Improvements in Protein Function Prediction Using Confidence in Protein Interactions

Characterizing protein function is a crucial part of understanding biological systems. Here we improve protein function prediction by accounting for data quality issues inherent in protein-protein interaction (PPI) databases. To accomplish this, we incorporate confidence information into the function prediction pipeline. The model pipeline uses weighted majority voting on the proteinprotein interaction network, with weights defined by shortest paths distance, confidence, diffusion state distance (DSD), or DSD with confidence. The result is that incorporating confidence weights in general significantly helps improve protein function prediction. Confidence with DSD performs especially well, improving by 11.9 pp over Majority Vote with ordinary shortest paths distance and no confi-