A Biophysical Prior-Based Hierarchical Graph Pooling Strategy for Protein Function Prediction.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning

Background:

• Accurate protein function prediction is crucial for understanding biological systems.
• Graph Neural Networks (GNNs) show potential but struggle with aggregating residue-level features.
• Existing pooling methods often ignore protein hierarchy, losing structural information and interpretability.

Purpose of the Study:

• To introduce BioPhysPool, a hierarchical pooling strategy based on biophysical priors.
• To enhance protein function prediction using GNNs by preserving structural information.
• To improve the interpretability of GNNs in biological applications.

Main Methods:

• Developed BioPhysPool, utilizing secondary structure elements for deterministic graph coarsening.
• Created a hierarchical GNN architecture with cross-granularity attention fusion.
• Integrated local chemical features with macro-topological patterns.

Main Results:

• BioPhysPool outperformed classical pooling methods on protein function prediction benchmarks.
• The method demonstrated superior performance in both binary and multi-class classification tasks.
• Interpretability analysis showed attention allocation driven by structural context, aligning with known residue importance.

Conclusions:

• Embedding biophysical priors into graph learning frameworks offers a robust solution for protein function prediction.
• BioPhysPool provides an interpretable and effective approach by respecting protein hierarchical organization.
• The strategy successfully integrates structural and chemical information for improved predictive accuracy.