Abstract: In the prediction of main components in a macrolide antibiotics-roxithromycin by the near infrared spectrometry,the algorithm of back-ward interval partial least square (biPLS) was used in the pretreatment of selection of wavelength interval from the whole near infrared spectrum,and in obtaining the principal components,a prediction model based on these selected data,was then established by using the BP artificial neural network (BP-ANN).Amounts of the main components of roxithromycin itself and lactose were predicted by the established model.The specialties of this prediction model were its efficient resistance against the instant disturbances due to accidental factors,such as the noise signal and etc.,and its high precision of prediction with effective decrease of wavelength number and less time for calculation simultaneously.It was shown by experimental results that by the use of this model,not only the predicted results were quite satisfactory,but also the presented method was feasible to be used in the rapid on-line analysis of multi-componental pharmaceuticals in the process of manufacturing.