High-resolution transcription factor binding sites prediction improved performance and interpretability by deep...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Transcription factors (TFs) regulate gene expression by binding to specific DNA sequences (TFBSs).
• Accurate TFBS prediction is vital for understanding gene regulation and advancing precision medicine.
• Current deep learning models for TFBS prediction have limitations in integrating positional and semantic information and lack interpretability.

Purpose of the Study:

• To develop a novel deep learning model, D-AEDNet, for high-resolution prediction of TFBSs.
• To improve the interpretability of deep learning models in TFBS prediction.
• To introduce a method, TF-MoDSW, for discovering TF-DNA binding motifs from predicted TFBSs.

Main Methods:

• Developed the Deep Attentive Encoder-Decoder Neural Network (D-AEDNet) incorporating a Skip Architecture and Attention Gate.
• Utilized Attention Gate to leverage nucleotide position information and reduce noisy responses.
• Proposed Transcription Factor Motif Discovery based on Sliding Window (TF-MoDSW) for motif analysis.

Main Results:

• D-AEDNet demonstrated superior performance compared to existing methods on ChIP-exo datasets.
• Attention Gate visualization confirmed improved model interpretability.
• D-AEDNet effectively learned TFs-DNA binding motifs, outperforming state-of-the-art methods on ChIP-seq datasets.

Conclusions:

• D-AEDNet offers enhanced accuracy and interpretability for TFBS prediction.
• The TF-MoDSW method facilitates biological interpretation of TFBS predictions.
• These advancements contribute to a better understanding of TFs-DNA interactions and gene regulation.