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

Predicting drug–target interactions is crucial for accelerating modern drug discovery. Although deep learning-based approaches have achieved significant progress, effectively integrating biochemical features originating from heterogeneous data sources and multilevel representations remains a major challenge. To address this limitation, this paper proposes MFDR-DTI, a DTI prediction framework based on Multisource Feature Decoupling Representation and a Dual-level dynamic Reorganization Fusion Mechanism. The framework innovatively decouples features into three orthogonal spaces─topological structure, physicochemical properties, and contextual semantics─and integrates them through the dual-level dynamic reorganization fusion mechanism, which consists of homogeneous feature integration and a shared-weight cross-property attention network. To tackle the imbalance in contribution among different feature branches during training, we design an adaptive collaborative optimization strategy based on gradient statistics, which dynamically adjusts the weight of each branch’s loss term according to its real-time impact on the global objective, thereby improving convergence efficiency and training stability. Comprehensive experiments on multiple widely used data sets demonstrate that MFDR-DTI consistently outperforms various state-of-the-art baseline models across multiple evaluation metrics. Ablation studies further verify that each proposed component positively contributes to improving prediction accuracy and enhancing generalization capability.