Capsule enclosed coordinate attention based dual batch depthwise convolutional knowledge distillation model for...

Area of Science:

• Pharmacology and Cheminformatics
• Computational Drug Discovery
• Artificial Intelligence in Medicine

Background:

• Drug-drug interactions (DDIs) pose significant risks to patient safety and research efficiency.
• Accurate DDI prediction is crucial for safe drug co-administration.
• Integrating diverse data sources can improve the precision of DDI prediction models.

Purpose of the Study:

• To propose a novel model, CC-DBDKD, for enhanced drug-drug interaction prediction.
• To leverage capsule networks, coordinate attention, and knowledge distillation for improved DDI identification.
• To develop a scalable and robust model for predicting complex drug relationships.

Main Methods:

• Utilized DrugBank dataset and RDKit for data preprocessing and SMILES standardization.
• Generated molecular fingerprints (ECFPs, MACCS, PubChem, 3D, MD) and calculated Structural Similarity Profiles (SSP).
• Developed the CC-DBDKD model incorporating capsule networks, coordinate attention, dual-batch depthwise convolutions, and knowledge distillation.

Main Results:

• The CC-DBDKD model achieved superior accuracy (0.987, 0.989) and an F1-score of 0.986.
• Demonstrated improved performance over existing models like CNN, CNN-LSTM, Autoencoder, and D-CNN.
• The model showed strong scalability for large datasets.

Conclusions:

• The CC-DBDKD model offers a powerful and scalable solution for accurate drug-drug interaction prediction.
• This approach enhances drug discovery efficiency and patient safety by identifying potential DDIs.
• The integration of advanced deep learning techniques provides a robust framework for future DDI research.