The use of long-term animal studies for human and environmental toxicity estimation is
more discouraged than ever before. Alternative models for toxicity prediction, including QSAR
studies, are gaining more ground. A recent approach is to combine in vitro chemical profiling and in
silico chemical descriptors with the knowledge about toxicity pathways to derive a unique signature for toxicity endpoints.
In this study we investigate the ToxCast™ Phase I data regarding their ability to predict long-term animal toxicity. We
investigated thousands of models constructed in an effort to predict 61 toxicity endpoints using multiple descriptor
packages and hundreds of in vitro assays. We investigated the use of in vitro assays and biochemical pathways on model
performance. We identified 10 toxicity endpoints where biologically derived descriptors from in vitro assays or pathway
perturbations improved the model prediction ability. In vivo toxicity endpoints proved generally challenging to model.
Few models were possible to readily model with a balanced accuracy (BA) above 0.7.
We also constructed in silico models to predict the outcome of 144 in vitro assays. This showed better statistical metrics
with 79 out of 144 assays having median balanced accuracy above 0.7. This suggests that the in vitro datasets have a
better modelability than in vivo animal toxicities for the given datasets.
Moreover, we published an online platform (http://iprior.ochem.eu) that automates large-scale model building and
Keywords: Alternative testing, computational toxicology, iPRIOR, QSAR, REACH, ToxCast.
Rights & PermissionsPrintExport