Other researchers are attempting to use computational methods to uncover undocumented protein-protein interactions. Recently, Califano collaborated with Columbia biochemist Barry Honig to search for potential interactions between proteins using the tertiary structure of the proteins (4). That data alone is not context-specific, Calfano notes, so the researchers took advantage of additional data sets, such as gene expression analysis, to refine their findings. Using this approach, protein interactions in adenocarcinomas were analyzed. Although they've shown that it's accurate, the method only shows some overlap with known interactions. As a result, Califano says, “we know that there is a tremendous amount of work to do.”
Even if researchers haven't found the connections between particular proteins or genes from a known biochemical interaction, data patterns can help point to how genes might be connected. Here, the idea of mutual exclusivity can be a huge help. If two proteins interact to form an essential complex within a cell, mutations in one of the proteins could disrupt the complex, leading to disease. But in diseased cells you'll rarely see mutations to both genes. Raphael and his colleagues have developed an algorithm called Dendrix (de novo driver exclusivity) to look for statistical relationships in genomic data that could reflect these patterns. He has already used this algorithm with TCGA data for samples of acute myeloid leukemia, finding clear patterns of mutual exclusivity (5). And they are now analyzing data from other cancer types as well.
In addition to finding distinctive features and classifying tumor types, networks can also help researchers identify new patterns of mutations occuring in multiple cancers, similar to the BRCA1 and BRCA2 mutations that occur in both breast and ovarian tumors. These hotspots could suggest new targets for therapies that target multiple tumor types.Looming challenges
Tumor samples present unique challenges in network analysis. They are heterogeneous, cautions Lincoln Stein of the Ontario Institute for Cancer Research, producing yet another layer of biological complexity.
A patient sample can include normal and tumor cells, with the tumor cells possessing multiple subtypes in some instances. Understanding that heterogeneity could be incredibly important for patients, Stein adds. In some cases tumors generate different subtypes from a common ancestor, while other tumors come from multiple independent tumor modules. Heterogeneity can affect the clinical outcome. For example, if a patient has one tumor subclone with an EGFR mutation but another subclone that doesn't include that mutation, a targeted inhibitor might kill only some of the cells in the tumor. Computational methods can help researchers tease apart some of these complications, Stuart says, but biochemical techniques that produce data from single cells would provide cleaner data to start with.
For Ideker, moving network models into a healthcare setting remains a major priority as such networks would assist clinicians in making sense of the rare mutations that show up in their patients. “No one is unique. These mutations are hitting the same regions of the network.” And this is one time that it is good to be like everyone else.