Revisiting drug-protein interaction prediction: a novel global-local perspective

Area of Science:

• Computational biology
• Bioinformatics
• Drug discovery

Background:

• Accurate drug-protein interaction (DPI) inference is crucial for understanding drug mechanisms and developing new therapeutics.
• Current deep learning models face limitations in node representation accuracy for DPI prediction.

Purpose of the Study:

• To develop a novel computational framework for efficient and accurate inference of potential drug-protein interactions (DPIs).
• To improve node representation by integrating global and local features within a drug-protein bipartite graph.

Main Methods:

• Utilized pre-trained models for initial drug and protein feature extraction.
• Employed MinHash and HyperLogLog for local feature estimation (similarity and set cardinality).
• Integrated an energy-constrained diffusion mechanism within a transformer architecture for global feature extraction and node interdependency analysis.

Main Results:

• The proposed framework effectively fuses local and global node features for precise DPI prediction.
• Experimental validation confirmed the model's accuracy and reliability in identifying potential DPIs.
• Molecular docking results demonstrated the model's ability to discover novel DPIs beyond existing databases.

Conclusions:

• The developed computational framework offers a significant advancement in DPI prediction through comprehensive node representation.
• This approach is expected to provide valuable insights for drug repurposing and personalized medicine research.
• The model's capability to identify unknown DPIs holds promise for accelerating therapeutic development.