Development and Validation of the PREMM₅ Model for Comprehensive Risk Assessment of Lynch Syndrome

Abstract

Purpose

Current Lynch syndrome (LS) prediction models quantify the risk to an individual of carrying a pathogenic germline mutation in three mismatch repair (MMR) genes: MLH1, MSH2, and MSH6. We developed a new prediction model, PREMM₅, that incorporates the genes PMS2 and EPCAM to provide comprehensive LS risk assessment.

Patients and Methods

PREMM₅ was developed to predict the likelihood of a mutation in any of the LS genes by using polytomous logistic regression analysis of clinical and germline data from 18,734 individuals who were tested for all five genes. Predictors of mutation status included sex, age at genetic testing, and proband and family cancer histories. Discrimination was evaluated by the area under the receiver operating characteristic curve (AUC), and clinical impact was determined by decision curve analysis; comparisons were made to the existing PREMM_1,2,6 model. External validation of PREMM₅ was performed in a clinic-based cohort of 1,058 patients with colorectal cancer.

Results

Pathogenic mutations were detected in 1,000 (5%) of 18,734 patients in the development cohort; mutations included MLH1 (n = 306), MSH2 (n = 354), MSH6 (n = 177), PMS2 (n = 141), and EPCAM (n = 22). PREMM₅ distinguished carriers from noncarriers with an AUC of 0.81 (95% CI, 0.79 to 0.82), and performance was similar in the validation cohort (AUC, 0.83; 95% CI, 0.75 to 0.92). Prediction was more difficult for PMS2 mutations (AUC, 0.64; 95% CI, 0.60 to 0.68) than for other genes. Performance characteristics of PREMM₅ exceeded those of PREMM_1,2,6. Decision curve analysis supported germline LS testing for PREMM₅ scores ≥ 2.5%.

Conclusion

PREMM₅ provides comprehensive risk estimation of all five LS genes and supports LS genetic testing for individuals with scores ≥ 2.5%. At this threshold, PREMM₅ provides performance that is superior to the existing PREMM_1,2,6 model in the identification of carriers of LS, including those with weaker phenotypes and individuals unaffected by cancer.

INTRODUCTION

Nearly 1 million individuals in the United States have Lynch syndrome (LS) but most are unaware of their diagnosis.¹ LS is caused by germline alterations in DNA mismatch repair (MMR) genes (MLH1, MSH2, MSH6, or PMS2) or EPCAM (which causes epigenetic silencing of MSH2) and confers a 40% to 80% lifetime risk of colorectal cancer (CRC).^2-4 In addition to CRC, mutation carriers are at increased risk for cancers of the endometrium, ovaries, stomach, small intestine, pancreas, urinary tract, brain, and cutaneous sebaceous glands.^5-8 Personal and family histories of these component cancers can guide genetic testing to identify mutation carriers. If unidentified, these individuals miss the opportunity to pursue interventions known to effectively reduce the risk of Lynch-associated cancers, such as frequent colonoscopies, prophylactic surgeries, and chemoprevention.^9-14

Prediction models, such as PREMM_1,2,6, are evidence-based tools that can help identify carriers of LS by using the personal and family history of Lynch-associated malignancies in an individual to quantify the likelihood of carrying a germline mutation in the MLH1, MSH2, and MSH6 genes.¹⁵ Clinical practice guidelines from various national organizations recommend that individuals with ≥5% likelihood of LS by PREMM_1,2,6 undergo genetic testing.^2,3,16 However, current LS prediction models do not assess for PMS2 or EPCAM mutations.^17,18 Our aim was to develop and validate a new prediction model to quantify the risk of an individual carrying a pathogenic mutation in any of the five genes associated with LS (MLH1, MSH2, MSH6, PMS2, and EPCAM).

PATIENTS AND METHODS

Patients

Development cohort.

