AxialTAD: an axial contrast architecture for TAD boundary prediction from Hi-C.

AxialTAD predicts TAD boundaries from Hi-C in four human cell lines with F1 comparable to deepTAD while using 76%–80% fewer trainable parameters (113K–137K vs 571K).

• Why it matters: Boundary detection from Hi-C is still challenging, and prior deep learning gains may depend on very large models, leaving uncertainty about whether boundary-focused encoding can reduce capacity needs.
• What they did: The compact Axial Contrast Module encodes directional contrast along row and column axes of local Hi-C patches, and AxialTAD was evaluated on GM12878, HeLa-S3, IMR-90, and K562 with leave-one-cell-out generalization.
• The result: Within the tested architecture range, AxialTAD matched deepTAD in human 25 kb experiments, outperformed controls and deepTAD in two mouse datasets, and its predicted boundaries aligned with CTCF and cohesin ChIP-seq signals.

AI summary of the abstract · may contain errors · not medical advice · read the paper.