scHiC-Diff

scHiC-Diff is a conditional diffusion model for scHi-C imputation that reconstructs chromatin contact maps through iterative denoising. Using a Diffusion Transformer to model long-range dependencies, scHiC-Diff consistently improves interaction recovery and downstream analyses across nine simulated datasets and three real scHi-C datasets. It also enhances chromatin organization reconstruction and chromatin loop detection on developmental scHi-C data.

Environment summary

•	Python: 3.9
•	CUDA runtime: 11.6
•	PyTorch: 1.12.1 + cu116
•	PyTorch Lightning: 1.9.0
•	Bib ecosystem: Scanpy / scvi-tools / rapids-singlecell / etc.

scdiff Environment Setup (micromamba + CUDA 11.6)

This repository provides the recommended Python environment configuration for running scdiff under Python 3.9 with CUDA 11.6 support.

1. Prerequisites

•	Linux recommended (HPC/Ubuntu).
•	NVIDIA GPU + driver installed.
•	micromamba installed and available in your shell.
•	CUDA runtime is provided via cudatoolkit=11.6 (Conda), so no need to install system CUDA toolkit manually.

2. Create a micromamba environment

micromamba create -n scdiff python=3.9 pip=24.0
micromamba activate scdiff

3. Install CUDA runtime (Conda)

Install CUDA runtime libraries (CUDA 11.6):

micromamba install -c conda-forge cudatoolkit=11.6

Note: cudatoolkit provides CUDA runtime libraries inside the environment, but your machine still needs a working NVIDIA driver.

4. Install PyTorch (CUDA 11.6 build)

pip install torch==1.12.1+cu116 torchvision==0.13.1+cu116 torchaudio==0.12.1 \
  --extra-index-url https://download.pytorch.org/whl/cu116

5. Install core training libraries

pip install pytorch-lightning==1.9.0 pyro-ppl==1.8.6

6. Install full Python dependencies by requirements.txt

pip install -r requirements.txt

7. Optional: bitsandbytes + triton

If your workflow requires bitsandbytes and triton:

pip install bitsandbytes==0.39.1 triton==2.3.0

Warning: bitsandbytes can be sensitive to CUDA runtime and driver versions. If it fails to import, try removing it or installing a compatible build for your GPU/driver setup.

8. Verify installation

Run:

python -c "import torch; print(torch.__version__); print(torch.cuda.is_available())"