We analyzed data from 18,734 patients (probands) who underwent germline testing of the MLH1, MSH2, MSH6, PMS2, and EPCAM genes by Myriad Genetic Laboratories (Salt Lake City, UT) after May 2011. Clinical data were obtained from the test order form completed by health care professionals who ordered germline testing, as previously described,^15,19 including proband age at genetic testing, sex, personal and family cancer histories (including ages at diagnoses), and germline testing results. Cancer history included Lynch-associated malignancies of the colorectum, endometrium, stomach, ovaries, urinary tract, small intestine, pancreas, bile ducts, brain, or sebaceous glands. Probands without personal history of these cancers were defined as unaffected. Family history was limited to first-degree relatives (FDRs) and second-degree relatives (SDRs) on the affected side.

Validation cohort.

Validation of the model was performed in an independent cohort of 1,058 patients with CRC who were recruited to an institutional sample registry at Dana-Farber Cancer Institute from 2008 to 2014, without preselection for high-risk features of LS (eg, age at diagnosis, personal or family cancer histories, or tumor microsatellite instability/MMR deficiency). All participants underwent germline testing for multiple genes, including MLH1, MSH2, MSH6, PMS2, and EPCAM.

Laboratory Methods

Germline analysis was uniform for probands in both cohorts and was performed by Myriad Genetics. MLH1, MSH2, and MSH6 analyses were performed as previously described.¹⁵ PMS2 testing involved sequencing of all exons and adjacent intronic regions as well as large rearrangement testing by multiplex ligation-dependent probe amplification. EPCAM analysis involved large rearrangement testing by microarray comparative genomic hybridization for the 3-prime region. Individuals with deleterious/suspected deleterious mutations were named mutation positive. Patients with polymorphisms, unclassified variants, no alterations, or missense mutations for which clinical significance is not established were named mutation-negative.¹⁵

Statistical Methods

Variables in the previous PREMM_1,2,6 and PREMM_1,2 models were considered for the new PREMM₅ model.^15,19,20 We examined the association of proband sex with mutation presence, because male sex was a predictor of mutations in previous analyses.¹⁵ Proband age at testing was considered for inclusion because of the high proportion of unaffected individuals who underwent testing. Summary statistics for each variable were catalogued by gene type for probands and relatives, and ages at diagnosis of CRC and endometrial cancer were truncated at the lower and upper one percentile to stabilize the resulting model. Patients were excluded if all cancer-related data were missing, if sex was unreported (n = 11), or if two pathogenic mutations (other than EPCAM and MSH2) were detected (n = 5). Multiple imputation was applied for missing values; five completed data sets were analyzed, and results were combined by Rubin’s rules.²¹ The imputation model included personal and family cancer characteristics, and the outcome was presence of a specific gene mutation. Among the 18,734 probands, missing data were imputed for ages at CRC (n = 28), endometrial cancer (n = 28), other cancers (n = 50), and genetic testing (n = 14). Among the 49,237 relatives, missing ages of CRC (n = 1,995), endometrial cancer (n = 407), and other cancers (n = 4,581) was imputed. Associations of overall and gene-specific mutation statuses were analyzed with F tests (continuous variables) and χ² tests (categoric variables). A two-sided P value of less than .05 indicated statistical significance.

Development of the PREMM₅ model.

PREMM₅ predictions of any pathogenic mutation were based on a logistic regression model derived from the full cohort. The associations of predictors with mutation status were reported as odds ratios (ORs) with 95% CIs. Interaction terms were added to test for sex or age-specific effects. Different models were explored and adjusted for unaffected individuals and age at testing. MSH2 and EPCAM were combined into one category (MSH2), because there were few EPCAM mutations and because these mutations induce epigenetic inactivation of MSH2.²² Polytomous logistic regression was used to assess the associations of clinical features with mutation status by the specific gene, in which the response variable was a categoric variable with five levels: (1) MLH1, (2) MSH2 or EPCAM, (3) MSH6, (4) PMS2, and (5) mutation-negative status. For each individual, the PREMM₅ score was calculated as the predicted probability of a mutation within the resulting multivariable polytomous logistic regression model. Internal validation was performed by bootstrap resampling that used 200 random samples drawn with replacement. Predictive models were developed in each bootstrap sample and evaluated in the entire cohort to quantify the optimism in the estimated apparent performance.²³

Assessment of Model Performance.

