Single-cell analysis of gene expression in metastatic cells from distinct human breast tumour models shows that early metastatic cells possess basal, stem and mesenchymal cell properties, whereas advanced metastatic cells have more proliferative properties and are more mature, enabling them to be targeted with an anti-proliferative compound. An understanding the dynamics of metastasis is critical for the development of new cancer treatments. In an effort to characterize metastatic cell properties and correlate them with tumour burden, Zena Werb and colleagues used single-cell genomics tools to investigate cellular differentiation in individual human metastatic breast cancer cells from distinct breast tumour models. They find that early metastatic cells possess basal and mesenchymal properties, and carry markers of dormant tumour cells, while expressing pluripotent markers. In contrast, late metastatic cells have more proliferative properties and are less stem-cell like, expressing markers of differentiation. Given these differences, authors were also able to target the late metastatic cells with an anti-proliferative compound to reduce metastatic burden in a mouse model. Despite major advances in understanding the molecular and genetic basis of cancer, metastasis remains the cause of >90% of cancer-related mortality1. Understanding metastasis initiation and progression is critical to developing new therapeutic strategies to treat and prevent metastatic disease. Prevailing theories hypothesize that metastases are seeded by rare tumour cells with unique properties, which may function like stem cells in their ability to initiate and propagate metastatic tumours2,3,4,5. However, the identity of metastasis-initiating cells in human breast cancer remains elusive, and whether metastases are hierarchically organized is unknown2. Here we show at the single-cell level that early stage metastatic cells possess a distinct stem-like gene expression signature. To identify and isolate metastatic cells from patient-derived xenograft models of human breast cancer, we developed a highly sensitive fluorescence-activated cell sorting (FACS)-based assay, which allowed us to enumerate metastatic cells in mouse peripheral tissues. We compared gene signatures in metastatic cells from tissues with low versus high metastatic burden. Metastatic cells from low-burden tissues were distinct owing to their increased expression of stem cell, epithelial-to-mesenchymal transition, pro-survival, and dormancy-associated genes. By contrast, metastatic cells from high-burden tissues were similar to primary tumour cells, which were more heterogeneous and expressed higher levels of luminal differentiation genes. Transplantation of stem-like metastatic cells from low-burden tissues showed that they have considerable tumour-initiating capacity, and can differentiate to produce luminal-like cancer cells. Progression to high metastatic burden was associated with increased proliferation and MYC expression, which could be attenuated by treatment with cyclin-dependent kinase (CDK) inhibitors. These findings support a hierarchical model for metastasis, in which metastases are initiated by stem-like cells that proliferate and differentiate to produce advanced metastatic disease.