Exploiting epigenomic and sequence-based features for predicting enhancer-promoter interactions

Jian‐Guo Zhou, Renyang Liu, Zifeng Wu, Jintao Zhang, Junhui LiuE3S Web of Conferences 2020 · DOI

Abstract

Identifying distal regulatory elements and their target genes is a fundamental challenge in understanding gene regulation and the etiology of complex diseases. Enhancers are distal elements that interact with target gene promoters to regulate expression. While various machine learning approaches have been developed to predict these enhancer-promoter interactions (EPIs), achieving global and precise genomic-level prediction remains an ongoing area of research. This paper introduces EpPredictor, an integrated prediction method that leverages the LightGBM algorithm. By combining histone modifications, transcription factor binding sites, and DNA sequence motifs, EpPredictor demonstrates improved performance across six human cell lines, achieving superior F1-scores and AUC compared to established methods like TargetFinder and PEP.

Introduction

In the human genome, enhancers act as distal "volume knobs," often located far from the promoters they regulate. Mapping these enhancer-promoter interactions (EPIs) is crucial for deciphering the regulatory landscape. However, the complexity of these interactions—where one gene may be controlled by multiple enhancers across vast distances—makes prediction difficult. Existing models often lack precision or fail to integrate the full spectrum of available genomic and epigenomic data. EpPredictor addresses these gaps by utilizing a high-performance gradient boosting framework and a comprehensive feature set.

Methods

EpPredictor utilizes LightGBM (Light Gradient Boosting Machine), chosen for its computational efficiency and predictive accuracy. The methodology centers on three key pillars:

1. Region Definition

Unlike models that focus solely on the enhancer or promoter, EpPredictor analyzes four distinct genomic areas: the Enhancer Region (ER), the Promoter 2k Region (PR), the Window Region (WR) between them, and the Full Region (FR), which encompasses the entire span from the enhancer start to the promoter.

Schematic of the genomic regions (Enhancer, Window, Promoter, and Full Region) and associated ChIP-seq data tracks used for feature extraction.

2. Feature Fusion

The model integrates two data types:

  • Epigenomic Features: Signal values and peaks from histone modifications and transcription factor binding sites (ChIP-seq data).
  • Sequence Features: DNA motifs identified via Position Weight Matrices (PWM), specifically using 20-base-pair sequence segments for optimal motif detection.

3. Data Balancing

To handle the natural imbalance where non-interactions far outnumber true EPIs, the model was trained using a balanced 1:1 ratio of positive to negative samples.

Key Insights and Results

The study evaluated EpPredictor across six cell lines: GM12878, K562, IMR90, HeLa-S3, HUVEC, and NHEK.

  • The Power of the Full Region: The research found that the Full Region (FR) provides the most robust predictive power when sequence and epigenomic data are combined, challenging previous assumptions that the window region alone was the most informative.
  • Sequence-Epigenomic Synergy: While DNA sequence data is a poor predictor in isolation, its inclusion alongside epigenomic tags significantly boosts performance, suggesting a functional tandem between the physical DNA code and chemical modifications.
  • Superior Performance: EpPredictor achieved an average F1-score of 0.88, representing a 5% improvement over TargetFinder and PEP. In the HUVEC cell line, the improvement reached 10-12%. The model also demonstrated high reliability, with AUC scores reaching 0.98 in the NHEK cell line.
Performance comparison of F1, AUC, and MCC metrics across GM12878 and HeLa S3 cell lines, illustrating the competitive advantage of the integrated approach.

Limitations and Future Work

While EpPredictor excels in F1 and AUC metrics, the authors noted discrepancies in the Matthews Correlation Coefficient (MCC) for specific cell lines like IMR90, where it trailed slightly behind TargetFinder. This highlights that no single metric can fully capture classifier performance across all biological contexts. Future research will focus on refining feature selection and investigating cell-line-specific responses to sequence-epigenomic integration.

Summary updated

Abstract

How to discriminate distal regulatory elements to a gene target is challenging in understanding gene regulation and illustrating causes of complex diseases. Among known distal regulatory elements, enhancers interact with a target gene’s promoter to regulate its expression. Although the emergence of many machine learning approaches has been able to predict enhancer-promoter interactions (EPIs), global and precise prediction of EPIs at the genomic level still requires further exploration.In this paper, we develop an integrated EPIs prediction method, called EpPredictor with improved performance. By using various features of histone modifications, transcription factor binding sites, and DNA sequences among the human genome, a robust supervised machine learning algorithm, named LightGBM, is introduced to predict enhancer-promoter interactions (EPIs). Among six different cell lines, our method effectively predicts the enhancer-promoter interactions (EPIs) and achieves better performance in F1-score and AUC compared to other methods, such as TargetFinder and PEP.
BibTeX
@inproceedings{zhou2020exploiting,
  author = {Jian‐Guo Zhou and Renyang Liu and Zifeng Wu and Jintao Zhang and Junhui Liu},
  title = {{Exploiting epigenomic and sequence-based features for predicting enhancer-promoter interactions}},
  booktitle = {E3S Web of Conferences},
  year = {2020},
  doi = {10.1051/e3sconf/202021803046},
}
Figure