Algorithms for the analysis of protein interaction networks

Abstract

In the decade since the human genome project, a major research trend in biology has been towards understanding the cell as a system. This interest has stemmed partly from a deeper appreciation of how important it is to understand the emergent properties of cellular systems (e.g., they seem to be the key to understanding diseases like cancer). It has also been enabled by new high-throughput techniques that have allowed us to collect new types of data at the whole-genome scale. We focus on one sub-domain of systems biology: the understanding of protein interactions. Such understanding is valuable: interactions between proteins are fundamental to many cellular processes. Over the last decade, high-throughput experimental techniques have allowed us to collect a large amount of protein-protein interaction (PPI) data for many species. A popular abstraction for representing this data is the protein interaction network: each node of the network represents a protein and an edge between two nodes represents a physical interaction between the two corresponding proteins. This abstraction has proven to be a powerful tool for understanding the systems aspects of protein interaction. We present some algorithms for the augmentation, cleanup and analysis of such protein interaction networks: 1. In many species, the coverage of known PPI data remains partial. Given two protein sequences, we describe an algorithm to predict if two proteins physically interact, using logistic regression and insights from structural biology. We also describe how our predictions may be further improved by combining with functional-genomic data. 2. We study systematic false positives in a popular experimental protocol, the Yeast 2-Hybrid method. Here, some "promiscuous" proteins may lead to many false positives. We describe a Bayesian approach to modeling and adjusting for this error. 3. Comparative analysis of PPI networks across species can provide valuable insights. We describe IsoRank, an algorithm for global network alignment of multiple PPI networks. The algorithm first constructs an eigenvalue problem that encapsulates the network and sequence similarity constraints. The solution of the problem describes a k-partite graph that is further processed to find the alignment. 4. For a given signaling network, we describe an algorithm that combines RNA-interference data with PPI data to produce hypotheses about the structure of the signaling network. Our algorithm constructs a multi-commodity flow problem that expresses the constraints described by the data and finds a sparse solution to it.