Searching for sequence features that control DNA flexibility

Abstract

Modern genomics experiments measure functional behaviors for many thousands of DNA sequences. Using correlation functions between sequences and measured behaviors, we developed a simple physical model for interpreting such experimental outputs. Analysis of recent high throughput data on DNA mechanics shows that this is highly effective, leading directly to the extraction of distinct features for DNA flexibility and predictions comparable to more complex machine learning models. Our approach follows the conventional use of correlation functions in statistical physics and connects the search for relevant sequence features to the search for relevant stimulus features of sensory neurons.