Cystic fibrosis transmembrane conductance regulator (CFTR) mutation is associated with a phenotypic spectrum that includes cystic fibrosis (CF). The disease liability of some common CFTR mutations is known, but rare mutations are seen in too few patients to categorize unequivocally, making genetic diagnosis difficult. Computational methods can predict the impact of mutation, but prediction specificity is often below that required for clinical utility. Here, we present a novel supervised learning approach for predicting CF from CFTR missense mutation. The algorithm begins by constructing custom multiple sequence alignments called phenotype-optimized sequence ensembles (POSEs). POSEs are constructed iteratively, by selecting sequences that optimize predictive performance on a training set of CFTR mutations of known clinical significance. Next, we predict CF disease liability from a different set of CFTR mutations (test-set mutations). This approach achieves improved prediction performance relative to popular methods recently assessed using the same test-set mutations. Of clinical significance, our method achieves 94% prediction specificity. Because databases such as HGMD and locus-specific mutation databases are growing rapidly, methods that automatically tailor their predictions for a specific phenotype may be of immediate utility. If the performance achieved here generalizes to other systems, the approach could be an excellent tool to help establish genetic diagnoses. Hum Mutat 33:1267–1274, 2012. © 2012 Wiley Periodicals, Inc.