Predicting drug-drug interactions by graph convolutional network with multi-kernel.

Area of Science:

• Pharmacology
• Computational Biology
• Drug Discovery

Background:

• Drug repositioning identifies new uses for existing medications.
• Predicting drug-drug interactions (DDIs) is crucial for understanding drug pharmacology and discovering novel therapeutics.
• Current DDI prediction models often integrate diverse interactions into a single network, potentially losing valuable information.

Purpose of the Study:

• To propose a novel Graph Convolutional Network with Multi-Kernel (GCNMK) for enhanced DDI prediction.
• To leverage distinct DDI types by employing separate graph kernels for 'increase'- and 'decrease'-related interactions.
• To identify potential drug candidates for novel treatments through accurate DDI prediction.

Main Methods:

• Developed a GCNMK model utilizing two specialized DDI graph kernels (increased and decreased interactions).
• Employed learned drug features, with target features demonstrating superior performance over other feature types.
• Compared GCNMK against three existing DDI prediction methods.

Main Results:

• GCNMK achieved superior performance in DDI prediction, evidenced by higher area under the receiver operating characteristic curve (AUC-ROC) and area under the precision-recall curve (AUC-PR).
• Target drug features were identified as the most effective for DDI prediction.
• Case studies successfully identified numerous potential DDIs, with many validated by existing evidence.

Conclusions:

• The proposed GCNMK model offers a significant advancement in predicting drug-drug interactions.
• Distinguishing between different types of DDIs enhances prediction accuracy and aids in drug repositioning efforts.
• GCNMK demonstrates potential for identifying novel therapeutic strategies and supporting drug discovery pipelines.