Unsupervised Methods to Identify Cellular Signaling Networks from Perturbation Data

Unsupervised Methods to Identify Cellular Signaling Networks from Perturbation Data

Madhusudan Natarajan (Pfizer, USA)
Copyright: © 2013 |Pages: 18
DOI: 10.4018/978-1-4666-3604-0.ch030
OnDemand PDF Download:
No Current Special Offers


The inference of cellular architectures from detailed time-series measurements of intracellular variables is an active area of research. High throughput measurements of responses to cellular perturbations are usually analyzed using a variety of machine learning methods that typically only work within one type of measurement. Here, summaries of some recent research attempts are presented–these studies have expanded the scope of the problem by systematically integrating measurements across multiple layers of regulation including second messengers, protein phosphorylation markers, transcript levels, and functional phenotypes into signaling vectors or signatures of signal transduction. Data analyses through simple unsupervised methods provide rich insight into the biology of the underlying network, and in some cases reconstruction of key architectures of the underlying network from perturbation data. The methodological advantages provided by these efforts are examined using data from a publicly available database of responses to systematic perturbations of cellular signaling networks generated by the Alliance for Cellular Signaling (AfCS).
Chapter Preview


How does cellular machinery function? What is the network of molecular entities that governs equally well resting basal homeostasis, as well as specific and precise response to cellular stimuli that drive a diverse range of cellular function? The dissection of cellular architectures and accompanying function has followed those precise routes – to first identify molecular entities, followed by a reductionist approach to place it in a simplified physiological context. With the advent of high-throughput experimentation and rapid technological advances, focused biochemical experiments are being replaced with high content experiments. These experiments permit the sampling of large numbers of molecular entities with sufficient resolution and accuracy to develop a comprehensive parts list. However, quantitative descriptions of both (relative) amounts and activity levels of a large number of molecular entities within cells pose novel problems. First, identification of multiple causal relationships between vast numbers of measured molecular entities is needed to define the architecture of the interactions between molecules. Second, the relative contribution of a molecule in the context of a network of interactions is needed to delineate functional consequences.

The inferencing of cellular architectures and functions from biological measurements has rich precedents in engineering - especially in systems identification in the fields of systems engineering and control systems, but mostly in non-biological areas (Ljung, 1999). The goal in systems identification is to be able to build dynamical models from measured data. Therefore, the characterization of cellular networks using mathematical methods is an extension of systems identification theories in that it is simply an attempt to formalize and summarize the system’s resting behavior and its response to perturbation. In an ideal scenario, the cellular architecture inferred from the data is isomorphic (i.e., structurally identical) and is identifiable – given sufficient observations from the system, it is possible to uniquely infer the parameters of the model producing the data. In reality, these are almost impossible to achieve – first, cellular architectures are vastly more complex than engineered systems such that even if resource limitations are not considered, sufficient observations may not be available to satisfy the identifiability criterion. Second, while we have made remarkable progress in the inference of biological networks and functions from data, the algorithmic methods are still far from perfect. The availability of additional data is currently not a bottleneck. Thus, state of the art is still focused on developing better methods to deal with existing data sets.

One classical method of exposition of biological function is through perturbation analysis. Classical biological perturbation studies have typically attempted to make incremental change to biological systems through tools such as pharmacological agents, environmental stressors, or more recently anti-sense technologies and measure functional performance after each change. With high-throughput data available to monitor responses to each such perturbation, studies have shown that responses to perturbation are directly amenable to mapping network topology and can yield significant insight into network architecture (Tegner, Yeung, Hasty, & Collins, 2003; Yeung, Tegner, & Collins, 2002).

Complete Chapter List

Search this Book: