A Biophysical Model for Analysis of Transcription Factor Interaction and Binding Site Arrangement from Genome-Wide Binding Data

Abstract

BackgroundHow transcription factors (TFs) interact with cis-regulatory sequences and interact with each other is a fundamental, but not well understood, aspect of gene regulation.Methodology/Principal FindingsWe present a computational method to address this question, relying on the established biophysical principles. This method, STAP (sequence to affinity prediction), takes into account all combinations and configurations of strong and weak binding sites to analyze large scale transcription factor (TF)-DNA binding data to discover cooperative interactions among TFs, infer sequence rules of interaction and predict TF target genes in new conditions with no TF-DNA binding data. The distinctions between STAP and other statistical approaches for analyzing cis-regulatory sequences include the utility of physical principles and the treatment of the DNA binding data as quantitative representation of binding strengths. Applying this method to the ChIP-seq data of 12 TFs in mouse embryonic stem (ES) cells, we found that the strength of TF-DNA binding could be significantly modulated by cooperative interactions among TFs with adjacent binding sites. However, further analysis on five putatively interacting TF pairs suggests that such interactions may be relatively insensitive to the distance and orientation of binding sites. Testing a set of putative Nanog motifs, STAP showed that a novel Nanog motif could better explain the ChIP-seq data than previously published ones. We then experimentally tested and verified the new Nanog motif. A series of comparisons showed that STAP has more predictive power than several state-of-the-art methods for cis-regulatory sequence analysis. We took advantage of this power to study the evolution of TF-target relationship in Drosophila. By learning the TF-DNA interaction models from the ChIP-chip data of D. melanogaster (Mel) and applying them to the genome of D. pseudoobscura (Pse), we found that only about half of the sequences strongly bound by TFs in Mel have high binding affinities in Pse. We show that prediction of functional TF targets from ChIP-chip data can be improved by using the conservation of STAP predicted affinities as an additional filter.Conclusions/SignificanceSTAP is an effective method to analyze binding site arrangements, TF cooperativity, and TF target genes from genome-wide TF-DNA binding data.

Spatiotemporal clustering of the epigenome reveals rules of dynamic gene regulation

“…TET1's interacting partner MLL contains a zinc finger CXXC domain as well. Theoretically, the MLL-TET1-u CpGs three way interaction can be a lot more stable than a two-way interaction of a protein and its DNA recognition site (He et al 2009). This MLL-TET1-u CpGs interaction is reinforced by MLL's roles to methylate H3K4 and interact with methylated H3K4.…”

Comprehensive prediction in 78 human cell lines reveals rigidity and compactness of transcription factor dimers

“…Thus, it differs from several existing bioinformatics approaches that aim to identify fuzzily spaced cobinding of TF pairs, i.e., indirect cooperativity (Qian et al 2005;He et al 2009;Bais et al 2011;Myšičková and Vingron 2012). Recently, Whitington et al (2011) described a method that, similarly to ours, predicts TF-TF dimerization based on enrichment of rigidly spaced motif pairs.…”

Nucleosomes in gene regulation: Theoretical approaches

“…The recurrent relations method, which belongs to the class of dynamic pro gramming algorithms, proved the fastest in the case when nucleosome unwrapping is not considered. Therefore, it is currently used in many works on calcu lating the arrangement of nucleosomes and regulatory proteins in chromatin [20,35,36,[74][75][76][77][78]. Interest ingly, a dynamic programming method that considers the general case of interactions between proteins sep arated by several base pairs along the DNA was pro posed in 1978 by Gurskii and Zasedatelev [79].…”

“…On the other hand, today the main challenge is to calculate protein binding maps for large genomic regions, and the calculation time is becoming the bottleneck. The calculation time of the Zasedatelev-Gurskii approach increases linearly with the DNA length, which makes it probably the fastest algorithm for the given task [20,35,36,[74][75][76][77][78].…”