HiCDiff: single-cell Hi-C data denoising with diffusion models

Yanli Wang, Jianlin ChengBriefings in Bioinformatics 2024 · OpenReview

TL;DR

This paper introduces HiCDiff, a novel deep learning method that uses diffusion models to denoise single-cell Hi-C data, a notoriously noisy and sparse dataset. HiCDiff significantly improves data quality, outperforming many existing methods and achieving state-of-the-art results in some benchmarks, making single-cell Hi-C data more reliable for biological analysis.

What Problem Does This Paper Address?

Understanding the three-dimensional (3D) organization of the genome within a cell is crucial for deciphering its function, including gene regulation and interactions. Hi-C is a powerful technique for mapping these 3D contacts across the genome. However, single-cell Hi-C (scHi-C) data is particularly challenging due to its extreme sparsity and high noise levels. This noise arises from various experimental steps like DNA amplification and restriction enzyme digestion, which introduce biases and reduce the signal-to-noise ratio. These limitations make it difficult to reliably analyze scHi-C data for downstream biological insights. While several deep learning methods exist for denoising bulk Hi-C data, scHi-C presents a greater challenge due to its inherent data scarcity.

Methods

The core of HiCDiff is a generative diffusion model, inspired by Denoising Diffusion Probabilistic Models (DDPMs) and Denoising Diffusion Restoration Models (DDRMs). These models work by gradually adding noise to data in a "forward process" and then learning to reverse this process to recover the original, clean data. HiCDiff adapts this framework to chromosomal contact matrices.

Key methodological aspects include:

  • Diffusion Framework: HiCDiff employs a Markov chain model for both the forward (adding Gaussian noise) and reverse (denoising) processes. The reverse process aims to predict and remove the noise at each step.
  • Deep Residual Network Architecture: Unlike the standard U-Net architecture used in many diffusion models, HiCDiff utilizes a deep residual network. This architecture, proven effective in image super-resolution, is adapted for Hi-C data denoising.
  • Unsupervised and Supervised Learning Modes: HiCDiff can be trained in two modes:
  • Unsupervised (Conditional Diffusion): Trained only on clean data, learning to denoise by reversing the diffusion process.
  • Supervised (Conditional Diffusion): Trained on pairs of noisy and clean data, using the noisy data as a condition to guide the denoising process.
  • Noise Modeling: The paper specifically uses Gaussian noise for corruption, citing its favorable mathematical properties and its use in prior Hi-C denoising methods.

Datasets

The study utilized several scHi-C and bulk Hi-C datasets from human and Drosophila cell lines, downloaded from the Gene Expression Omnibus (GEO).

  • Human Datasets:
    • Human frontal cortex cells (single-cell and bulk).
  • Data from chromosomes 1-22, X, and Y.
  • Drosophila Datasets:
    • Drosophila male Dm-BG3c2 (BG3) cells (single-cell and bulk).
  • Data from chromosomes chr2L, chr2R, chr3L, chr3R, chr4, chrX, and chrM.

For training and validation, chromosomal contact matrices were generated at a 40 kilobase (kb) resolution. To create noisy input data, Gaussian noise was added to the clean (ground truth) matrices at two levels: 0.1 and 0.5. The paper also mentions using a clean Hi-C dataset of yeast as an additional benchmark.

Key Insights and Ideas

  • Diffusion Models are Powerful for Sparse Data: The success of HiCDiff demonstrates that diffusion models, despite their computational intensity, are highly effective at handling the extreme sparsity and noise characteristic of scHi-C data.
  • Residual Networks Outperform U-Nets: The choice of a deep residual network architecture for the generative model proved superior to the standard U-Net in this application, leading to better denoising performance.
  • Unsupervised Learning is Highly Competitive: Surprisingly, the unsupervised HiCDiff often performed as well as, or even better than, its supervised counterpart and many supervised non-diffusion methods, highlighting the power of learning from clean data alone.
  • Generalizability Across Cells and Species: HiCDiff showed remarkable generalization capabilities, performing well on unseen cells within the same species and even across different species (human to Drosophila), suggesting robustness.
  • Bulk Hi-C Benefits Too: The method isn't limited to scHi-C; it also shows strong performance in denoising bulk Hi-C data, indicating broader applicability.

Key Results

  • Substantial Quality Improvement: HiCDiff significantly improved image-based metrics like PSNR, SSIM, and SNR, and reduced MSE on scHi-C test datasets. For example, on human_cell_1_test_data at noise level 0.1, HiCDiff1 increased PSNR from 28.9701 to 42.4171 and SSIM from 0.1885 to 0.9662.
  • State-of-the-Art Performance: In terms of image-based metrics, unsupervised HiCDiff (HiCDiff1) outperformed most supervised non-diffusion methods and was comparable to the state-of-the-art supervised method, SCHICEDRN.
  • HiCRep Score Excellence: HiCDiff achieved state-of-the-art performance on the HiCRep metric, which assesses the reproducibility of Hi-C data features, particularly at higher noise levels.
  • Robustness to Noise Levels: The model performed well across different noise levels (0.1 and 0.5), with performance generally improving as noise decreased.
  • Generalization: HiCDiff generalized well to unseen cells and species, demonstrating its robustness.

Limitations and Future Work

  • Computational Cost: While effective, diffusion models are known for their computational demands during training and inference. The paper notes that HiCDiff takes longer to converge than some other methods.
  • Noise Types: The study primarily focused on Gaussian noise. The authors suggest that exploring other noise types, such as Poisson or exponential noise, could be a future direction.
  • Resolution Impact: While the study explored different resolutions, further investigation into optimal resolutions for denoising could be beneficial.
  • Other Noise Characteristics: The paper briefly touches upon symmetric vs. asymmetric noise, suggesting that while HiCDiff performs well, further exploration of noise characteristics could refine the models.

Summary updated

BibTeX
@article{wang2024hicdiff,
  author = {Yanli Wang and Jianlin Cheng},
  title = {{HiCDiff: single-cell Hi-C data denoising with diffusion models}},
  journal = {Briefings in Bioinformatics},
  year = {2024},
  url = {https://openreview.net/forum?id=nmZ3LNu8sb},
}
Figure