Abstract
Gene expression is controlled by networks of transcription factors and regulators, but the structure of these networks is as yet poorly understood and is thus inferred from data. Recent work has shown the efficacy of information theoretical approaches for network reconstruction from single cell transcriptomic data. Such methods use information to estimate dependence between every pair of genes in the dataset, then edges are inferred between top-scoring pairs. Dependence, however, does not indicate significance, and the definition of “top-scoring” is often arbitrary and a priori related to expected network size. This makes comparing networks across datasets difficult, because networks of a similar size are not necessarily similarly accurate. We present a method for performing formal hypothesis tests on putative network edges derived from information theory, bringing together empirical Bayes and work on theoretical null distributions for information measures. Thresholding based on empirical Bayes allows us to control network accuracy according to how we intend to use the network. Using single cell data from mouse pluripotent stem cells, we recover known interactions and suggest several new interactions for experimental validation (using a stringent threshold) and discover high-level interactions between sub-networks (using a more relaxed threshold). Furthermore, our method allows for the inclusion of prior information. We use in-silico data to show that even relatively poor quality prior information can increase the accuracy of a network, and demonstrate that the accuracy of networks inferred from single cell data can sometimes be improved by priors from population-level ChIP-Seq and qPCR data.