We quantified the overall ability of the PREMM₅ model to discriminate carriers from noncarriers by the area under the receiver operating characteristic curve (AUC), which also was used for each specific gene in the development cohort.²⁴ Decision curve analysis was used to determine the clinical usefulness of the model, and the true-positive (TP) and false-positive (FP) classifications were considered at increasing decision thresholds. This methodology evaluates prediction models for their potential to improve clinical decision making^24-28 and, in this case, to refer individuals for genetic testing. A decision curve shows the net benefit (NB) of using a model at different thresholds. The NB sums the TPs minus a weighted number of FPs: NB = (TP − wFP) / n, in which n is the total sample size and w is the relative weight of the harm of unnecessary testing versus the benefit of identification of a carrier. The relative weight—w—is defined by the threshold probability to define at-risk patients who need genetic testing. The NB of PREMM₅ and two reference strategies—test none or test all—was calculated. Thresholds between 0% (test all) and 10% (test high-risk probands) were considered, and 5% was the focus, as recommended by national guidelines.^2,3,16 The number needed to test (NNT), which represents the number of patients who have PREMM₅ scores greater than a given threshold who should undergo testing to identify one mutation carrier, was calculated. The same performance metrics were analyzed with the validation cohort.

Comparison of the PREMM₅ and PREMM_1,2,6 Models.

PREMM_1,2,6 predictions were calculated for all patients and were compared with PREMM₅ predictions by using receiver operating characteristic curves and reclassification analysis. Under- or overprediction of each model was quantified by the calibration intercept and was converted to a ratio of observed to expected (O/E) results: O/E = exp (intercept).²⁹ Statistical analysis was performed using SAS (version 9.4) for data management and univariable analysis, and R software (version 3.1.2) was used for multivariable analysis, bootstrap and external validation, and decision curve and reclassification analyses.

The study was investigator-initiated for model development; data collection and germline analyses occurred at Myriad Genetics, and anonymized data sets were provided to investigators at Dana-Farber Cancer Institute and Columbia University. Statistical analyses were conducted by researchers and independent statisticians, all of whom were unaffiliated with Myriad. Investigational review board approval was obtained for this study.

RESULTS

Pathogenic mutations were found in 1,000 (5%) of 18,734 probands in the development cohort: MLH1 (n = 306), MSH2 (n = 354), MSH6 (n = 177), PMS2 (n = 141), and EPCAM (n = 2; Table 1). A total of 15,363 (82%) of 18,734 participants were women. Pathogenic mutations were more frequent among men (370 [11%] of 3,371) than women (630 [4%] of 15,363; P < .001).

Table 1.

Clinical Characteristics of Gene Mutation Carriers Stratified by Gene

Carriers of gene mutations were younger at the time of genetic testing than noncarriers (47.0 v 49.4 years of age), and carriers of MLH1, MSH2, and EPCAM were younger at the time of testing (45.9 v 46.2 v 46.3 years of age, respectively) than carriers of MSH6 and PMS2 (49 years of age for both; P < .001). Forty-six percent of probands (8,590 of 18,734) were unaffected by cancer, including 200 (20%) of 1,000 mutation carriers.

Multiple CRCs and other LS-associated malignancies were more frequent among carriers of MLH1, MSH2, and EPCAM than carriers of MSH6 and PMS2. Carriers of MSH6 and PMS2 were older than other carriers diagnosed with CRC (P < .001) or endometrial cancer (P = .034) (Appendix Table A3, online only). Carriers of PMS2 had fewer relatives with CRC (P < .001) and fewer FDRs with endometrial cancer (P < .001) compared with other carriers.

In the validation cohort, the mean CRC age was 55.7 years (standard deviation, 12.6 years), and 587 (55.5%) of 1,058 were men (Appendix Table A4, online only). Pathogenic gene mutations were detected in 33 (3.1%) of 1,058 individuals: 13 with MLH1, seven with MSH2, six with MSH6, seven with PMS2, and none with EPCAM.

