Comparison of aggregation methods for multiphenotype exomic variant prioritization

The identification of disease-causing genes in Mendelian disorders has been facilitated by the detection of rare disease-causing variation through exome sequencing experiments. These studies rely on population databases to filter a majority of the putatively neutral variation in the genome and additional filtering steps using either cohorts of diseased individuals or familial information to narrow down the list of candidate variants. Recently, new computational methods have been proposed to prioritize variants by scoring them not only based on their potential impact on protein function but also on their relevance given the available information on the disease under study. Usually these diseases comprise several phenotypic presentations, which are separately prioritized and then aggregated into a global score. In this study we compare several simple (e.g. maximum and mean score) and more complex aggregation methods (e.g. order statistics, parametric modeling) in order to obtain the best possible prioritization performance. We show that all methods perform reasonably well (median rank below 20 out of more than 8000 variants) and that the selection of an optimal aggregation method depends strongly on the fraction of uninformative phenotypes. Finally, we propose guidelines as to how to select an appropriate aggregation method based on knowledge of the phenotype under study.