A Unified Differential Denoising Learning Framework With a Pre-Trained Model and Fuzzy Graph Networks for Drug-Drug Interaction Prediction.

Abstract

Combination therapy has become increasingly prevalent in modern clinical practice, yet the concomitant issue of drug-drug interactions (DDIs) poses significant challenges to medication safety. Accurate DDI prediction is therefore crucial for both drug development and clinical regulation. However, existing methods exhibit limitations in handling the inherent uncertainty of molecular data and attention noise in interaction modeling. This study proposes UDPF-DDI, a unified differential denoising learning framework with a pre-trained model and fuzzy graph networks for DDI prediction. First, multi-source feature representations were constructed for drug molecules: high-dimensional feature representations were extracted using the large-scale molecular pre-trained model Uni-Mol2, while low-dimensional graph structures were obtained via RDKit. Second, a parallel three-channel encoder was employed. This utilized a fuzzy graph convolution neural network (FGCNN) to capture high-order fuzzy features reflecting internal chemical environments and external network topologies of molecules. Simultaneously, the high-dimensional 3-D atomic features generated by the molecular pre-trained model were refined and denoised using a differential Transformer. Subsequently, the outputs of the three channels are weighted and integrated through graph readout and adaptive fusion operations to obtain the interaction score for the drug pair. Finally, the interaction probability is computed via a task-specific activation function. Results from fivefold cross-validation on multiple real-world binary and multi-class datasets demonstrated that UDPF-DDI achieved significant performance improvements across various evaluation metrics compared to several state-of-the-art baseline models. It exhibited exceptional capability, particularly in handling data uncertainty and denoising drug features.