Modeling Chromatin States

Models of chromatin have evolved from a DNA scaffolding structure to a gene regulation platform with exquisite complexity. Traditionally divided in euchromatin and heterochromatin, accumulating exceptions imposed the need to update the classification. With the popularization of genome-wide technologies and bioinformatic methods, several approaches to divide the chromatin into functionally coherent units were proposed. Here we review the use of unsupervised learning methods to model chromatin states. Using hidden Markov models as an example, we highlight the scope, benefits, and limitations of those approaches. We discuss how systematic studies based on unsupervised learning have changed the vision of chromatin, which involves two major findings: (1) More than half of the genome is transcriptionally silent, and devoid of both chromatin proteins and histone marks. (2) Short DNA regions called highly occupied targets (HOTs) accumulate most of the chromatin factors and show the highest diversity of histone marks. Inherently static and descriptive, unsupervised learning methods revealed the big picture but did not provide an explanation for it. We argue that a systems view may be a better framework to understand how the chromatin state of a locus is set. In particular, we suggest that chromatin states should be interpreted as the attractors in a dynamical system.