ScHiCAtt

Hi-C measures all-versus-all contact frequency between genomic loci, and at the single-cell level it exposes cell-to-cell variation in chromatin folding that bulk experiments average away. The cost is sparsity: a single nucleus yields a tiny fraction of a bulk library's read depth, so its contact matrix is mostly zeros and its effective resolution is poor. Deep learning super-resolution methods built for bulk Hi-C — HiCPlus, HiCNN, hicGAN, DeepHiC — were not designed for this regime, and the single-cell-specific methods that followed struggle with long-range dependencies and with cross-cell-line transfer.

ScHiCAtt, from Rohit Menon, H. M. A. Mohit Chowdhury and Oluwatosin Oluwadare at the University of Colorado Colorado Springs with a joint appointment at the University of Colorado Anschutz Medical Campus, attacks that gap with attention. The architecture is a cascading residual generator trained adversarially, and the paper's core experiment is a systematic search over which attention mechanism — self, local, global, or dynamic attention-in-attention — to place at which depth in the network. Self-attention at the second cascading block wins across every metric, and that configuration is what ships as ScHiCAtt.

It joins a small group of cataloged chromatin-contact models including TruHiC, Mix-HiC and Evo2Hic, and is the single-cell-focused member of that group.

#Key Features

• Attention placement as a design variable: Self-, local-, global- and dynamic attention were each evaluated at layers 2, 3 and 5 of the cascading generator. Placement changed output quality as much as the choice of mechanism.
• Composite four-term loss: Pixel-level MSE is combined with a VGG perceptual loss, a total-variation smoothness term, and an adversarial term, weighted 0.5, 0.3, 0.1 and 0.1 respectively.
• Holds up under extreme sparsity: The margin over prior methods is largest at the harshest downsampling ratio tested, which is the regime that actually matches real single-cell libraries.
• Cross-cell-line and cross-species transfer: A model trained on one human cell is applied to a second human cell and to Drosophila chromosomes without retraining, the transfer setting the paper identifies as the main failure mode of earlier methods.
• Biological reproducibility, not just pixel accuracy: Evaluation includes GenomeDISCO concordance and downstream TAD calling with deDoc2, alongside SSIM, PSNR, MSE and SNR.

#Technical Details

The generator begins with an entry convolution over the raw contact map, passes through cascading residual blocks with skip connections that progressively upscale the map, and ends with an exit convolution. Local attention uses depthwise and pointwise convolutions over a sliding window; global attention aggregates across the whole map after several cascading blocks; dynamic attention adaptively weights an attention branch against a non-attention branch via global pooling. Inputs are 40x40 contact patches.

Datasets follow the ScHiCEDRN preparation: seven Drosophila melanogaster chromosomes and human frontal cortex single-cell Hi-C, with low-resolution maps produced by downsampling to 75%, 45%, 10% and 2% of raw reads. Human cell 1 supplies 14 chromosomes for training and four for validation; chromosomes 2, 6, 10 and 12 from both human cell 1 and a second human cell serve as test sets, with chr2L and chrX used for Drosophila. Against ScHiCEDRN, Loopenhance and DeepHiC, ScHiCAtt leads on PSNR, SSIM, MSE, SNR and GenomeDISCO across all tested chromosomes and downsampling ratios. Combining several attention mechanisms in one network gave slight gains at mild downsampling but underperformed single self-attention at aggressive ratios, which the authors attribute to architectural complexity interfering with long-range modelling.

#Applications

The intended use is a preprocessing step ahead of single-cell 3D genome analysis: enhance sparse per-cell contact maps, then call TADs, compute A/B compartments, or reconstruct structures on the improved matrices. The repository provides training and inference scripts and a Docker image, and the authors host a public web server that runs the model on uploaded data, which is the practical entry point for groups without a GPU. The scope is deliberately narrow — this is contact-matrix super-resolution, not a general chromatin foundation model.

#Impact

ScHiCAtt is peer-reviewed in Computational and Structural Biotechnology Journal, and its attention-placement ablation is a useful negative result for the field: stacking attention mechanisms is not free, and where a mechanism sits in a cascading generator matters as much as which one it is. Reproducibility has one significant gap. The trained checkpoint the inference script expects, schicatt.pth, is not committed to the GitHub repository, which contains only architecture, training, utility and analysis code; a user who clones the repo must train from scratch, obtain the checkpoint from the published Docker image, or use the hosted web server. The README states an MIT license, but the repository has no LICENSE file, so reuse terms are undefined for the code and unstated for the weights. The parameter count is not reported.