Predictive features of gene expression variation reveal a mechanistic link between expression variation and differential expression

Abstract

For most biological processes, organisms must respond to extrinsic cues, while maintaining essential gene expression programs. Although studied extensively in single cells, it is still unclear how variation is controlled in multicellular organisms. Here, we used a machine-learning approach to identify genomic features that are predictive of genes with high versus low variation in their expression across individuals, using bulk data to remove stochastic cell-to-cell variation. Using embryonic gene expression across 75 Drosophila isogenic lines, we identify features predictive of expression variation, while controlling for expression level. Genes with low variation fall into two classes, indicating they employ different mechanisms to maintain a robust expression. In contrast, genes with high variation seem to lack both types of stabilizing mechanisms. Applying the framework to human tissues from GTEx revealed similar predictive features, indicating that promoter architecture is an ancient mechanism to control expression variation. Remarkably, expression variation features could also predict differential expression upon stress in both Drosophila and human. Differential gene expression signatures may therefore be partially explained by genetically encoded gene-specific features, unrelated to the studied treatment.