the term “familial colorectal cancer, type X” be used for patients from HNPCC families who have tumors that are proficient in DNA mismatch repair ( 6 ). Tumors from patients with Lynch syndrome contain DNA with microsatellite instability (MSI). MSI is also observed in DNA from 10% to 15% of sporadic (noninherited) colorectal cancers, in which it arises from the somatic inactivation of the mismatch repair gene, MLH1 , by methylation of its promoter. In the clinical setting, it is important to identify patients with Lynch syndrome because their clinical management differs from that of patients with sporadic colorectal cancer and because members of their extended family should receive genetic counseling, clinical screening, and DNA testing to identify presymptomatic mutation carriers.

The Amsterdam criteria ( 2 ) were designed as a research tool to identify HNPCC families with high specificity, but these criteria lacked sufficient sensitivity to be useful in the clinical setting. The criteria were modified as the Amsterdam II criteria ( 3 ) to increase the sensitivity of the criteria somewhat by including certain extracolonic cancers (such as cancers of the endometrium, small intestine, ureter, and renal pelvis). Subsequently, the Bethesda guidelines ( 7 ) and then the revised Bethesda guidelines ( 8 ) were introduced to identify patients whose tumors warranted molecular testing to determine whether the MSI phenotype was present. However, because molecular testing of all colorectal cancer patients as a screening tool would be too expensive—and not particularly efficient—and because the revised Bethesda guidelines lack adequate specificity, several models have been developed to aid in calculating the probability that a particular patient with colorectal cancer has Lynch syndrome. These models include the Leiden model (9), the MMRpredict model ( 10 ), the PREMM 1,2 model ( 11 ), the MMRpro model ( 12 ), and the AIFEG model ( 13 ).

All five models were developed with data from high-risk patients who were ascertained because they had a strong family history of colorectal cancer or were diagnosed at a young age. It is, however, not known whether these models can predict the risk of Lynch syndrome in colorectal cancer patients from the general population. We tested the diagnostic utility of four models (Leiden, MMRpredict, PREMM 1,2 , and MMRpro) in distinguishing colorectal cancer patients in the general population who had Lynch syndrome (as defined by an inherited DNA mismatch repair gene mutation) from patients with sporadic colorectal cancer. We also compared the performance of each model with that of the Amsterdam and Bethesda criteria and investigated whether variations in family size lead to biases in risk prediction.

Subjects and Methods

Subjects

The colorectal cancer patients and their families who contributed data to this analysis were identified through the Newfoundland Colorectal Cancer Registry. This registry contains a series of 748 consecutive patients with colorectal cancer who were younger than 75 years at diagnosis during the 5-year period from January 1, 1999, through December 31, 2003 ( 14 ). There were no other inclusion or exclusion criteria. We obtained detailed family histories for 725 of the 748 colorectal cancer patients who consented to participate in this study (more than 98% with north-west European ancestry). Eighteen (2.5%) of the 725 patients carried a mutation in a DNA mismatch repair gene.

All family pedigrees were constructed by use of Progeny data management software (Progeny Software, South Bend, IN), whereby information on each family member, as well as information on the family structure, was maintained in a relational database that facilitated the subsequent analyses. Information in the pedigree database was easily linked to demographic, clinical, and molecular information in other databases. This research was approved by the Research Ethics Board of Memorial University, and written informed consent was obtained from all participants.

Tumor Analysis

The methods of MSI and immunohistochemistry analysis of tumors were described previously ( 15 ). Briefly, DNA extracted from tumor tissue was tested for MSI at up to eight loci, and tissue slices that contained both tumor and normal tissue were subjected to immunohistochemical analysis for MSH2, MLH1, and MSH6 proteins. Tumors that demonstrated the MSI-high phenotype (ie, showed instability at ≥30% of loci tested) but had detectable levels of MSH2, MLH1, and MSH6 proteins by immunohistochemistry were immunohistochemically tested for PMS2. Methylation of the MLH1 promoter was detected by use of the MS-MLPA kit ME001B (MRC-Holland, Amsterdam, the Netherlands). Use of this kit generates two methylation-dependent signals from the MLH1 promoter. A 166–base pair (bp) fragment is produced if the Hha I site at position −7 (relative to the ATG start codon) is methylated and a 463-bp fragment is generated if the Hha I sites at positions −378 and −401 are both methylated. We scored the tumor DNA sample as methylated if either of these fragments was present at a normalized ratio of 0.15 or more of the peak area of the sample before digestion with Hha I. To identify the c.1799T>A variant in the BRAF gene in tumor DNA, we used a protocol as described previously ( 16 ). This method uses an allele-specific polymerase chain reaction assay that includes a set of primers for the GAPDH gene as an internal positive control.

Mutation Detection

All patients from families that met the Amsterdam II criteria and who had DNA mismatch repair–deficient tumors (ie, an MSI-high phenotype or the absence of MSH2, MLH1, MSH6, or PMS2 protein by immunohistochemistry) were subjected to screening of their leukocyte DNA for mutations in DNA mismatch repair genes. In addition, leukocyte DNA from all other patients with colorectal tumors that were deficient in mismatch repair, showed no evidence of MLH1 promoter methylation, and lacked a mutation in the BRAF gene was also screened for DNA mismatch repair gene mutations as follows. All exons and flanking intron sequences of the MSH2 , MLH1 , and MSH6 genes were directly sequenced. Large insertions and deletions were detected in the MSH2 and MLH1 genes by use of the SALSA P003 kit, which uses the multiplex ligation-dependent probe amplification method, as described in the manufacturer's instructions (MRC-Holland, Amsterdam, the Netherlands). The PMS2 gene was sequenced ( 17 ) in leukocyte DNA from patients who had tumors with the MSI-high phenotype and lacked the PMS2 protein but expressed the other mismatch repair proteins.

Risk Prediction Models

We evaluated four prediction models ( 9–12 )—three (Leiden, MMRpredict, and PREMM 1,2 ) that use a logistic regression approach and one (MMRpro) that uses a Bayesian analysis. The three models that use regression analysis give no weighting to the presence of unaffected relatives. A fifth model, AIFEG, which also uses Bayesian–Mendelian analysis, has been described ( 13 ), but software was not readily available for this model and so we did not include it in this analysis.

