Molecular Substructure-Aware Network With Reinforcement Pooling and Deep Attention Mechanism for Drug-Drug Interaction Prediction

Abstract

Predicting potential drug-drug interactions (DDIs) can effectively mitigate unforeseen interactions throughout the entire drug development process, playing a pivotal role in ensuring drug safety. However, traditional methods are laborious and require specific expert knowledge. This paper proposes RPDAnet, a novel molecular substructure-aware network based on Reinforced Pooling and Deep Attention mechanism, to investigate the interactive relationships between drugs and predict the potential DDIs. Particularly, RPDAnet leverages reinforcement learning to dynamically select informative molecular fragments, thus enhancing its generalization capacity without relying on prior knowledge. Subsequently, RPDAnet develops Communicative Message Massing Neural Network (CMPNN) to enhance the representation of molecular structures by reinforcing message interactions between nodes and edges through a communicative kernel. Finally, RPDAnet aggregates the interactions between substructures of drugs to predict the DDI between a pair of drugs. The experimental results on two real-world datasets demonstrate that our proposed RPDAnet outperforms the state-of-the-art methods with more than 5% performance gains in DDI prediction.

Keywords

• Drug-drug interactions
• reinforcement learning
• deep attention neural networks
• learning latent representations
• graph algorithms