Rutgers team predicts toxicity by mining PubChem data

A US team has created an algorithm to predict the toxicity of unknown chemicals by mining bioassay data held in a US National Institutes of Health database.Many studies use computers to compare untested chemicals with structurally similar compounds whose toxicity is already known. But the results can be confounded by the fact that structurally similar chemicals may have very different levels of toxicity.The new method uses fragments of chemical structures and biological information from in vitro tests. The approach not only predicts acute oral toxicity classification, but also hints at biological mechanisms, the researchers suggest.A team led by Hao Zhu from Rutgers University, and including Thomas Hartung from Johns Hopkins Bloomberg School of Public Health, first built a training database of more than 7,000 compounds with associated rat acute oral toxicity data. They then created "bioprofiles" for the chemicals based on in vitro data in PubChem, a public database that is updated daily.