Development, Validation, and Performance of PREMM₅

In multivariable analysis (Table 2), younger age at testing was associated with the presence of any gene mutation (OR, 0.69 per decade; 95% CI, 0.66 to 0.73) and added into PREMM₅ as a new predictor. After analysis was adjusted for all other predictors, male sex was associated with the presence of any mutation (OR, 2.22; 95% CI, 1.88 to 2.61); other predictors included personal history of CRC with one occurrence (OR, 6.18, 95% CI, 5.22 to 7.32), or multiple occurrences (OR, 8.54; 95% CI, 5.65 to 12.93), and relatives with CRC (OR, 3.02; 95% CI, 2.76 to 3.31). Personal and family CRC diagnoses and corresponding ages were less predictive of PMS2 mutations (OR, 2.69; 95% CI, 1.78 to 4.07 and OR, 1.00; 95% CI, 0.76 to 1.32, respectively) compared with other genes. Personal history of endometrial cancer was associated with any mutation (OR, 5.42; 95% CI, 4.39 to 6.68), but age was only predictive of MSH6 mutations. Personal history of other LS-associated cancers had an OR of 3.16 (95% CI, 2.49 to 4.02) for any mutation but was not predictive of MSH6 or PMS2 mutations. Family history of endometrial cancer was predictive of all gene mutations except PMS2. Family history of other LS-associated cancers was weakly predictive of any mutation (OR, 1.50; 95% CI, 1.29 to 1.74) but not of MSH6 or PMS2. The PREMM₅ prediction model is available online³⁰ and in Appendix Tables A1 and A2 (online only).

Table 2.

Multivariable Logistic Regression Analyses for the Presence of Lynch Syndrome Gene Mutations

PREMM₅ performed well in discriminating carriers from noncarriers; the optimism-corrected AUC was 0.81 (95% CI, 0.79 to 0.82; Fig 1). The polytomous multivariable model showed good discrimination for MLH1 (AUC, 0.89; 95% CI, 0.87 to 0.91), MSH2/EPCAM (AUC, 0.84; 95% CI, 0.82 to 0.86), and MSH6 (AUC, 0.76; 95% CI, 0.73 to 0.79) but less discrimination for PMS2 (AUC, 0.64; 95% CI, 0.60 to 0.68; Appendix Table A5, online only). PREMM₅ had similar discrimination on external validation (AUC, 0.83; 95% CI, 0.75 to 0.92; Fig 1).

Fig 1.

Receiver operating characteristic (ROC) curves to discriminate mismatch repair mutation carriers from noncarriers for the PREMM₅ and the PREMM_1,2,6 models. AUC, area under the ROC curve.

The median PREMM₅ score for carriers was 9.8% versus 2.6% for noncarriers. At the recommended ≥ of 5% threshold, 721 of 1,000 carriers were identified (Fig 2) with variability by specific gene (Fig 3). A threshold ≥ 2.5% identified 894 of 1,000 carriers, but with lower specificity. The NNT to identify one carrier at 5% and 2.5% was seven and 11 individuals, respectively; a decrease from 19 needed to test if PREMM₅ was not used. Between 2.5% and 10%, the negative predictive value was 97% to 99% of individuals correctly identified as mutation negative. Similar sensitivity, specificity, and NNT data were observed at external validation with 2.5% and 5.0% thresholds (Appendix Fig A1, online only).

Fig 2.

Performance characteristics of PREMM₅. Decreasing sensitivity and increasing specificity are shown for increasing risk thresholds for the PREMM₅ score, with a histogram for the distribution of predicted risks. FN, false negative; FP, false positive; NaN, not a number; NNT, number needed to test; NPV, negative predictive value; PPV, positive predictive value; TN, true negative; TP, true positive. FN, FP, TN, and TP are reported for a given cutoff value on the PREMM₅ score. NNT= (TP + FP) / TP.

Fig 3.

