TriCvT-DTI: Predicting Drug-Target Interactions Using Trimodal Representations and Convolutional Vision Transformers.

Area of Science:

• Computational chemistry
• Bioinformatics
• Drug discovery

Background:

• Predicting drug-target interactions (DTI) is crucial for drug discovery and repositioning.
• Traditional methods are time-consuming; deep learning offers efficiency but often uses limited drug representations.
• Existing deep learning models may not fully capture local and global drug information essential for DTI tasks.

Purpose of the Study:

• To propose TriCvT-DTI, a novel deep learning approach for accurate drug-target interaction prediction.
• To comprehensively represent drugs by integrating molecular images, chemical sequence features, and graph representations.
• To enhance feature learning between drugs and targets using a bidirectional multi-head attention mechanism.

Main Methods:

• Developed TriCvT-DTI, combining Convolutional Vision Transformers (CvTs) for image feature extraction with sequence and graph data.
• Implemented a bidirectional multi-head attention mechanism for interactive feature learning between drug and target data.
• Evaluated the model on Human, C. elegans, and Davis datasets, comparing uni-modality and bi-modality training.

Main Results:

• TriCvT-DTI significantly outperformed existing state-of-the-art methods on DTI prediction across multiple datasets.
• The model demonstrated strong performance on both balanced and unbalanced datasets.
• Experimental results highlighted the effectiveness of integrating diverse data modalities for improved prediction accuracy.

Conclusions:

• TriCvT-DTI provides a robust and effective framework for drug-target interaction prediction.
• The comprehensive feature representation and attention mechanism contribute to superior model performance.
• The approach shows promising generalization capabilities, advancing drug discovery and repositioning efforts.