atlas-deconvolution-references

Snakemake workflow to generate single-cell reference dataset for bulk RNA-seq deconvolution in Atlas

It contains rules for:

• SeuratObject creation
• Spliting SeuratObjects into tissues
• Reducing cell type label granularity
• Creation of references for deconvolution tools
• UMAP plots for quality control
• Creating summary file of references

Prerequisites

• Snakemake
• A batch scheduler, LSF or SLURM
• Set up configuration variables at run_deconvolution_reference_creation.sh
• Config.yaml with sc experiment accessions as input and tissues that are excluded for accessions (look at example.yaml how that should look like)
• Download ontology files for UBERON and CL ontology and put in folder files

Run pipeline

bash run_deconvolution_reference_creation.sh

This file contains:

• definition of variables
• runs snakemake like this:

snakemake \
  -s $SNAKEFILE  \
  --profile [cluster profile 'lsf' or 'slurm'] \
  --rerun-incomplete \
  --retries 3 \ # if job fails try with more memory
  --use-conda --conda-frontend mamba --keep-going \
  --config atlas_prod=[path/to/atlas_prot_dir]\
  anndata_prod=[path/to/anndata_prot_dir]\
  deconv_ref=[path/to/desination/of/reference_library] \

For SLURM profile, add --slurm.

• checks if references for organims parts are generated more than once and returns duplicated organism parts which can then be included in config.yaml.

example: For these UBERON ids multiple references exist:

UBERON_0001159

If this message is shown it is neccessary to go look at /UMAP/UBERON_0001159_*.png and decide Then either:

• delete in $DECONV_REF the three references that belong to the reference we do not want the reference from and include UBERON_id_scxa_accession in config.yaml file under "exclude_tissues_from_accessions:" OR
• decide which reference to delete, add them to config.yaml and reurn whole generation of deconvolution reference library

Output

The output of this pipeline are :

• deconvolution reference lib with three files per scxa accession + organism part (CO_scaled, C1 and phenData)
• one summary file of all the references created
• one png file per scxa accession + organism part with two UMAP plots (old + new cell type labels) for quality control

Contributions

The following scripts are taken from the CATD_snakemake pipeline which benchmarks deconvolution methods:

• scaleTable.R
• genRef_1.R