Improved half-maximal inhibitory concentration regression model using amyotrophic lateral sclerosis data

Abstract

The current research addresses the critical need for precise half-maximal inhibitory concentration regression in the neurodegenerative condition amyotrophic lateral sclerosis (ALS). Unavailable drug-induced gene expressions and irrelevant molecular descriptors have yielded regression models with less accuracy using traditional machine learning (ML). Drugs can be converted to graph format and integrated with gene expressions to learn drug-gene interactions better thereby producing precise half-maximal inhibitory concentration regression models. To accomplish this, three variants of graph neural networks (GNN) namely graph attention networks (GAT), message passing neural networks, and graph isomorphism networks are utilized in the proposed work. The gene expression profiles of ALS drugrelated genes were retrieved from the DepMap PRISM drug repurposing hub, and the drug graphs with their accompanying half-maximal inhibitory concentration values were obtained from the ChEMBL databases. The graph is constructed for ninety approved drugs connected to 32 key protein targets of ALS and its related conditions. The half-maximal inhibitory concentration regression model trained with optimized hyperparameters in GAT performs well with an R2 score of 0.92, a mean absolute error (MAE) of 0.20, and a root mean square error (RMSE) of 0.17. This model produced better results than other ML and deep learning models.