MFDR-DTI: Multisource Feature Decoupling and Dual-Level Reorganization Fusion for Accurate Drug-Target Interaction...

Area of Science:

• Bioinformatics
• Computational Chemistry
• Drug Discovery

Background:

• Accurate drug-target interaction (DTI) prediction is vital for efficient drug discovery.
• Current deep learning methods struggle to integrate heterogeneous biochemical features from multiple sources.
• Challenges include handling multilevel representations and feature imbalances during model training.

Purpose of the Study:

• To propose MFDR-DTI, a novel framework for DTI prediction.
• To effectively integrate diverse biochemical features using a unique decoupling and fusion mechanism.
• To enhance model training stability and convergence efficiency through adaptive optimization.

Main Methods:

• Developed MFDR-DTI framework utilizing Multisource Feature Decoupling Representation and Dual-level dynamic Reorganization Fusion.
• Decoupled features into topological structure, physicochemical properties, and contextual semantics.
• Implemented adaptive collaborative optimization strategy based on gradient statistics to manage feature branch contributions.

Main Results:

• MFDR-DTI consistently outperformed state-of-the-art baseline models on multiple datasets and evaluation metrics.
• Ablation studies confirmed the positive contribution of each proposed component to prediction accuracy.
• The framework demonstrated enhanced generalization capability in DTI prediction.

Conclusions:

• MFDR-DTI offers a robust and effective approach for predicting drug-target interactions.
• The proposed feature decoupling and fusion mechanism significantly improves prediction performance.
• The adaptive optimization strategy enhances training stability and efficiency in DTI prediction models.