Evolutionary optimization, backpropagation, and data preparation issues in QSAR modeling of HIV inhibition by HEPT derivatives

Artificial neural networks (ANNs) can be utilized to generate predictive models of quantitative structure–activity relationships between a set of molecular descriptors and activity. Evolutionary computation provides a means to appropriately search for the set of weights and bias terms associated with artificial neural networks that minimize selected functions of the error between the actual and desired outputs. This method is demonstrated by evolutionary training of artificial neural networks capable of predicting anti-HIV activity for a set of 1-[(2-hydroxyethoxy)methyl]-6-(phenylthio)thymine (HEPT) derivatives. The results of this work further confirm the growing indication that evolutionary computation can outperform backpropagation as a method of artificial neural network training. The results also indicate the degree to which bias in the initial training and testing data can affect performance and the importance of bootstrapping.