Critical Assesment of Transcriptomic Deconvonvolution-snakemake pipeline
\`*-.
) _`-. Hi
. : `. .
Hey : _ ' \
.-=-. .--. ; *` _. `*-._
__ .' '. / " ) `-.-' `-.
_ .' '. / .-. \ / .-'\ ; ` `.
( \ / .-. \ / / \ \ / / ^ :. . \
\ `-` / \ `-' / \ `-` / . \ . : .-' .
`-.-` '.____.' `.____.' ' `+.; ; ' :
: ' | ; ;-.
https://ascii.co.uk/art/snake ; ' : :`-: _.`* ;
.*' / .*' ; .*`- +' `*'
`*-* `*-* `*-*'
https://ascii.co.uk/art/cat
Maintainer contact : Anna Vathrakokoili Pournara [email protected]
Software Authors ✍🏼 ✍🏼: Chichao Miao, Anna Vathrakokoili Pournara & Ozgur Beker
Snakemake Pipeline implementation 👨💻 : Ozgur Beker [email protected]
Many Thanks to Nadja Nolte for testing and further developping.
- This pipeline has been inspired from and extends the functionalities in deconv_benchmark, which was published at: Nature Communications as well as paper_deconvBenchmark which is published at:Genome Biology
If you are interested in our pipeline or you would like to include your method in the CATD pipeline please contact us or create a pull request.
Please use the issue tracker if you encounter a problem
The CATD pipeline is a benchmarking pipeline meant to facilitate the assessment of cell-type deconvolution methods(29 now) across different simulation scenarios in a standardised way. It also allows the deconvolution of real bulk samples with many user-unput parameters so that users can deconvolute their own in-house data following our proporsed guidelines. The pipeline includes:
- Pseudobulk generation methods that allow to create diverse bulk samples and compare deconvolution methods across different scenarios.
- 17 normalization methods implemented in the pipeline for the normalization of the input single-cell reference and the (pseudo)bulk samples
- 4 transformation methods implemented
- 9 DE tests for the selection of marker genes from single-cell reference data(Seurat)
- 29 deconvolution methods
- 7 metrics to assess the results when we test deconvolution methods on pseudo-bulks or when ground truth proportions from real data are available.
For more details check our preprint:
Vathrakokoili Pournara, A., Miao, Z., Beker, O. Y., Brazma, A. & Papatheodorou, I. Power analysis of cell-type deconvolution methods across tissues. http://biorxiv.org/lookup/doi/10.1101/2023.01.19.523443 (2023) doi:10.1101/2023.01.19.523443.
The github repository includes:
- The CATD pipeline with the driver snakefile and main pipeline scripts
/bin
.
├── basicSetup.sh #script for the basic set up
├── config.yaml #config file with all the user-defined parameters
├── getDag.sh #get DAG graphic aka a directed acyclic graph (DAG) of jobs where the edges represent dependencies
├── runPip.sh #script to run the pipeline
└── Snakefile #Driver script for the deconvolution assessment pipeline- All the modules used by the pipeline.
Each Module consists of an enviroment file, a snakefile with the rule(s) used in this module & the script to run the method e.g
/Modules/EpiDISH
.
├── env.yaml
├── EpiDISH_run.R
└── Snakefile-
A README file containing between other things :
-
Instructions/Demo for running self-reference, cross-reference & real bulk scenarios(tasks)
-
examples datasets deposited on google drive
-
Except for file format conversion and results generation, it is NOT recommended to run the pipeline on personal computers.
Memory: File format conversion and results generation parts of the pipeline require relatively less memory (enough to load the dataset into the session, ~4-5 GBs for single cell data of ~30k cells). The memory requirement for pseudo-bulk generation scales with cores provided, so numThreads*memData amount of memory should be allocated, with again a good baseline being ~4-5 GBs per thread allocated. For deconvolution with CDSeq, Bisque, CPM, DWLS, MuSiC, SCDC, TIMER, bseqsc and MOMF the memory requirement should be around the same as pseudo-bulk generation. For all other intermediate operations, minimal memory is needed. Note that the memory requirements will definitely change with the size of data, so users are recommended to check logs for specific runs/modules and scale the baseline values given in config.yaml accordingly.
Threads: It is highly recommended to run the pipeline with at least 16 threads to avoid long runtimes, especially for pseudo-bulk generation. For bseqsc and CIBERSORT, 3 should be the maximum amount of threads. For DWLS, threads should be equal (ideally) or smaller than the number of cell types. Allocating more threads will not cause any speed-ups for these methods.
For a general comparison of methods and their memory requirements, users can consult Supplementary Figure 9 which was generated with the default values in config.yaml.
