Multiview Deep Learning Framework for Precise Prediction of Transcription Factor Binding Sites.

Area of Science:

• Genomics and Molecular Biology
• Computational Biology and Bioinformatics
• Systems Biology

Background:

• Transcription factors (TFs) regulate gene expression by binding to specific DNA sites (TFBSs).
• Accurate TFBS prediction is vital for understanding gene regulation, disease mechanisms, and synthetic biology.
• Current machine learning methods face challenges in modeling DNA structure, long-range dependencies, and integrating diverse data.

Purpose of the Study:

• To develop an advanced deep learning framework for enhanced Transcription Factor Binding Prediction (TFBS prediction).
• To address limitations in existing models by incorporating multiple data views and DNA sequence properties.
• To improve the accuracy and interpretability of TFBS prediction models.

Main Methods:

• Proposed MDNet-TFP, a multiview deep learning model for TFBS prediction.
• Introduced a bidirectional reverse complement module (BiRC-Mamba) to capture DNA sequence properties.
• Developed a multiscale convolutional recurrent attention network (MCRAN) for feature extraction and data integration.

Main Results:

• MDNet-TFP achieved superior performance across 165 ChIP-seq datasets, with average ACC of 88.13%, ROC-AUC of 93.72%, and PR-AUC of 93.40%.
• The model demonstrated high performance across a broader set of 690 ChIP-seq datasets.
• Motif visualization confirmed that model attention aligns with known TFBS motifs, indicating biological relevance.

Conclusions:

• MDNet-TFP effectively addresses limitations in current TFBS prediction methods.
• The model offers enhanced accuracy, interpretability, and generalization capabilities for genomic data analysis.
• This work advances research in transcriptional regulation and has implications for biomedical applications.