A　novel　algorithm　for　gene　selection　is　proposed　based　on　multiple　principal　component　analysis　with　sparsity　algorithm(MSPCA).　Specifically,　we　apply　MSPCA　to　normal　and　disease　samples　respectively　and　set　those　component　loadings　to　zero　if　they　are　smaller　than　a　threshold　for　sparse　solutions.　Next,　we　remove　genes　with　zero　loading　elements　across　all　samples　(normal　and　disease)　and　extract　as　＇feature　genes＇.　The　feature　genes　are　essentially　genes　that　contribute　differentially　to　variations　in　normal　and　disease　samples　and　thus　can　be　used　as　features　for　classification.　We　apply　our　method　to　two　commonly　used　microarray　data　to　select　feature　genes,　and　use　the　linear　support　vector　machine　to　evaluate　the　performance　of　our　algorithm.　The　results　show　that　MSPCA　for　gene　selection　has　a　high　classification　accuracy　and　model　stability.