The Leiden model was developed by use of a high-risk population in which the mutation prevalence was 26% ( 9 ) and only MSH2 and MLH1 mutations were considered. The following three factors are incorporated in the model: meeting the Amsterdam I criteria, mean age at diagnosis of colorectal cancer of all affected members of a family, and family history of endometrial cancer. Calculations with the Leiden model were performed with the MMRpro software (part of the CancerGene program package) ( 18 ).

The MMRpredict model was developed by use of data from colorectal cancer patients who were younger than 55 years at diagnosis. The mutation prevalence in this cohort was 4.4% ( 10 ). Mutations in MSH2 , MLH1 , and MSH6 were considered. Variables in the model include age at diagnosis of colorectal cancer, sex, location of tumor (proximal vs distal), multiple colorectal cancers (synchronous or metachronous), occurrence of and age at diagnosis of colorectal cancer in first-degree relatives, and occurrence of endometrial cancer in any first-degree relative.

The PREMM 1,2 model was developed by use of data from a population in which the mutation prevalence was 14.5% ( 11 ), and only MSH2 and MLH1 mutations were considered. Proband-specific variables in this model include the occurrence and age at diagnosis of colorectal cancer, colonic adenomas, endometrial cancer, and other Lynch syndrome–associated cancers, including cancers of the ovary, stomach, kidney, ureter, bile duct, small bowel, brain (glioblastoma multiforme), pancreas, or sebaceous gland. Variables related to the relatives include the number of relatives with colorectal cancer, endometrial cancer, or other Lynch syndrome–associated cancers; the relationship to the proband (first degree vs second degree); the minimum age at diagnosis of each cancer in the family; and the presence of a relative with more than one Lynch syndrome–associated cancer.

In the MMRpro model, unlike in the other three models, risk estimation depends on the application of Bayes’ theorem and the Mendelian laws of inheritance. The model was developed a priori by use of published values for mutation prevalence and penetrance and was then validated by use of data from a high-risk population of colorectal cancer patients in which the mutation prevalence was 43% ( 12 ). Mutations in the MSH2 , MLH1 , and MSH6 genes were considered. Required data for the proband and for each first- and second-degree relative include age at diagnosis of colorectal cancer, age at diagnosis of endometrial cancer, and current age or age at last follow-up for those unaffected by colorectal or endometrial cancer. The MMRpro model does not take into account the occurrence of synchronous or metachronous colorectal cancers. The MMRpro software included in the CancerGene package ( 18 ) was used to calculate the MMRpro risk score. To use MMRpro software, preparation of the input files required software coding that was much more extensive and time consuming than the steps required for the other three models. In addition, it took 6.5 hours on a desktop Pentium IV computer to calculate the risk scores for all 725 patients with the MMRpro model, compared with less than 20 seconds with the PREMM 1,2 or MMRpredict models.

Because the MMRpro model relies on penetrance calculations, it requires that the age at diagnosis or age at last follow-up must be available for each first- and second-degree relative. For relatives without such age information, we used estimated ages as follows. For those diagnosed with cancer, we used the mean age at diagnosis from all patients in our dataset (namely, colorectal cancer, 68 years; endometrial, 56 years; and ovarian, 55 years). For those unaffected by cancer, we estimated their date of birth as follows: for siblings, the proband's date of birth; for children, nieces, and nephews, 30 years later than the proband's date of birth; for parents, aunts, and uncles, 25 years before the proband's date of birth; and for grandparents, 50 years before the proband's date of birth. We had the true age at diagnosis or age at last follow-up for 100% of the index patients with colorectal cancer and for 85% of their first-degree relatives.

The MMRpro software in the CancerGene software package also calculates the risk score by using the Leiden model. Web pages for both the PREMM 1,2 ( 19 ) and the MMRpredict ( 20 ) models are available and can be used to calculate the risk score for an individual patient and family.

All four models depend on the history of certain cancers in members of the patient's family. However, it is impossible to have a substantial family history of cancer if one does not come from a sufficiently large and informative family. For all four models, we used the so-called sum of information on 70-year-old equivalents (SISE) coefficient to correct the risk scores for the effect of family size. We previously developed the SISE coefficient ( 21 ) to quantify the informativeness of a family by use of the number of relatives in a kinship, their relationship to the proband, their ages, and the age-dependent penetrance of HNPCC mutations. A large SISE coefficient reflects a large family that contains many close relatives who survived into old age, whereas a small SISE coefficient reflects a small family with few family members of advanced age. In this study, we used the SISE coefficient to normalize risk scores on the basis of family informativeness by dividing the risk score for an individual patient by the SISE coefficient of that patient's family.

Statistical Analyses

The capability of a quantitative diagnostic test to discriminate between individuals with and without disease is often evaluated by use of the receiver operating characteristics curve. For this study, the presence of a germline mutation in a DNA mismatch repair gene in the proband was modeled as the disease and the individual risk score was used as the test result. Receiver operating characteristics curves were constructed by plotting the true-positive rate (ie, the sensitivity) against the false-positive rate (ie, 100 − specificity) for each different value of the risk score. The area under the receiver operating characteristics curve (AUC) was used as an overall measure of the discriminatory value of the test.

Statistical analyses were conducted with MedCalc for Windows, version 9.5 (MedCalc Software, Mariakerke, Belgium). Median values for skewed datasets were compared by use of the Mann–Whitney test. For pairwise comparisons of the proportion of patients whose data exceeded a fixed sensitivity cutoff, we used the McNemar test for correlated proportions on matched-pair samples. The comparison of the mean SISE coefficients—for families with and without mismatch repair mutations—was made by use of the independent samples t test, for which equal variances were assumed, and the Kolmogorov–Smirnov test used for normal distribution. All statistical tests were two-sided, and statistical significance for all tests was defined as a P value of less than .05.

Results

Patient and Family Characteristics

We obtained detailed family histories for 725 of the 748 colorectal cancer patients who consented to participate in this study (more than 98% of whom had north-west European ancestry). These 725 patients formed the basis of this study and were from 684 different families. Of these 684 families, 30 had two affected individuals who were diagnosed within the 5-year ascertainment period, four had three affected individuals, and one had four affected individuals. The mean age at diagnosis of the 725 patients was 60.5 years (SD = ±9.5 years), the median age was 62 years (interquartile range = 54–68), and 447 (61.7%) were male ( Table 1 ). Of the 725 patients, 435 (60%) had a tumor distal to the splenic flexure, and 71 tumors (9.8%) from 69 different patients had an MSI-high phenotype.

