ABSTRACT
Avoidance of structural alerts (SAs) might reduce the risk of failure in drug discovery. However, there are still some marketed drugs containing SA, which indicates that SA should be analyzed carefully to avoid their excessive uses. Several detection systems, including automatic mining methods and expert systems, have been developed to identify SA. These methods only focus on toxic compounds that support the SA without consideration of nontoxic ones. Here, we proposed a frequency-based substructure detection protocol that learns from the nontoxic compounds containing SA to get nontoxic substructures (NTSs), whose appearance will reduce the probability of a compound becoming toxic. Kazius and Hansen's Ames mutagenicity dataset was used as an example to demonstrate the protocol. SARpy and ToxAlerts were first employed to obtain the potential SA. Then 2 kinds of NTS were exploited: reverse effect substructures (RESs) and conjugate effect substructures. Contribution and prediction performance of the substructures were evaluated via neural network and rule-based methods. We also compared substructure-based methods with the conventional machine learning-based methods. The results demonstrated that most substructures contributed as supposed and substructure-based methods performed better in the resistance of overfitting. This work indicated that the protocol could effectively reduce the false positive rate in prediction of chemical mutagenicity, and possibly extend to other endpoints.