Probabilistic methods for predicting protein functions in protein-protein interaction networks

2005/03/12 by Christoph Best, Ralf Zimmer, Joannis Apostolakis
Biochemistry, Genetics and Molecular Biology · #q-bio.MN

paper · pdf

published as in: R. Giegerich, J. Stoye (eds.), German Conference on Bioinformatics 2004, Lecture Notes in Informatics, Ges. f. Informatik, Bonn, Germany, 2004 · 11 pages, 3 figures. Paper presented at the German Conference on Bioinformatics, 2004, Oct 4-6, Bielefeld, Germany

arxiv created 2005/03/12 · arxiv updated 2009/12/01

Abstract

We discuss probabilistic methods for predicting protein functions from protein-protein interaction networks. Previous work based on Markov Randon Fields is extended and compared to a general machine-learning theoretic approach. Using actual protein interaction networks for yeast from the MIPS database and GO-SLIM function assignments, we compare the predictions of the different probabilistic methods and of a standard support vector machine. It turns out that, with the currently available networks, the simple methods based on counting frequencies perform as well as the more sophisticated approaches.

Related

• Predicting protein functions with message passing algorithms
• Global protein function prediction in protein-protein interaction networks
• Conserved network motifs allow protein-protein interaction prediction
• Mixed membership analysis of high-throughput interaction studies: Relational data
• Global Snapshot of Protein Interaction Network -- A Percolation Based Approach