T-cell acute lymphoblastic leukemia (T-ALL) is associated with a significant risk of disease relapse, but the biological basis for relapse is poorly understood. Here, we identify leukemiainitiating cells (L-ICs) on the basis of functional assays and prospective isolation and report a role for L-ICs in T-ALL disease and relapse. Long-term proliferation in response to NOTCH1 activating signals in OP9-DL1 coculture system or capacity to initiate leukemia in xenografts by the CD7(+)CD1a(-) subset of primary T-ALL samples was superior to other subsets, refining the identity of T-ALL L-ICs. T-ALL engraftment was improved in nonobese diabetic/severe combined immunodeficiency (NOD/scid)IL2Rγ(null) (NSG) mice compared with NOD/scid with anti-CD122 treatment (NS122), but both showed changes in leukemia immunophenotype. Clonal analysis of xenografts using the TCRG locus revealed the presence of subclones of T-ALL L-ICs, some of which possess a selective growth advantage and correlated with the capacity of CD7(+)CD1a(+) xenograft cells to engraft secondary NSG mice. Treatment of high-risk T-ALL xenografts eliminated CD1a(+) T-ALL cells, but CD1a(-) cells were resistant and their number was increased. Our results establish that primary CD1a(-) T-ALL cells are functionally distinct from CD1a(+) cells and that the CD7(+)CD1a(-) subset is enriched for L-IC activity that may be involved in mediating disease relapse after therapy.