TRIO RVEMVS: A Bayesian framework for rare variant association analysis with expectation-maximization variable selection using family trio data

by Duo Yu, Matthew Koslovsky, Margaret C. Steiner, Kusha Mohammadi, Chenguang Zhang, Michael D. Swartz

It is commonly reported that rare variants may be more functionally related to complex diseases than common variants. However, individual rare variant association tests remain challenging due to low minor allele frequency in the available samples. This paper proposes an expectation maximization variable selection (EMVS) method to simultaneously detect common and rare variants at the individual variant level using family trio data. TRIO_RVEMVS was assessed in both large (1500 families) and small (350 families) datasets based on simulation. The performance of TRIO_RVEMVS was compared with gene-level kernel and burden association tests that use pedigree data (PedGene) and rare-variant extensions of the transmission disequilibrium test (RV-TDT). At the region level, TRIO_RVEMVS outperformed PedGene and RV-TDT when common variants were included. TRIO_RVEMVS performed competitively with PedGene and outperformed RV-TDT when the analysis was only restricted to rare variants. At the individual variants level, with 1,500 trios, the average true positive rate of individual rare variants that were polymorphic across 500 datasets was 12.20%, and the average false positive rate was 0.74%. In the datasets with 350 trios, the average true and false positive rates of individual rare variants were 13.10% and 1.30%, respectively. When applying TRIO_RVEMVS to real data from the Gabriella Miller Kids First Pediatric Research Program, it identified 3 rare variants in q24.21 and q24.22 associated with the risk of orofacial clefts in the Kids First European population.