We’ve built and validated QSAR models of seven key toxicity endpoints, based on data made available by the US Environmental Protection Agency as part of its Toxicity Evaluation Software Tool (T.E.S.T.) package. The models were built with StarDrop’s Auto-Modeller module and are available to all StarDrop users free-of-charge.

The chemical and biological mechanisms leading to toxicities are complex and varied, making prediction of toxicities from chemical structure a major challenge. Even the best models of toxicity have a high degree of uncertainty in the predictions they make. Therefore, when considering the results from toxicity predictions, it is essential to give them an appropriate degree of weight in the decision-making process; rejecting an area of chemistry purely on the basis of a predicted toxicity would run the risk of missing valuable opportunities, unless viable alternatives are available that are more likely to be benign.

StarDrop provides an ideal environment for application of toxicity models to the selection of compounds. StarDrop’s probabilistic scoring algorithm allows predictive models of toxicity to be used, while explicitly taking into consideration the uncertainty in each prediction. This ensures that uncertain predictions are not given undue weight, relative to other data, when prioritising compounds that are more likely to have an appropriate balance of properties.

The white paper “Toxicity Prediction in StarDrop” describes the data and methods used to build the models and their validation.

Please note that due to the fact that the EPA data sets are comprised of mainly small compounds, many ‘drug-like’ compounds fall outside of the ‘chemical spaces’ covered by the models. This can be identified by high uncertainty in the predicted values in StarDrop, e.g. ‘inf’ for the standard deviation or a class probability of 0.5. In these circumstances, the results should be interpreted with additional caution.

Installation

Download and save the zipped archive (StarDrop Tox Models.zip)
Right click on the file and select “Extract All” and choose a convenient place to save the models
The ReadMe.pdf file in the resulting directory provides detailed instructions on loading the models into StarDrop.

We’d welcome your comments and feedback. In silico prediction of toxicity is an important field and there is much more work to be done.