CCTAD: A topologically associating domains detection method integrating convolutional autoencoder and hierarchical...

More accurate identification of topologically associating domains (TADs) is crucial for understanding chromatin spatial organization and gene regulation. Although many computational methods have been developed for TAD detection, existing approaches still exhibit several key limitations, such as difficulty in simultaneously modeling local and global interaction patterns, reliance on manually tuned parameters that lead to unstable boundary granularity, and high sensitivity to resolution and cell type. These issues result in unstable boundaries and high sensitivity to noise, thereby affecting the overall biological interpretability of TAD structures.

To address these challenges, we propose a novel unsupervised TAD identification method, CCTAD, which for the first time integrates a one-dimensional convolutional autoencoder (1D-CAE) with connectivity-constrained hierarchical clustering. The 1D-CAE automatically extracts high-quality, low-dimensional feature representations from Hi-C contact matrices in an unsupervised manner, effectively capturing both local and global patterns of chromatin interactions. These learned features are then partitioned using a hierarchical clustering strategy augmented with genomic adjacency constraints, which simultaneously preserves the continuity of TADs along the linear genome and adaptively optimizes clustering granularity. This design overcomes the limitations of traditional methods in terms of boundary precision and stability. As a result, CCTAD produces TAD delineations that are more biologically meaningful and robust across different resolutions. The source code is available on GitHub at https://github.com/ruiping-Feng/CCTAD.

Evaluation of CCTAD across multiple cell lines and resolutions demonstrates its advantages in boundary identification of key regulatory elements.