Quantitative structure-property relationships for drug solubility prediction using evolved neural networks

Preclinical in vivo studies of small molecule compound libraries can be enhanced using a model of specific quantitative structure-property relationships. This may include toxicological or solubility measures such as prediction of drug solubility in mixtures of polyethylene glycol and/or water. Here we examine the utility of both multiple linear regressions and evolved neural networks for the prediction of drug solubility in aqueous solution. Initial results suggest that modeling requires compound libraries with high similarity. Clustering approaches can be used to group compounds by similarity with models built for each cluster. Linear and nonlinear models can be used for modeling, however evolved neural networks can be used to simultaneously reduce the feature space as well as optimize models for solubility prediction. With these approaches it is also possible to identify ldquohuman interpretablerdquo features from the best models that can be used by chemists during preclinical drug development.