In silico prediction of adverse drug reactions of antidepressant drugs

Adverse drug reactions (ADRs) are one of the leading health risks throughout the world, and a major handicap in drug advancement. Therefore, early and accurate prediction of ADRs by computer-aided methods is extremely important for decreasing health threats, along with preventing product withdrawal after clinical trials which need long time and high cost. In this study, we built a model to predict 329 known ADRs of 27 approved antidepressant drugs, and we examined three machine learning algorithms in order to find which one is better for this duty. We integrated the known ADRs of a drug with the drug's chemical and biological properties, including drug targets, enzymes and transporters for prediction of ADR in our predictive model. A total of 338 chemical properties such as atom counts, formal charge, zagreb index and 144 biological properties such as cytochrome P450 3A4, sodium-dependent serotonin transporter and p-glycoprotein 1 were used. In addition, 329 known ADRs were grouped according to their system organ classes (SOCs) and compared the performances of ADR predictions. This assessment, based on a ten-fold cross-validation, showed that the multilayer perceptron algorithm surpassed the others. ADRs which have F-measure value zero and below 50% accuracy were considered as unsuccessful. Depending on these evaluation criterias, 248 out of 329 ADRs were successfully predicted. The best performing attribute group was the chemical plus biological, followed by the chemical and biological. The mean AUC values are 0.623, 0.676, and 0.695 for chemical, biological, and all attributes, respectively. Since imbalanced dataset is present, the F-measure is adopted as the primary metric to evaluate the prediction models. Among ADRs of the 21 SOCs, only ADRs belonging to 5 SOCs were predicted successfully by chemical, biological or all attributes. Also, the ADRs related to withdrawal of indalpine, pheniprazine, medifoxamine, zimelidine and amineptine were successfully predicted by our model. The results showed that the combination of attributes can significantly improve the ADR prediction performance. Furthermore, for validation, we predicted ADRs related to withdrawn antidepressants. The results showed the validity of our approach, predicting satisfactory amount of previously known ADRs from the literature. In conclusion, the proposed approach is an efficient and promising tool for predicting ADRs of antidepressants.