POLAT: Protein function prediction based on soft mask graph network and residue-Label ATtention

Area of Science:

• Computational Biology
• Biochemistry
• Genetics
• Molecular Biology

Background:

• Accurate protein function prediction is crucial due to the gap between sequence and functional data.
• Protein structure, strongly linked to function, enables structure-based prediction methods.
• Current methods fail to leverage residue-specific contributions and correlations with function.

Purpose of the Study:

• To develop an advanced computational method for protein function prediction.
• To effectively integrate sequence, predicted structure, and function-level information.
• To address limitations of existing structure-based methods by considering residue-function relationships.

Main Methods:

• Proposed POLAT (Protein function prediction based on Soft Mask Graph Networks and Residue-Label Attention).
• Utilized soft mask graph networks for adaptive extraction of function-relevant residues.
• Employed a residue-label attention mechanism for encoding protein-level features.

Main Results:

• POLAT achieved superior performance on the PDB cdhit test set (MFO, BPO, CCO domains) compared to the state-of-the-art GAT-GO method.
• Demonstrated competitive results against sequence-based and multimodal methods.
• Achieved state-of-the-art performance in three out of six evaluated metrics.

Conclusions:

• POLAT effectively combines diverse data types for accurate protein function prediction.
• The method shows significant improvements over existing structure-based approaches.
• POLAT represents a advancement in computational protein function prediction.