Computational methods are used to predict the molecular consequences of amino-acid substitutions on the basis of evolutionary conservation or protein structure, but their utility in clinical diagnosis or prediction of disease outcome has not been well validated. We evaluated three popular computer programs, namely, PANTHER, SIFT and PolyPhen, by comparing the predicted clinical outcomes for a group of known CFTR missense mutations against the diagnosis of cystic fibrosis (CF) and clinical manifestations in cohorts of subjects with CF-disease and CFTR-related disorders carrying these mutations. Owing to poor specificity, none of tools reliably distinguished between individual mutations that confer CF disease from mutations found in subjects with a CFTR-related disorder or no disease. Prediction scores for CFTR mutations derived from PANTHER showed a significant overall statistical correlation with the spectrum of disease severity associated with mutations in the CFTR gene. In contrast, PolyPhen- and SIFT-derived scores only showed significant differences between CF-causing and non-CF variants. Current computational methods are not recommended for establishing or excluding a CF diagnosis, notably as a newborn screening strategy or in patients with equivocal test results.