In　this　paper　we　proposed　an　approach　for　tumor　informative　genes　selection　by　analysis　of　gene　sensitivity　based　on　SVM.　We　analyzed　the　gene　expression　profiles　of　colon　and　recursively　eliminated　the　genes　which　have　lower　sensitivity　to　SVM,　then　a　set　of　candidate　nested　feature　subsets　were　generated.　Support　Vector　Machines　were　employed　to　classify　the　samples　using　these　candidate　feature　subsets,　and　the　feature　subset　with　a　minimum　error　was　chosen　as　a　set　of　colon　informative　genes.　The　results　show　that　this　feature　subset　contains　more　tumor　classification　information　than　other　feature　subsets　identified　in　the　literatures.　The　method　proposed　in　this　paper　is　feasible　and　effective.