Sequence-Only Prediction of Super-Enhancers in Human Cell Lines Using Transformer Models

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Super-enhancers (SEs) are crucial for gene regulation but challenging to predict.
• Previous methods often rely on epigenetic markers, limiting their application.
• Understanding SEs is vital for cancer research and gene expression studies.

Purpose of the Study:

• To develop and evaluate a transformer-based deep learning model for predicting super-enhancers using sequence-only features.
• To compare the model's performance against existing methods like SENet.
• To investigate the correlation between sequence features and epigenetic landscapes in SEs.

Main Methods:

• Utilized the GENA-LM transformer model for classifying super-enhancers versus enhancers.
• Focused on sequence-only features from human genomic DNA.
• Trained and tested the model on diverse human cell line datasets (HeLa, HEK293, K562, etc.) using H3K36me, H3K4me1, H3K4me3, and H3K27ac data for validation.
• Fine-tuned the model on relevant sequence data for analyzing long genomic sequences.

Main Results:

• The proposed SE-prediction method achieved high balanced accuracy, outperforming SENet, especially in HEK293 and K562 cell lines.
• Demonstrated that super-enhancers frequently co-localize with epigenetic marks (H3K4me3, H3K27ac).
• The model's attention mechanism revealed correlations between sequence-only features and epigenetic landscapes, providing insights into SE classification.

Conclusions:

• Transformer models, like GENA-LM, are effective for super-enhancer prediction using sequence data alone.
• This approach offers a powerful alternative to methods requiring epigenetic markers.
• The findings support the use of transformer models in genomic sequence analysis for enhancer characterization and understanding gene regulation.