BP neural network's excellence in non-linear modeling is shown with two examples. The first is the non-linear calibration for the relationship between the infrared reflectance rates and contents of protein. The second is QSAR, predicting the physical property of some compounds by using the structural parameters. By employing a new mode of BP neural network, with two hidden layers and direct connections from input to output layer, better results than those of multivariate linear regression and normal BP neural network are achieved.

Key words: QUANTITATIVE STRUCTURE ACTIVITY RELATIONSHIP