A central question in biology is whether variability between genetically identical cells exposed to the same culture conditions is largely stochastic or deterministic. Using image-based transcriptomics in millions of single human cells, we find that while variability of cytoplasmic transcript abundance is large, it is for most genes minimally stochastic and can be predicted with multivariate models of the phenotypic state and population context of single cells. Computational multiplexing of these predictive signatures across hundreds of genes revealed a complex regulatory system that controls the observed variability of transcript abundance between individual cells. Mathematical modeling and experimental validation show that nuclear retention and transport of transcripts between the nucleus and the cytoplasm is central to buffering stochastic transcriptional fluctuations in mammalian gene expression. Our work indicates that cellular compartmentalization confines transcriptional noise to the nucleus, thereby preventing it from interfering with the control of single-cell transcript abundance in the cytoplasm.

Fluorescence in situ hybridization (FISH) is widely used to obtain information about transcript copy number and subcellular localization in single cells. However, current approaches do not readily scale to the analysis of whole transcriptomes. Here we show that branched DNA technology combined with automated liquid handling, high-content imaging and quantitative image analysis allows highly reproducible quantification of transcript abundance in thousands of single cells at single-molecule resolution. In addition, it allows extraction of a multivariate feature set quantifying subcellular patterning and spatial properties of transcripts and their cell-to-cell variability. This has multiple implications for the functional interpretation of cell-to-cell variability in gene expression and enables the unbiased identification of functionally relevant in situ signatures of the transcriptome without the need for perturbations. Because this method can be incorporated in a wide variety of high-throughput image-based approaches, we expect it to be broadly applicable.