A framework for predicting tissue-specific 1 effects of rare genetic variants 2

17 Despite the abundance of rare genetic variants—variants carried by less than one percent of the population—in human genomes, the impact of these variants on specific tissues has been largely uncharacterized. Population-level test statistics, while effective in understanding the impact of common variants—variants carried by at least five percent of the population, have had limited success in characterizing the effect of rare variants mainly due to limited statistical power. In addition, the effect of each rare variant can vary greatly between specific tissues. This heterogeneity coupled with limited sample sizes and a lack of known disease-causing rare variants makes predicting tissue-specific cellular consequences of rare variants a difficult task. To make these predictions, we propose a new method called SPEER (SPecific tissuE variant Effect predictoR): a hierarchical Bayesian model that uses transfer learning, allowing separate predictions in each tissue while flexibly sharing signal across tissues to improve power. Our probabilistic model capitalizes on a growing body of rich epigenetic annotations to inform the consequences of a variant in specific tissues. These annotations are integrated with tissue-specific RNA expression levels and common variants. We show our method improves prediction accuracy in simulations and in genomic data from the Genotype-Tissue Expression (GTEx) project. 18