Diff-SE: A Diffusion-Augmented Contrastive Learning Framework for Super-Enhancer Prediction

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Super-enhancers (SEs) are critical cis-regulatory elements controlling gene expression.
• SEs are implicated in diseases like cancer and Alzheimer's.
• Current identification methods (ChIP-seq) are resource-intensive, and computational methods struggle with data imbalance and generalization.

Purpose of the Study:

• To develop an advanced computational framework for accurate and robust super-enhancer prediction.
• To overcome limitations of existing methods, including class imbalance and poor cross-species performance.

Main Methods:

• Proposed Diff-SE, a deep learning framework integrating diffusion-based data augmentation and contrastive learning.
• Diffusion module generates synthetic SE data to balance training sets.
• Contrastive learning enhances feature representations for improved discrimination.

Main Results:

• Diff-SE achieved 10%-30% improvement in precision, MCC, and F1-score across eight datasets compared to baseline models.
• Demonstrated superior generalization capabilities in cross-species validation (human and mouse).

Conclusions:

• Diff-SE offers a significant advancement in computational super-enhancer prediction.
• The framework provides a more accurate, generalizable, and efficient alternative to traditional methods.
• Available code and data facilitate further research in SEs and related diseases.