Table 1

Characteristics of the 725 patients *

Characteristic Value 
Mean age, y (±SD) 60.5 (±9.5) 
Median age, y (interquartile range) 62.0 (54–68) 
No. of patients (%) 
    Diagnosed at <50 y 93 (12.8) 
    Male 447 (61.7) 
    Tumor distal to splenic flexure 435 (60.0) 
    Synchronous or metachronous CRC 61 (8.4) 
    Clinical TNM stage † 
        Stage I 110 (15.6) 
        Stage II 223 (31.8) 
        Stage III 229 (32.7) 
        Stage IV 139 (19.8) 
        Insufficient stage data 47 
    MSI-high tumor 69 (9.5) 
    Case meets revised Bethesda guidelines ‡ 364 (50.2) 
        CRC in ≥1 first-degree relative 241 (33.2) 
        Meets AM1 criteria 28 (3.9) 
        Meets AM2 criteria § 31 (4.3) 
        Meets clinical criteria for FAP ‖ 7 (1) 
Characteristic Value 
Mean age, y (±SD) 60.5 (±9.5) 
Median age, y (interquartile range) 62.0 (54–68) 
No. of patients (%) 
    Diagnosed at <50 y 93 (12.8) 
    Male 447 (61.7) 
    Tumor distal to splenic flexure 435 (60.0) 
    Synchronous or metachronous CRC 61 (8.4) 
    Clinical TNM stage † 
        Stage I 110 (15.6) 
        Stage II 223 (31.8) 
        Stage III 229 (32.7) 
        Stage IV 139 (19.8) 
        Insufficient stage data 47 
    MSI-high tumor 69 (9.5) 
    Case meets revised Bethesda guidelines ‡ 364 (50.2) 
        CRC in ≥1 first-degree relative 241 (33.2) 
        Meets AM1 criteria 28 (3.9) 
        Meets AM2 criteria § 31 (4.3) 
        Meets clinical criteria for FAP ‖ 7 (1) 
*

MSI-high = microsatellite instability high; CRC = colorectal cancer; AM1 = Amsterdam I criteria; AM2 = Amsterdam II criteria; FAP = familial adenomatous polyposis.

Percentage represents proportion of all patients who were adequately staged.

Bethesda criterion 3 was not evaluated.

§

Includes families meeting AM1 criteria.

Known FAP family, or patient had more than 100 adenomatous polyps.

Pathogenic DNA mismatch repair gene mutations were found in leukocyte DNA from 18 patients (2.5%): 13 patients had mutations in MSH2 , one had a mutation in MLH1 , two had mutations in MSH6 , and two had mutations in PMS2 . Data from 11 of these 18 patients met the Amsterdam I criteria, one more patient met the Amsterdam II criteria, five others met only the Bethesda criteria, and one met none of these criteria.

Calculation of Risk Scores

We calculated a risk score for each patient from each of the models. The risk score represents the probability that the patient with colorectal cancer is carrying a pathogenic mutation in a mismatch repair gene. The risk scores were not normally distributed but rather were skewed to the left, with many individuals at low risk and only a few at high risk ( Figure 1 ). Table 2 shows the median risk scores that were calculated by each of the four models for patients who met various criteria. Among all patients, the median risk scores ranged from 0.86% for the Leiden model to 6.32% for the PREMM 1,2 model. Among patients who met the Amsterdam I criteria, the median risk scores for carriers of mismatch repair gene mutation (ie, those with Lynch syndrome) were statistically significantly greater than those for noncarriers (ie, those with familial colorectal cancer, type X) (for this difference, P ≤ .001 for all models).

Table 2

Probabilities that patients meeting certain criteria are carrying a DNA mismatch repair gene mutation, as revealed by the risk scores from the Leiden, MMRpredict, PREMM 1,2 , and MMRpro models *

  Median risk score, % (95% CI) 
Criteria No. of patients Leiden MMRpredict  PREMM 1,2 MMRpro 
All patients 725 0.86 (0.78 to 0.90) 0.91 (0.74 to 1.01) 6.32 (5.96 to 6.90) 0.17 (0.12 to 0.24) 
AM1, all 28 19.6 (13.8 to 36.6) 53.3 (5.6 to 79.9) 32.1 (17.1 to 70.1) 22.7 (5.5 to 96.6) 
    AM1 with a MMR mutation 11 37.3 (23.2 to 81.6) 94.9 (52.6 to 100) 81.1 (45.9 to 96.2) 97.2 (27.2 to 99.5) 
    AM1—FCCX † 17 14.6 (11.7 to 20.4) 5.6 (3.8 to 58.7) 17.4 (14.1 to 33.6) 6.3 (0.8 to 66.9) 
AM2 31 17.3 (13.6 to 36.4) 57.6 (8.1 to 86.5) 28.0 (19.1 to 68.7) 22.5 (5.9 to 96.3) 
Bethesda 364 1.2 (1.1 to 1.5) 2.37 (2.0 to 3.5) 11.0 (9.7 to 11.8) 0.86 (0.5 to 1.2) 
MSI-high 69 1.34 (0.9 to 2.5) 6.69 (1.9 to 15.9) 9.89 (6.6 to 13.2) 0.56 (0.2 to 3.8) 
MMR mutation 18 32.2 (4.0 to 71.6) 80.2 (38.5 to 99.2) 78.4 (17.5 to 90.0) 96 (22.4 to 99.1) 
  Median risk score, % (95% CI) 
Criteria No. of patients Leiden MMRpredict  PREMM 1,2 MMRpro 
All patients 725 0.86 (0.78 to 0.90) 0.91 (0.74 to 1.01) 6.32 (5.96 to 6.90) 0.17 (0.12 to 0.24) 
AM1, all 28 19.6 (13.8 to 36.6) 53.3 (5.6 to 79.9) 32.1 (17.1 to 70.1) 22.7 (5.5 to 96.6) 
    AM1 with a MMR mutation 11 37.3 (23.2 to 81.6) 94.9 (52.6 to 100) 81.1 (45.9 to 96.2) 97.2 (27.2 to 99.5) 
    AM1—FCCX † 17 14.6 (11.7 to 20.4) 5.6 (3.8 to 58.7) 17.4 (14.1 to 33.6) 6.3 (0.8 to 66.9) 
AM2 31 17.3 (13.6 to 36.4) 57.6 (8.1 to 86.5) 28.0 (19.1 to 68.7) 22.5 (5.9 to 96.3) 
Bethesda 364 1.2 (1.1 to 1.5) 2.37 (2.0 to 3.5) 11.0 (9.7 to 11.8) 0.86 (0.5 to 1.2) 
MSI-high 69 1.34 (0.9 to 2.5) 6.69 (1.9 to 15.9) 9.89 (6.6 to 13.2) 0.56 (0.2 to 3.8) 
MMR mutation 18 32.2 (4.0 to 71.6) 80.2 (38.5 to 99.2) 78.4 (17.5 to 90.0) 96 (22.4 to 99.1) 
*