Box plots of PREMM₅ scores by specific gene mutation at model development. C-statistic: area under the receiver operating characteristic curve. Median values are represented by the horizontal lines in the boxes; 25th and 75th percentiles are represented by the lower and upper lines of the boxes, respectively. NA, not available.

By decision curve analysis, the clinical impact of PREMM₅ to identify individuals for germline testing was observed at thresholds ≥ 2.5%; maximal utility occurred at 5%. A threshold ≥ 2.5% to guide testing was superior compared with testing of all patients (Fig 4). PREMM₅ had minimal clinical impact at less than 2.5%, because few individuals would be excluded from genetic testing at this threshold, so it was similar to a test-all approach.

Fig 4.

Net benefit curves for PREMM₅ compared with PREMM_1,2,6. The y-axis measures net benefit, which is calculated by summing the benefits (true positives) and subtracting the harms (false positives), in which the latter are weighted by a factor related to the relative harm of a missed mutation carrier compared with the harm of unnecessary genetic testing. A model is considered of clinical value if it has the highest net benefit compared with other models and simple strategies, such as performing genetic testing in all patients (gold line) or no patients (horizontal blue line) across the full range of threshold probabilities at which a patient would undergo genetic testing. For example, the net benefit of using PREMM₅ to selectively test for mutation carriers exceeds that of PREMM_1,2,6, and testing all at risk threshold ≥ 2.5%.

Performance Comparison of PREMM₅ and PREMM_1,2,6

PREMM_1,2,6 had an AUC of 0.83 (95% CI, 0.82 to 0.84) for the identification of carriers of MLH1, MSH2, or MSH6 among the 18,734 probands. Extending prediction to all five genes decreased discrimination to 0.79 (95% CI, 0.78 to 0.81), whereas PREMM₅ prediction yielded an AUC of 0.81 (95% CI, 0.79 to 0.82). PREMM_1,2,6 overpredicted mutation-positive status (ie, the observed mutation fraction for carriers of MLH1, MSH2, and MSH6 was 4.5% but the average PREMM_1,2,6 prediction was 8.0% [O/E ratio, 0.557]). Reclassification plots confirmed overprediction by PREMM_1,2,6 and showed considerable differences between PREMM₅ and PREMM_1,2,6 predictions (Appendix Fig A2, online only).

DISCUSSION

To improve the identification of carriers of LS mutations, we developed and validated the PREMM₅ model, a risk assessment tool that uses readily ascertained clinical features (age, sex, and personal and family cancer history) to provide accurate and comprehensive risk estimation for all five associated genes. PREMM₅ is a simple and efficient online tool for a diverse array of health care providers (eg, oncologists, gynecologists, gastroenterologists, and primary care practitioners) to rapidly identify individuals for germline testing for LS in routine clinical practice. The performance of the model is robust in quantifying an individual’s overall risk of carrying any pathogenic gene mutation, and its performance exceeds that of the previous model, PREMM_1,2,6, which only predicted MLH1, MSH2, and MSH6 gene mutation status. We propose that PREMM₅ should replace PREMM_1,2,6 and that individuals with a ≥ 2.5% likelihood of LS undergo genetic testing.

Beyond the use of prediction models such as PREMM₅, the primary strategy to identify individuals with LS involves screening CRC and endometrial cancer tumor specimens for microsatellite instability or deficient MMR protein expression, followed by germline testing of individuals who have suggestive tumor testing results.^2,3 The limitation of tumor testing is that it is only relevant to patients with cancer, so the opportunity for cancer prevention in the proband has inherently been lost. Prior versions of PREMM and other LS models (eg, MMRpro, MMRpredict) were developed and validated in cohorts predominated by patients with cancer,^17,18,31-35 and the performance of these models in unaffected patients is poorly understood. In this study, 46% of the development cohort (including 20% of LS mutation carriers) were unaffected but had a family history of LS-associated cancer, which supports the potential use of PREMM₅ as a risk assessment tool for unaffected individuals.