There are no specifications on OS requirements, any OS which can run shell scripts within the pipeline and adhere to remaining requirements should work. The pipeline was tested on Linux with Ubuntu 20.04.
Mamba package manager (Mambaforge is recommended) should be installed, with instructions available at the respective repository.
-
Snakemake must be installed to run the pipeline, which can be achieved through Mamba with:
mamba create -n snakemake snakemake
Excluding the initial Snakemake install, conda environments for each module of the pipeline will be handled automatically by Snakemake / Mamba. To achieve this, the setup script will prompt for the absolute path of your conda.sh file. The only exception is the CIBERSORT.R file which is required for CIBERSORT and bseqsc deconvolution. Users should register and download the code manually, then place it under Modules/CIBERSORT and Modules/bseqsc respectively.
-
Download the pipeline to your directory of choice(also install git-lfs in your system before cloning, you can use conda)
git clone https://github.com/Functional-Genomics/CATD_snakemake.git -
Set up Conda environment with snakemake, pipeline strictly uses mamba for much faster environment setup.
mamba create -c conda-forge -c bioconda -n snakemake snakemake=7.14.0
#downgrade tabulate=0.8.10 : mamba install tabulate=0.8.10
-
Build the pipeline.
make
If you are running the pipeline for the first time you should use few samples to run in order for all the environments to be installed fast( each module/method in our pipeline has its own environment.
-
Place the input file in the
Inputdirectory that has been created
follow the link to the example file - use either XinY_4Decon.h5ad or XinY_4Decon_seurat.rds ) -
Adjust settings in
config.yamllike suggested: (familiarise yourself with the config.yaml file and all the parameters that you can choose)
Specifically:
sampleNames: #Enter the part BEFORE file extension (ex. 'data.h5ad' should be entered as 'data')
- XinY_4Decon # in this case
pbulkParam:
pbulkMode: 2 #Different mode for building pbulks
cellCount: 100 #How many cells should be used to create a bulk sample
nSamples: 10 #How many different samples should be generated #first time only generate one sample
propVar: 500 #Variances inbetween proportions, check vignette (EFFECTIVE IN MODE 2 ONLY, enter negative value to switch to min max prop mode)
sampleCT: 0 #Sampling for cell types allowed or not (1 or 0) (EFFECTIVE IN MODE 2 ONLY) -
(Optional) Run
getDag.shto generate the updated DAG after adjusting config.
This will output a dag.png where you can observe the jobs that will be executed -
(Optional) If on cluster set up cluster profile if you haven't, instructions available here.
-
(Optional) Dry run:
conda activate snakemakesnakemake -n -
Run the pipeline using
bsub < runPip.sh(LSF) or throughsnakemake --use-conda --cores [N]if on local.
The steps(1-2-3) should take few seconds to complete.
If all the environments are included in the first run it will take around 25-30 minutes,roughly 1min/environment is a good estimate.(step 9)
When the job(step 9) finishes all the environments will have been installed and users are ready to run actual experiments and test the pipeline
IMPORTANT:As mentioned above, if running the pipeline for the first time, use very few samples (e.g 10) only as all the environments are installed throughout the workflow.
Uses one single-cell reference to generate both the pseudobulks and the reference for deconvolution benchmarking. Important assumptions:
- The cell types should be annotated in the .h5ad or seurat object in the metadata strictly under the column name cellType (written in camelCase).
- In the metadata, there needs to be cellID and sampleID columns, strictly under those names. If they don't exist, simply assign rownames to those columns.
Input/{sampleName}.h5ad
OR
Input/{sampleName}_seurat.rds
- Evaluation of selected methods based on selected metrics in config, found in: Metrics/
- Plots describing the metrics, found in: Plots/
- Resource usage of each step, summarized in: {sampleName}_benchmarks_summarized.png
- Individual benchmarks for steps, found in: Benchmarks/
- Predictions produced by methods, found in: Results/
- Start from step 4 from above
- Edit again the 'config.yaml' file : This time you can change/tune the parameters e.g increase the nSamples to 100 or more , increase cellCount to 1000 etc based on the pseudobulks you want to create, you can also edit tranformation/normalization/deconvolution methods parameters
- step 6 and 8 from above are optional again but recommended.
- run the self-reference task: step 9
Uses two single-cell references to generate the pseudobulks and references for deconvolution benchmarking. Important assumptions are:
- All assumptions in the self-reference part
- The levels (i.e unique list) of cell types must be the same in both references provided.
Input/Cell_splits/{sampleName}_gen.rds (Will be used to generate psuedobulks)
Input/Cell_splits/{sampleName}_C0.rds (Will be used to generate references)
(a)
- BaronHuman_Sege_cross_gen.rds #Segerstope dataset from where pseudobulks will be created
- BaronHuman_Sege_cross_C0.rds #BaronHuman dataset from where the reference will be generated\
(b)
- Sege_BaronHuman_cross_gen.rds #BaronHuman dataset from where pseudobulks will be created
- Sege_BaronHuman_cross_C0.rds #Segerstope dataset from where the reference will be generated
Same as self-reference.
