Person:

Gunawardena, Jeremy

Cellular physiology is implemented by formidably complex biochemical systems with highly nonlinear dynamics, presenting a challenge for both experiment and theory. Time-scale separation has been one of the few theoretical methods for distilling general principles from such complexity. It has provided essential insights in areas such as enzyme kinetics, allosteric enzymes, G-protein coupled receptors, ion channels, gene regulation and post-translational modification. In each case, internal molecular complexity has been eliminated, leading to rational algebraic expressions among the remaining components. This has yielded familiar formulas such as those of Michaelis-Menten in enzyme kinetics, Monod-Wyman-Changeux in allostery and Ackers-Johnson-Shea in gene regulation. Here we show that these calculations are all instances of a single graph-theoretic framework. Despite the biochemical nonlinearity to which it is applied, this framework is entirely linear, yet requires no approximation. We show that elimination of internal complexity is feasible when the relevant graph is strongly connected. The framework provides a new methodology with the potential to subdue combinatorial explosion at the molecular level.

The DNA microstates that regulate transcription include sequence-specific transcription factors (TFs), coregulatory complexes, nucleosomes, histone modifications, DNA methylation, and parts of the three-dimensional architecture of genomes, which could create an enormous combinatorial complexity across the genome. However, many proteins and epigenetic marks are known to colocalize, suggesting that the information content encoded in these marks can be compressed. It has so far proved difficult to understand this compression in a systematic and quantitative manner. Here, we show that simple linear models can reliably predict the data generated by the ENCODE and Roadmap Epigenomics consortia. Further, we demonstrate that a small number of marks can predict all other marks with high average correlation across the genome, systematically revealing the substantial information compression that is present in different cell lines. We find that the linear models for activating marks are typically cell line-independent, while those for silencing marks are predominantly cell line-specific. Of particular note, a nuclear receptor corepressor, transducin beta-like 1 X-linked receptor 1 (TBLR1), was highly predictive of other marks in two hematopoietic cell lines. The methodology presented here shows how the potentially vast complexity of TFs, coregulators, and epigenetic marks at eukaryotic genes is highly redundant and that the information present can be compressed onto a much smaller subset of marks. These findings could be used to efficiently characterize cell lines and tissues based on a small number of diagnostic marks and suggest how the DNA microstates, which regulate the expression of individual genes, can be specified.