ScHiCAtt: Enhancing Single-Cell Hi-C Data Resolution Using Attention-Based Models

---

OluwadareLab, University of Colorado, Colorado Springs

---

Developers:

Rohit Menon
Department of Computer Science
University of Colorado, Colorado Springs Email: rmenonch@uccs.edu

Contact:

Oluwatosin Oluwadare, PhD Department of Computer Science University of Colorado, Colorado Springs Email: ooluwada@uccs.edu

---

Overview:

ScHiCAtt is a deep learning model designed to enhance the resolution of Single-Cell Hi-C contact matrices using various attention mechanisms, such as self-attention, local attention, global attention, and dynamic attention. The model leverages GAN-based training to optimize the quality of Hi-C contact maps through a composite loss function consisting of MSE, perceptual, total variation, and adversarial losses.

---

Build Instructions:

ScHiCAtt runs in a Docker-containerized environment. Follow these steps to set up ScHiCAtt.

1. Clone this repository:

git clone https://github.com/OluwadareLab/ScHiCAtt.git && cd ScHiCAtt

2. Pull the ScHiCAtt Docker image:

docker pull oluwadarelab/schicatt:latest

3. Run the container and mount the present working directory to the container:

docker run --rm --gpus all -it --name schicatt -v ${PWD}:${PWD} oluwadarelab/schicatt

4. You can now navigate within the container and run the model.

---

Dependencies:

All necessary dependencies are bundled within the Docker environment. The core dependencies include:

• Python 3.8
• PyTorch 1.10.0 (CUDA 11.3)
• NumPy 1.21.1
• SciPy 1.7.0
• Pandas 1.3.1
• Scikit-learn 0.24.2
• Matplotlib 3.4.2
• tqdm 4.61.2

Note: GPU usage for training and testing is highly recommended.

---

Training

To train the ScHiCAtt model on your Hi-C data, navigate to the Training folder and use the following command:

python3 train.py \
  --train_data path/to/train_data.npz \
  --valid_data path/to/valid_data.npz \
  --epochs 50 \
  --batch_size 32 \
  --lr 0.0001 \
  --save_path checkpoints/schicatt.pth

Input Format

Your .npz files must contain the following keys:

• 'data': Low-resolution Hi-C patches (shape: [N, 1, 40, 40])
• 'target': Ground truth high-resolution patches (same shape)

Arguments

Argument	Description	Default
--train_data	Path to the training .npz dataset	Required
--valid_data	Path to the validation .npz dataset	Required
--epochs	Number of training epochs	1
--batch_size	Batch size for training	64
--lr	Learning rate for the optimizer	0.0003
--save_path	Path to save the best model checkpoint	schicatt.pth

Output

The script saves the best-performing model (based on validation loss) to the specified path.

By default, the model will be saved to schicatt.pth.

---

Inference

After training the model, you can run inference (present in the Training folder) using the saved checkpoint on test data using:

python3 infer.py \
  --input path/to/test_data.npz \
  --checkpoint path/to/schicatt.pth \
  --output path/to/output_directory \
  --cuda 0

Input Format

The input .npz file must contain the following keys:

• 'data': Low-resolution Hi-C patches (shape: [N, 1, 40, 40])
• 'inds': Index array indicating patch positions and chromosome IDs (shape: [N, 4])

Arguments

Argument	Description	Default
--input	Path to input .npz test dataset	Required
--checkpoint	Path to the trained model checkpoint .pth file	Required
--output	Directory to save reconstructed full matrices	Required
--multi-chrom	(Optional) Enable multi-chromosome handling if needed	False
--cuda	CUDA device ID (e.g., 0 for GPU, -1 for CPU)	-1

Output

The script reconstructs chromosome-wise Hi-C matrices and saves them as compressed .npz files inside the specified output directory.

Each output file is named as:

chr<chromosome_id>_schicatt.npz

and contains the key 'schicatt' with the predicted high-resolution matrix.

Example

python3 infer_schicatt.py \
  --input data/test_chr11.npz \
  --checkpoint schicatt.pth \
  --output results/ \
  --cuda 0

This will generate the file results/chr11_schicatt.npz containing the inferred matrix.

---

Analysis

All the analysis scripts are available at analysis folder

TAD Plots

TAD detection with deDoc2

• We used https://github.com/zengguangjie/deDoc2 for TAD detection from scHiC data. We used lower TAD-like domains.
• Download the doDoc2
• Edit the necessary variables in call_tads.py script such as INPUT_FILEPATH and adjust other things as necessary.
• Run python3 call_tads.py

TAD plot

• We used https://xiaotaowang.github.io/TADLib/index.html to produce the TAD figures.
• Here, we provided a sample python script draw_tad_plots.py to produce the plots. Update INPUT_FILEPATH, MATRIX_FILEPATH, FILENAMES, CHROMOSOMES, ALGORITHMS, OUTPUT_PATH.
• It takes matrix and TADs as input. TAD file structure should be (without heading):

Chromosome	Start Position	End Position
chr12	80000	1960000
chr12	2000000	2720000
chr12	2760000	4040000
chr12	4080000	5480000
chr12	5520000	5720000

• Run python3 draw_tad_plots.py

L2 norm

• Edit draw_l2_norm.py with your filenames and paths including MATRIX_FILEPATH, FILENAMES, CHROMOSOMES, ALGORITHMS.
• It takes matrix as input.
• Run python3 draw_l2_norm.py

---

Data availability

The ScHiCAtt project is publicly available at https://github.com/OluwadareLab/ScHiCAtt. ScHiCAtt web-server is publicly available at http://schicatt.hicrobin.online/. Drosophila Hi-C Data is publicly available at https://doi.org/10.5281/zenodo.10535486. Human Cell HiC data is available at https://salkinstitute.app.box.com/s/fp63a4j36m5k255dhje3zcj5kfuzkyj1.

---

Cite:

If you use ScHiCAtt in your research, please cite the following:

Menon, R., Chowdhury, H. M., & Oluwadare, O. (2025). ScHiCAtt: Enhancing single-cell Hi-C data resolution using attention-based models. Computational and Structural Biotechnology Journal, 27, 978-991. https://doi.org/10.1016/j.csbj.2025.02.031

---

License:

This project is licensed under the MIT License. See the LICENSE file for details.