The mechanisms whereby some tumour cells detach from the primary lesion to colonize distant sites are still largely unknown. Pro-metastatic events common to the majority of solid tumours might include the reversible transition of tumour cells from an epithelial to a mesenchymal state as well as their interactions with stromal components or tumour-activated stromal cells1–17. Some tumours also secrete metastasis-promoting exosomes that contain proteins, mRNAs and microRNAs to establish a distant pro-metastatic niche9,13,18,19. However, whether a subpopulation of metastasis-initiating cells exists among primary tumour-initiating cells is not clear. LRCs express lipid metabolism genes When cell lines and patient-derived cells (PDCs) arising from human oral carcinomas (Methods) were pulsed with a lipophilic fluorescent dye (DiD) that non-specifically binds to membranes and is diluted upon cell division20, and were orthotopically injected into the oral cavity of NOD.Cg-PrkdcscidIl2rgtm1Wjl/SzJ (NSG) mice, we observed a small percentage of slow-cycling CD44bright/dye+ long-term labelretaining cells (LRCs) within oral lesions (Fig. 1a, b and Extended Data Fig. 1a–k). Thus, the CD44bright population, which have been shown to have the highest tumour-initiating potential in oral squamous cell carcinomas (OSCCs), displayed cell cycle heterogeneity in vivo21–23. Although the transcriptomes of LRCs (CD44bright dye+) and nonLRCs (dye−) sorted by fluorescence-activated cell sorting (FACS) from orthotopic tongue tumours derived from the OSCC cell line SCC-25 were more similar to each other than to the differentiated CD44dim population, they still displayed a number of differentially expressed genes (Extended Data Fig. 2a and Supplementary Table 1a). Gene ontology analysis indicated that the CD44bright dye− signature was associated with chromosomal instability, cell transformation and neoplasm and genes involved in the cell cycle, as expected from a proliferative tumour population (Extended Data Fig. 2b, c and Supplementary Table 1b). On the other hand, the signature of LRCs included an over-representation of genes associated with lymphatic metastasis, neoplasm metastasis, response to lipids and lipid metabolic process (Fig. 1c, d and Extended Data Fig. 2c, d); this was confirmed in LRCs sorted from a second OSCC tumour line (Detroit-562; Extended Data Fig. 2e and Supplementary Table 1c, d). We validated the differential expression of several of these genes by quantitative PCR with reverse transcription (RT–qPCR) in five biological replicates (Extended Data Fig. 2f). Notably, 69 genes were found in the dye+ signatures from both the SCC-25 and the Detroit-562 cell lines; the main functions of these genes were related to neoplasm metastasis, lymphatic metastasis, response to stimulus, lipid distribution and translocation and response to lipid, underscoring their relevance in defining the LRC population (Extended Data Fig. 2g and Supplementary Table 1e). LRCs expressed genes involved in different aspects of fatty acid metabolism, including lipid uptake and transport, fatty acid β- and α-oxidation, lipid biosynthesis and intracellular lipid storage (Extended Data Fig. 2d). Among the products of these genes, the receptor CD36 is at the top of the signalling cascade that takes lipids up from the extracellular environment, allowing cells to obtain ATP energy through lipid β-oxidation24–26. As CD36 is a cell surface receptor, we used it as a surrogate marker to detect and isolate LRCs from tumours, circumventing fluorescent dyes. Strikingly, LRCs corresponded to the CD36bright CD44bright cells within primary oral lesions derived from all cell lines and PDCs tested (Extended Data Figs 2h, 3a).