Somatic single nucleotide variations(SNV)identification is key for studying cancer genome and evolution.However,previous methods had high called number due to sequencing errors.Here,we presented a machine-learning classifier,CASpoint,which learned an optimized decision rule by selecting maximum feature combination from the feature space.We evaluated CASpoint with real data of whole genome/exome sequencing of solid and blood tumors and synthetic dataset in ICGC-TCGA DREAM Somatic Mutation Calling Challenge.Compared with other callers,CASpoint achieved higher F1 score and smaller number of somatic SNV calls.