DEEPSIDE: ENHANCING DRUG SIDE EFFECT PREDICTION WITH DEEP LEARNING

Unexpected adverse effects during clinical trials may cause drug failures, endangering the participants' health and causing large financial losses. The creation of novel medications may be guided by algorithms that forecast adverse effects. The LINCS L1000 dataset provides a plethora of information on cell line gene expression that has been impacted by different drugs and lays the groundwork for understanding context-specific traits. The state-of-the-art approach, which aims to employ context-specific information, only uses the high-quality experiments in LINCS L1000 and discards a large portion of the trials. Our goal in this study is to use this data to its fullest potential in order to maximise prediction performance. Our experiments involve five deep learning architectures. We find that among multi-layer perceptron-based designs, a multi-modal architecture provides the best prediction performance when drug chemical structure (CS) and the whole set of drug changed gene expression profiles (GEX) are used as modalities. Overall, we discover that the CS offers more details than the GEX. A convolutional neural network-based model that only uses the SMILES string representation of the drugs produces the greatest results; it beats the state-of-the-art by 13:0% macro-AUC and 3:1% micro-AUC. We also show that the model can predict drug-side effect pairings that are mentioned in the literature but are not present in the ground truth side effect dataset.