Despite existing recommendations for universal molecular tumor testing of all colorectal and endometrial cancers to screen for LS, the majority of the nearly 1 million carriers in the United States remain unidentified.¹ Use of the PREMM₅ model can ameliorate the gap that exists between those identified through tumor testing and the vast majority of undiagnosed carriers of LS, many of whom are unaffected by cancer. Therefore, systematic implementation of PREMM₅ can be considered by providers in routine clinical care, including primary and preventive health care, for individuals with a personal or family history of LS-associated cancers. Also, the PREMM₅ model can be used in individuals who have colorectal and endometrial cancer when molecular tumor testing is unavailable or when resources are limited and universal tumor testing cannot be adopted.

The cohort used to develop PREMM₅ was far larger than those used for older prediction models, including PREMM_1,2,6. We examined data on 1,000 mutation carriers, in which 32% carried pathogenic PMS2 and MSH6 mutations, the less penetrant but more prevalent MMR genes associated with LS.³⁶ Although PREMM₅ performs well estimating individuals’ overall risk of carrying any MMR gene mutation, variability exists in estimates for each individual gene. Reliable prediction of PMS2 was more difficult than other genes because of a weaker phenotype—carriers of PMS2 were older at CRC diagnoses and had fewer cancers among relatives compared with other families with LS.³⁷ This is in line with studies of carriers of the PMS2 gene mutation, in which retention of mRNA expression resulted in a milder phenotype in patients who were older at the time of their CRC diagnosis and/or had no family history of CRC.³⁸ The performance measures associated with PREMM₅ were comparable in the cohort used for external validation. A strength of the validation data set is that it minimized selection bias; patients had CRC but were not selected for genetic testing for LS because of features suggestive of the condition, such as young age of diagnosis, fulfillment of clinical criteria, or results of MMR tumor testing.

The large size of the development cohort allowed us to reassess previous predictors of mutation carrier status, such as sex of the individual tested. The frequency of pathogenic mutations differed between men and women (11% v 4% mutation prevalence, respectively), despite the presence of more women (73%) in the data set. This is consistent with prior reports⁶ and may be due to unmeasured selection bias in the development cohort. The sex distribution was more even in the validation cohort and did not alter the ability of the model to discern mutation carriers from noncarriers.

Incorporation of age at genetic testing improved risk estimation with PREMM₅; with every decade increase in age, the likelihood an individual would carry a pathogenic mutation decreased by 32%. This refinement compared with PREMM_1,2,6 is relevant to individuals unaffected by cancer, because the lack of such history in a 75-year-old patient who undergoes genetic evaluation is strong evidence against LS, whereas the absence of personal cancer history is less reassuring in a 25-year-old patient.

At the currently recommended threshold of 5%, the PREMM₅ model identifies fewer individuals than PREMM_1,2,6 for genetic evaluation and has a higher likelihood than PREMM_1,2,6 of mutation carrier detection. Compared with PREMM₅, PREMM_1,2,6 overpredicts an individual’s need to obtain genetic evaluation. With PREMM₅ at a ≥ 5% threshold, seven patients would need to be tested to identify one mutation carrier, with a negative predictive value of 98% (Fig 1). However, a ≥ 2.5% PREMM₅ score cutoff to guide LS genetic testing markedly increased the number of identified mutation carriers and preserved a high negative predictive value of 99% compared with the ≥ 5% cutoff endorsed by national guidelines for PREMM_1,2,6.

With ongoing advances in next-generation DNA sequencing technologies, many commercial laboratories currently offer multigene hereditary cancer panels that provide simultaneous germline analysis of dozens of cancer risk genes. PREMM₅ provides accurate and comprehensive risk assessment specifically for LS, so its performance must be examined when genetic testing includes an expanded panel of genes. Recent data to examine such panels have shown that individuals with PREMM_1,2,6 scores ≥ 5% often have mutations in cancer susceptibility genes beyond those linked to LS, including APC, MUTYH, BRCA1, and BRCA2.³⁹ These results suggest that prediction models such as PREMM₅ may identify individuals who may have underlying mutations in a wide spectrum of syndromes, rather than just LS.

