4.1 Introduction

4.1.1 Biological Networks and Computational Challenges

The huge interest in complex networks across many research areas has also found application in biological studies, where associations between genes, proteins, and metabolites deserve further investigation particularly due to the underlying regulative or interactive dynamics. Here the proposed work addresses protein interactome networks (PIN) [1] from an integrative dynamic perspective, and aims to establish a better definition of their modular configurations.

There are currently reasons of concern in relation to the computational analysis of PIN, and they mainly refer to three problems. First, there is a limited interactome coverage [2] that depends on the organism under study [3] and on the available data-generating methodologies (yeast two-hybrid, co-IP, text mining and literature mining, DB curation, orthology, etc.). Consequently, data integration is often needed to ensure a better data uncertainty control and validation quality.

Second, there is also limited measurement accuracy as a limiting factor, and refers to the uncertainty inherent to both experimentally measured and predicted interactions (due to various sources of errors, biases, etc.). For example, evidence was recently provided [4] with regard to literature-curated interactome data about the necessity of careful quality control for reliable inference. Notably, scoring systems have been proposed to assign reliability to the interactions, thus leading to common classification into low-and high-confidence PIN.

Third, detecting modularity is a very complicated task that offers only approximate solutions. Various different principles and methods (see Ref. 5 for comparative evaluations) can be applied for network partitioning, but without guarantee of achieving the best possible approximation quality due to the so-called “network resolution limit” problem [6,7]. As the truly informative module sizes may not match the algorithmically retrievable ones, we can observe suboptimal configurations, either sparse (with a few dense modules) or highly redundant (with many small-to-intermediate overlapping modules) maps.

As a result, such incompleteness and inaccuracy of representation calls for both new inference methods and better use of the currently adopted ones. These are challenging tasks, further complicated by the fact that the available protein interaction map consists of a mix of real interactions and false positives, and correspondingly non-interactions and false negatives. Therefore, while such map represents static entities subject to limitations and constraints, they actually refer to underlying associations that dynamically change depending on experimental conditions, system perturbations, and so on. Accordingly, a better control of the degree of uncertainty embedded in the protein maps requires that differential network features might be considered together with a variety of modular structures.

In order to control the uncertainty level, data integration is adopted in many systems biology applications; for instance, in PIN applications the use of gene coexpression and pathway information sources can complement the observed interactions. Another common strategy involves the analysis of topological properties [8,9]. Further refinement of interactome data can rely on similarity (dissimilarity) measures to allow for comparative analysis of PIN, and for the assignment of confidence levels to each interaction depending on biological and computational aspects.

4.1.2 Outline

The structure of this paper is as follows. We describe in Section 4.2 our methodological approach in both general and particular aspects related to the PIN setting; then, we present our results in Section 4.3; and finally, concluding remarks with a discussion follows in Section 4.4.