Probabilities are expressed as median risk scores. MMR = mismatch repair; CI = confidence interval; AM1 = Amsterdam I criteria; FCCX = familial colorectal cancer, type X; AM2 = Amsterdam II criteria; MSI-high = microsatellite instability high.

P ≤ .001 for all models, compared with AM1 with a MMR mutation (Mann–Whitney test). All statistical tests were two-sided.

Figure 1

Distribution of risk scores for the Leiden, MMRpredict, PREMM 1,2 , and MMRpro models. The mean, median, minimum (Min), and maximum (Max) risk scores, expressed as percentages, for each model are shown. To reduce the size of the figures, data for risk scores between 20% and 80% have been excluded. The frequency distribution of the missing data is similar to that on either side of the break. Supplementary Figure 1 , showing the complete range of data, is available online.

Figure 1

Distribution of risk scores for the Leiden, MMRpredict, PREMM 1,2 , and MMRpro models. The mean, median, minimum (Min), and maximum (Max) risk scores, expressed as percentages, for each model are shown. To reduce the size of the figures, data for risk scores between 20% and 80% have been excluded. The frequency distribution of the missing data is similar to that on either side of the break. Supplementary Figure 1 , showing the complete range of data, is available online.

Accuracy of the Risk Assessment

We next compared the predicted frequency of DNA mismatch repair gene mutations (as given by the risk scores) with the observed frequency. We divided the dataset for each model into four arbitrarily selected groups according to the calculated risk scores and compared the predicted number of mutation carriers with the observed number ( Table 3 ). All models overestimated the frequency of mutations. The ratio of estimated to observed mutations ranged from 1.2 for the Leiden model to 4.3 for the PREMM 1,2 model.

Table 3

Observed and expected numbers of carriers of DNA mismatch repair gene mutations predicted by the Leiden, MMRpredict, PREMM 1,2 , and MMRpro models for various ranges of predicted risk *

Model and risk score range † , %  No. of patients Mean risk score, % (95% CI) No. of MMR mutation carriers 
Expected ‡ Observed Expected/observed 
Leiden 
    0–10 686 1.4 (1.2 to 1.5) 9.6 1.9 
    10–40 31 19.6 (16.5 to 22.6) 6.1 0.9 
    40–80 57.9 (37.4 to 78.4) 2.9 1.0 
    80–100 83.5 (77.9 to 89.1) 2.5 0.8 
    Total 725 2.9 (2.3 to 3.5) 20.7 18 1.2 
MMRpredict 
    0–10 628 1.5 (1.3 to 1.6) 9.4 4.7 
    10–40 54 21.5 (19.2 to 23.8) 11.6 5.8 
    40–80 28 56.8 (49.2 to 61.9) 15.9 3.2 
    80–100 15 94.2 (91.0 to 97.3) 14.1 1.6 
    Total 725 7.1 (5.8 to 8.3) 51.2 18 2.8 
PREMM 1,2 
    0–10 524 5.5 (5.3 to 5.6) 28.7 9.6 
    10–40 174 17.4 (16.4 to 18.4) 30.2 15.1 
    40–80 18 53.7 (48.3 to 59.0) 9.7 1.9 
    80–100 89.3 (84.5 to 94.1) 8.0 1.0 
    Total 725 10.6 (9.7 to 11.6) 77.1 18 4.3 
MMRpro 
    0–10 654 0.9 (0.7 to 1.0) 5.6 2.8 
    10–40 42 19.3 (16.9 to 21.6) 8.1 2.7 
    40–80 50.3 (41.2 to 67.8) 4.5 2.3 
    80–100 20 95.4 (92.7 to 98.0) 19.1 11 1.7 
    Total 725 5.2 (4.0 to 6.4) 37.7 18 2.1 
Model and risk score range † , %  No. of patients Mean risk score, % (95% CI) No. of MMR mutation carriers 
Expected ‡ Observed Expected/observed 
Leiden 
    0–10 686 1.4 (1.2 to 1.5) 9.6 1.9 
    10–40 31 19.6 (16.5 to 22.6) 6.1 0.9 
    40–80 57.9 (37.4 to 78.4) 2.9 1.0 
    80–100 83.5 (77.9 to 89.1) 2.5 0.8 
    Total 725 2.9 (2.3 to 3.5) 20.7 18 1.2 
MMRpredict 
    0–10 628 1.5 (1.3 to 1.6) 9.4 4.7 
    10–40 54 21.5 (19.2 to 23.8) 11.6 5.8 
    40–80 28 56.8 (49.2 to 61.9) 15.9 3.2 
    80–100 15 94.2 (91.0 to 97.3) 14.1 1.6 
    Total 725 7.1 (5.8 to 8.3) 51.2 18 2.8 
PREMM 1,2 
    0–10 524 5.5 (5.3 to 5.6) 28.7 9.6 
    10–40 174 17.4 (16.4 to 18.4) 30.2 15.1 
    40–80 18 53.7 (48.3 to 59.0) 9.7 1.9 
    80–100 89.3 (84.5 to 94.1) 8.0 1.0 
    Total 725 10.6 (9.7 to 11.6) 77.1 18 4.3 
MMRpro 
    0–10 654 0.9 (0.7 to 1.0) 5.6 2.8 
    10–40 42 19.3 (16.9 to 21.6) 8.1 2.7 
    40–80 50.3 (41.2 to 67.8) 4.5 2.3 
    80–100 20 95.4 (92.7 to 98.0) 19.1 11 1.7 
    Total 725 5.2 (4.0 to 6.4) 37.7 18 2.1 
*

CI = confidence interval; MMR = mismatch repair.

