MIRAGE: a Bayesian rare variant association analysis method incorporating functional information of variants

Many biobanks and consortiums have generated large whole-exome and whole-genome sequencing datasets, making it possible to perform rare variant association studies. Compared to common variants, rare variants tend to have larger effect sizes and generally have much lower linkage disequilibrium (LD), making it easier to identify causal variants. These potential benefits make association analysis with rare variants a priority for human genetics researchers. Because of the low frequencies of these variants, however, the power of single-variant association is often low. To improve the power, numerous methods have been developed to aggregate information of all variants of a gene to better identify risk genes. These gene-based methods, however, often make unrealistic assumptions, e.g. all rare variants in a risk gene have the same effects. In practice, current gene-based analysis methods often fail to show any advantage over simple single-variant analysis. In this work, we develop a Bayesian method: MIxture model based Rare variant Analysis on GEnes (MIRAGE). MIRAGE captures the heterogeneity of variant effects by treating all variants of a gene as a mixture of risk and non-risk variants, and models the prior probabilities of being risk variants as function of external information of variants, such as allele frequencies and predicted deleterious effects. MIRAGE uses an empirical Bayes approach, combining information across genes to estimate these prior probabilities. We demonstrate in both simulations and analysis of an exome-sequencing dataset of Autism, that MIRAGE significantly outperforms current methods for rare variant analysis. The top genes identified by MIRAGE are highly enriched with known or plausible Autism risk genes. Our results highlight several novel Autism genes. MIRAGE is available at https://xinhe-lab.github.io/mirage .