A hierarchical Bayesian mixture model for distinguishing active gene expression from transcriptional noise

Abstract

Transcriptomes are key to understanding the relationship between genotype and phenotype. The ability to infer the expression state (active or inactive) of genes in the transcriptome offers unique benefits for addressing this issue. For example, qualitative changes in gene expression may underly the origin of novel phenotypes, and expression states are readily comparable between tissues and species. However, inferring the expression state of genes is a surprisingly difficult problem, owing to the complex biological and technical processes that give rise to observed transcriptomic datasets. Here, we develop a hierarchical Bayesian mixture model that describes this complex process, and allows us to infer expression state of genes from replicate transcriptomic libraries. We explore the statistical behavior of this method with analyses of simulated datasets--where we demonstrate its ability to correctly infer true (known) expression states--and empirical-benchmark datasets, where we demonstrate that the expression states inferred from RNA-seq datasets using our method are consistent with those based on independent evidence. The power of our method to correctly infer expression states is generally high and, remarkably, approaches the maximum possible power for this inference problem. We present an empirical analysis of primate-brain transcriptomes, which identifies genes that have a unique expression state in humans. Our method is implemented in the freely-available R package zigzag.\n\nSignificance StatementHow do the cells of an organism--each with an identical genome--give rise to tissues of incredible phenotypic diversity? Key to answering this question is the transcriptome: the set of genes expressed in a given tissue. We would clearly benefit from the ability to identify qualitative differences in expression (whether a gene is active or inactive in a given tissue/species). Inferring the expression state of genes is surprisingly difficult, owing to the complex biological processes that give rise to transcriptomes, and to the vagaries of techniques used to generate transcriptomic datasets. We develop a hierarchical Bayesian mixture model that--by describing those biological and technical processes--allows us to infer the expression state of genes from replicate transcriptomic datasets.