MF-DTA: Predicting drug-target affinity with multi-modal feature fusion model.

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Predicting drug-target interactions (DTIs) and binding affinities (DTAs) is crucial for drug discovery.
• Existing methods often underutilize multimodal information from molecular structures.

Purpose of the Study:

• To develop a multimodal feature fusion model (MF-DTA) for accurate DTI and DTA prediction.
• To leverage novel molecular representations and advanced deep learning architectures.

Main Methods:

• Introduced molecular fragment graphs (via BRICS decomposition) as a new drug modality.
• Applied deformable convolutions to protein contact maps for enhanced feature extraction.
• Utilized a mixture-of-experts (MoE) multihead attention and dual-decoder architecture for feature fusion and cross-modal interaction.

Main Results:

• MF-DTA significantly outperformed state-of-the-art methods on benchmark datasets (Davis, KIBA, BindingDB).
• Achieved notable improvements in concordance index (CI) and excelled in MSE and R ^{m 2} metrics.
• Model visualization confirmed its ability to learn meaningful drug-target interaction patterns.

Conclusions:

• MF-DTA provides accurate and robust binding affinity predictions.
• The model's interpretability makes it a valuable tool for drug design and target identification.
• Demonstrated practical utility by screening natural products for tubulin targets.