PEB-DDI: A Task-Specific Dual-View Substructural Learning Framework for Drug-Drug Interaction Prediction.

Area of Science:

• Computational chemistry
• Pharmacology
• Artificial intelligence in drug discovery

Background:

• Adverse drug-drug interactions (DDIs) are a significant risk in polypharmacy, often stemming from uncharacterized physicochemical incompatibilities.
• Current graph neural network models for DDI prediction excel at atom-level features but neglect crucial chemical bond information and lack adaptability for novel drug scenarios.

Purpose of the Study:

• To develop an enhanced DDI prediction framework, PEB-DDI, that incorporates chemical bond information and employs adaptive strategies for both known and novel drug interactions.
• To improve the accuracy and generalization capabilities of DDI prediction models, addressing limitations in existing substructural frameworks.

Main Methods:

• PEB-DDI integrates chemical bond information synchronously with atomic nodes in molecular graphs.
• Employs distinct dual-view strategies: Molecular fingerprint-Molecular graph view for transductive tasks and Bipartite graph-Molecular graph view for inductive tasks, adapting to the presence of novel drugs.
• Utilizes enhanced substructure extraction incorporating bond representations.

Main Results:

• PEB-DDI demonstrates superior performance on benchmark datasets, achieving 98.18% accuracy on DrugBank for predicting previously unknown interactions among approved drugs.
• The framework exhibits strong generalization capabilities for novel drugs, reaching an accuracy rate of 88.06%.

Conclusions:

• PEB-DDI effectively enhances DDI prediction by integrating chemical bond information and employing adaptive dual-view strategies.
• The proposed method significantly improves prediction accuracy for both known and novel drug-drug interactions, offering a more robust solution for polypharmacy safety.