Risk score ranges were chosen arbitrarily.

Expected = (no. of patients × mean risk score)/100.

The overestimation of the frequency of mutations by the MMRpro model was relatively consistent across all four risk categories, with the ratio of estimated to observed mutations ranging from 1.7 to 2.8. However, the PREMM 1,2 model greatly overestimated the risk for the low-risk categories, with a maximum overestimation in mutation frequency of 15.1-fold in the 10%–40% risk category. Risk distribution histograms for all four models are shown in Figure 1 . The minimum possible risk that could be calculated by the PREMM 1,2 algorithm was 2.7%, whereas the risk that was calculated by the three other models (which have no such limitations) was less than 2% for most patients. This inability of the PREMM 1,2 algorithm to calculate a risk of less than 2.7% would account for the overestimation of the risk by the PREMM 1,2 model in low-risk patients.

Receiver Operating Characteristics

The AUCs from the four models ranged from 0.91 to 0.96 ( Figure 2 ). The AUCs from the MMRpro and MMRpredict models were virtually the same, whereas those from the Leiden and PREMM 1,2 models were somewhat lower, but the difference was not statistically significant. For comparison, the performance of the Amsterdam II and Bethesda criteria is also indicated in Figure 2 .

Figure 2

Receiver operating characteristics curves for the Leiden, MMRpredict, PREMM 1,2 , and MMRpro models. Black dots = the sensitivity and specificity of the Amsterdam II criteria (AM2) and the revised Bethesda (Beth) criteria, as indicated. The diagonal line represents the so-called line of no-discrimination. AUC = area under the curve; CI = confidence interval.

Figure 2

Receiver operating characteristics curves for the Leiden, MMRpredict, PREMM 1,2 , and MMRpro models. Black dots = the sensitivity and specificity of the Amsterdam II criteria (AM2) and the revised Bethesda (Beth) criteria, as indicated. The diagonal line represents the so-called line of no-discrimination. AUC = area under the curve; CI = confidence interval.

Discrimination Between Mutations in Different DNA Mismatch Repair Genes

Most mutations that have been reported in Lynch syndrome patients are in the MSH2 or MLH1 genes (81% combined), whereas those in the MSH6 gene accounted for 12% of reported mutations and PMS2 accounts for 7% ( 22 ). The penetrance of mutations in MSH6 or PMS2 is lower than that of MSH2 or MLH1 mutations ( 17 , 23 )—that is, carriers of an MSH6 or PMS2 mutation are less likely to develop cancer than carriers of an MSH2 or MLH1 mutation. We therefore examined how the different models performed in patients with mutations in different mismatch repair genes ( Figure 3 ). For all four models, the difference between the median risk scores for MSH2 and MLH1 mutation carriers combined was statistically significantly higher than that for MSH6 and PMS2 mutation carriers combined ( P < .015). For example, the MMRpredict model gave median scores of 94% and 24% for these two groups, respectively, of mutation carriers ( P = .015). The differences between the median risk scores for patients with no mutation in a DNA mismatch repair gene and those carrying an MSH6 or PMS2 mutation were also statistically significant as calculated by MMRpredict ( P = .01) and MMRpro ( P = .03); this difference approached statistical significance for the Leiden model ( P = .05). Figure 4 shows the distribution of individual risk scores for patients with mutations in different DNA mismatch repair genes and also for patients with APC mutations or biallelic MYH mutations. Mutations in APC or MYH genes, which are not involved in DNA mismatch repair, also lead to inherited forms of colorectal cancer. In general, patients with biallelic MYH mutations had lower risk scores than those with other mutations, reflecting the recessive nature of the MYH mutation and the consequent lack of affected individuals in consecutive generations. Carriers of dominant APC mutations had risk scores that were intermediate between those for the MSH2–MLH1 and MSH6–PMS2 groups of mismatch repair mutations, reflecting the high penetrance of APC mutations (which tends to raise the risk score), combined with the paucity of extracolonic tumors (which has the opposite effect).

Figure 3

Median probability of the index patient carrying a DNA mismatch repair gene mutation, for the Leiden, MMRpredict, PREMM 1,2 , and MMRpro models. Error bars = 25th–75th percentiles (the interquartile range).

Figure 3

Median probability of the index patient carrying a DNA mismatch repair gene mutation, for the Leiden, MMRpredict, PREMM 1,2 , and MMRpro models. Error bars = 25th–75th percentiles (the interquartile range).

Figure 4

Distribution of individual risk scores as calculated by the Leiden, MMRpredict, PREMM 1,2 , and MMRpro models. Each symbol represents the risk score of an individual patient.

Figure 4

Distribution of individual risk scores as calculated by the Leiden, MMRpredict, PREMM 1,2 , and MMRpro models. Each symbol represents the risk score of an individual patient.

In a typical clinical setting, these models would be used to stratify patients and their families according to the calculated risk scores. Low-risk patients should receive care appropriate for noninherited colorectal cancers, whereas those with high risk scores should have their tumors tested for evidence of mismatch repair deficiency. If such a deficiency were found, additional molecular testing would be warranted, possibly leading to a search for mutations in DNA mismatch repair genes. In this context, we compared the performance of the four models by choosing a cut point to produce either a positive or a negative test result. For each model, we used the receiver operating characteristics to choose a cut point that would produce a sensitivity of 94% (which was also the sensitivity achieved by the Bethesda criteria in our dataset). In this approach, one of the 18 patients who carried a mutation in a DNA mismatch repair gene (ie, 6% of the 18 patients) would be designated as having a false-negative result. In all models, the excluded mutation carrier (ie, the carrier—of a DNA mismatch repair gene mutation—with the lowest risk score) carried either a PMS2 or a MSH6 mutation ( Figure 4 ). However, the patient with the lowest risk score is not necessarily the same individual in all four models. Table 4 shows the results from this analysis. At a sensitivity cut point of 94%, the specificities of the four models ranged from 61% to 88%; the specificity of the Bethesda criteria was only 51% (95% confidence interval [CI] = 47.2% to 54.7%). The relationship between the sensitivity cut points and the specificities can be visualized from the receiver operating characteristics curves in Figure 2 . The best-performing model was MMRpredict, in which a sensitivity of 94% corresponded to a cutoff risk score of 9.3% that would eliminate 86% of all patients from the need for further molecular testing.

