HiCARN

A Hi-C map's usable resolution is bought with sequencing reads. Binning contacts at 10 kb — the scale at which subTADs, enhancer–promoter contacts and CTCF-anchored loops become visible — takes billions of read pairs, so most published contact matrices are sequenced too shallowly to support the structure a study wants. Deep learning offers a cheaper route: treat the sparse matrix as a degraded image and learn the mapping back to the full-depth one.

HiCARN, from Parker Hicks at Concordia University Irvine and Oluwatosin Oluwadare's lab at the University of Colorado Colorado Springs, adapts the cascading residual network from natural-image super-resolution to this problem. The cascade is the distinguishing element: each block's output is concatenated forward into every later block, and the same pattern repeats at the network level, so features extracted early stay reachable at the end rather than being re-derived. Those concatenations double as shortcuts for forward and backward propagation, which is why a five-block network converges in fewer epochs and on less GPU memory than the heavier stacks it is measured against.

Two variants ship from the one paper and share one generator. HiCARN-1 is that generator trained on a pixel, perceptual and total-variation loss; HiCARN-2 wraps it in a GAN with a seven-block convolutional discriminator and adds an adversarial term. The authors conclude that neither the CNN-versus-GAN choice nor the number of cascading blocks materially changes accuracy — a useful negative result in a subfield that has cycled through both designs.

#Key Features

• Cascading feature reuse: Concatenating each block's output into all downstream blocks keeps low-level contact texture available at the final layers and gives gradients a short path back.
• One checkpoint across cell lines and species: The published weights were fit on GM12878 and applied without retraining to human K562 and the mouse CH12-LX line, neither of which the model saw during training.
• Checkpoints for four sparsity regimes: Separate weights are released for maps downsampled to 1/16, 1/32, 1/64 and 1/100 of their reads.
• Validated on chromatin structures, not just pixels: Enhanced maps were passed to TopDom for TAD calling, GenomeFlow for 3D chromosome reconstruction, and HiCExplorer for loop and CTCF-anchor detection, so the evaluation reaches features a biologist would act on.
• Containerized inference: A prebuilt Docker image carries the PyTorch environment and the prediction script loads a released checkpoint directly, so enhancing a matrix needs no training.

#Technical Details

The generator is five cascading blocks, each holding residual units of two 3×3 convolutions with ReLU activations and local skip connections, separated by 1×1 convolutions that condense the concatenated channels. Training used KR-normalized 10 kb intrachromosomal matrices from Rao et al. (2014, GSE63525), cut into 40×40 sub-matrices at stride 40 within a 201-bin genomic-distance bound: fourteen GM12878 chromosomes for training, four for validation, and chromosomes 4, 14, 16 and 20 held out for test. Optimization ran 100 epochs of Adam at batch size 64, initial learning rate 1.0×10⁻³ decayed by 0.1 every 30 epochs. Averaged over those test chromosomes at 1/16 downsampling, HiCARN-1 reaches PSNR 35.07, MSE 0.0003, SSIM 0.9119 with a GenomeDISCO reproducibility score of 0.9173 ± 0.0092; HiCARN-2 trails it by 0.005 SSIM, 0.16 PSNR and 0.003 GenomeDISCO. On the unseen cell lines the same checkpoints score 0.8301 ± 0.0070 GenomeDISCO for K562 at 1/64 downsampling and 0.5971 ± 0.0569 for mouse CH12-LX at 1/32. TAD concordance over a 2.5 Mb window of chromosome 4 is 0.8697, against 0.7834 for the strongest baseline there. Each released checkpoint is a 5.5 MB float32 state dict.

#Applications

The immediate use is rescuing Hi-C libraries that were never sequenced deeply enough — older public datasets, limited primary material, pilot experiments — so that TAD boundaries, loop anchors and 3D chromosome models can be extracted from them instead of resequencing. Inference is a single forward pass over 40×40 sub-matrices with released weights, so a lab can enhance a genome and hand the output straight to a TAD caller, a loop caller or a structure-reconstruction tool. The envelope is narrow: one training cell line, a 10 kb target, and the four downsampling ratios above. The cohesin-depletion experiment on HCT-116 in the paper involved retraining rather than zero-shot transfer.

#Impact

HiCARN became a standard comparison point for Hi-C resolution enhancement, appearing as a baseline in later work including TRUHiC and HiCFoundation. Its codebase also seeded the Oluwadare lab's own follow-ups — DiCARN-DNase adopts the cascading-residual design and adds dilated convolutions with DNase-seq augmentation, and ScHiCAtt carries a cascading generator into single-cell Hi-C with self-attention. The repository is unusually complete for the subfield, committing the authors' retrained checkpoints for the competing methods and the processed data splits alongside its own weights, which keeps the comparison reproducible. That a lightweight five-block network matched much deeper ensembles pushed attention away from depth and toward what the connections between layers do.