Résumé : Haploidentical transplant is now established as a procedure of choice for patients who lack a compatible donor. However, they are still referred too late, heavily pretreated, at very advanced stages. We initiated a three-step phase I study trying improve transplant-related mortality, relapse rate, and immunity: G-CSF + DLI, GM-CSF + DLI, patient- and disease-adapted strategy. Thirty-three consecutive leukemia patients, aged 18-55, were investigated (20 very poor risk, 11 poor risk, and 2 better risk). GvH type NK alloreactivity was chosen when possible (18/33) and balanced across the three groups. In the first nine patients, G-CSF was used and escalated prophylactic DLI started at month 1. Thus, G-CSF and 1-3 DLI (10(4) CD3/kg) is safe. It results in faster CD4 recovery and a low rate of infections. However, it was insufficient to induce a GVL effect. In the next 12 patients, GM-CSF was used plus 1 DLI (10(4) CD3/kg) at day 30 unless aGVHD (3 patients). The comparison between the two first groups can be summarized as follows: G-CSF + DLI: TRM at day 100: 0, RR: 6/9, severe aGVHD: 0. GM-CSF + 1 DLI group: RR: 1/12, TRM at day 100: 3, aGVHD > 1: 9/12, price to pay: GVHD resulting in five deaths in total. Step 3 (13 patients) consists of a patient-adapted strategy: no more aspecific DLI (selected anti-CMV and aspergillus DLI planned in all patients); in myeloid disorders with NK alloreactivity: no GF. In the other cases, GM-CSF (at a reduced total dose of 500 mug) is given the follow-up of these 13 patients, although promising is currently short (median 5 months). Overall, TRM at day 100 is 3/29, reflecting the good tolerance of the conditioning in a heavily pretreated population (median age: 43). NRR mortality (8/26) at 1 year is greater in the GM-CSF + DLI group, reflecting the impact of severe aGVHD. We conclude that the third strategy might improve the outcome without exposing patients to unnecessary severe GVHD.