DEEP SIDE- A Novel Deep Learning Approach to Anticipating Drug Reactions

Both the participants' health and the sponsors' bottom lines are jeopardised when drugs fail clinical trials owing to unanticipated side effects. Algorithms that can forecast side effects might help direct medication development. In order to build a knowledge base for context specific aspects, the LINCS L1000 dataset compiles extensive data on cell line gene expression as it is affected by various medications. The current gold standard for context-specific data uses just the high-quality LINCS L1000 studies and ignores the rest. We want to maximise the use of this data in order to improve the prediction performance in this research. We look at five different deep learning architectures. When using drug chemical structure (CS) and the whole collection of drug altered gene expression profiles (GEX) as modalities, we demonstrate that a multi-modal design offers the greatest prediction performance among multi-layer perceptron-based architectures. We find the CS to be more illuminating than the GEX on the whole. With gains of 13% in macro-AUC and 3% in micro-AUC compared to the state-of-the-art, the greatest results were achieved by a model based on convolutional neural networks that only used SMILES string representations of the medicines. Also, we demonstrate that the model can anticipate medication-side effect couples that have been described in the literature but were absent from the ground truth side effect dataset.