TAWFN: a deep learning framework for protein function prediction.

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Bioinformatics

Background:

• Protein function prediction is crucial for biological research and applications.
• High-throughput techniques generate vast protein sequence data, but functional annotation remains challenging.
• Existing structure-based methods often use Convolutional Neural Networks (CNN) or Graph Convolutional Networks (GCN) independently.

Purpose of the Study:

• To develop an integrated framework for protein function prediction that leverages protein structural information.
• To combine the strengths of CNN and GCN in a unified model for enhanced prediction accuracy.
• To introduce the Two-Model Adaptive Weight Fusion Network (TAWFN) for predicting protein functions.

Main Methods:

• Extracted amino acid contact maps and sequences from protein structures.
• Employed adaptive graph convolutional networks (AGCN) and multi-layer convolutional neural networks (MCNN) modules.
• Developed an adaptive weight computation network to fuse predictions from AGCN and MCNN for final classification.

Main Results:

• TAWFN achieved promising performance on PDBset and AFset datasets.
• Achieved Area Under the Precision-Recall Curve (AUPR) values of 0.718 (molecular function), 0.385 (biological process), and 0.488 (cellular component).
• Outperformed existing protein function prediction methods in experimental evaluations.

Conclusions:

• The proposed TAWFN framework effectively integrates CNN and GCN for protein function prediction.
• Leveraging protein structures with the TAWFN model offers a significant advancement over sequence-based methods.
• TAWFN demonstrates superior performance and potential for practical applications in bioinformatics.