It　is　very　important　but　difficult　to　identify　which　genes　in　gene　expression　data　can　contribute　most　to　tumor　subtype　classification.　An　approach　to　select　a　small　subset　of　genes　for　leukemia　subtype　classification　from　large　scale　gene　expression　profile　is　proposed　in　this　paper.　Having　removed　the　noisy　genes　with　little　relevance　to　the　classification　task,　the　＇sequential　floating　forward　search＇　method　was　employed　to　generate　candidate　feature　subsets　consisting　of　informative　genes,　and　then,　a　support　vector　machine　was　employed　as　a　classifier　to　select　the　optimal　feature　subset　with　minimum　classification　errors.　The　results　of　our　experiment　showed　that　all　the　samples　can　be　correctly　classified　without　any　error　with　only　five　genes.