Cellformer

An implementation of Cellformer from our publication: Berson et al. "Whole genome deconvolution unveils Alzheimer’s resilient epigenetic signature"

Installation from source

The lastest source version of Cellformer can be accessed by running the following command:

git clone https://github.com/elo-nsrb/Cellformer.git
cd Cellformer

Requirements

• Python 3
• PyTorch (2.0.0)
• ArchR (R version 4.2.2)
• Scikit-learn (1.2.2)
• Asteroid (0.5.2)
• Pytorch-lightning (2.0.4)
• comet_ml (3.32.8)

In order to install package dependencies, you will need Anaconda. After installing Anaconda, please run the following command to create two conda environnements with R and Pytorch dependencies:

.\setup.sh

Usage

1. Peak calling and peak matrix creation

To create a peak matrix from single-cell ATAC-seq fragment files, please use the following commands:

conda activate R_env
./createPeakMatrix.sh --path_data ./data/ --input_dir ./scDATA/ --metadata ./data/scATAC_cell_annotations.csv

usage: createDataset -p | --path_data PATH_DATA
                     -i | --input_dir INPUT_DIR
                     -m | --metadata METADATA
                     [ -h | --help]
positional arguments:
-p, --path_data         Path to save the peak Matrix
-i, --input_dir         Directory with single cell arrow files
-m, --metadata          Metadata with cell annotations

optional arguments:
-h, --help              show help message and exit

2. Synthetic dataset generation

Synthetic dataset can be created from snATAC-seq peak matrix in AnnData format with celltype and Sample_num columns in obs(see example in data):

./createDataset.sh -p ../data/ -n 500

usage: createDataset -p | --path_data PATH_DATA
                     -n | --nbSamplesPerCase NBSAMPLESPERCASE
                     [ -f | --matrixfilename MATRIXFILEMANE]
                     [ -h | --help]
positional arguments:
-p, --path_data         Path to directory with the peak matrix
-n, --nbCellsPerCase    Number of synthetic samples per individual

optional arguments:
-h, --help              Show help message and exit
-f, --matrixfilename    Name of peak matrix file, default=adata_peak_matrix.h5

3. Pretrained model inference and bulk deconvolution

We provided the pretrained model used in the manuscript in cellformer. The pretrained model can be used to deconvolute bulk peak matrix by running:

conda activate pytorch_env
./deconvolution --model_path cellformer/ --peak_matrix ./data/CTRL_CAUD_AD.peak_countMatrix.txt

Usage: deconvolution  -p | --model_path MODEL_PATH
                      -m | --peak_matrix PEAK_MATRIX
                      [ -h | --help  ]"
positional arguments:
-p, --model_path        Path to model directory with train.yml
-m, --peak_matrix       Peak matrix to deconvolute

optional arguments:
-h, --help              Show help message and exit

You can find an example of the expected peak matrix format CTRL_CAUD_AD.peak_countMatrix.txt in data.

4. Model training

Cellformer can be trained from scratch using a synthetic dataset and configuration file train.yml (see an example in cellformer) by running:

conda activate pytorch_env
./trainModel.sh --model_path cellformer/

usage: trainModel -p | --model_path MODEL_PATH
                  [ -h | --help]
positional arguments:
-p, --model_path        Path to model directory with train.yml

optional arguments:
-h, --help              Show help message and exit

Please modify the path to the data folder in train.yml.

5. Model testing using trained model

Model testing can done using bulk or pseudobulk data by running:

./validation.sh --model_path cellformer/ 

Usage: validationModel  -p | --model_path MODEL_PATH
                        [ -h | --help  ]
positional arguments:
-p, --model_path        Path to model directory with train.yml

optional arguments:
-h, --help              Show help message and exit

Please note that in the new version, the mask is computed on the cross validation performance matrix. We recommend to remove cell type specific peaks with a spearman correlation <0.2 across samples.

Licence

This project is covered under the GNU General Public License v3.0