BCDB: A dual-branch network based on transformer for predicting transcription factor binding sites

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Transcription factor binding sites (TFBSs) are crucial for gene expression regulation.
• Accurate TFBS prediction aids in understanding transcription factor mechanisms.
• Existing deep learning models for TFBS prediction struggle with limited data and lack transparency.

Purpose of the Study:

• To develop a robust and interpretable deep learning framework for TFBS prediction.
• To improve TFBS prediction accuracy, particularly under data-scarce conditions.
• To enhance the transparency of deep learning models in TFBS prediction.

Main Methods:

• Developed the BCDB framework integrating DNABERT, CNNs, and multi-head attention.
• Employed a dual-branch output strategy to balance global and local DNA information.
• Utilized transfer learning for cross-cell line TFBS prediction.

Main Results:

• BCDB significantly outperforms existing deep learning methods on 165 ChIP-seq datasets.
• The framework demonstrates improved accuracy and interpretability.
• Achieved successful cross-cell line TFBS prediction using transfer learning.

Conclusions:

• BCDB offers a superior approach to TFBS prediction, addressing limitations of current methods.
• The model's interpretability through attention mechanisms provides insights into predictions.
• BCDB's transfer learning capability enables effective predictions across different cell lines.