Table 4

Comparison of the discriminatory power of the Leiden, MMRpredict, PREMM 1,2 , and MMRpro models at a sensitivity of 94% *

 Uncorrected risk score Risk corrected for SISE 
Model  Cutoff criterion † , %  No. of patients exceeding cutoff (%) Specificity, % (95% CI)  Cutoff criterion ‡ No. of patients exceeding cutoff (%) Specificity,% (95% CI) 
MMRpredict >9.30 104 (14) 88 (85 to 90) >1.66 82 (11) 91 (88 to 93) 
Leiden >1.71  190 (26) § 76 (72 to 79) >0.32  127 (18) ‖ 84 (82 to 87) 
MMRpro >1.27  187 (26) § 76 (73 to 79) >0.18  191 (26) § 76 (73 to 79) 
PREMM 1,2 >7.40  295 (41) § 61 (57 to 64) >1.03  265 (37) § 65 (62 to 69) 
 Uncorrected risk score Risk corrected for SISE 
Model  Cutoff criterion † , %  No. of patients exceeding cutoff (%) Specificity, % (95% CI)  Cutoff criterion ‡ No. of patients exceeding cutoff (%) Specificity,% (95% CI) 
MMRpredict >9.30 104 (14) 88 (85 to 90) >1.66 82 (11) 91 (88 to 93) 
Leiden >1.71  190 (26) § 76 (72 to 79) >0.32  127 (18) ‖ 84 (82 to 87) 
MMRpro >1.27  187 (26) § 76 (73 to 79) >0.18  191 (26) § 76 (73 to 79) 
PREMM 1,2 >7.40  295 (41) § 61 (57 to 64) >1.03  265 (37) § 65 (62 to 69) 
*

For each model, a cutoff score was chosen that excluded the MMR mutation carrier with the lowest risk score. This procedure results in a sensitivity of 94% for all models. SISE = sum of information on 70-year-old equivalents; MMR = mismatch repair; CI = confidence interval.

Risk score.

Risk score divided by the SISE coefficient.

§

P < .001, compared with the value for MMRpredict (two-sided McNemar test).

P = .001, compared with the value for MMRpredict (two-sided McNemar test).

Because the four models relied on the presence of a positive family history of colorectal and other tumors, we tested the effect of normalizing the risk scores for each patient to the SISE coefficient of their respective families. This procedure should reduce the bias of the non-Bayesian models to assign a higher risk to larger, more informative families. The value of the SISE coefficient for individual families ranged from 1.0 to 23.6, and the mean SISE coefficient for families that carried mismatch repair mutations (9.76) was similar to that for noncarrier families (9.82) ( P = .95, independent samples t test, for which equal variances were assumed). We observed that correcting for the effect of the family SISE coefficient enhanced the performance of the three logistic regression models but had no effect on the performance of MMRpro, which inherently compensates for family size and age structure ( Table 4 ). After correcting for family size, the best-performing model was MMRpredict, which achieved a sensitivity of 94% (95% CI = 73% to 99%) and a specificity of 91% (95% CI = 88% to 93%). This model identified 11% of all patients who should have additional molecular or immunohistochemical testing. By comparison, the revised Bethesda criteria (sensitivity = 94% and specificity = 51%) identified 50% of all patients for additional testing.

Discussion

Our study indicates that all four models showed higher specificity than the revised Bethesda criteria in identifying patients who should receive molecular evaluations to investigate a potential genetic basis for their colorectal cancer. In particular, the MMRpredict model performed the best on our dataset, especially when corrected by the SISE coefficient for the effect of family size and age structure.

The identification of carriers of DNA mismatch repair gene mutations among patients with colorectal cancer is important because the first-degree relatives of such patients have a 50% risk of having inherited the mutation and because clinical screening can prevent cancer and cancer-related mortality in mutation carriers. Consequently, the development of predictive models that are able to identify patients whose tumors should be tested for deficient DNA mismatch repair is a logical approach. In this report, we compared the diagnostic utility of four different models, adjusted the models for family size, and then applied the models to a population-based cohort of patients with colorectal cancer. The four models that we evaluated were originally designed to assist in the management of high-risk populations of patients who are typical of those referred to a cancer genetics clinic. Of the four models, MMRpredict was developed by use of the lowest risk population, in which the sole inclusion criterion was that patients were diagnosed with colorectal cancer when they were younger than 55 years, and this model performed best with our dataset. The models are targeted primarily toward detection of mutations in the MSH2 and MLH1 genes, and no model took into account mutations in the PMS2 gene. Nevertheless, we conclude that all of the models have diagnostic utility in identifying colorectal cancer patients from the general population who should receive further evaluation for a possible genetic basis for their cancer. All models assigned statistically significantly higher scores to patients with Lynch syndrome than to patients with familial colorectal cancer, type X ( P ≤ .001, for all models; Table 2 ). The AUC provides a global assessment of the utility of each predictive model over the entire range of sensitivity and specificity. The relatively high AUC values that we obtained in this study indicate that despite overestimating the risk for individual patients, all four models were capable of providing clinically useful information for colorectal cancer patients from the general population. In constructing receiver operating characteristics curves, it is the rank order of the estimated risk scores that is important, not the actual values of the risk scores. In typical clinical use, the risk estimated for a patient or a family will be compared with a previously established risk cutoff value. If the risk estimation exceeds this cut point, then additional diagnostic testing should be initiated. The PREMM 1,2 ( 19 ) and MMRpredict ( 20 ) models have web pages that allow the risk to be calculated quickly after questions about the personal and family history of cancer are answered.

