The human genome contains frequent single-basepair variants that may or may not cause genetic disease. To characterize benign vs. pathogenic missense variants, numerous computational algorithms have been developed based on comparative sequence and/or protein structure analysis. We compared computational methods that use evolutionary conservation alone, amino acid (AA) change alone, and a combination of conservation and AA change in predicting the consequences of 254 missense variants in the CDKN2A (n = 92), MLH1 (n = 28), MSH2 (n = 14), MECP2 (n = 30), and tyrosinase (TYR) (n = 90) genes. Variants were validated as either neutral or deleterious by curated locus-specific mutation databases and published functional data. All methods that use evolutionary sequence analysis have comparable overall prediction accuracy (72.9-82.0%). Mutations at codons where the AA is absolutely conserved over a sufficient evolutionary distance (about one-third of variants) had a 91.6 to 96.8% likelihood of being deleterious. Three algorithms (SIFT, PolyPhen, and A-GVGD) that differentiate one variant from another at a given codon did not significantly improve predictive value over conservation score alone using the BLOSUM62 matrix. However, when all four methods were in agreement (62.7% of variants), predictive value improved to 88.1%. These results confirm a high predictive value for methods that use evolutionary sequence conservation, with or without considering protein structural change, to predict the clinical consequences of missense variants. The methods can be generalized across genes that cause different types of genetic disease. The results support the clinical use of computational methods as one tool to help interpret missense variants in genes associated with human genetic disease.

Variants in the CDKN2A tumor suppressor are associated with Familial Melanoma (FM), although for many variants the linkage is weak. The effects of missense variants on protein function and pathogenicity are often unclear. Multiple methods (e.g., laboratory, computational, epidemiological) have been developed to analyze whether a missense variant is pathogenic or not. It is not yet clear how to integrate these data types into a strategy for variant classification. We studied 51 CDKN2A missense variants using a cell cycle arrest assay. There was a continuum of results ranging from full wild-type effect through partial activity to complete loss of arrest. A reproducible decrease of 30% of cell cycle arrest activity correlated with FM association. We analyzed missense CDKN2A germline variants using a Bayesian method to combine multiple data types and derive a probability of pathogenicity. When equal to or more than two data types could be evaluated with this method, 22 of 25 FM-associated variants and 8 of 15 variants of uncertain significance were classified as likely pathogenic with 495% probability. The other 10 variants were classified as uncertain (probability 5-95%). For most variants, there were insufficient data to draw a conclusion. The Bayesian model appears to be a sound method of classifying missense variants in cancer susceptibility genes. Hum Mutat 32:900-911,