Most of the human genetic variations are single nucleotide polymorphisms (SNPs), and among them, non-synonymous SNPs, also known as SAPs, attract extensive interest. SAPs can be neural or disease associated. Many studies have been done to distinguish deleterious SAPs from neutral ones. Since many previous studies were based on both structural and sequence features of the SAP, these methods are not applicable when protein structures are not available. In the current paper, we developed a method based on UMDA and SVM using protein sequence information to predict SAP’s disease association. We extracted a set of features that are independent of protein structure for each SAP. Then a SVM-based machine-learning classifier that used grid search to tune parameters was applied to predict the possible disease associa-tion of SAPs. The SVM method reaches good prediction accuracy. Since the input data of SVM contain irrelevant and noisy features and parameters of SVM also affect the prediction performance, we introduced UMDA-based wrapper approach to search for the ‘best’ solution. The UMDA-based method greatly improved prediction performance. Com-pared with current method, our method achieved better performance.