Seminar Series: Tissue specific transcriptome prediction and gene-level association mapping and fine-mapping

Tissue-specific gene expressions have direct relevance to disease phenotypes, and knowing gene expression in disease-relevant tissues is advantageous in both detecting novel associations and fine-mapping known genetic associations. Unfortunately, almost all genome-wide association study (GWAS) datasets have no companion gene expression assay, let alone tissue-specific ones. The idea of predicting gene expression to perform gene-level association mapping was pioneered by others. Their software PrediXcan uses penalized regression to predict gene expression. We present a Bayesian approach to impute tissue-specific gene expression imple- mented in our software package fastBVSR. We use Bayesian variable selection regres- sion (BVSR) to perform prediction, and our innovative iterative complex factorization algorithm, which efficiently solves ridge-regression in the context of BVSR, makes the computation feasible to analyze tens of thousands gene expressions overnight. BVSR has inherent advantage from Bayesian model averaging and outperforms penalized regression in prediction. When using Genotype-Tissue Expression (GTEx, 338 whole blood sample) as training and Depression Genes and Networks (DGN, 922 whole blood samples) as testing datesets, fastBVSR significantly out-performed PrediXcan in out-of-sample prediction. Using predicted gene expressions in a specific tissue type as surrogate genotypes, we can test gene-level associations. Through jointly analyzing expressi