Constructing and comparing gene regulatory networks (GRNs) from single-cell RNA sequencing (scRNAseq) data present formidable computational challenges. Many existing solutions lack effectiveness or efficiency, due to technical and analytical issues of scRNAseq such as random dropout, uncharacterized cell states, and heterogeneous samples. Here, we present a robust, unsupervised machine learning workflow, called scTenifoldNet, to improve on existing solutions. The scTenifoldNet workflow combines principal component regression, low-rank tensor approximation, and manifold alignment. It constructs and compares transcriptome-wide single-cell GRNs (scGRNs) from different samples to identify gene expression signatures shifting with cellular activity changes such as pathophysiological processes and responses to environmental perturbations. We used simulated data to benchmark the performance of scTenifoldNet. Application of scTenifoldNet on three real data sets shows it to be a powerful tool to reveal biological insights, which are otherwise difficult to obtain. In particular, scTenifoldNet identified highly specific shifts in transcriptional regulation associated with aging and acute morphine responses in mouse cortical neurons, as well as those associated with doublestranded RNA-induced immune responses in human dermal fibroblasts.