A unified framework for using neural networks to build QSARs.

We propose a new neural network architecture that explicitly separates linear and nonlinear contributions to the biological activity. To facilitate the use of neural networks as a regular tool we demonstrate that (1) a perceptron with linear output units is equivalent to multiple linear regression and (2) one hidden unit at a time can be added to the network so that QSAR data can be modeled by everything from the simplest linear hypersurfaces to complicated ones. The significant improvements accrued by the use of weight decay are demonstrated. We conclude that models built without attempting weight decay may not be reliable either for interpretation or extrapolation. Finally we compare models generated by neural networks, rank regression, and standard regression on non-normally distributed data and conclude that neural networks like rank regression bring out many facets of the data that are inaccessible to multiple linear regression. All the experiments were done on either triazine inhibition of pure DHFR from L1210 leukemia cells and on the inhibition of intact L1210 leukemia cells sensitive and resistant to methotrexate or on steroid binding to progesterone.