Section: Application Domains
Construction of Biological Networks
Comparative genomics provides the means to identify the set of protein-coding genes that comprise the components of a cell, and thus the set of individual functions that can be assured, but a more comprehensive view of cell function must aim to understand the ways that those components work together. In order to predict how genomic differences influence function differences, it is necessary to develop representations of the ways that proteins cooperate.
One such representation are networks of protein-protein interactions . Protein-protein interactions are at the heart of many important biological processes, including signal transduction, metabolic pathways, and immune response. Understanding these interactions is a valuable way to elucidate cellular function, as interactions are the primitive elements of cell behavior. One of the principal goals of proteomics is to completely describe the network of interactions that underly cell physiology.
As networks of interaction data become larger and more complex, it becomes more and more important to develop data mining and statistical analysis techniques. Advanced visualization tools are necessary to aid the researcher in the interpretation of these relevant subsets. As databases grow, the risk of false positives or other erroneous results also grows, and it is necessary to develop statistical and graph-theoretic methods for excluding outliers. Most importantly, it is necessary to build consensus networks , that integrate multiple sources of evidence. Experimental techniques for detecting protein-protein interactions are largely complementary, and it is reasonable to have more confidence in an interaction that is observed using a variety of techniques than one that is only observed using one technique.
The ProViz software tool [42] addresses the need for efficient visualization tools, and provides a platform for developing interactive analyses. But the key challenge for comparative analysis of interaction networks is the reliable extrapolation of predicted networks in the absence of experimental data.
A complementary challenge to the network prediction is the extraction of useful summaries from interaction data. Existing databases of protein-protein interactions mix different types too freely, and build graph representations that are not entirely sensible, as well as being highly-connected and thus difficult to interpret. We have developed a technique called policy-directed graph extraction that provides a framework for selecting observations and for building appropriate graph representations. A concrete example of graph extraction is subtractive pathway modeling , which uses correlated gene loss to identify loss of biochemical pathways.
