deepTAD: an approach for identifying topologically associated domains based on convolutional neural network and...

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Topologically associated domains (TADs) are crucial for genome 3D organization and function.
• Accurate TAD detection is essential for linking genomic structure to function.
• Current methods struggle with Hi-C contact matrix complexities for precise TAD identification.

Purpose of the Study:

• To introduce deepTAD, a novel computational method for accurate TAD boundary detection.
• To leverage deep learning, specifically CNNs and transformer models, for improved feature extraction from Hi-C data.
• To enhance the understanding of genome organization by accurately identifying hierarchical TADs.

Main Methods:

• DeepTAD employs a Convolutional Neural Network (CNN) to extract features directly from Hi-C contact matrices.
• A transformer model analyzes boundary variation features to determine TAD boundaries.
• Wilcoxon rank-sum test and cosine similarity are used to refine boundary identification and assemble hierarchical TADs.

Main Results:

• DeepTAD identified TAD boundaries show significant enrichment of biological features like structural proteins and histone modifications.
• The method demonstrates strong performance in completeness and accuracy compared to existing TAD identification tools.
• Experimental validation confirms the biological relevance of deepTAD-identified TAD boundaries.

Conclusions:

• DeepTAD offers a robust and accurate method for identifying TADs and hierarchical TADs.
• The approach effectively overcomes limitations of previous methods in analyzing Hi-C data complexities.
• DeepTAD provides valuable insights into genome 3D organization and function.