Sparse Bayesian latent factor decompositions for identifying trans-eQTLs. V. Hore1, J. Marchini1,2 1) Department of Statistics, University of Oxford, Oxford OX1 3TG, UK; 2) Wellcome Trust Centre for Human Genetics, Oxford OX3 7BN, UK.

   Expression quantitative trait loci (eQTL) mapping aims to uncover the role of genetic variation in gene regulation. Many approaches to eQTL mapping employ mass univariate regression between nearby genes and SNPs, and although these techniques have been successful in finding cis-eQTLs, it is harder to identify trans-eQTLs due to their non-local nature. These techniques do not explicitly model the joint effect of genetic variation across networks of genes, nor do they naturally extend to allow for simultaneous analysis of multiple tissues.
   We have developed a method for identifying trans-eQTLs in multiple tissues by jointly modeling correlations between genes and tissues. Our approach is a general framework for decomposing matrices and tensors into sparse latent factors, where latent factors consist of networks of co-varying genes. The model also determines the subset of tissues in which each latent factor is active. We fit our model using variational Bayes, which allows for relatively fast inference on large data sets, and imputation of missing data. In addition, the method is capable of integrating other information to inform correlations in gene expression, such as genotypes (both directly or through a kinship matrix) and measured covariates.
   Via simulations we are able to demonstrate that the method can reliably identify trans-eQTLs, in the presence of cis-eQTLs, multiple confounding factors and realistic levels of experimental noise. We will also report results of applying our method to the GTEx Project data, which consists of gene expression in up to 30 tissues for each individual.

You may contact the first author (during and after the meeting) at