Gated Recurrent Unit with SMILES2Vec-based Descriptor for Predicting Drug Side Effects: Case Study of Hepatobiliary Disorders

The unique chemical properties of drugs enable them to interact with specific proteins, effectively treating or preventing diseases by disrupting various molecular processes. However, the accurate prediction of drug side effects is of paramount importance due to the associated risks involved. In this regard, machine learning algorithms offer promising avenues for early-stage identification of potential side effects, with in-silico approaches proving valuable for drug prediction. A notable in-silico method is SMILES2Vec, which utilizes word embedding techniques to encode a compound’s molecular structure and generate vector representations of molecules. To evaluate its performance, this study employed a dataset from the SIDER (Side Effect Resource), specifically focusing on Hepatobiliary disorder side effects. The investigation involved assessing different layer combinations and optimizers within the SMILES2Vec model. In particular, the study explored the impact of introducing a convolution layer before the GRU layer and a dense layer after it, aiming to enhance the model’s performance. The results revealed that a model consisting of l-layer convolution, 2-layer GRU, and 1-layer dense achieved the highest accuracy and F1-Score, reaching 64.39% and 67.39%, respectively. Furthermore, the Adagrad optimizer played a significant role by dynamically adjusting the learning rate for each parameter. This optimization technique not only improved the F1-Score to 68.62%, but also effectively prevented overfitting during the model training process. The evaluation of SMILES2Vec model’s layer combinations and optimizers provided valuable insights for predicting drug side effects. Selecting appropriate architectural components and optimization techniques is crucial to enhance accuracy, benefiting patient care and well-being.