Saikat Banerjee

Genetic variants regulating distant target genes are called trans-acting expression quantitative trait loci (trans-eQTLs). Many genetic variants are believed to mediate disease risk via the trans-eQTLs. It is crucial to discover trans-eQTLs and understand their mechanism to reveal the genetic variants' link to disease phenotypes. It is challenging to identify trans-eQTLs due to small effect sizes, tissue specificity, and a severe multiple-testing burden.

Our goal was to develop a reliable method of identifying trans-eQTLs. We proposed a new model and created open-source software. Applying our method to the eQTL data from the Genotype-Tissue Expression Project (GTEx) proved its performance is significantly better than the state-of-the-art.

Logistic regression is the method of choice to analyze binary outcomes. Multiple logistic regression uses numerous variables in a logistic model. Bayesian multiple logistic regression offers several benefits, including variable selection, prediction, easier interpretation of results, and leveraging prior information. However, Bayesian multiple logistic regression requires costly and technically challenging Markov Chain Monte Carlo (MCMC) sampling or approximations that significantly reduce the logistic model's flexibility.

We proposed a methodology using the point-normal prior for faster and more accurate Bayesian multiple logistic regression, developing open-source software for the project. Applying our method to human genetics data, we proved it outperforms state-of-the-art variable selection and prediction for sparse multiple logistic regression problems of high dimension (n >> p problems.)