Strength of stabilizing selection on the amino-acid sequence is associated with the amount of non-additive variance in gene expression.

Contemporary populations are unlikely to respond to natural selection if much of their genetic variance is non-additive. Understanding the evolutionary and genomic factors that drive amounts of non-additive variance in natural populations is therefore of paramount importance. Here, we use a quantitative genetic breeding design to separate the additive from the non-additive components of expression variance in 17,657 gene transcripts of the outcrossing plant Arabidopsis lyrata. We partition the expressed genes according to their predominant variance components in a set of half- and full-sib families obtained by crossing individuals from different populations. As expected, a population-genetic simulation model shows that when divergent alleles segregate in the population, our ability to detect non-additive genetic variance is enhanced. Variation in its relative contribution can thus be analyzed and compared across transcribed genes. We find that most of the genetic variance in gene expression represents non-additive variance, especially among long genes or genes involved in epigenetic gene regulation. Genes with the most non-additive variance in our design not only display markedly lower rates of synonymous variation, they have also been exposed to stronger purifying selection compared to genes with high additive variance. Our study demonstrates that both the genomic architecture and the past history of purifying selection impacts the composition of genetic variance in gene expression.