DiCARN-DNase

The resolution of a Hi-C contact map is set by sequencing depth: halving the bin size quarters the read pairs in each cell of the matrix, so TAD boundaries and enhancer-promoter loops wash out unless the library was sequenced deeper than most budgets allow. A decade of "super-resolution" networks treat the contact matrix as an image and sharpen it, and they share a weakness — a network trained on one cell line degrades on another, because chromatin folding is cell-type specific while the pixel statistics the network memorized are not.

DiCARN-DNase, from the Bioinformatics Lab at the University of Colorado Colorado Springs with the University of Colorado Anschutz Medical Campus, attacks that transfer gap from two directions. The architecture — DiCARN, a Dilated Cascading Residual Network — widens the receptive field with dilated convolutions and inserts spatial self-attention into the early cascade, so the network weighs loci unevenly rather than treating every bin as equally informative. The training set is then augmented with a signal that is cell-type specific and cheap to obtain: interaction frequencies imputed from DNase-seq chromatin accessibility.

That second idea is not tied to this architecture: applied to four existing enhancers — HiCSR, HiCARN, HiCNN and DFHiC — the same augmentation improved cross-cell-line reproducibility in ten of twelve test scenarios. The work extends HiCARN from the same lab, and sits beside that lab's single-cell counterpart ScHiCAtt and contemporaries such as TRUHiC.

#Key Features

• Dilated convolutions for a wider receptive field: A dilation rate of 2 in every residual block and in the exit stack widens the genomic neighborhood the network sees around each contact, without adding filter weights.
• Spatial self-attention in the early cascade: Query/key/value 1×1 convolutions form an attention map over the feature grid after the first two cascading blocks — a hyperparameter search found that placement better than attention everywhere.
• DNase-seq interaction frequencies as augmentation: A linear regression maps accessibility signal and 1D genomic distance to 3D distance, which converts back to interaction frequencies concatenated with the Hi-C training data.
• Two augmentation regimes, both retrained per target cell line: The source variant (DiCARN-DNase-S) imputes frequencies from the DNase signal of the GM12878 training chromosomes; the target variant (DiCARN-DNase-T) imputes them from the target cell line's test chromosomes. Both concatenate the result with the GM12878 Hi-C data and retrain, so each ships separate K562, HMEC and NHEK checkpoints — only vanilla DiCARN is a single checkpoint applied unchanged elsewhere.

#Technical Details

The generator is a fully convolutional network of roughly 1.5 million parameters at 64 feature channels: a 3×3 entry convolution, two dilated convolutions, a stack of cascading blocks each holding three residual blocks fused by 1×1 convolutions with self-attention after the first two, global cascading connections concatenating every block output, and a dilated exit stack down to one channel. Training minimizes mean squared error with Adam at a 3e-4 learning rate decayed tenfold every 30 epochs, batch size 64, for 100 epochs on a single RTX 4090.

Data comes from the Rao et al. 2014 Hi-C maps (GEO accession GSE63525), KR-normalized, with 10 kb as the target resolution and 40 kb input downsampled at 1/16 and 1/64, cut into 40 × 40 submatrices under a 201-bin distance bound. Training used GM12878 chromosomes 1–22 excluding 4, 14, 16 and 20, validating on 2, 6, 10 and 12. On the held-out chromosomes at 1/16 downsampling, DiCARN averaged SSIM 0.9161, PSNR 35.31, GenomeDISCO 0.9191 and HiCRep 0.9037, ahead of HiCSR and DFHiC on all four; the same checkpoint applied to K562, HMEC and NHEK reached SSIM 0.9498, 0.9757 and 0.9739, and averaged rank 2.3 among five methods on HiCRep across those lines. Downstream, TopDom TAD calls scored by Measure of Concordance and 3DMax reconstructions of a K562 chromosome 20 region scored by Spearman correlation both improved with DNase augmentation over the vanilla model.

#Applications

The immediate beneficiaries are groups holding shallow Hi-C libraries for a cell type nobody has sequenced deeply — primary tissues, patient samples, comparative panels — where re-sequencing is impractical but a matched accessibility track is cheap. The enhanced matrices are ordinary .npz contact maps that drop straight into existing TAD callers and 3D reconstruction tools, and the augmentation step is a standalone R script that can feed any of the enhancers it was tested on.

#Impact

DiCARN-DNase is a bioRxiv preprint awaiting peer review, and its evaluation envelope is narrow: one training cell line, one resolution pair, two downsampling ratios and four test chromosomes. Its durable idea is the one the authors generalized — that a cheaply measured epigenomic signal can stand in for the cell-type specificity a Hi-C-only enhancer cannot learn, and that the benefit survives transplant into competing architectures. The repository ships training and prediction code, the imputation script, trained .pytorch checkpoints for every variant and a public Docker image, with preprocessed project data on Zenodo — but no license file, so reuse terms are unstated even though the preprint and the Zenodo deposit are both CC BY 4.0.