A protocol that could be applied to unselected patients with colorectal cancer and that indicates the number of patients at each stage in this study is shown in Figure 5 . In addition to testing the tumor for DNA mismatch repair deficiency by microsatellite analysis and/or immunohistochemistry for DNA mismatch repair proteins, testing for MLH1 promoter methylation and the presence of BRAF mutations will further reduce the number of patients who should be sent for detailed mutation analysis by DNA sequencing ( 16 , 24 ). However, these tests are expensive and time consuming to conduct. It may also take many months before results are available, and even then, the results may be inconclusive. While waiting for molecular test results, family members at high risk—as determined by one of the four models—should be offered clinical screening tests to detect colorectal cancer and other neoplasms. Thus, risk estimation with one of the models would not be a substitute for MSI testing or immunohistochemistry, but it would serve as a valuable initial screen ( Figure 5 ).

Figure 5

Flow chart for identifying Lynch syndrome mutations, including use of the MMRpredict model. MSI-high indicates that 30% or more of microsatellite loci tested were unstable. MSI = microsatellite instability; CRC = colorectal cancer. *Cutoff for the ratio of MMRpredict score to SISE coefficient > 1.66. †The tumor carried a methylated MLH1 promoter or a BRAF gene with the V600E mutation.

Figure 5

Flow chart for identifying Lynch syndrome mutations, including use of the MMRpredict model. MSI-high indicates that 30% or more of microsatellite loci tested were unstable. MSI = microsatellite instability; CRC = colorectal cancer. *Cutoff for the ratio of MMRpredict score to SISE coefficient > 1.66. †The tumor carried a methylated MLH1 promoter or a BRAF gene with the V600E mutation.

The choice of an appropriate cutoff value for risk estimation is a critical step in the risk management process. A cutoff that is too high will lead to carriers of mismatch repair mutations being missed, with a subsequent increase in mortality and morbidity among these false-negative cases. On the other hand, choosing a cutoff value that is too low leads to more patients being subjected to detailed molecular investigations, with increased cost to the health-care system, increased exposure to invasive clinical screening tests, and increased anxiety in these families with false-positive results. Table 4 shows the cutoff values for risk estimation that result in a sensitivity of 94%. To achieve this sensitivity, the MMRpredict model identified only 104 patients (14%) who would require additional analyses, whereas the PREMM 1,2 model identified nearly three times that number of patients. It is perhaps not a coincidence that the MMRpredict model was developed on a patient population with the fewest restrictions and with a prevalence of DNA mismatch repair gene mutations of 4.4%, which is closest to that previously observed in unselected patients with colorectal cancer ( 4 , 5 ) and was also closest to the prevalence of 2.5% observed in our dataset. The MMRpredict model is also fairly simple to use, and calculating the risk score requires fewer steps than does any other model. All models outperformed the Bethesda criteria. Although the areas under the entire receiver operating characteristics curves did not differ substantially between models, there were statistically significant differences between the curves in the clinically significant region of high sensitivity, as shown in Table 4 . Additionally, we found that the three non-Bayesian models discriminated better between patients with Lynch syndrome and patients without Lynch syndrome, after correcting the risk scores for the size and age structure of each family, as defined by the SISE coefficient.

Because all models tended to assign lower risks to patients who carried PMS2 or MSH6 mutations, such patients may turn out to have false-negative results if a sensitivity of less than 100% is used as the cutoff. However, these families are also likely to have a lower incidence of Lynch syndrome–related tumors than families carrying the more penetrant mutations of the MSH2 and MLH1 genes.

The actual probability that a patient or family harbors a mutation in a DNA mismatch repair gene cannot now be accurately calculated because the models overestimate the probability that an individual or family carries a mutation across all risk categories, especially in low-risk categories ( Table 3 ). This overestimation can be attributed to the fact that the models were developed by use of data from high-risk populations in which the prior probability of carrying a mutation in a mismatch repair gene is much greater than it is in the general population.

A limitation of this study is that the cohort contained only 18 carriers of a DNA mismatch repair gene mutation. Given that the prevalence of Lynch syndrome in unselected patients with colorectal cancer appears to be less than 3%, the small number of mutation carriers is likely to be a limitation for any study investigating Lynch syndrome. The optimum cutoff value of the risk score to identify patients who require molecular testing will become clearer as more centers report their experiences with these models. In this study, we did not enroll patients who were older than 74 years when diagnosed with colorectal cancer. It is possible that the optimum cutoff values could differ between a younger cohort and an older cohort. Risk factors should be investigated in patients who are older than 74 years when diagnosed with colorectal cancer. Because the prevalence of Lynch syndrome in our population is similar to that reported in other analyses ( 14 ), our data should provide useful guidance for other populations.

We conclude that for colorectal cancer patients who were younger than 75 years at diagnosis, MMRpredict appears to be an excellent screening tool that identifies an at-risk group that should undergo further diagnostic workup. In the general Newfoundland population, the proportion of patients with colorectal cancer requiring additional workup has been reduced from 50%, as identified by the Bethesda criteria, to 11% as identified by the MMRpredict model after adjustment for family size.

After submission of the manuscript for this article, Balmaňa et al. ( 25 ) published a article comparing use of the PREMM 1,2 and MMRpredict models on a cohort of 1222 Spanish patients. However, mismatch repair mutations were found in only eight patients and no statistical comparisons were made.

Funding

Supported by the Canadian Institutes of Health Research (CIHR) Interdisciplinary Health Research Team award CRT-43821 and CIHR Team in Interdisciplinary Research on Colorectal Cancer FRN-79845 (to PSP and HBY) and by Genome Canada, Atlantic Medical Genetics and Genomics Initiative.

