AptRank: an adaptive PageRank model for protein function prediction on   bi-relational graphs

Area of Science:

• Bioinformatics
• Computational Biology
• Systems Biology

Background:

• Diffusion-based network models are effective for protein function prediction.
• Integrating Gene Ontology (GO) hierarchy significantly enhances prediction accuracy.
• Previous methods have not jointly modeled protein networks and GO hierarchy as a two-layer network.

Purpose of the Study:

• To propose novel diffusion-based methods that leverage a two-layer network model combining protein-protein associations and GO hierarchy.
• To develop BirgRank and AptRank, utilizing PageRank for information diffusion on this integrated network.
• To evaluate the performance of these new methods against existing approaches using multiple validation strategies.

Main Methods:

• Constructed a Bi-relational graph (Birg) model integrating protein-protein association and GO function-function hierarchy.
• Developed BirgRank (fixed decay PageRank) and AptRank (adaptive diffusion PageRank) for information diffusion.
• Evaluated methods on yeast, fly, and human protein datasets using four distinct validation strategies.

Main Results:

• Both BirgRank and AptRank demonstrated superior performance compared to four existing methods (GeneMANIA, TMC, ProteinRank, clusDCA).
• The proposed methods showed particular strength in missing function prediction, even with limited training data (10%).
• Comprehensive evaluation across multiple prediction tasks confirmed the effectiveness of the two-layer network approach.

Conclusions:

• The proposed two-layer network model integrating protein networks and GO hierarchy is highly effective for protein function prediction.
• AptRank, with its adaptive diffusion, offers improved performance over BirgRank and other existing methods.
• These findings advance the field of computational protein function prediction by offering more accurate and robust tools.