Predicting potential side effects of drugs by recommender methods and ensemble learning

Drugs provide help and promise for human health, but they usually come with side effects. Predicting side effects of drugs is a critical issue for the drug discovery. Although several machine-learning methods were proposed to predict the drug side effects, it remains the space for the improvement. To the best of our knowledge, many side effects are not detectable in clinical trials until drugs are approved, thus predicting potential or missing side effects based on the known side effects is important for the postmarketing surveillance. In order to solve this specific problem, we formulate approved drugs, side effect terms and drug-side effect associations as a recommender system, and transform the problem of predicting side effects into a recommender task. Two recommender methods, i.e. the integrated neighborhood-based method and the restricted Boltzmann machine-based method, are designed to make predictions. Further, in order to achieve better performances, we combine proposed methods and existing methods of the same type to develop ensemble models. Compared with benchmark methods, the proposed methods and the ensemble method lead to better performances, and the statistical analysis demonstrates the improvements are significant (p-value<0.05). In conclusion, the integrated neighborhood-based method, the restricted Boltzmann machine-based method and the ensemble method are promising tools for the side effect prediction. The source codes and datasets are provided as the supplementary.