DeepDrug

DeepDrug predicts drug-drug interactions (DDIs) and drug-target interactions (DTIs) using residual graph convolutional networks (RGCNs) and convolutional neural networks (CNNs) to learn structural and sequential representations of chemicals and proteins for interaction prediction.

Key Features:

• Residual Graph Convolutional Networks (RGCNs): Learns comprehensive structural representations of chemical structures and enables visualization of learned structural features that reveal patterns in chemical properties and drug categories.
• Convolutional Neural Networks (CNNs): Learns sequential representations of proteins for integration with chemical structure representations.
• Representation Learning: Jointly learns structural and sequential representations of drugs and proteins to improve prediction accuracy across tasks.
• Supported Prediction Tasks: Performs binary-class DDIs, multi-class/multi-label DDIs, binary-class DTI classification, and DTI regression.
• Benchmarking and Performance: Systematically evaluated on multiple datasets, demonstrating improved accuracy and robustness compared with state-of-the-art methods.
• Drug Repositioning Application: Applied to the DrugBank database to prioritize candidates against SARS-CoV-2, with three of five top-ranked drugs reported as viable repurposed treatments.

Scientific Applications:

• DDI prediction: Predicts binary and multi-class/multi-label drug-drug interactions using learned chemical structure representations.
• DTI prediction: Predicts drug-target interactions via binary classification and regression using combined chemical and protein representations.
• Drug repositioning: Prioritizes repurposing candidates from DrugBank for SARS-CoV-2.
• Interpretation of chemical properties: Visualizes RGCN-learned features to reveal patterns linking chemical properties and drug categories for insight into biochemical interactions.

Methodology:

Applies residual graph convolutional networks (RGCNs) to learn structural representations of chemical structures, convolutional neural networks (CNNs) to learn sequential representations of proteins, visualizes structural features learned by the RGCN module, and evaluates performance through systematic benchmarking on binary-class DDIs, multi-class/multi-label DDIs, binary-class DTI classification, and DTI regression tasks, including drug repositioning experiments on DrugBank for SARS-CoV-2.