Cancer stem cells have been hypothesized to drive the growth and metastasis of tumors. Because they need to be targeted for cancer treatment, they have been isolated from many solid cancers. However, cancer stem cells from primary human gastric cancer tissues have not been isolated as yet. For the isolation, we used two cell surface markers: the epithelial cell adhesion molecule (EpCAM) and CD44. When analyzed by flow cytometry, the EpCAM(+)/CD44(+) population accounts for 4.5% of tumor cells. EpCAM(+)/CD44(+) gastric cancer cells formed tumors in immunocompromised mice; however, EpCAM(-)/CD44(-), EpCAM(+)/CD44(-) and EpCAM(-)/CD44(+) cells failed to do so. Xenografts of EpCAM(+)/CD44(+) gastric cancer cells maintained a differentiated phenotype and reproduced the morphological and phenotypical heterogeneity of the original gastric tumor tissues. The tumorigenic subpopulation was serially passaged for several generations without significant phenotypic alterations. Moreover, EpCAM(+)/CD44(+), but not EpCAM(-)/CD44(-), EpCAM(+)/CD44(-) or EpCAM(-)/CD44(+) cells grew exponentially in vitro as cancer spheres in serum-free medium, maintaining the tumorigenicity. Interestingly, a single cancer stem cell generated a cancer sphere that contained various differentiated cells, supporting multi-potency and self-renewal of a cancer stem cell. EpCAM(+)/CD44(+) cells had greater resistance to anti-cancer drugs than other subpopulation cells. The above in vivo and in vitro results suggest that cancer stem cells, which are enriched in the EpCAM(+)/CD44(+) subpopulation of gastric cancer cells, provide an ideal model system for cancer stem cell research.