In conclusion, PREMM₅ provides comprehensive LS risk estimation for all five genes, including PMS2 and EPCAM. PREMM₅ identifies fewer high-risk individuals for genetic evaluation and testing who have a higher likelihood of carrying a pathogenic mutation compared with PREMM _1,2,6. These analyses support the use of a ≥ 2.5% threshold to identify individuals suitable for genetic evaluation for LS; this threshold optimizes identification of carriers of gene mutations, including those who carry MMR genes associated with a weaker phenotype or who are unaffected by cancer but have a family history suggestive of an inherited CRC syndrome.

Appendix

Equation for the PREMM₅ Model

We encourage researchers who use this equation to contact the PREMM investigators for guidance on the appropriate application. Variables are explained in detail in Appendix Table A1. The use of the term “PREMM” is protected by a servicemark.

Predicted probability of any mismatch repair gene mutation: p(any) = predicted probability of MLH1 mutation + predicted probability of MSH2/EPCAM mutation + predicted probability of MSH6 mutation + predicted probability of PMS2 mutation

• Predicted probability of a mutation in MLH1: p(MLH1)

• Predicted probability of a mutation in MSH2 or EPCAM: p(MSH2/EPCAM)

• Predicted probability of a mutation in MSH6: p(MSH6)

• Predicted probability of a mutation in PMS2: p(PMS2)

• Predicted probability of no mutation: p(none)

• p(none) = 1 − [p(MLH1) + p(MSH2/EPCAM) + p(MSH6) + p(PMS2)]

• p(MLH1) = exp (lp(MLH1)) / [(1 + exp(lp(MLH1)) + exp(lp(MSH2/EPCAM)) + exp (lp(MSH6)) + exp(lp(PMS2))]

• p(MSH2/EPCAM) = exp(lp(MSH2)) / [(1 + exp(lp(MLH1)) + exp(lp(MSH2)) + exp(lp(MSH6) + exp(lp(PMS2))]

• p(MSH6) = exp(lp(MSH6)) / [(1 + exp(lp(MLH1)) + exp(lp(MSH2)) + exp(lp(MSH6)) + exp(lp(PMS2))]

• p(PMS2) = exp(lp(PMS2)) / [(1 + exp(lp(MLH1)) + exp(lp(MSH2)) + exp(lp(MSH6)) + exp(lp(PMS2))]

• lp(MLH1) = −5.325 + (0.904 × V0) + (2.586 × V1) + (3.183 × V2) + (1.621 × V3) + (1.276 × V4) + (1.560 × V5) + (0.804 × V6) + (0.397 × V7) + (−0.0557 × V8) + (0.0115 × V9) + (−0.0476 × V10);

• lp(MSH2/EPCAM) = −4.427 + (0.937 × V0) + (1.799 × V1) + (2.593 × V2) + (1.924 × V3) + (1.585 × V4) + (1.337 × V5) + (0.670 × V6) + (0.607 × V7) + (−0.0441 × V8) + (0.0002 × V9) + (−0.0482 × V10);

• lp(MSH6) = −4.675 + (0.816 × V0) + (1.265 × V1) + (−53.205 × V2) + (1.759 × V3) + (0.538 × V4) + (0.545 × V5) + (0.923 × V6) + (0.313 × V7) + (−0.0095 × V8) + (0.0344 × V9) + (−0.0363 × V10);

• lp(PMS2) = −4.913 + (0.294 × V0) + (0.989 × V1) + (−0.354 × V2) + (0.739 × V3) + (0.395 × V4) + (−0.002 × V5) + (−0.426 × V6) + (−0.105 × V7) + (−0.0086 × V8) + (0.0008 × V9) + (−0.0074 × V10);

Table A1.

Definitions for Variables of the PREMM₅ Model Equation

Table A2.

Ages at Diagnoses of CRC and EC in Calculation of Overall Risk Estimate

Table A3.

