In silico toxicity prediction is a high priority for industry and government agencies. Yet, most of the available models continue to disappoint in terms of prediction quality. We describe special innovations in Artificial Neural Network modeling methodology that have resulted in improved toxicity predictions. Our logP and solubility models based on these methods have consistently been ranked at the top in published, independent comparisons.Moreover, we created a set of novel descriptors of atomic and molecular reactivities via rapid estimates of quantum mechanical indices, such as sigma and pi partial atomic charges, pi system HOMO/LUMO energies, chemical hardness and electronegativity, and Fukui reactivity indices. Electron densities for a diverse set of almost 950 organic molecules were calculated at the high B3LYP/6-311G** level of ab initio quantum chemistry as the training set for these estimates.We will describe a collection of 35 high-quality, diverse toxicity models based on this new approach.

Toxicity modeling done right with robust artificial neural network ensembles and descriptors of reactivity

243rd National Meeting (2012)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.