HiC2Self

HiC2Self: Self-supervised denoising for bulk and single-cell Hi-C contact maps

Overview of the HiC2Self framework.

• Manuscript: bioRxiv
• Poster: HiC2Self_poster.pdf

Requirements

We recomment creating the enrivonment using the provided environment.yml file:

mamba env create -f environment.yml

• Core dependencies
  • python=3.10
  • PyTorch, NumPy, SciPy, scikit-learn, h5py, anndata, UMAP-learn
• Visualization
  • matplotlib, seaborn
  • coolbox
• Other
  • torchvision
  • pytorch-ssim

A complete walkthrough for setting up the environment can be found in create_hic2self_env.sh.

Data Processing

HiC2Self does NOT require Hi-C data normalization. The HiC2Self framework consists of three main steps:

1. Data preprocessing

Raw Hi-C contact maps are first processed by cropping them along the diagonal into equally sized, diagonally symmetric submatrices. These submatrices serve as the basic training and inference units for the model. This step is implemented in: data_processing/input_data_preparation_full_chr_split.py

2. Model training

HiC2Self is trained using the set of diagonally symmetric submatrices generated in the preprocessing step. The model learns to denoise low-coverage contact maps by predicting high-quality contact patterns from masked corrupted inputs.

3. Inference

During inference, HiC2Self predicts each diagonal submatrix independently. Overlapping predictions along the diagonal are then averaged to reconstruct the full chromosome-scale contact map. Inference using trained models is implemented in the prediction_postprocessing directory.

For detailed implementation and usage instructions, please refer to the data_processing and prediction_postprocessing directories.

Keep connected

• Email: ruiyang204@gmail.com
• others: @RuiYang53922541, @ruiyang.bsky.social