-
Same as self-reference, except in 4th step, add the input files in the following directory:
Input/Cell_splits
Make sure that the inputs conform to the standards written in the 'Inputs' section above.
-
Edit the config.yaml file similarly to self-reference * if you would like to run two or more tasks at the same time(for example use both example (a) and (b) you can add a line in the sampleNames in the
config.yaml:**sampleNames**: #Enter the part BEFORE file extension (ex. 'data.h5ad' should be entered as 'data') - BaronHuman_Sege_cross - Sege_BaronHuman_cross -
Then continue with the 6th step.
Uses one reference single cell matrix with user-defined bulks and known proportions for deconvolution benchmarking. Assumptions are:
- All assumptions in the self reference part
- The rownames (cell types) in the proportions should be the same as the cell types annotated in the reference
Input/{sampleName}.h5ad / Input/{sampleName}_seurat.rds
OR
Input/Cell_splits/{sampleName}_C0.rds
AND
Input/Psuedobulks/{sampleName}_pbulks.rds
Input/Psuedobulks/{sampleName}_props.rds
The first two options will only use half of the data to generate references. The third will use all of the data to generate the reference. Alongside the reference, you need to input the pseudo-bulks inside the folder specified under those names.
FinotelloB_WilkR_props.rds #ground truth proportions table
FinotelloB_WilkR_pbulks.rds#real bulk matrix (Finotello et al)
FinotelloB_WilkR_C0.rds#WilkR single-cell expression matrix
Same as self-reference.
Same as self-reference, except in the 4th step, note the following directory:
Input/Psuedobulks
This is where the bulks and the ground truth proportions should go(FinotelloB_WilkR_pbulks.rds & FinotelloB_WilkR_props.rds). Also, enable realBulk in config.yaml
If you wish to use all the data for the reference, go to:
Input/Cell_splits
and place the reference file here under the name {sampleName}_C0.rds(FinotelloB_WilkR_C0.rds).
If you want to run your own real bulk data with a single-cell reference but you do not have ground truth proportions enable realBulk-noProp in config.yaml in this case you will not have a {sampleName}_props.rds and the pipeline will not calculate any metrics and will output only a heatmap
Same as self-reference.
- 1 Heatmap plot that aids to compare the results of different methods, found in: Plots/
- Resource usage of each step, summarized in: {sampleName}_benchmarks_summarized.png
- Predictions produced by methods, found in: Output/
Users need to download the CIBERSORT.R code manually and place it within the modules (Modules/CIBERSORT and Modules/bseqsc) to run these two methods.
Please register in the website https://cibersortx.stanford.edu/register.php first and access the CS Archive gallery to download CIBERSORT.R
The image below is generated by selecting only a couple of methods. If all methods are selected, snakemake cannot generate a pretty DAG as all deconvolution methods run in parallel.
The full list of jobs looks like this:
job count min threads max threads
-------------------------- ------- ------------- -------------
Autogenes_run 1 1 1
Bisque_run 1 1 1
CDSeq_run 1 16 16
CIBERSORT_run 1 3 3
CPM_run 1 16 16
CellmixInstall 1 1 1
DCQ_run 1 1 1
DSA_run 1 1 1
DWLS_run 1 16 16
DeconRNASeq_run 1 1 1
EPIC_run 1 1 1
EpiDISH_run 1 1 1
FARDEEP_run 1 1 1
Momf_run 1 1 1
MuSiC_run 1 1 1
NNLS_run 1 1 1
OLS_run 1 1 1
OmniInstall 1 1 1
RLR_run 1 1 1
SCDC_run 1 1 1
TIMER_run 1 1 1
all 1 1 1
bseqsc_run 1 3 3
convertAnndata 1 1 1
debCAM_C1 1 1 1
debCAM_marker 1 1 1
deconf_run 1 1 1
dtangle_run 1 1 1
elasticNET_run 1 1 1
exploreBenchmarks 1 1 1
exploreResults 1 1 1
generatePsuedobulks 1 32 32
generateReferences 1 32 32
lasso_run 1 1 1
proportionsInAdmixture_run 1 1 1
ridge_run 1 1 1
scaleTransform_T_C 1 32 32
splitCells 1 1 1
ssFrobenius_run 1 1 1
ssKL_run 1 1 1
sumRun 1 1 1
vioplotResults 1 1 1
visualizeResults 1 1 1
total 43 1 32
Use make clean to delete all outputs from a workflow. Keep in mind that this will delete ALL outputs including results.
