Studying the Explanatory Capacity of Artificial Neural Networks for Understanding Environmental Chemical Quantitative Structure−Activity Relationship Models

Although artificial neural networks (ANNs) have been shown to exhibit superior predictive power in the study of quantitative structure-activity relationships (QSARs), they have also been labeled a "black box" because they provide little explanatory insight into the relative influence of the independent variables in the predictive process so that little information on how and why compounds work can be obtained. Here, we have turned our interests to their explanatory capacities; therefore, a method was proposed for assessing the relative importance of variables indicating molecular structure, on the basis of axon connection weights and partial derivatives of the ANN output with respect to its input, which can identify variables that significantly contribute to network predictions, and providing a variable selection method for ANNs. We show that, by extending this approach to ANNs, the "black box" mechanics of ANNs can be greatly illuminated, thereby making it very useful in understanding environmental chemical QSAR models.