Diffusion-enhanced Fine-grained Cross Semantic Fusion for Drug-disease Association Prediction.

Abstract

Identifying novel drug-disease associations (DDAs) is critical for advancing drug discovery. While AI-driven approaches have shown promise, most still struggle to maintain semantic consistency between learned high-order representations and the original feature space. Additionally, they often neglect the varying importance of different views and fail to model the semantic gaps between heterogeneous drug and disease features, hindering cross-modal alignment. To overcome these challenges, we propose a novel diffusion-enhanced fine-grained cross-semantic fusion framework for DDA prediction, namely DFCDDA. First, a conditional diffusion-based decoder is leveraged to ensure semantic consistency between the high-order features learned by the model and the original features. Second, an attention-guided fine-grained graph convolutional network dynamically generates soft adjacency matrices from multi-view structural data, enabling precise feature aggregation. Third, a bidirectional cross-attention module aligns heterogeneous drug and disease features and captures their complementary interactions. These three modules work collaboratively to improve the learning of robust drug and disease embeddings. Experiments on three real-world datasets show that DFCDDA consistently outperforms existing approaches in DDA prediction. Case studies further demonstrate its effectiveness in uncovering plausible drug-disease associations.