References

1.
Rustgi
AK
The genetics of hereditary colon cancer
Genes Dev.
 , 
2007
, vol. 
21
 
20
(pg. 
2525
-
2538
)
2.
Vasen
HF
Mecklin
JP
Khan
PM
Lynch
HT
The International Collaborative Group on Hereditary Non-Polyposis Colorectal Cancer (ICG-HNPCC)
Dis Colon Rectum
 , 
1991
, vol. 
34
 
5
(pg. 
424
-
425
)
3.
Vasen
HF
Watson
P
Mecklin
JP
Lynch
HT
New clinical criteria for hereditary nonpolyposis colorectal cancer (HNPCC, Lynch syndrome) proposed by the International Collaborative group on HNPCC
Gastroenterology
 , 
1999
, vol. 
116
 
6
(pg. 
1453
-
1456
)
4.
De la Chapelle
A
The incidence of Lynch syndrome
Fam Cancer
 , 
2005
, vol. 
4
 
3
(pg. 
233
-
237
)
5.
Hampel
H
Frankel
WL
Martin
E
, et al.  . 
Screening for the Lynch syndrome (hereditary nonpolyposis colorectal cancer)
N Engl J Med
 , 
2005
, vol. 
352
 
18
(pg. 
1851
-
1860
)
6.
Lindor
NM
Rabe
K
Petersen
GM
, et al.  . 
Lower cancer incidence in Amsterdam-I criteria families without mismatch repair deficiency: familial colorectal cancer type X
JAMA
 , 
2005
, vol. 
293
 
16
(pg. 
1979
-
1985
)
7.
Rodriguez-Bigas
MA
Boland
CR
Hamilton
SR
, et al.  . 
A National Cancer Institute Workshop on Hereditary Nonpolyposis Colorectal Cancer Syndrome: meeting highlights and Bethesda guidelines
J Natl Cancer Inst
 , 
1997
, vol. 
89
 
23
(pg. 
1758
-
1762
)
8.
Umar
A
Boland
CR
Terdiman
JP
, et al.  . 
Revised Bethesda Guidelines for hereditary nonpolyposis colorectal cancer (Lynch syndrome) and microsatellite instability
J Natl Cancer Inst
 , 
2004
, vol. 
96
 
4
(pg. 
261
-
268
)
9.
Wijnen
JT
Vasen
HF
Khan
PM
, et al.  . 
Clinical findings with implications for genetic testing in families with clustering of colorectal cancer
N Engl J Med
 , 
1998
, vol. 
339
 
8
(pg. 
511
-
518
)
10.
Barnetson
RA
Tenesa
A
Farrington
SM
, et al.  . 
Identification and survival of carriers of mutations in DNA mismatch-repair genes in colon cancer
N Engl J Med
 , 
2006
, vol. 
354
 
26
(pg. 
2751
-
2763
)
11.
Balmana
J
Stockwell
DH
Steyerberg
EW
, et al.  . 
Prediction of MLH1 and MSH2 mutations in Lynch syndrome
JAMA
 , 
2006
, vol. 
296
 
12
(pg. 
1469
-
1478
)
12.
Chen
S
Wang
W
Lee
S
, et al.  . 
Prediction of germline mutations and cancer risk in the Lynch syndrome
JAMA
 , 
2006
, vol. 
296
 
12
(pg. 
1479
-
1487
)
13.
Marroni
F
Pastrello
C
Benatti
P
, et al.  . 
A genetic model for determining MSH2 and MLH1 carrier probabilities based on family history and tumor microsatellite instability
Clin Genet.
 , 
2006
, vol. 
69
 
3
(pg. 
254
-
262
)
14.
Green
RC
Green
JS
Buehler
SK
, et al.  . 
Very high incidence of familial colorectal cancer in Newfoundland: a comparison with Ontario and 13 other population-based studies
Fam Cancer
 , 
2007
, vol. 
6
 
1
(pg. 
53
-
62
)
15.
Woods
MO
Hyde
AJ
Curtis
FK
, et al.  . 
High frequency of hereditary colorectal cancer in Newfoundland likely involves novel susceptibility genes
Clin Cancer Res.
 , 
2005
, vol. 
11
 
19
(pg. 
6853
-
6861
)
16.
Loughrey
MB
Waring
PM
Tan
A
, et al.  . 
Incorporation of somatic BRAF mutation testing into an algorithm for the investigation of hereditary non-polyposis colorectal cancer
Fam Cancer
 , 
2007
, vol. 
6
 
3
(pg. 
301
-
310
)
17.
Clendenning
M
Senter
L
Hampel
H
, et al.  . 
A frame-shift mutation of PMS2 is a widespread cause of Lynch syndrome
J Med Genet.
 , 
2008
, vol. 
45
 
6
(pg. 
340
-
345
)
18.
CancerGene
 
Dallas, TX
The University of Texas Southwestern Medical Center
 
19.
PREMM 1,2 Calculator
 
Boston, MA
Dana-Farber Cancer Institute
 
20.
MMRpredict Calculator
 
Edinburgh, Scotland
University of Edinburg and MRC Human Genetics Unit
 
http://hnpccpredict.hgu.mrc.ac.uk/ . Accessed January 9, 2009
21.
Green
RC
McLaughlin
JR
Younghusband
HB
SISE matters: the sum of information on seventy-yr-old equivalents measures pedigree information content when assessing the risk of HNPCC in a family
Fam Cancer
 , 
2005
, vol. 
4
 
2
(pg. 
169
-
175
)
22.
Woods
MO
Williams
P
Careen
A
, et al.  . 
A new variant database for mismatch repair genes associated with Lynch syndrome
Hum Mutat
 , 
2007
, vol. 
28
 
7
(pg. 
669
-
673
)
23.
Hendriks
YM
Wagner
A
Morreau
H
, et al.  . 
Cancer risk in hereditary nonpolyposis colorectal cancer due to MSH6 mutations: impact on counseling and surveillance
Gastroenterology
 , 
2004
, vol. 
127
 
1
(pg. 
17
-
25
)
24.
Bettstetter
M
Dechant
S
Ruemmele
P
, et al.  . 
Distinction of hereditary nonpolyposis colorectal cancer and sporadic microsatellite-unstable colorectal cancer through quantification of MLH1 methylation by real-time PCR
Clin Cancer Res.
 , 
2007
, vol. 
13
 
11
(pg. 
3221
-
3228
)
25.
Balmana
J
Balaguer
F
Castellvi-Bel
S
, et al.  . 
Comparison of predictive models, clinical criteria and molecular tumor screening for the identification of patients with Lynch syndrome in a population-based cohort of colorectal cancer patients
J Med Genet.
 , 
2008
, vol. 
45
 
9
(pg. 
557
-
563
)
Financial interests or other conflicts of interest: none. Neither the Canadian Institutes of Health Research nor Genome Canada had any role in the design or conduct of the study or in the preparation of the manuscript.
We gratefully acknowledge the work of Chris Hatcher, database administrator with the Newfoundland and Labrador Centre for Health Information, who did the computer programming that facilitated the calculation of risk for the different models.