BERT-TFBS: a novel BERT-based model for predicting transcription factor binding sites by transfer learning.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Transcription factors (TFs) regulate gene transcription by binding to specific DNA sequences (TFBSs).
• Accurate TFBS prediction is crucial for understanding gene regulation and designing synthetic biological systems.
• Current deep learning models for TFBS prediction require performance enhancements.

Purpose of the Study:

• To develop an advanced deep learning model, BERT-TFBS, for accurate TFBS prediction using only DNA sequences.
• To leverage transfer learning and attention mechanisms for improved feature extraction in TFBS prediction.

Main Methods:

• The BERT-TFBS model integrates a pre-trained BERT module (DNABERT-2) for long-term dependency learning.
• Convolutional Neural Network (CNN) and Convolutional Block Attention Module (CBAM) are employed for high-order local feature extraction.
• The model was trained and validated on 165 ENCODE ChIP-seq datasets.

Main Results:

• BERT-TFBS demonstrated superior performance in predicting TFBSs compared to existing deep learning models.
• Experimental results confirmed the model's effectiveness and generalization capabilities across different datasets.
• Cross-cell-line validation further supported the robustness of the BERT-TFBS model.

Conclusions:

• BERT-TFBS represents a significant advancement in TFBS prediction accuracy and efficiency.
• The model's architecture effectively captures complex sequence dependencies and local features.
• The proposed method offers a powerful tool for genomic research and synthetic biology applications.