Physics-Diffusion-Driven Multiscale Aggregation for Drug-Target Interaction Prediction

Abstract

Drug-target interaction prediction is an important task in computational drug discovery. To address the limitations of existing graph neural network approaches in propagation depth and robustness, we propose PDDMA-DTI (Physics-Diffusion-Driven Multiscale Aggregation for Drug-Target Interaction Prediction), a physics-inspired, multiscale adaptive aggregation framework. PDDMA-DTI injects a physics-motivated diffusion operator into a heterogeneous drug-target network: global diffusion models cross-modal information propagation via the graph Laplacian, while local diffusion performs neighborhood smoothing in the low-dimensional embedding space. Both processes are solved with the implicit Euler scheme to ensure numerical stability. For each node, we introduce an adaptive stopping rule based on steady-state convergence and a weighted fusion strategy. This strategy leverages hop counts and node degrees to aggregate multi-step diffusion snapshots into a final representation that preserves original features while capturing multi-scale context. To emphasize strong interaction pathways, we further propose a Physical Interaction Summation Enhancement module that augments pairwise interaction features with a linear plus quadratic energy term, thereby amplifying highly correlated drug-target pairs. Extensive experiments on multiple benchmark datasets show that PDDMA-DTI consistently outperforms representative baselines in accuracy and robustness.