One　of　the　most　important　link　in　improves　diagnostic　accuracy　and　disease　cure　rate　is　accurate　classification　of　disease.　The　current　gene　chip＇s　development　and　widely　applications　making　the　diagnosis　based　on　tumor　gene　expression　profiling　expected　to　be　on　a　fast　and　effective　clinical　diagnostic　method.　But　the　sample　of　gene　is　small　and　the　expression　data　is　multi-variable.　In　this　article,　we　uses　three　data　sets　on　gene　expression　profiles　of　gastric　cancer　for　the　construction　of　classification　model,　First,　screened　the　gene　which　significantly　changed　in　expression　pattern,　and　use　these　genes　as　a　set　of　the　feature　to　reduce　the　number　of　variables,　and　then　using　genetic　algorithms　and　bayesian　network　model　to　build　the　classifier,　the　build　process　uses　these　three　gene　expression　data　to　learn　classifier.　Classification　accuracy　is　calculated　by　leave-one　cross-validation　(LOOCV)　and　it　reached　99.8%.　Last　we　use　the　GO　and　pathway　to　analysis　the　classifier＇s　network　structure.　©　2012　IEEE.