Dependencies within CTCF binding sites

by Ivo Grosse, 
Institute of Computer Science, 
Martin Luther University,
Halle, Germany

The identification of DNA binding sites has been a challenge since the
early days of computational biology, and its importance has been
increasing with the development of new experimental techniques and the
ensuing flood of large-scale genomics and epigenomics data yielding
approximate regions of binding.  The question to which extent
dependencies within binding sites exist and how much they help in the
computational identification of DNA binding sites has been debated
intensively, in particular in the context of analyzing in-vitro data
from universal protein-binding microarrays.  Here, we address this
question based on in-vivo data from different ChIP-seq experiments of
the human enhancer-blocking insulator protein CTCF, and we find that
the sensitivity of de-novo motif discovery increases from 72% to 84%
by taking into account dependencies within CTCF binding sites of up to
order four.  We find that CTCF binding sites are poorly represented by
a position weight matrix model, which neglects dependencies within
binding sites, and that these dependencies are particularly strong in
the unconserved region at the 3'end of the CTCF motif.