In　order　to　discover　informative　gene　of　cancer,　a　view　to　regard　different　tumors　as　a　single　class　was　presented　in　this　paper.　The　purpose　is　to　find　informative　genes　that　can　classify　tumor　tissues　from　normal　tissues,　which　can　be　used　for　gene　expression　research　of　biomedicine　and　biotechnology.　We　used　the　correlation　coefficient　for　each　gene　as　the　criterion　for　classification,　and　remove　the　noise-genes　with　smaller　correlation　coefficient　values.　A　statistical　method　called　＇nearest　shrunken　centroids＇　is　applied　in　order　to　find　informative　gene　with　good　ability　of　classifying　and　clustering　the　samples　corresponding　to　their　tissues　types.　We　correctly　clustered　87.7%　samples　and　classify　the　testing　samples　with　an　accuracy　of　81.1%　using　the　informative　genes.　The　results　show　that　both　the　performance　of　clustering　and　classifying　are　improved　after　the　feature　selection.