DEEPSEN: a convolutional neural network based method for super-enhancer prediction

Hongda Bu, Jiaqi Hao, Yanglan Gan, Shuigeng Zhou, Jihong GuanBMC Bioinformatics 2019 · OpenReview

Abstract

Super-enhancers (SEs) are clusters of transcriptional active enhancers, which dictate the expression of genes defining cell identity and play an important role in the development and progression of tumors and other diseases. Many key cancer oncogenes are driven by super-enhancers, and the mutations associated with common diseases such as Alzheimer’s disease are significantly enriched with super-enhancers. Super-enhancers have shown great potential for the identification of key oncogenes and the discovery of disease-associated mutational sites. In this paper, we propose a new computational method called DEEPSEN for predicting super-enhancers based on convolutional neural network. The proposed method integrates 36 kinds of features. Compared with existing approaches, our method performs better and can be used for genome-wide prediction of super-enhancers. Besides, we screen important features for predicting super-enhancers. Convolutional neural network is effective in boosting the performance of super-enhancer prediction.

Introduction

In the landscape of our genome, enhancers are short regions of DNA that act like volume knobs for gene expression. Super-enhancers (SEs) are essentially "power strips" of these knobs—dense clusters of enhancers that drive the expression of genes defining a cell’s identity. Because SEs are often hijacked by cancer cells to drive oncogenes, identifying them is crucial for precision medicine.

Traditionally, SEs are identified using biological assays like ChIP-seq to find high concentrations of specific proteins (like Mediator complex subunit Med1). However, these experiments are expensive and time-consuming. While previous researchers tried using standard machine learning (like Random Forests or SVMs), these models struggled to capture the complex, non-linear relationships between different chromatin marks and transcription factors. DEEPSEN addresses this by using Convolutional Neural Networks (CNNs) to integrate 36 different biological features, providing a more accurate and automated way to identify these genomic "master switches."

Methods

The DEEPSEN framework follows a structured pipeline from data acquisition to performance evaluation.

Workflow of the DEEPSEN framework, from data preprocessing to performance evaluation.

The authors introduce a CNN-based architecture designed to handle the multi-dimensional nature of genomic data. They tested three versions (2, 3, and 4 convolutional layers), with the final model, DEEPSEN-4L, utilizing a series of convolution and max-pooling layers to extract hierarchical patterns.

Detailed architecture of the DEEPSEN convolutional neural network.
  • Feature Integration: DEEPSEN integrates 36 distinct features, including DNA sequence composition, histone modifications (e.g., H3K27ac, H3K4me1), transcription factor binding sites (e.g., Oct4, Sox2, Nanog), and chromatin regulators (e.g., p300, Brd4).
  • Optimization: The model uses the Adam optimizer and a cross-entropy loss function to refine its predictions across 140–150 training epochs. Dropout is employed in the fully connected layers to prevent overfitting.

Datasets

The study utilized 32 publicly available ChIP-seq and DNase-seq datasets from mouse embryonic stem cells (mESCs), sourced from the Gene Expression Omnibus (GEO). The training set consisted of 11,100 total samples. Reflecting the actual rarity of SEs in the genome, the dataset was highly imbalanced, containing 1,119 positive samples (super-enhancers) and 9,981 negative samples (normal enhancers).

Results and Discussion

DEEPSEN was benchmarked against Improse, the previous state-of-the-art machine learning method for SE prediction. DEEPSEN achieved a precision of 0.92 (vs. 0.88 for Improse) and a recall of 0.88 (vs. 0.81). The overall F1-score was 0.90 for DEEPSEN compared to 0.84 for the baseline. While both models achieved an AUC of 0.97, DEEPSEN provided more reliable classifications at specific thresholds.

The core advantage of deep learning in this context is its ability to perform "feature fusion." While traditional models might look at features in isolation, CNN filters recognize how specific histone marks combined with transcription factor binding create the unique signature of a super-enhancer. Through Pearson correlation analysis, the authors found that Med12, Cdk8, and Brd4 were the most significant predictors. Interestingly, H3K27ac, the "gold standard" for manual identification, ranked 6th, suggesting that other proteins are more reliable indicators for computational models.

Conclusion

While DEEPSEN represents a significant advancement, it currently requires 36 input features, which may be unrealistic in many clinical settings. Future work will focus on "feature screening" to identify a smaller, more economical subset of features that maintain high predictive power. Additionally, while validated on mouse embryonic stem cells, the model's generalizability to human tissues and specific cancer types remains a key area for future investigation.

Summary updated

Abstract

Super-enhancers (SEs) are clusters of transcriptional active enhancers, which dictate the expression of genes defining cell identity and play an important role in the development and progression of tumors and other diseases. Many key cancer oncogenes are driven by super-enhancers, and the mutations associated with common diseases such as Alzheimer’s disease are significantly enriched with super-enhancers. Super-enhancers have shown great potential for the identification of key oncogenes and the discovery of disease-associated mutational sites. In this paper, we propose a new computational method called DEEPSEN for predicting super-enhancers based on convolutional neural network. The proposed method integrates 36 kinds of features. Compared with existing approaches, our method performs better and can be used for genome-wide prediction of super-enhancers. Besides, we screen important features for predicting super-enhancers. Convolutional neural network is effective in boosting the performance of super-enhancer prediction.
BibTeX
@article{bu2019deepsen,
  author = {Hongda Bu and Jiaqi Hao and Yanglan Gan and Shuigeng Zhou and Jihong Guan},
  title = {{DEEPSEN: a convolutional neural network based method for super-enhancer prediction}},
  journal = {BMC Bioinformatics},
  year = {2019},
  url = {https://openreview.net/forum?id=L6AeFcqhj6},
}
Figure