INForum'2017 - Sérgio Matos

ABSTRACT. The capability to predict the pathogenicity of genetic vari- ants is of utmost importance for prioritizing the results of high-throughput analysis. This paper compares the performance of different classifiers in predicting the pathogenicity of single nucleotide variants. Furthermore, we analyse different encoding schemes for training deep learning classi- fiers for this problem. The results on five standard datasets show that Support Vector Machines, Logistic Regression, and AdaBoost achieve high classification accuracy, with area-under-the-curve between 0.862 and 0.898. The deep learning approaches, based on convolutional neural networks, performed worst than classical classifiers.