Computational methods for identifying enhancer-promoter interactions

Haiyan Gong, Zhengyuan Chen, Yuxin Tang, Minghong Li, Sichen Zhang, Xiaotong Zhang, Yang ChenQuant. Biol 2023 · DOI

TL;DR

This review provides a comprehensive overview of computational methods for identifying enhancer-promoter interactions (EPIs), a crucial mechanism for gene regulation. It categorizes existing approaches, highlights key data sources and prediction strategies (from traditional machine learning to deep learning), and discusses their applications in understanding diseases like cancer.

What Problem Does This Paper Address?

Understanding how enhancers, distant DNA elements, interact with promoters, located near genes, is fundamental to deciphering gene regulation. These interactions, known as enhancer-promoter interactions (EPIs), play a vital role in controlling gene expression, and their dysregulation is implicated in various diseases. While numerous computational methods have emerged for identifying enhancers, promoters, and their interactions, a systematic review that consolidates these approaches, discusses their underlying data sources, and guides researchers in selecting appropriate tools has been lacking. This paper aims to fill that gap by providing a structured overview of the field.

Methods

The paper categorizes EPI identification methods into two main types:

  • Screening EPIs from High-Throughput Sequencing Experiments: This involves using data from techniques like ChIP-seq, Hi-C, HiChIP, and ChIA-PET to directly identify interactions. Methods here often focus on calling chromatin loops and then annotating them as EPIs based on co-occurring epigenetic marks.
  • Identifying EPIs from DNA Sequences and Epigenomic Data using Machine Learning or Deep Learning: This approach leverages various data features, including DNA sequences, histone modifications, chromatin accessibility, and gene expression data.
  • Traditional Machine Learning: Methods like Support Vector Machines (SVMs), Random Forests, and Gradient Boosting are employed. These often involve feature extraction techniques (e.g., k-mers, pseudo-nucleotide compositions) followed by classification.
  • Deep Learning: This includes Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and Transformer-based models. Deep learning approaches are increasingly used to directly predict EPIs from DNA sequences, reducing reliance on complex epigenomic data and potentially speeding up training. Transfer learning is also highlighted as a promising technique to improve generalization across different cell types.

The paper also outlines a general framework for EPI identification, which involves processing DNA sequences, extracting features, generating training sets, and then applying prediction models.

Datasets

The paper extensively references and discusses the use of various public datasets for training and evaluating EPI identification methods. Key data types include:

  • Epigenomic data: ChIP-seq data for histone modifications (e.g., H3K4me1, H3K27ac, H3K4me3), chromatin accessibility data (e.g., DNase-seq, ATAC-seq), and chromosome conformation capture data (e.g., Hi-C, ChIA-PET, HiChIP).
  • Gene expression data: RNA-seq.
  • DNA sequence data: The raw DNA sequences themselves.
  • Databases for Enhancers and Promoters: The paper lists several public repositories for obtaining enhancer and promoter data, including SEDB, PReMod, VISTA, EPD, and others.
  • Benchmarking EPI Datasets: The paper mentions datasets used in benchmarking studies, such as those from GM12878, HUVEC, Hela-S3 cell lines, often derived from tools like TargetFinder.

Key Insights and Ideas

  • EPIs are fundamental to gene regulation: The paper strongly emphasizes that EPIs are not just incidental DNA contacts but are crucial for orchestrating gene transcription, with implications for cell differentiation and disease.
  • Data integration is key: While DNA sequence alone can be used, incorporating epigenomic features significantly improves the accuracy of EPI prediction.
  • Deep learning is advancing the field: Deep learning methods, particularly those leveraging transfer learning, are showing promise in directly predicting EPIs from DNA sequences, potentially overcoming limitations of traditional methods and reducing computational burden.
  • Tissue specificity is a major challenge: The inherent tissue-specificity of enhancers and EPIs makes it difficult to develop universally applicable prediction models, highlighting the need for cell-type-specific approaches or methods that can adapt.
  • The "why" behind EPIs is crucial: Beyond just identifying interactions, understanding how these interactions lead to specific gene expression outcomes, especially in disease contexts, is a critical area of research.

Key Results

The paper is a review and does not present novel experimental results. Instead, it synthesizes the findings of numerous studies. Key qualitative findings include:

  • Advancement in prediction accuracy: Over the past decade, computational methods have significantly improved in their ability to predict enhancers, promoters, and EPIs.
  • Dominance of ML/DL: Machine learning and deep learning methods are the predominant computational strategies for EPI identification.
  • Increasing reliance on deep learning: Deep learning models, especially those that can learn directly from DNA sequences, are becoming increasingly popular due to their potential for higher accuracy and reduced data requirements.
  • Practical guidance provided: The paper includes tables summarizing various prediction methods, their data sources, and performance metrics (implicitly through citation counts), offering practical guidance for researchers.

Limitations and Future Work

Limitations:

  • Data scarcity and validation: Despite advancements, there is still a lack of experimentally validated enhancer-promoter interaction datasets, which limits the training and evaluation of predictive models.
  • Species specificity: Enhancers are often species-specific, making it challenging to develop models that perform well across different species.
  • Spatial proximity in sequence encoding: Current sequence encoding methods (like one-hot encoding) may not adequately preserve the spatial proximity of DNA sequences, which is crucial for understanding interactions.
  • Complexity of epigenomic data: While epigenomic data improves accuracy, its acquisition can be costly, and its interpretation complex.

Future Work:

  • Novel sequence coding methods: Developing sequence coding methods that better capture spatial proximity and sequence features is urged.
  • Single-cell EPI analysis: Leveraging single-cell sequencing technologies to study EPIs at a single-cell level to address cellular heterogeneity.
  • Optimizing models for sparse data: Adapting EPI identification methods for the sparsity inherent in single-cell sequencing data.
  • Applications in disease research: Further exploring tumor-specific EPIs, the impact of mutations on EPIs, and the role of EPIs in gene expression regulation using advanced technologies like CRISPR.
  • Transfer learning for cross-cell line prediction: Prioritizing research on transfer learning to enable prediction of EPIs across different cell lines more effectively.

Summary updated

BibTeX
@article{gong2023computational,
  author = {Haiyan Gong and Zhengyuan Chen and Yuxin Tang and Minghong Li and Sichen Zhang and Xiaotong Zhang and Yang Chen},
  title = {{Computational methods for identifying enhancer-promoter interactions}},
  journal = {Quant. Biol.},
  year = {2023},
  doi = {10.15302/j-qb-022-0322},
}
Figure