ABSTRACTGene expression data has been widely used to infer gene regulatory networks (GRNs). Recent single-cell RNA sequencing (scRNAseq) data, containing the expression information of the individual cells (or status), are highly useful in blindly reconstructing regulatory mechanisms. However, it is still not easy to understand transcriptional cascade from large amount of expression data. Besides, the reconstructed networks may not capture the major regulatory rules.Here, we propose a novel approach called TENET to reconstruct the GRNs from scRNAseq data by calculating causal relationships between genes using transfer entropy (TE). We show that known target genes have significantly higher TE values. Genes with higher TE values were more affected by various perturbations. Comprehensive benchmarking showed that TENET outperformed other GRN prediction algorithms. More importantly, TENET is uniquely capable of identifying key regulators. Applying TENET to scRNAseq during embryonic stem cell differentiation to neural cells, we show that Nme2 is a critical factor for 2i condition specific stem cell self-renewal.