GRAPH CONVOLUTIONAL NEURAL NETWORK FOR IC50 PREDICTION MODEL WITH DRUG SMILES GRAPHS AND GENE EXPRESSIONS OF AMYOTROPHIC LATERAL SCLEROSIS

Abstract

IC50 prediction for neurodegenerative disorders like Amyotrophic Lateral Sclerosis is crucial in biomedical studies. Traditional machine learning models that use molecular descriptors and gene expression for building IC50 prediction models produce less accuracy and also most of the descriptors created by different tools are irrelevant and undefined. In this paper, a Graph Convolutional Neural Network, a deep learning algorithm, is employed for constructing a more precise IC50 prediction model. The model leverages the structural properties of drug molecules represented in graph format, and incorporates gene expression data as global features. So, the model is able to learn drug-gene interactions better. The drug-gene interactivity is learned by the model without drug-induced gene expressions as it is not found for most of the diseases. The work is implemented with well-known and most relevant 80 drugs related to ALS based on the pIC50 values of 32 protein targets of ALS disorder. The Canonical Smiles graph and their corresponding IC50 values of 80 drugs have been derived from the ChEMBL databases. Based on information from the Repurposing Hub in the Depmap database gene expression data for drug-related genes connected with ALS-related conditions is collected. The predictive results show that the proposed GCNN model with fine-tuned hyperparameters achieves MAE of 0.18, RMSE of 0.16 and R2 Score of 0.90.