BertSNR: an interpretable deep learning framework for single-nucleotide resolution identification of transcription...

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Transcription factors (TFs) regulate gene expression by binding to specific DNA sequences.
• Accurate identification of transcription factor binding sites (TFBSs) is essential for understanding gene regulation.
• Current computational methods for TFBS identification often lack high resolution and interpretability.

Purpose of the Study:

• To develop an interpretable deep learning framework for high-resolution TFBS identification.
• To improve the accuracy and interpretability of TFBS prediction methods.
• To facilitate the study of gene regulation in promoter regions.

Main Methods:

• Proposed BertSNR, an interpretable deep learning framework for TFBS identification at single-nucleotide resolution.
• Integrated sequence-level and token-level information using multi-task learning.
• Utilized pre-trained DNA language models for enhanced performance.
• Employed attentional weight visualization and motif analysis for model interpretability.

Main Results:

• BertSNR outperformed existing state-of-the-art methods in TFBS predictions.
• Demonstrated enhanced model interpretability through visualization and motif analysis.
• Discovered a relationship between attention weights and sequence motifs.
• Successfully identified TFBSs in promoter regions, aiding gene regulation studies.

Conclusions:

• BertSNR offers a powerful and interpretable solution for high-resolution TFBS identification.
• The framework advances computational biology by improving TFBS prediction accuracy.
• BertSNR facilitates deeper insights into the mechanisms of gene regulation.