Benchmarking and refining probability-based models for nucleosome-DNA interaction

Abstract

Background: In investigations of nucleosome positioning preferences, a model that assigns an affinity to a given sequence is necessary to make predictions. One important class of models, which treats a nucleosome sequence as a Markov chain, has been applied with success when informed with experimentally measured nucleosomal sequence preferences. Results: We find that we can also use such models as a fast approximative scheme for computationally expensive biophysical models, vastly increasing their reach. Employing these models in this way also allows us to benchmark them for the first time. Doing so for the approximative in silico models indirectly tells us about the accuracy we can expect of them when applied to real data. Conclusion: We find that models presented in the literature should perform well, but this performance depends on factors such as the order of the Markov model, the preprocessing of the probability distributions on which the model is based, and the size and quality of the sequence ensemble from which those distributions are calculated.