A　leukemia　molecular　prediction　model　is　constructed　by　using　bioinformatics　and　machine　learning　methods　with　gene　expression　profile.　Firstly,　three　methods　including　relief,　classification　information　index　and　information　gain　index　are　used　to　select　candidate　feature　gene　set　from　the　leukemia　gene　expression　profile.　Secondly,　intersection　of　three　candidate　feature　gene　sets　is　generated,　and　then　the　best　classification　performance　of　intersection　genes　which　is　tested　by　SVM　is　selected　as　feature　genes.　Thirdly,　the　classification　rule　sets　are　extracted　from　these　feature　genes　by　using　decision　tree　method.　Finally,　the　leukemia　molecular　prediction　model　is　constructed　with　these　classification　rules.　The　results　show　that　the　model　is　helpful　to　cancer　clinical　diagnosis　and　cancer　gene　biological　experiments.　Also,　the　two　key　genes　(CD33,　MPO)　are　biomarkers　of　leukemia　clinically.