Allele-specific expression (ASE) reveals the mechanism of expression variation between two differentially expressed alleles. Expression variation can be classified as trans-acting when the variation is due to diffusible elements such as transcription factors or cis-acting when it is caused by mutations in regulatory sequences such as enhancer elements. Knowing the relative contribution of cis- and trans-acting variation to total expression variation is crucial in understanding the evolution of gene regulation. We predict that poorly replicated experiments and experiments employing an unsuitable model underestimate the variance among independent biological replicates, and ignore the effect of hybrid-specific expression due to novel cis-trans interaction, leading to high rates of false positives for inferring differential ASE. We propose a new model which is statistically well-justified both from first principles as well as from empirical work. This method takes advantage of higher replication and appropriate reference genomes.
We introduce an R package which implements our model based on beta-binomial sampling of expression at two alleles. With this package, we test both our model and a previous binomial model using yeast expression data from recent work permitting testing of 48-fold replication. We show that our model estimates cis/trans expression differences with a type I error rate of approximately 5% when the significance level is set to 0.05. The binomial model underestimates the variance in the data and exhibits a much larger type I error rate of 0.15-0.45. This model will significantly revise conclusions from previous allele-specific expression studies; a large proportion of previous differential expression results are likely false positives, and we probably still do not have a very accurate picture of the nature of allele specific expression even in model systems.