Can a GAN sharpen cheap, low-coverage chromosome contact maps?
A generative adversarial network rebuilt sharp high-resolution chromosome contact maps from as little as 1% of the sequencing reads, beating earlier CNNs that produced blurry results.
Source
DeepHiC: A generative adversarial network for enhancing Hi-C data resolution
Study at a glance
- Design
- Computational / modelling — cGAN trained on chromosomes 1-14 of GM12878 Hi-C data downsampled from 10-kb matrices, tested on held-out chromosomes and other cell lines against HiCPlus, HiCNN and Boost-HiC.
- N
- No single N; training and testing use Hi-C matrices from several cell lines (GM12878, its replicate, K562, IMR90) split by chromosome, evaluated on 1 Mb sub-regions.
- Population
- Public human (and mouse) Hi-C chromatin contact datasets
- Outcome
- Structural similarity (SSIM) and Pearson correlation to real high-resolution matrices; accuracy of downstream loop and TAD-boundary detection
Structured fields used in claim comparison tables when every cited study has a complete layer.
What they did
The authors built DeepHiC, a conditional GAN whose generator turns a low-coverage Hi-C contact matrix into a high-resolution one, trained with adversarial, perceptual and total-variation losses instead of plain mean squared error. They made low-coverage inputs by randomly downsampling reads from deeply sequenced data, trained on chromosomes 1-14 of one cell line and tested on the remaining chromosomes and on other cell lines. They compared it with the CNN methods HiCPlus and HiCNN and with Boost-HiC, and checked whether the enhanced maps improved downstream detection of chromatin loops and domain boundaries.
What they found
Genome-wide SSIM on the GM12878 test data averaged 0.89 for DeepHiC versus 0.71 for HiCPlus, 0.66 for HiCNN and 0.15 for the raw downsampled data. Correlations with the real data were about 5% higher than HiCPlus at every genomic distance, and even from 1% of reads the output matched the real data about as well as an independent experimental replicate. Loops called on DeepHiC output agreed better with real high-resolution calls, and separation of CTCF-mediated interactions reached an average AUC of 0.825. A generator trained without the adversarial part gave unstable test SSIM.
The limits
What it doesn't show
Low-coverage inputs were simulated by downsampling reads, which may not capture the biases of genuinely shallow experiments. Training targets were themselves experimental data, so the model can be no better than the deepest available datasets, and separate models were needed for different downsampling ratios. The authors note inputs need more than 10% non-zero entries. The mouse embryo application has no deep ground truth, so improvements there are judged only by indirect enrichment at promoters and open chromatin.
Key terms
- Generative adversarial network (GAN)
- A pair of networks trained against each other: a generator produces outputs and a discriminator tries to tell them from real data, pushing the generator toward realistic outputs.
- Conditional GAN
- A GAN whose generator is given an input (here a low-coverage matrix) and must produce the matching output rather than a random sample.
- Perceptual loss
- A loss that compares features extracted by a pretrained network rather than raw pixel values, rewarding similar structure and texture.
- Structural similarity index (SSIM)
- An image-quality score between two images that emphasises matching local structure, contrast and brightness rather than just average error.
- Super-resolution
- Predicting a detailed, high-resolution version of a signal from a coarse or noisy one.
Flashcards
0 of 10 answers reviewed
Research intelligence for this paper
See its role on concept claims, tensions it is part of, placement history, and related discoveries.
Quiz yourself
What problem with MSE-trained CNNs such as HiCPlus did DeepHiC aim to fix?
Common questions
Why not just train a CNN with mean squared error?
MSE rewards averaging over plausible answers, which blurs sharp features like loops and domain edges; the adversarial and perceptual losses push the model to keep those sharp structures.
How was overfitting checked?
Training and test losses fell together, results were consistent across different train/test splits, and the model generalised to cell lines it was not trained on.
Why does resolution matter for Hi-C?
Fine features such as enhancer-promoter loops and small domains are invisible at coarse bin sizes, and deep sequencing to reach fine resolution is expensive.
More on Generative models