Does a map-level insulation constraint improve sequence-based prediction of 3D genome folding? A Stage 1 Registered Report with a reverse-complement-equivariant state-space model (HiFold-Mamba)

HiFold-Mamba predicts 448×448 2,048 bp contact maps from 1,048,576 bp DNA using a reverse-complement-equivariant state-space model, and is pre-registered to test insulation losses on DNALongBench.

• Why it matters: Pixel-wise training may miss that insulation, boundaries, and loops are map summaries, so it is unclear whether insulation-based losses improve sequence-to-structure prediction beyond training-derived annotations.
• What they did: The study uses DNALongBench with paired bootstrap over 237 one-megabase test blocks, comparing HiFold-Mamba to reproduced Akita, a parameter-matched Transformer, and a parameter-matched CNN, and testing an insulation-plus-boundary loss against independent 4DN and ENCODE annotations.
• The result: The pre-registered hypotheses require SCC gains of at least 0.01 versus baselines, and insulation loss is expected to improve boundary AUROC and loop AP without reducing SCC, with few-shot adaptation outperforming training from scratch on four held-out cell lines.

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