SpatialPPIv2: Enhancing protein-protein interaction prediction through graph neural networks with protein language...

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Biology

Background:

• Protein-protein interactions (PPIs) are essential for cellular processes.
• Accurate PPI prediction aids biological mechanism understanding and drug discovery.
• Existing methods often rely on experimentally determined structures or complex prediction pipelines.

Purpose of the Study:

• To introduce SpatialPPIv2, an improved graph-neural-network-based model for predicting PPIs.
• To enhance the specificity and robustness of PPI prediction.
• To develop a standalone PPI prediction tool independent of protein structure prediction algorithms.

Main Methods:

• Utilized large language models for sequence feature embedding.
• Employed graph attention networks to capture structural information.
• Leveraged the PINDER dataset, integrating data from RCSB PDB and AlphaFold databases.

Main Results:

• SpatialPPIv2 demonstrates superior accuracy and reliability compared to state-of-the-art PPI predictors.
• The model achieves robust performance even when using predicted structures from AlphaFold3, AlphaFold2, and ESMFold.
• SpatialPPIv2 can predict protein interactions independently, without requiring protein structure prediction.

Conclusions:

• SpatialPPIv2 offers a powerful and versatile solution for accurate PPI prediction.
• The model provides valuable insights into protein function, supporting drug discovery and synthetic biology.
• SpatialPPIv2 represents a significant advancement in computational approaches to understanding protein interactions.