MCANet: shared-weight-based MultiheadCrossAttention network for drug-target interaction prediction.

Area of Science:

• Computational biology
• Pharmacology
• Machine learning

Background:

• Drug-target interaction (DTI) prediction is crucial for efficient drug development.
• Deep learning models for DTI require robust feature representations and struggle with class imbalance and overfitting.
• Computational efficiency is a key consideration in DTI prediction models.

Purpose of the Study:

• To develop a novel attention mechanism for improved DTI prediction.
• To create accurate and computationally efficient deep learning models for DTI.
• To address class imbalance and overfitting in drug-target datasets.

Main Methods:

• Proposed a shared-weight-based MultiheadCrossAttention mechanism to model drug-target associations.
• Developed MCANet, utilizing cross-attention for feature extraction and PolyLoss to mitigate data challenges.
• Introduced MCANet-B, an ensemble of MCANet models to enhance robustness and prediction accuracy.

Main Results:

• Achieved state-of-the-art performance on six public drug-target datasets.
• MCANet demonstrated significant computational resource savings while maintaining high accuracy.
• MCANet-B further boosted prediction accuracy through model ensembling, balancing performance and resource usage.

Conclusions:

• The proposed MultiheadCrossAttention mechanism enhances DTI prediction accuracy and efficiency.
• MCANet and MCANet-B offer effective solutions for challenges in DTI prediction, including data imbalance and overfitting.
• These models represent a significant advancement in computational approaches to drug discovery.