Use of estimated evolutionary strength at the codon level improves the prediction of disease-related protein mutations in humans

Predicting the functional impact of protein variation is one of the most challenging problems in bioinformatics. A rapidly growing number of genome-scale studies provide large amounts of experimental data, allowing the application of rigorous statistical approaches for predicting whether a given single point mutation has an impact on human health. Up until now, existing methods have limited their source data to either protein or gene information. Novel in this work, we take advantage of both and focus on protein evolutionary information by using estimated selective pressures at the codon level. Here we introduce a new method (SeqProfCod) to predict the likelihood that a given protein variant is associated with human disease or not. Our method relies on a support vector machine (SVM) classifier trained using three sources of information: protein sequence, multiple protein sequence alignments, and the estimation of selective pressure at the codon level. SeqProfCod has been benchmarked with a large dataset of 8,987 single point mutations from 1,434 human proteins from SWISS-PROT It achieves 82% overall accuracy and a correlation coefficient of 0.59, indicating that the estimation of the selective pressure helps in predicting the functional impact of single-point mutations. Moreover, this study demonstrates the synergic effect of combining two sources of information for predicting the functional effects of protein variants: protein sequence/profile-based information and the evolutionary estimation of the selective pressures at the codon level. The results of large-scale application of SeqProfCod over all annotated point mutations in SWISS-PROT (available for download at http://sgu.bioinfo.cipf.es/services/Omidios/; last accessed: 24 August 2007), could be used to support clinical studies.