Abstract

Gene co-expression networks inferred from microarray and sequencing data offer valuable information on the functional organization of genes. The advancement of single cell RNA-sequencing technology enables researchers to study co-expression networks at the individual cell type level. However, the high noise and heterogeneity of single cell data present great challenges in recovering true expression levels from the observed counts, and lack of attention to these issues in existing single cell network inference methods may lead to biased estimates, inflated type I error and reduced power. In this talk, we describe CS-CORE, a statistical method that is built on an expression-measurement model tailored to single cell data to explicitly model the technical noises. We show that CS-CORE can decouple co-expression from measurement noises for single cell co-expression estimation and testing in both simulations and real data.