TransSE: A Transfer Learning-Based Predictive Model for Distinguishing Super Enhancers and Typical Enhancers.

Area of Science:

• Genomics and Bioinformatics
• Computational Biology
• Molecular Biology

Background:

• Super-enhancers (SEs) are crucial for gene expression and cell fate.
• Existing computational methods for SE identification lack accuracy and cross-species generalizability.

Purpose of the Study:

• To develop TransSE, a deep learning framework for accurate SE prediction.
• To improve cross-species generalizability in SE identification.
• To provide a tool for investigating regulatory elements and disease-associated variants.

Main Methods:

• TransSE combines convolutional and recurrent neural networks with cross-species transfer learning.
• A two-phase strategy involves pre-training on human/mouse data and species-specific fine-tuning.
• Model performance was evaluated against existing methods using AUC metrics on human and mouse datasets.

Main Results:

• TransSE achieved superior accuracy (AUC 0.828 human, 0.832 mouse) compared to SENet, DeepSE, DNABERT, and Enformer.
• Ablation studies confirmed the importance of convolutional blocks, transfer learning, and recurrent layers.
• The model demonstrated effective cross-species prediction (AUC > 0.79) and distinguished SEs from typical enhancers.

Conclusions:

• TransSE offers a robust and accurate method for SE prediction with improved cross-species capabilities.
• The framework integrates conserved and species-specific regulatory features for enhanced performance.
• TransSE provides a user-friendly web interface for automated SE prediction, motif analysis, and variant impact assessment, aiding disease-associated variation studies.