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

Area of Science:

• Computational biology
• Artificial intelligence in medicine
• Pharmacogenomics

Background:

• Identifying novel drug-disease associations (DDAs) is crucial for drug discovery.
• Existing AI methods often lack semantic consistency and fail to align heterogeneous drug and disease features.

Purpose of the Study:

• To propose a novel diffusion-enhanced fine-grained cross-semantic fusion framework (DFCDDA) for improved DDA prediction.
• To address limitations in semantic consistency and cross-modal alignment in current AI-driven DDA prediction models.

Main Methods:

• Utilized a conditional diffusion-based decoder to ensure semantic consistency between learned representations and original features.
• Employed an attention-guided graph convolutional network for precise feature aggregation from multi-view structural data.
• Implemented a bidirectional cross-attention module for aligning heterogeneous drug and disease features and capturing interactions.

Main Results:

• DFCDDA demonstrated superior performance in DDA prediction compared to existing approaches across three real-world datasets.
• The framework effectively learned robust drug and disease embeddings through collaborative module interactions.
• Case studies validated DFCDDA's capability in identifying plausible drug-disease associations.

Conclusions:

• The proposed DFCDDA framework offers a significant advancement in AI-driven DDA prediction.
• DFCDDA's ability to maintain semantic consistency and align heterogeneous features enhances drug discovery efforts.
• The model shows promise for uncovering novel and clinically relevant drug-disease relationships.