Abstract
Toxicity is a concern with many chemicals currently in commerce, and with new chemicals that are introduced each year. The standard approach to testing chemicals is to run studies in laboratory animals (e.g. rats, mice, dogs), but because of the expense of these studies and concerns for animal welfare, few chemicals besides pharmaceuticals and pesticides are fully tested. Over the last decade there have been significant developments in the field of computational toxicology which combines in vitro tests and computational models. The ultimate goal of this field is to test all chemicals in a rapid, cost effective manner with minimal use of animals. One of the simplest measures of toxicity is provided by high-throughput in vitro cytotoxicity assays, which measure the concentration of a chemical that kills particular types of cells. Chemicals that are cytotoxic at low concentrations tend to be more toxic to animals than chemicals that are less cytotoxic. We employed molecular characteristics derived from density functional theory (DFT) and predicted values of log(octanol–water partition coefficient) (log P) to construct a design variable space, and built a predictive model for cytotoxicity based on U.S. EPA Toxicity ForeCaster (ToxCast) data tested up to 100 μM using a Näive Bayesian algorithm. External evaluation showed that the area under the curve (AUC) for the receiver operating characteristic (ROC) of the model to be 0.81. Using this model, we provide probabilistic design rules to help synthetic chemists minimize the chance that a newly synthesized chemical will be cytotoxic.
- This article is part of the themed collection: Molecular Design for Reduced Toxicity