Identifying true negative drugs to support research in drug safety

Abstract: The lack of high-quality reference data of drug side effects is a major limitation for drug safety and drug discovery science. Unreliable standards prohibit the use of supervised learning methods and make evaluation of algorithms difficult. While some data is available positive examples (e.g. what side effects drugs are associated) there are no systematic resources of negative examples. To solve this issue, we introduced a computational method that identifies negative drugs from all unreported drugs. We applied the method to predict negative drugs for 890 side effects from SIDER. Our predictions decreased the false negative rate by one-third according to a validation study using AEOLUS data. Three sets of predicted drugs with different levels of precision were provided, and can be accessed at XXX. This new reference standard will be important in chemical biology, drug development, and pharmacoepidemiology.

Download the negative drug data.