In Silico prediction of transcription factor binding sites by probabilistic models

The characterization of in silico detected transcription factor binding sites represents a fundamental problem in the field of regulatory gene expression analysis. Several approaches have been proposed to model DNA-protein-interactions, composed by two main classes: qualitative models considering a consensus sequence and quantitative models providing a measure of binding affinity. The latter can be further subdivided in models assuming an independent contribution of the nucleotides forming a potential binding site and more flexible ones implicating a positional interdependence.

In this work the applicability of three probabilistic models to predict transcription factor binding sites has been investigated: (i) the simple position weight matrix (PWM), assuming independence, and two flexible models capturing positional interdependencies represented by a (ii) Chow-Liu Tree and (iii) Ensemble of Trees model. The training and validation of the models on the Mus musculus subset of the UniPROBE database revealed that complex models provide a better predictive power suggesting a high amount of transcription factors binding motifs being affected by positional interdependencies. Additionally, numerous transcription factors were detected, for which the Ensemble of Trees model outperformed both, the Chow-Liu Tree and PWM model.

The UniPROBE-based trained models have been applied in a biological context - the prediction of differential binding profiles in five different ChIP-seq samples, followed by the detection of causative regulatory SNPs. The chosen set-up involved family trio data, meaning genotype data from a family composed of father, mother and daughter, providing internal validation. The models provide strong power to correctly classify true negatives in an independent biological sample, represented by a high specificity. The applied approach to detect causative regulatory SNPs, resulted in a candidate list of 20 SNPs. Those gain strong support by epigenetic markers and both, model-based predicted binding affinity of the comprising binding site and significant p-values, describing the effect of the nucleotide exchange.