Gaining confidence in high-throughput protein interaction networks

Abstract

Although genome-scale technologies have benefited from statistical measures of data quality, extracting biologically relevant pathways from high-throughput proteomics data remains a challenge. Here we develop a quantitative method for evaluating proteomics data. We present a logistic regression approach that uses statistical and topological descriptors to predict the biological relevance of protein-protein interactions obtained from high-throughput screens for yeast. Other sources of information, including mRNA expression, genetic interactions and database annotations, are subsequently used to validate the model predictions without bias or cross-pollution. Novel topological statistics show hierarchical organization of the network of high-confidence interactions: protein complex interactions extend one to two links, and genetic interactions represent an even finer scale of organization. Knowledge of the maximum number of links that indicates a significant correlation between protein pairs (correlation distance) enables the integrated analysis of proteomics data with data from genetics and gene expression. The type of analysis presented will be essential for analyzing the growing amount of genomic and proteomics data in model organisms and humans.

A simple physical model for scaling in protein-protein interaction networks

“…It has long been noted, however, that Y2H screens are rather inaccurate and can lead to relatively ''noisy'' sets of interactions (19)(20)(21)(22). Indeed, when the two major S. cerevisiae proteinprotein interaction (PPI) experiments are compared with one another, one finds that only Ϸ150 of the thousands of interactions identified in each experiment are recovered in the other experiment (22).…”

Structure, function, and evolution of transient and obligate protein–protein interactions

“…With increasing amounts of data on protein-protein interactions for several species as well as the emphasis on representing and understanding basic biological processes in terms of networks of interactions, it is important to focus on the precise definition and classification of these underlying interactions. Some computational analyses tend to group together disparate datasets originating from different experimental methods to get more robust answers (5), which sometimes tends to blur the definitions of the nodes and edges of the merged networks. Although the simplest approach to networks as sets of binary interactions provides some rudimentary understanding of the data, the more realistic and nuanced view in terms of modular complexes and subcomplex structures is needed, and such characterizations have recently started appearing in the literature (6).…”

Probabilistic Paths for Protein Complex Inference

“…We then compared algorithm performance for confidence-weighted edges taken from LR [10]. The PROPATH-EXP and PROPATH-ALG algorithms perform the best and are comparable, followed closely by BESTPATH (Fig.…”

“…1A and 1B are from edge weights using NB [9], Fig. 1C and 1D are from edge weights using LR [10] and Fig. 1E and 1F are from edge weights using DT [12].…”

“…The second network, by Bader and coworkers, predicts the probability of a direct or indirect physical interaction using a logistic regression model (LR) [10], …”