Direct Reconstruction of Gene Regulatory Networks underlying Cellular state Transitions without Pseudo-time Inference

Abstract

Nowadays the advanced technology for single-cell transcriptional profiling enables people to routinely generate thousands of single-cell expression data, in which data from different cell states or time points are derived from different samples. Without transferring such time-stamped cross-sectional data into pseudo-time series, we propose COSLIR (COvariance restricted Sparse LInear Regression) for directly reconstructing the gene regulatory networks (GRN) that drives the cell-state transition. The differential gene expression between adjacent cell states is modeled as a linear combination of gene expressions in the previous cell state, and the GRN is reconstructed through solving an optimization problem only based on the first and second moments of the sample distributions. We apply the bootstrap strategy as well as the clip threshold method to increase the precision and stability of the estimation. Simulations indicate the perfect accuracy of COSLIR in the oracle case as well as its good performance and stability in the sample case. We apply COSLIR separately to two cell lineages in a published single-cell qPCR dataset during mouse early embryo development. Nearly half of the inferred gene-gene interactions have already been experimentally reported and some of them were even discovered during the past decade after the dataset was published, indicating the power of COSLIR. Furthermore, COSLIR is also evaluated on several single-cell RNA-seq datasets, and the performance is comparable with other methods relying on the pseudo-time reconstruction.