In　this　paper,　the　problem　of　informative　gene　selection　from　broad　patterns　of　gene　expression　data　was　studied　using　the　method　of　sensitivity　analysis　based　on　neural　networks　with　single　hidden　layer.　First,　noise　genes　were　removed　with　small　weighted　Bhattacharyya　distance.　Secondly,　the　sensitivity　analysis　with　neural　networks　was　employed　to　generate　candidate　feature　subsets.　Finally,　the　candidate　feature　subsets　were　evaluated　by　cross　validation　and　independent　test　and　the　candidate　feature　subset　with　minimum　errors　was　selected　as　the　informative　gene　set.　The　proposed　method　was　applied　in　classifying　four　subtypes　of　small　round　blue　cell　tumor　gene　expression　profiles.　The　result　demonstrated　the　feasibility　and　effectiveness　of　the　method.