Abstract: A new pattern discovery method based on the best individual feature selection and BP neural network was proposed to select characteristic metabolites in urine which were most correlative with breast cancer.Four nucleosides(orotidine,1-methyladenosine,S-adenosylmethionine,and N²-methylguanosine),which were identified by using HPLC/MS/MS,were selected out and composed a characteristic pattern for diagnosis of breast cancer.Subsequently,BP neural network was investigated as potential tools to diagnose breast cancer by using those four nucleosides as the input features.The results of Leave-One-Out and independent cross validation show that the prediction rate of the model built with BP neural network is higher than 90%.As a consequence,those four selected nucleosides could be considered as a characteristic pattern for the diagnosis of breast cancer.

Key words: Metabonomics/Metabolomics, Feature selection, HPLC/MS/MS, Diagnosis of breast cancer, Nucleoside