مشخصات پژوهش

Drug design is a complex and resource-intensive process, partly due to the challenge of adverse drug reactions (ADRs)[(Yang and Kar 2023)], which impacts drug safety and only becomes evident after clinical trials on a drug has already began. In this study, we developed machine learning (ML) methodologies aimed at predicting ADRs by leveraging data from SIDER database [(Kuhn, Letunic et al. 2016)], which contain ADR information for approved drugs. ADR data was collected from 1,430 approved drugs. Molecular descriptors, such as polar surface area and molecular weight, were extracted from drug SMILES strings which were obtained from Chembl [(Nowotka, Gaulton et al. 2017)], and RdKit [(Bento, Hersey et al. 2020)] was used for molecular fingerprinting. We employed several machine learning algorithms, including Support Vector Machine (SVM), Random Forest (RF), and Gradient Boosted Trees (GBT), for ADR classification tasks. To ensure robust evaluation and optimization of these ML methods, we utilized metrics such as accuracy, precision, recall, and F1-score, after addressing class imbalance using synthetic minority over-sampling technique-nominal continuous (SMOTE-NC)[(Gök and Olgun 2021)]. Our results demonstrated that no single algorithm outperformed others in all cases; for example, the best balance between precision and recall for predicting common ADRs might be different from those algorithms for rare ADRs, or some algorithms perform better for some tissues and worse for the others. We suggest the use of ensemble learning to combine the strengths of different algorithms for improved ADR prediction in drug discovery. Future work should focus on optimizing ensemble models and extending the approach to other drug classes.