Loading…
Chromatin loop caller for Hi-C, Micro-C, DNA SPRITE, and single-cell contact maps, pairing axial attention with a U-Net to work at very low coverage.
No providers recorded yet. Browse all providers
Chromatin loops bring distant regulatory elements into contact, and calling them from a Hi-C contact map is the entry point to nearly every 3D genome analysis. The established tools look for local enrichment: a pixel brighter than its immediate surroundings. That works on deeply sequenced bulk maps and fails on sparse ones — a problem, because single-cell Hi-C, the assay that could reveal cell-to-cell heterogeneity in genome folding, produces exactly the sparse maps these methods cannot handle. The workarounds either impute missing contacts, degrading accuracy, or require a panel of reference samples, restricting the method to species where such panels exist.
Polaris, from Yanlin Zhang's group at HKUST (Guangzhou) with collaborators at McGill, starts from a different observation about what a loop looks like. Loops are not isolated bright dots: they co-occur along rows and columns, sit at the corners of topologically associating domains, and overlap architectural stripes — axial features spanning kilobases to megabases, invisible to a detector examining a small window. Polaris therefore combines axial attention, applied separately along the width and height of the contact matrix, with a U-Net encoder–decoder, giving it a receptive field large enough to see the stripe a candidate loop belongs to while retaining the resolution to place it.
The second half of the design addresses labels. High-confidence loop annotations are scarce while unlabeled Hi-C data is abundant, so Polaris pretrains by knowledge distillation from the reference-panel method RefHiC before fine-tuning on orthogonally validated loops. The result inherits RefHiC-level accuracy without its dependence on reference panels at inference, making the model portable to zebrafish and mouse as readily as to human.
The encoder begins with a shallow convolutional feature extractor and then applies three down-convolution blocks, each halving spatial dimensions and doubling channels, with an axial-attention transformer block after each. The decoder mirrors it with skip connections, ending in a convolutional head that emits a per-pixel loop probability. Discrete loops come from a density-based clustering step over pixels scoring above 0.6, accelerated by a KD-tree, with singletons discarded as likely false positives.
Pretraining used 25 human Hi-C contact maps at 5 kb resolution, restricted to contacts within 3 Mb and divided into 157,456 submatrices, with soft labels derived from RefHiC logits. Fine-tuning used 74,855 positive loci pooled from CTCF ChIA-PET, RAD21 ChIA-PET, SMC1 HiChIP, and H3K27ac HiChIP experiments against 256,609 negatives, with chromosomes 11 and 12 held out for validation and 15 through 17 for test.
On GM12878 Hi-C downsampled to 500 million valid read pairs, the top 1,700 Polaris predictions included 1,271 CTCF-supported, 759 RAD21-supported, 557 SMC1-supported, and 213 H3K27ac-supported loops, comparable to RefHiC while other tools recovered 17–53% fewer. On cohesin-depleted HCT-116 cells, where loops should largely disappear, Polaris returned fewer than 50 predictions — a direct false-positive check. Across Micro-C it detected 3,746 loops with 1,107 of the top 3,500 CTCF-supported, more than any alternative.
Polaris fits wherever loop annotation is the first analysis step and coverage is the constraint: single-cell Hi-C studies of folding heterogeneity, low-input bulk experiments, and comparative 3D genome work in non-model species. It installs from pip and runs as a command-line tool with documented walkthroughs and worked example data, so it drops into existing Hi-C pipelines without a training step. Its multi-resolution behavior — validated at 5, 10, and 25 kb — also suits studies analyzing the same maps at several scales.
Loop calling has been a crowded field of specialized tools, most tuned for one assay at one depth. Polaris's contribution is showing that a single model trained on diverse contact maps generalizes across assays, species, depths, and resolutions well enough that retraining on specific conditions gave negligible gains — an argument for treating loop calling as a pretrained-model problem rather than a per-dataset one. Weights are published on Hugging Face under an MIT license alongside an MIT-licensed repository and a Zenodo reproducibility archive. The work remains a preprint awaiting peer review, and neither the parameter count nor the full composition of the pretraining corpus is stated. Fine-tuning was performed on GM12878 at 5 kb, so the strongest claims are anchored to one human cell line even though evaluation ranges well beyond it.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.