CellContrast: Reconstructing Spatial Relationships in Single-Cell RNA Sequencing Data via Deep Contrastive Learning
Contact: Yuanhua Huang, Ruibang Luo, Shumin Li
Email: [email protected], [email protected], [email protected]
CellContrast reconstructs the spatial relationships for single-cell RNA sequencing (SC) data. Its fundamental assumption is that gene expression profiles can be projected into a latent space, where physically proximate cells demonstrate higher similarities. To achieve this, cellContrast employs a contrastive learning framework of an encoder-projector structure. The model was trained with spatial transcriptomics (ST) data and applied to SC data for obtaining the spatially-related representations. The produced results of cellContrast can be used in multiple down-stream tasks that requires spatial information, such as cell-type co-localization and cell-cell communications.
CellContrast's paper describing its algorithms and results is at Cell Patterns.
- v0.1 (Sep, 2023): Initial release.
To install CellContrast, python 3.9 is required and follow the instruction
- Install Miniconda3 if not already available.
- Clone this repository:
git clone https://github.com/HKU-BAL/CellContrast- Navigate to
CellContrastdirectory:
cd CellContrast- (5-10 minutes) Create a conda environment with the required dependencies:
conda env create -f environment.yml- Activate the
cellContrastenvironment you just created:
conda activate cellContrast- Install pytorch: You may refer to pytorch installtion as needed. For example, the command of installing a cpu-only pytorch is:
conda install pytorch torchvision torchaudio cpuonly -c pytorchCellContrast contains 3 main moduels: train, eval and inference, for training model, benchmarking evaluation and inference of spatial relationships, respectively. In addition, We also provide reconstruct module for integrating train and inference. To check available modules, run:
python cellContrast.py -h
python cellContrast.py reconstruct \
--train_data_path train_ST.h5ad ## required, use your ST h5ad file here\
--query_data_path query_sc.h5ad ## path of query SC h5ad file\
--parameter_file parameters/parameters_spot.json ## optional. use the our default for spot or single-cell ST, or your customized parameters here\
--save_folder cellContrast_models/ ## optional, model output path\
--enable_denovo ## optional, run MDS to leverage the SC-SC pairwise distance to 2D pseudo space
--save_path spatial_reconstructed_sc.h5ad \ ## path of of the spatial reconstructed SC data- Adopt the predefined parameters for imaging-based ST data by setting
--single_cell.
python cellContrast.py reconstruct \
--train_data_path train_ST.h5ad ## required, use your ST h5ad file here\
--query_data_path query_sc.h5ad ## path of query SC h5ad file\
--single_cell \
--parameter_file parameters/parameters_singleCell.json ## optional. use the our default for spot or single-cell ST, or your customized parameters here\
--save_folder cellContrast_models/ ## optional, model output path\
--enable_denovo ## optional, run MDS to leverage the SC-SC pairwise distance to 2D pseudo space
--save_path spatial_reconstructed_sc.h5ad \ ## path of of the spatial reconstructed SC data
CellContrast model was trained based on ST data (which should be in AnnData format, with truth locations in .obs[['x','y']]). The model can be trained with the following command:
‼️ Default parameters are defined for sequencing-based ST, adopt the predefined parameters for imaging-based ST data by setting--single_cell.
python cellContrast.py train \
--train_data_path train_ST.h5ad \ ## required, use your ST h5ad file here
--save_folder cellContrast_models/ \ ## optional, model output path
--single_cell # defaut: not enabled. Set this flag to switch to our prefined parameters for imaging-based ST.
--parameter_file parameters/parameters_singleCell.json ## optional. use the our default for spot or single-cell ST, or your customized parameters here\
## Output file: cellContrast_models/epoch_3000.ptThe peformance of benchmarking can be evaluated with the following command, and three metrics are included: nearest neighbor hit, Jessen-Shannon distance, and Spearman's rank correlation.
python cellContrast.py eval \
--ref_data_path ref_ST.h5ad \ ## path of refernece ST h5ad file
--query_data_path query_ST.h5ad \ ## path of testing h5ad file with truth locations
--model_foldercellContrast_models\ ## folder of trained model
--parameter_file parameters/parameters_singleCell.json ## Take the parameter file you used in the training phase.\
--save_path results.csv \ ## evaluation result path
## Output file: result.csv with neighbor hit, JSD, spearman's rank correlation for each testing sample.
The spatial relationships of SC data can be obtained with the following command:
python cellContrast.py inference \
--ref_data_path train_ST.h5ad \ ## path of refernece ST h5ad file
--query_data_path query_sc.h5ad \ ## path of query SC h5ad file
--model_folder \ ## folder of trained model
--parameter_file parameters/parameters_singleCell.json ## Take the parameter file you used in the training phase.\
--save_path spatial_reconstructed_sc.h5ad \ ## path of of the spatial reconstructed SC data
--enable_denovo \ ## optional, run MDS to leverage the SC-SC pairwise distance to 2D pseudo space
## Output file: spatially reconstructed h5ad file of annData- what will be newly added in
sptial_reconstructed_sc.h5ad:.uns[['cosine sim of rep','representation','referenced x','referenced y','de novo x','de novo y']]