Big data mining predicts toxicity better than animal tests

Algorithms derived from large databases of chemical structure and toxicity relationships may be better at predicting the toxicity of chemicals than individual animal tests, according to a study by scientists at Johns Hopkins Bloomberg School of Public Health (Toxicol. Sci. 2018, DOI: 10.1093/toxsci/kfy152). The study relies on the world’s largest machine-readable chemical database, created by the team two years ago. The researchers used machine-learning algorithms to read across structure and toxicity information for about 10,000 chemicals from 800,000 toxicology tests. They then created software to predict whether any given chemical is likely to cause skin irritation, DNA damage, or other toxic effects. The software was on average about 87% accurate in predicting consensus results for nine common toxicity tests that use animals. The actual animal tests averaged only about 81% accuracy, the authors report. This is “big news for toxicology,” principal investigator Thomas Hartung, a professor in the Department of Environmental Health & Engineering at the Bloomberg School, says in a press release. “These results are a real eye-opener—they suggest that we can replace many animal tests with computer-based prediction and get more reliable results.”