What Can Networks Do for You?

Abstract

This chapter aims at demonstrating the utility of network approaches in classification and outlier detection tasks in the context of stem cell biology and related fields. With modern high-through-put methods it has now become easier and cheaper to accurately measure thousands of features on a genome-wide scale than to define a low number of markers that can be tested, for example with low throughput RT-PCR assays. Typically the number of potential markers exceeds the number of experiments by several orders of magnitude. Therefore the significance – let alone mechanistic involvement – of each possible feature cannot be guaranteed from the data alone. Fortunately, easy-to-use implementations of many powerful network based algorithms have been made freely available so one can readily employ these advanced algorithms on new high-content datasets.

We will exemplify how network information and structure can be used to improve the prediction of biological phenotypes, and discuss methodological considerations pertinent to enabling reliable and biologically meaningful inferences from in silico network studies. We will touch upon difficulties inferring “true” (i.e. mechanistic) networks from biological data and note that, from a practical standpoint, in silico networks need not to fully reflect observable biological phenomena for real-world predictability and utility. We have found that a particularly successful strategy is to use statistical learning theory as a stringent framework for comparative evaluation of alternative network methods. This pragmatic and evolutionary approach can be adopted in several biological realms and makes optimal use of todays sophisticated network modeling methodologies. We observe that only such a rigorous workflow can guarantee reproducibility of network-based findings.