Age of Diagnosis Among Gene Mutation Carriers and Noncarriers in the Development Cohort

Table A4.

Clinical Characteristics of Individuals With Colorectal Cancer in the External Validation Cohort

Table A5.

Discrimination of PREMM₅ for Any MMR Gene and by Specific Gene

Fig A1.

Performance characteristics of PREMM₅: validation cohort. FN, false negative; FP, false positive; NNT, number needed to test; NPV, negative predictive value; PPV, positive predictive value; TN, true negative; TP, true positive.

Fig A2.

Reclassification plots for the comparison of PREMM_1,2,6 with PREMM₅. The scatterplot diagrams show the added predictive value provided by the PREMM₅ model in identification of (A) mutation carriers (mutation in any gene) and (B) noncarriers (no mutation in any gene) on the basis of the data from this study. (A) PREMM₅ predicts mutation carriers, as shown by patients (blue circles) above the line. This is regardless of the mutation carrier status of the patient. (B) Compared with PREMM_1,2,6, the extended PREMM₅ model better predicts noncarriers, as shown by more patients (blue circles) below the line of identity (diagonal line). Gold circles indicate reclassification by PREMM₅, in which, on average, the extended model provided lower risk estimates than PREMM_1,2,6.

AUTHOR CONTRIBUTIONS

Conception and design: Fay Kastrinos, Matthew H. Kulke, Deborah Schrag, Jeffrey A. Meyerhardt, Robert J. Mayer, Kimmie Ng, Sapna Syngal

Provision of study materials or patients: Charles S. Fuchs

Collection and assembly of data: Fay Kastrinos, Chinedu Ukaegbu, Carmelita Alvero, Ashley McFarland, Matthew H. Kulke, Deborah Schrag, Jeffrey A. Meyerhardt, Charles S. Fuchs, Robert J. Mayer, Kimmie Ng

Data analysis and interpretation: Fay Kastrinos, Hajime Uno, Chinedu Ukaegbu, Carmelita Alvero, Ashley McFarland, Matthew B. Yurgelun, Ewout W. Steyerberg, Sapna Syngal

Manuscript writing: All authors

Final approval of manuscript: All authors

Accountable for all aspects of the work: All authors

AUTHORS' DISCLOSURES OF POTENTIAL CONFLICTS OF INTEREST

Development and Validation of the PREMM₅ Model for Comprehensive Risk Assessment of Lynch Syndrome

The following represents disclosure information provided by authors of this manuscript. All relationships are considered compensated. Relationships are self-held unless noted. I = Immediate Family Member, Inst = My Institution. Relationships may not relate to the subject matter of this manuscript. For more information about ASCO's conflict of interest policy, please refer to www.asco.org/rwc or ascopubs.org/jco/site/ifc.

Fay Kastrinos

No relationship to disclose

Hajime Uno

Consulting or Advisory Role: Ion Pharma

Chinedu Ukaegbu

No relationship to disclose

Carmelita Alvero

No relationship to disclose

Ashley McFarland

No relationship to disclose

Matthew B. Yurgelun

Research Funding: Myriad Genetics

Matthew H. Kulke

No relationship to disclose

Deborah Schrag

Consulting or Advisory Role: New Century Health, Ohio State University

Other Relationship: Journal of the American Medical Association

Jeffrey A. Meyerhardt

Consulting or Advisory Role: Genentech

Charles S. Fuchs

Consulting or Advisory Role: Genentech (a Roche group), Lilly, Sanofi, Bayer, Celgene, Merck, Bristol-Myers Squibb, Entrinic Health, Five Prime Therapuetics, Agios

Robert J. Mayer

Honoraria: Taiho Pharmaceutical

Consulting or Advisory Role: CASI Pharmaceuticals

Kimmie Ng

Consulting or Advisory Role: Genentech, Lilly, Tarrex

Research Funding: Gilead Sciences, Travegene, Tarrex

Ewout W. Steyerberg

Patents, Royalties, Other Intellectual Property: Royalties from Springer for book on prediction models

Sapna Syngal

No relationship to disclose