MuSE: A deep learning model based on multi-feature fusion for super-enhancer prediction.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Accurate identification of Super-enhancers (SEs) is crucial for understanding gene regulation.
• Current bioinformatics methods for SE identification are limited by feature design.
• Developing novel computational approaches for robust SE prediction is essential.

Purpose of the Study:

• To propose MuSE (Multi-Feature Fusion for Super-Enhancer), a deep learning model for improved SE identification.
• To evaluate the efficacy of multi-feature fusion strategies for DNA sequence representation in SE prediction.
• To assess the impact of different feature encoding methods on SE prediction performance.

Main Methods:

• Developed MuSE, a deep learning model employing multi-feature fusion for SE prediction.
• Utilized one-hot encoding and DNA2Vec (k-mer based) for DNA sequence representation.
• Trained and validated the model on human and mouse species datasets.

Main Results:

• MuSE demonstrated improved F1 scores compared to baseline methods, with a maximum improvement exceeding 0.05 on mouse datasets.
• DNA2Vec-based k-mer representations were identified as the most impactful features for prediction.
• Removing species-specific features enhanced cross-species prediction performance, achieving an AUC of nearly 0.8.

Conclusions:

• MuSE offers an effective deep learning framework for Super-enhancer identification through multi-feature fusion.
• Integrating diverse sequence features, particularly k-mer based representations, significantly enhances prediction accuracy.
• The model's generalization ability can be improved by addressing species-specific feature impacts, paving the way for robust cross-species SE prediction.