CNN-Siam: multimodal siamese CNN-based deep learning approach for drug‒drug interaction prediction

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning

Background:

• Drug-drug interactions (DDIs) are critical in drug development and clinical practice.
• Experimental DDI studies are infeasible for vast drug combinations.
• Deep learning offers a promising computational approach for DDI prediction.

Purpose of the Study:

• To develop a deep learning model for predicting drug-drug interactions (DDIs).
• To leverage multimodal drug data, including chemical substructures, targets, and enzymes.
• To enhance the accuracy and efficiency of DDI prediction.

Main Methods:

• A novel convolutional neural network (CNN) algorithm, CNN-Siam, employing a Siamese network architecture.
• Utilizing twin CNNs to learn feature representations from multimodal drug data.
• Predicting DDI types using advanced optimization algorithms (RAdam and LookAhead).

Main Results:

• CNN-Siam achieved an Area Under the Precision-Recall Curve (AUPR) score of 0.96.
• Demonstrated a prediction accuracy rate of 92%, a significant improvement over state-of-the-art methods.
• Ablation experiments confirmed the model's robustness and the effectiveness of the optimization algorithms.

Conclusions:

• The multimodal siamese convolutional neural network accurately predicts DDIs.
• The Siamese network architecture effectively learns drug pair feature representations.
• CNN-Siam shows superior performance, though generalization and training time require further improvement.