J Med Genet 50:785-793 doi:10.1136/jmedgenet-2013-101803
  • Review

Criteria and prediction models for mismatch repair gene mutations: a review

Editor's Choice
  1. Mark A Jenkins1
  1. 1Centre for Molecular, Environmental, Genetic and Analytic Epidemiology, The University of Melbourne, Parkville, Victoria, Australia
  2. 2Cancer Epidemiology Centre, Cancer Council Victoria, Carlton, Victoria, Australia
  1. Correspondence to Dr Aung Ko Win, Centre for Molecular, Environmental, Genetic and Analytic Epidemiology, Melbourne School of Population and Global Health, Level 3, 207 Bouverie Street, The University of Melbourne, Parkville, VIC 3010, Australia; awin{at}
  • Received 8 May 2013
  • Revised 2 July 2013
  • Accepted 17 July 2013
  • Published Online First 16 August 2013


One of the strongest predictors of colorectal cancer risk is carrying a germline mutation in a DNA mismatch repair (MMR) gene. Once identified, mutation carriers can be recommended for intensive screening that will substantially reduce their high colorectal cancer risk. Conversely, the relatives of carriers identified as non-carriers can be relieved of the burden of intensive screening. Criteria and prediction models that identify likely mutation carriers are needed for cost-effective, targeted, germline testing for MMR gene mutation. We reviewed 12 criteria/guidelines and 8 prediction models (Leiden, Amsterdam-plus, Amsterdam-alternative, MMRpro, PREMM1,2,6, MMRpredict, Associazione Italiana per lo studio della Familiarità ed Ereditarietà dei tumori Gastrointestinali (AIFEG) and the Myriad Genetics Prevalence table) for identifying mutation carriers. While criteria are only used to identify individuals with colorectal cancer (yes/no for screening followed by germline testing), all prediction models except MMRpredict and Myriad tables can predict the probability of carrying mutations for individuals with or without colorectal cancer. We conducted a meta-analysis of the discrimination performance of 17 studies that validated the prediction models. The pooled estimate for the area under curve was 0.80 (95% CI 0.72 to 0.88) for MMRpro, 0.81 (95% CI 0.73 to 0.88) for MMRpredict, 0.84 (95% CI 0.81 to 0.88) for PREMM, and 0.85 (95% CI 0.78 to 0.91) for Leiden model. Given the high degree of overlap in the CIs, we cannot state that one model has a higher discrimination than any of the others. Overall, the existing statistical models have been shown to be sensitive and specific (at a 5% cut-off) in predicting MMR gene mutation carriers. Future models may need to: provide prediction of PMS2 mutations, take into account a wider range of Lynch syndrome-associated cancers when assessing family history, and be applicable to all people irrespective of any cancer diagnosis.