Predicting protein complex membership using probabilistic network reliability

Abstract

Evidence for specific protein-protein interactions is increasingly available from both small- and large-scale studies, and can be viewed as a network. It has previously been noted that errors are frequent among large-scale studies, and that error frequency depends on the large-scale method used. Despite knowledge of the error-prone nature of interaction evidence, edges (connections) in this network are typically viewed as either present or absent. However, use of a probabilistic network that considers quantity and quality of supporting evidence should improve inference derived from protein networks. Here we demonstrate inference of membership in a partially known protein complex by using a probabilistic network model and an algorithm previously used to evaluate reliability in communication networks.

Figure 1

Probabilistic versus binary networks. (A) Schematic illustration of a probabilistic network, with higher edge weight (probability) represented by darker coloring. (B, C) Binary networks randomly sampled from the probabilistic network in A.

Figure 2

Results of several examples using MIPS complexes (Mewes et al. 2002) as “core complex” queries. Probabilistic interaction subgraphs are visualized by the software Pajek (Batagelj and Mrvar 1998). Query proteins are marked in red, and the top 50 proteins returned are colored in grayscale according to rank, with lighter coloring indicating better rank. Each edge is given thickness proportional to its posterior probability. Shown are SAGA complex (A), NOT complex (B), replication factor C complex (C), and the Arp2/Arp3 complex (D).

Figure 3

Success rate versus rank for ProNet and SPE methods. Success rate is the number of correct predictions found at or above the threshold rank R in all the cross-validation trials divided by the total number of predictions above the specified rank.