Graphical Models for Gene Regulation

Seminar Series

Wednesday, August 7, 2019 - 11:00

Abstract:   Probabilistic graphical models provide a flexible means of integrating multiple sources of information while accounting for known dependencies and other prior knowledge.  Implementing such models is now considerably easier than before, due to the availability of frameworks such as STAN that automate the inference task for a given model and data set.  In this talk I will describe recent work in developing such a model to identify genetic variants implicated in gene expression differences in high-throughput reporter experiments.  By decomposing the model and testing individual components, I show that the explicit modeling of experimental replicates and incorporation of a prior on effect size both contribute to improved predictive accuracy, allowing rare regulatory variants to be more reliably detected than with existing methods.  I will also describe current efforts to adapt this approach to the problem of quantifying allele-specific expression of endogenous genes with multiple heterozygous sites in the presence of phasing uncertainty.  In the second part of my talk I will describe work on predicting the effects of genetic variants that influence mRNA splicing, by incorporating splicing regulatory signals into an undirected graphical model of gene structure.  I will show that existing gene-structure prediction methods are poorly suited to this task, due to their reliance on protein-coding signals and their assumption that reading frames are maintained across exons.  By utilizing exon definition signals instead, gene structures can be predicted in the context of genetic variants that violate the translation reading frame.  I will describe a number of promising avenues for further enhancing the biological realism of the model.

Speaker:  William H. Majoros, Ph.D.    ** Faculty Candidate Seminar
Senior Bioinformatician
Center for Genomic and Computational Biology
Duke University