Three-dimensional Epigenome Statistical Model: Genome-wide Chromatin Looping Prediction

Abstract

This study aims to understand through statistical learning the basic biophysical mechanisms behind three-dimensional folding of epigenomes. The 3DEpiLoop algorithm predicts three-dimensional chromatin looping interactions within topologically associating domains (TADs) from one-dimensional epigenomics and transcription factor profiles using the statistical learning. The predictions obtained by 3DEpiLoop are highly consistent with the reported experimental interactions. The complex signatures of epigenomic and transcription factors within the physically interacting chromatin regions (anchors) are similar across all genomic scales: genomic domains, chromosomal territories, cell types, and different individuals. We report the most important epigenetic and transcription factor features used for interaction identification either shared, or unique for each of sixteen (16) cell lines. The analysis shows that CTCF interaction anchors are enriched by transcription factors yet deficient in histone modifications, while the opposite is true in the case of RNAP II mediated interactions. The code is available at the repository https://bitbucket.org/4dnucleome/3depiloop.

preciseTAD: A transfer learning framework for 3D domain boundary prediction at base-pair resolution

“…We further demonstrated that class imbalance hinders boundary prediction, but can be effectively addressed by a simple random under-sampling technique, an aspect of boundary prediction unaddressed in previous studies [42][43][44]. We showed that information about only four transcription factors (CTCF, SMC3, RAD21, ZNF143) is necessary and sufficient for accurate boundary prediction, outperforming histone modification-and BroadHMMbuilt models [43,44]. These are known components of the loop extrusion model, an established theory of how loops are made by a ring-shaped adenosine triphosphatase-driven complex called cohesin [18,[20][21][22][23][24].…”

“…Our machine learning framework yielded several interesting observations. We first demonstrated that RF models built using distance-type predictors outperformed models built on previously published feature engineering techniques, including signal strength, overlap counts, and overlap percents [42][43][44][45]. We further demonstrated that class imbalance hinders boundary prediction, but can be effectively addressed by a simple random under-sampling technique, an aspect of boundary prediction unaddressed in previous studies [42][43][44].…”

“…We first demonstrated that RF models built using distance-type predictors outperformed models built on previously published feature engineering techniques, including signal strength, overlap counts, and overlap percents [42][43][44][45]. We further demonstrated that class imbalance hinders boundary prediction, but can be effectively addressed by a simple random under-sampling technique, an aspect of boundary prediction unaddressed in previous studies [42][43][44]. We showed that information about only four transcription factors (CTCF, SMC3, RAD21, ZNF143) is necessary and sufficient for accurate boundary prediction, outperforming histone modification-and BroadHMMbuilt models [43,44].…”

A Comparative Study of Supervised Machine Learning Algorithms for the Prediction of Long-Range Chromatin Interactions

“…Therefore, in silico predictions that take advantage of the wealth of publicly available sequencing data emerges as a rational strategy to generate virtual chromatin interaction maps in new cell types for which experimental maps are still lacking. To date, several studies have been devoted to predict chromatin loops based on one-dimensional (1D) genomic information with accurate results [ 20 , 21 , 22 , 23 , 24 , 25 ]. In such works, authors have modeled loops using different designs and machine learning approaches.…”

Quantitative prediction of enhancer–promoter interactions

“…Several computational tools have been developed to address these challenges. Their task was to predict three-dimensional EP interactions, based on data on one-dimensional genetic and epigenetic marks (Fortin and Hansen 2015;Moore et al 2015;Chen et al 2016;Chiariello et al 2016;Whalen et al 2016;Zhu et al 2016;Di Pierro et al 2017;Al Bkhetan and Plewczynski 2018b;Buckle et al 2018;Kai et al 2018;Zeng et al 2018;Zhang et al 2018; Ibn-Salem and Andrade-Navarro 2019; Qi and Zhang 2019). All these tools fall into two categories: physical models and statistical approaches.…”