Identifying transcription factor-DNA interactions using machine learning

Abstract

Machine learning approaches have been applied to identify transcription factor (TF)-DNA interaction important for gene regulation and expression. However, due to the enormous search space of the genome, it is challenging to build models capable of surveying entire reference genomes, especially in species where models were not trained. In this study, we surveyed a variety of methods for classification of epigenomics data in an attempt to improve the detection for 12 members of the Auxin Response Factor (ARF) binding DNAs from maize and soybean as assessed by DNA Affinity Purification and sequencing (DAP-seq). We used the classification for prediction by minimizing the genome search space by only surveying unmethylated regions (UMRs). For identification of DAP-seq binding events within the UMRs, we achieved 93.54% accuracy, 6.2% false positive, and a 43.29% false negative rate across 12 members of ARFs of maize on average by encoding DNA with count vectorization for k-mer with a logistic regression classifier with up-sampling and feature selection. Importantly, feature selection helps to uncover known and potentially novel ARF binding motifs. This demonstrates an independent method for identification of transcription factor binding sites. Finally, we tested the model built with maize DAP-seq data and applied it directly to the soybean genome and found unacceptably high false positive rates, which accounted for more than 40% across the ARF TFs tested. The findings in this study suggest the potential use of various methods to predict TF-DNA interactions within and between species with varying degrees of success.