AUTOMATED DESCRIPTOR SELECTION FOR QUANTITATIVE STRUCTURE-ACTIVITY-RELATIONSHIPS USING GENERALIZED SIMULATED ANNEALING

The central steps in developing QSARs are generation and selection of molecular structure descriptors and development of the model. Recently, computational neural networks have been employed as nonlinear models for QSARs. Neural networks can be trained efficiently with a quasi-Newton method, but the results are dependent on the descriptors used and: the initial parameters of the network. Thus, two potential opportunities for optimization arise. The first optimization problem is the selection of the descriptors for use by the neural network. In this study, generalized simulated annealing (GSA) is employed to select an optimal set of descriptors. The cost function used to evaluate the: effectiveness of the descriptors is based on the performance of the neural network. The second optimization problem is selecting the starting weights and biases for the network. GSA is also used for this optimization. The result is an automated descriptor selection algorithm that is an optimization inside of an optimization. Application of the method to a QSAR problem shows that effective descriptor subsets are found, and they support models that are as good or better than those obtained using traditional linear regression methods.