NetQuilt: deep multispecies network-based protein function prediction using homology-informed network similarity.

Area of Science:

• Computational Biology
• Bioinformatics
• Systems Biology

Background:

• Transferring protein function knowledge across species is hindered by distinct proteomes and cellular architectures.
• Existing methods often rely on sequence similarity or single-species protein-protein interaction (PPI) networks, limiting accuracy for proteins without known homologues or species lacking PPI data.

Purpose of the Study:

• To develop a robust method for cross-species protein function prediction by integrating diverse biological data.
• To overcome limitations of single-species approaches and improve functional annotation accuracy, especially in data-scarce organisms.

Main Methods:

• Integrated sequence and multispecies PPI network data using IsoRank similarity to create a meta-network profile.
• Trained a maxout neural network using Gene Ontology (GO) terms as target labels on the multispecies meta-network.
• Evaluated performance against existing network-based, deep learning sequence-based, and BLAST annotation methods.

Main Results:

• The multispecies approach significantly improved protein function prediction performance compared to existing methods.
• The method demonstrated strong predictive power even when a species' own PPI network data was excluded.
• Leveraged increased training examples from multiple species for enhanced accuracy.

Conclusions:

• Integrating multispecies sequence and network information provides a powerful framework for accurate protein function prediction.
• This approach enhances functional annotation, particularly for organisms with limited available biological network data.
• The developed method offers a valuable tool for comparative genomics and functional genomics studies.