SCARlink

This repository contains the code for SCARlink described in the paper: Single-cell multi-ome regression models identify functional and disease-associated enhancers and enable chromatin potential analysis by Mitra, S et al., 2024: https://www.nature.com/articles/s41588-024-01689-8

Please visit wiki for details about installation and running SCARlink. Tutorial notebook describes how to run SCARlink on a small example dataset.

Single-cell ATAC+RNA linking (SCARlink) uses multiomic single-cell ATAC and RNA to predict gene expression from chromatin accessibility and predict regulatory regions. It uses tile-based chromatin accessibity data from a genomic locus (+/-250kb flanking gene body) and applies regularized Poisson regression to predict gene expression. The learned regression coefficients of the tiles are informative of regulatory regions in a gene-specific manner. The regression model by itself is cell-type-agnostic. Shapley values computed by grouping cells provide insight into the regulatory regions in a cell-type specific manner. Therefore, the trained regression model can be used multiple times to identify important tiles for different groupings without retraining.

Note: The equation for the loss function was incorrectly typeset in the original paper and should be: $$\frac{1}{N}\sum_{i=1}^{N} \left(e_i^{X_i\mathbf{w}+\epsilon} - Y_i(X_i\mathbf{w} + \epsilon)\right) + \alpha||\mathbf{w}||_2^2$$

Additionally, the default proportion of held-out data in model evaluation is 0.25 rather than 0.2 (one-fifth) as currently mentioned in the Methods.

Installation

It is optional to install SCARlink in a conda environment that also has R packages Seurat v4 and ArchR installed. If you already have the R packages installed and want to run SCARlink without conda, jump directly to step 2 of installation. Alternatively, for Docker setup, jump to step 3.

1. Conda setup

The following conda setup is linux compatible. For conda setup on Apple M1, visit wiki. To install SCARlink within conda, first create a conda environment:

conda create -n scarlink-env python=3.8
conda activate scarlink-env

Set the priority of conda channels

conda config --add channels defaults
conda config --add channels bioconda
conda config --add channels conda-forge

Install essential R packages^*

conda install -c conda-forge r-seurat r-devtools r-biocmanager
conda install -c bioconda bioconductor-rhdf5 \
                     bioconductor-genomeinfodbdata \
                     bioconductor-chromvar \
                     bioconductor-motifmatchr \
                     bioconductor-complexheatmap

Install ArchR in the conda environment inside R

devtools::install_github("GreenleafLab/ArchR", ref="master", repos = BiocManager::repositories())

2. SCARlink installation

Download SCARlink from GitHub and install

git clone https://github.com/snehamitra/SCARlink.git
pip install -e SCARlink

3. Docker setup

Skip steps 1 and 2 if you want to run SCARlink inside a Docker container. Visit wiki for additional details. First install Docker and then run SCARlink inside container as follows

docker run -it --rm --memory=50g mitrasneha/scarlink:latest

Usage

SCARlink requires preprocessed scRNA-seq (normalized counts) in a Seurat object and scATAC-seq (non-binary tile matrix) in ArchR object. The cell names in both the objects need to be identical. Please refer to the example notebook for generating Seurat and ArchR objects.

1. Preprocessing

Run scarlink_preprocessing to generate coasssay_matrix.h5 to use as input to SCARlink.

scarlink_processing --scrna scrna_seurat.rds --scatac scatac_archr -o multiome_out

2. Running locally

For small data sets with few genes, run SCARlink sequentially on the gene set in the same output directory multiome_out, as before, for all the remaining steps. Note that celltype needs to be present in either scrna_seurat.rds or scatac_archr.

scarlink -o multiome_out -g hg38 -c celltype

2a. Running locally without cell type information during training and computing cell type scores afterwards

SCARlink can also be run without providing cell type information using -c. In that case SCARlink only computes the gene-level regression model and does not estimate cell-type-specific Shapley values.

scarlink -o multiome_out -g hg38 

The cell-type-specific Shapley scores can be computed later on by running scarlink with the -c parameter. In that case the previously estimated gene-regression models are used.

scarlink -o multiome_out -g hg38 -c celltype

2b. Running locally with different cell type grouping

SCARlink can be run again with a different cell type annotation. For example, more granular annotations. Both celltype and celltype_granular scores would be retained.

scarlink -o multiome_out -g hg38 -c celltype_granular

2c. Running on cluster by parallelizing over gene set

To speed up computation, SCARlink can be run on a cluster in parallel. By default it parallelizes over 100 cores, -np 100 when -np is not provided.

scarlink -o multiome_out -g hg38 -c celltype -p $LSB_JOBINDEX

3. Get FDR corrected gene-linked tiles

Get table with tile-level siginificance for each gene and celltype.

scarlink_tiles -o multiome_out -c celltype 

4. Chromatin potential

Chromatin potential analysis can be performed after running steps 1 and 2. Please refer to example notebook.

5. Plot SCARlink output

SCARlink can create visualizations for the chromatin accessibility around a gene (250kb upstream and downstream of the gene body) and the gene expression for a given gene, grouped by user-provided celltype. The visualizations can help understand the connection between the accessibility at tile-level to gene expression. Note that SCARlink can generate visualizations for different cell type groupings on the fly. Example visulizations are provided in a notebook. The visualizations can also be generated at the command line.

scarlink_tiles -o multiome_out -c celltype --genes GENE1,GENE2

^* The manuscript version of SCARlink uses Seurat v4 and ArchR 1.0.2. The new Seurat v5 installation is compatible with SCARlink but Seurat v5 objects are not yet supported. We will add it soon!