AxialTAD: an axial contrast architecture for TAD boundary prediction from Hi-C.
- 1Department of Bioinformatics & Life Science, Soongsil University, 369, Sangdo-ro, Dongjak-gu, Seoul 06978, Republic of Korea.
- 2AI-Bio Convergence Research Institute, Soongsil University, 369, Sangdo-ro, Dongjak-gu, Seoul 06978, Republic of Korea.
Abstract
Topologically associating domains (TADs) are fundamental units of 3D chromatin organization, and accurate identification of TAD boundaries from Hi-C contact matrices remains an active challenge. Recent deep learning approaches have advanced substantially, but these rely on increasingly large architectures, raising the question of how much of this capacity boundary-centric analysis requires. Here we propose AxialTAD, a compact deep learning model centered on an Axial Contrast Module (ACM) that explicitly encodes directional contrast in contact intensity along the row and column axes of local Hi-C patches, an operator aligned with the geometric signature of TAD boundaries. Evaluated on four human cell lines (GM12878, HeLa-S3, IMR-90, K562) within cell lines and under leave-one-cell-out generalization, AxialTAD performed comparably with a recent convolutional neural network (CNN)+Transformer baseline (deepTAD) while using 76%-80% fewer trainable parameters (113K-137K versus 571K). A parameter-matched conventional CNN achieved similar F1 to AxialTAD in the human 25 kb experiments. When trained and evaluated within each of two mouse datasets, AxialTAD outperformed both this control and deepTAD. Orthogonal validation against CTCF and cohesin ChIP-seq supported the biological plausibility of the predicted boundaries. These results indicate that, within the range of architectures examined, a large architecture is not necessary to reach this level of boundary prediction.
Related
- CBAT-Net: a unified CNN–BiLSTM–MultiHead transformer architecture for joint Q3/Q8 protein secondary structure prediction
- A novel LLM-based GPT-neo transformer with machine learning architecture for promoters prediction in Agrobacterium
- ProCAS: Multi-Lysine Post-Translational Modification Site Prediction Based on the SPUS Algorithm and the ProCAS Architecture
- TriTower-m6Am: A triple-tower heterogeneous deep learning architecture integrating semantic, sequential, and structural information for mRNA N6,2′-O-dimethyladenosine site prediction
- Phylogenetic tree inference with tropical axial attention
BibTeX
@article{choi2026axialtad,
title = {{AxialTAD: an axial contrast architecture for TAD boundary prediction from Hi-C.}},
author = {Woo-Young Choi and Je-Keun Rhee},
journal = {Briefings in bioinformatics},
year = {2026},
month = nov,
doi = {10.1093/bib/bbag603},
url = {https://europepmc.org/article/MED/42969852},
note = {CC BY},
}
RIS
TY - JOUR
TI - AxialTAD: an axial contrast architecture for TAD boundary prediction from Hi-C.
AU - Woo-Young Choi
AU - Je-Keun Rhee
PY - 2026
DA - 2026/11/01
JO - Briefings in bioinformatics
DO - 10.1093/bib/bbag603
UR - https://europepmc.org/article/MED/42969852
AB - Topologically associating domains (TADs) are fundamental units of 3D chromatin organization, and accurate identification of TAD boundaries from Hi-C contact matrices remains an active challenge. Recent deep learning approaches have advanced substantially, but these rely on increasingly large architectures, raising the question of how much of this capacity boundary-centric analysis requires. Here we propose AxialTAD, a compact deep learning model centered on an Axial Contrast Module (ACM) that explicitly encodes directional contrast in contact intensity along the row and column axes of local Hi-C patches, an operator aligned with the geometric signature of TAD boundaries. Evaluated on four human cell lines (GM12878, HeLa-S3, IMR-90, K562) within cell lines and under leave-one-cell-out generalization, AxialTAD performed comparably with a recent convolutional neural network (CNN)+Transformer baseline (deepTAD) while using 76%-80% fewer trainable parameters (113K-137K versus 571K). A parameter-matched conventional CNN achieved similar F1 to AxialTAD in the human 25 kb experiments. When trained and evaluated within each of two mouse datasets, AxialTAD outperformed both this control and deepTAD. Orthogonal validation against CTCF and cohesin ChIP-seq supported the biological plausibility of the predicted boundaries. These results indicate that, within the range of architectures examined, a large architecture is not necessary to reach this level of boundary prediction.
SN - 1467-5463
SN - 1477-4054
ER -