Using an evolutionary framework to study robustness and predictive power of data-driven molecular regulatory networks
Living systems constantly process environmental signals and respond to changing conditions. Central to this information processing are transcriptional regulatory networks that control what genes are expressed and when. A fundamental question in understanding living systems is how regulatory networks drive the overall ability of organisms to respond and adapt to different environmental conditions. Two major directions of research have been undertaken towards addressing this question. The first direction focuses on what the networks are in a cell. Because most regulatory networks are not known, the ability to accurately reconstruct these networks is one of the major challenges in systems biology. The second direction asks what properties make networks robust and adaptable to external perturbations. Although both directions of research have contributed to our understanding of regulatory networks, we lack an integrated picture of the relationship between regulatory network changes to overall organism state. This is largely because these directions of research have been pursued independently. In particular, a large number of data-driven “network reconstruction” methods have been developed that aim to infer the unknown regulatory network from genome-wide gene expression measurements. Unfortunately, the networks inferred from these methods typically do not correspond to physical regulatory interactions between regulatory proteins and target genes. On the other hand, most studies of network robustness have used physical regulatory networks available in a model organism (e.g. yeast) or have used simulated networks. Because inferred networks have not been rigorously examined for robustness and evolvability, we currently do not know whether the networks that we learn from observational data do exhibit such properties or whether the inferred networks only reflect statistical correlations that do not have any mechanistic underpinnings in the cell. In particular, it is possible that there are many “equally likely or good” networks that could perform the same function and the expression data are inferring a different network than what is being measured using other experiments.
To understand how regulatory networks contribute to and determine overall organism function we turn to evolution. Evolution is the ultimate tinkerer of living systems. Therefore, examining how regulatory networks have evolved to their current day configurations in extant species and predicting changes in phenotypes from changes to network components, we can obtain a more complete understanding of how networks drive complex functions. However, this requires us to first, accurately infer regulatory networks and second, to examine how changes in regulatory networks drive overall organismal state change. We will use a probabilistic graphical model representation of regulatory networks to tackle these questions. We will develop methods to learn data-driven regulatory networks that exhibit topological properties of robustness. Our approach will impose structural constraints on the graph structure that have been associated with robust networks. Next we will develop a probabilistic framework of network evolution to detect selection on network components (e.g. network motifs). Our evolutionary approach will be developed and tested in collaboration to study three biological processes: stress response in yeast, tissue-specific expression in vertebrate species, and plant-microbe symbiotic interactions.