SERIOUS SIDE AN INTEGRATED LEARNING FRAMEWORK FOR DRUG SIDE EFFECT PREDICTION

Unexpected negative effects from drugs during clinical trials put participants’ health at danger and result in significant financial losses. Algorithms for side effect prediction may direct the process of developing new drugs. LINCS L1000 dataset generates a knowledge base for context-specific characteristics and offers a large collection of cell line gene expression data impacted by various medications. The most recent method that attempts to use context-specific information only uses the best trials in LINCS L1000 and discards a substantial number of the others. The objective of this work is to improve the prediction performance by making the most of this data. Five deep learning architectures are tested. We show that when drug chemical structure (CS) and the whole collection of drug altered gene expression profiles (GEX) are employed as modalities, a multi-modal architecture yields the greatest prediction performance among multi-layer perceptron-based designs. Generally speaking, we find that the CS is more instructive than the GEX. Best results are obtained by a convolutional neural network-based model that improves 13:0% macro-AUC and 3:1% micro-AUC over the state-of-the-art using just SMILES string representation of the medicines. We also demonstrate that while side effect-drug combinations that are documented in the literature but absent from the ground truth side effect dataset may be predicted by the model.