Despite advances in single-cell technologies and temporal sampling, the seamless connection of cell states over time and the inference of gene-gene relationships driving cancer cell state remains a formidable challenge. We present Cflows, a framework that combines neural ordinary differential equation network designed to unravel continuous cellular dynamics from timelapsed scRNAseq data with causality analysis to learn gene regulatory networks underlying cell state changes. When applied to a novel scRNAseq data from breast cancer in vitro models, we trace cell states back to their origins, allowing us to refine a marker profile for cancer stem cells (CSCs), and compute rich and complex gene-gene networks that drive pathogenic trajectories forward. Our comprehensive temporal regulatory networks reveal the dynamic transitions along the epithelial-to-mesenchymal (EMT) and the mesenchymal-to-epithelial (MET) trajectories. Newly, we identify estrogen-related receptor alpha (ESRRA) as a critical mediator of CSC plasticity and MET cell fate decisions. We extend Cflows to an in vivo xenograft model, demonstrating its potential for elucidating trajectories governing primary tumor metastasis to the lung. In summary, Cflows is an innovative algorithm that uncovers temporal molecular programs within dynamic cell systems from static single-cell data and has helped identify key drivers and transcriptional networks underlying cancer plasticity.