DualNetGO: a dual network model for protein function prediction via effective feature selection.

Area of Science:

• Bioinformatics
• Computational Biology
• Systems Biology

Background:

• Protein-protein interaction (PPI) networks are vital for understanding protein functions.
• Heterogeneous PPI networks present challenges for effective information utilization in protein function prediction.
• Current deep learning models often combine all network data, potentially increasing noise and reducing performance.

Purpose of the Study:

• To develop a novel dual-network model, DualNetGO, for accurate protein function prediction.
• To address the challenge of effectively selecting and integrating information from diverse PPI networks and protein attributes.
• To improve upon existing methods by judiciously harnessing information from multiple sources.

Main Methods:

• Developed DualNetGO, a dual-network architecture with a Classifier and a Selector.
• Integrated graph embeddings from PPI networks, protein domain information, and subcellular localization.
• Evaluated model performance on human and mouse datasets, and CAFA3 benchmark data.

Main Results:

• DualNetGO achieved significant improvements in Fmax scores across Gene Ontology categories (BP, MF, CC) for both human and mouse datasets compared to other network-based models.
• Demonstrated generalization capability on CAFA3 data and versatility with Esm2 embeddings.
• Showed insensitivity to graph embedding methods and efficiency in terms of time and memory.

Conclusions:

• A selective approach to integrating features from multiple PPI networks and protein attributes is superior to indiscriminate combination.
• DualNetGO offers a robust and efficient method for protein function prediction, outperforming existing approaches.
• The model's ability to select optimal features enhances the utility of complex biological network data.