Beyond position weight matrices: nucleotide correlations in transcription factor binding sites and their description

Abstract

The identification of transcription factor binding sites (TFBSs) on genomic DNA is of crucial importance for understanding and predicting regulatory elements in gene networks. TFBS motifs are commonly described by Position Weight Matrices (PWMs), in which each DNA base pair independently contributes to the transcription factor (TF) binding, despite mounting evidence of interdependence between base pairs positions. The recent availability of genome-wide data on TF-bound DNA regions offers the possibility to revisit this question in detail for TF binding {\em in vivo}. Here, we use available fly and mouse ChIPseq data, and show that the independent model generally does not reproduce the observed statistics of TFBS, generalizing previous observations. We further show that TFBS description and predictability can be systematically improved by taking into account pairwise correlations in the TFBS via the principle of maximum entropy. The resulting pairwise interaction model is formally equivalent to the disordered Potts models of statistical mechanics and it generalizes previous approaches to interdependent positions. Its structure allows for co-variation of two or more base pairs, as well as secondary motifs. Although models consisting of mixtures of PWMs also have this last feature, we show that pairwise interaction models outperform them. The significant pairwise interactions are found to be sparse and found dominantly between consecutive base pairs. Finally, the use of a pairwise interaction model for the identification of TFBSs is shown to give significantly different predictions than a model based on independent positions.

References in corpus (6)

• Identification of direct residue contacts in protein-protein interaction by message passing
• Searching for collective behavior in a network of real neurons
• Social interactions dominate speed control in driving natural flocks toward criticality
• From principal component to direct coupling analysis of coevolution in proteins: Low-eigenvalue modes are needed for structure prediction
• Dynamical maximum entropy approach to flocking
• Elements of Coevolution in Biological Sequences

Cited by in corpus (5)

• Inverse statistical problems: from the inverse Ising problem to data science
• Learning Maximum Entropy Models from finite size datasets: a fast Data-Driven algorithm allows sampling from the posterior distribution
• Random versus maximum entropy models of neural population activity
• A generalization of the maximum entropy principle for curved statistical manifolds
• Binding of transcription factors adapts to resolve information-energy trade-off