A conda environment containining all the required packages to run the pipeline is available (bigsnprenv_1911). Current versions for main packages (bigsnpr/bigstatsr) are reported below.
bigsnpr: 1.10.8
bigstatsr: 1.5.6git clone https://github.com/davidebolo1993/prs_pipeline
cd prs_pipelineThe pipeline consists in a single r script based on ldpred2.
#srun --nodes=1 --tasks-per-node=1 --partition cpu-interactive --cpus-per-task 8 --pty /bin/bash #this is to enter a computational node
conda activate bigsnprenv_1911
Rscript scripts/prs.r --help
#During startup - Warning message:
#Setting LC_CTYPE failed, using "C"
#Loading required package: bigstatsr
#Attaching package: 'dplyr'
#The following objects are masked from 'package:data.table':
# between, first, last
#The following objects are masked from 'package:stats':
# filter, lag
#The following objects are masked from 'package:base':
# intersect, setdiff, setequal, union
#R version 4.1.3 (2022-03-10)
#Platform: x86_64-conda-linux-gnu (64-bit)
#Running under: CentOS Linux 8 (Core)
#Matrix products: default
#BLAS/LAPACK: /project/alfredo/conda_envs/bigsnprenv_1911/lib/libopenblasp-r0.3.20.so
#locale:
# [1] LC_CTYPE=C LC_NUMERIC=C
# [3] LC_TIME=en_US.UTF-8 LC_COLLATE=en_US.UTF-8
# [5] LC_MONETARY=en_US.UTF-8 LC_MESSAGES=en_US.UTF-8
# [7] LC_PAPER=en_US.UTF-8 LC_NAME=C
# [9] LC_ADDRESS=C LC_TELEPHONE=C
#[11] LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C
#attached base packages:
#[1] tools stats graphics grDevices utils datasets methods
#[8] base
#other attached packages:
#[1] rjson_0.2.21 optparse_1.7.1 dplyr_1.0.9 fmsb_0.7.3
#[5] stringr_1.4.0 magrittr_2.0.3 data.table_1.14.2 bigsnpr_1.10.8
#[9] bigstatsr_1.5.6
#loaded via a namespace (and not attached):
# [1] Rcpp_1.0.8.3 bigsparser_0.6.1 pillar_1.7.0 compiler_4.1.3
# [5] bigparallelr_0.3.2 rngtools_1.5.2 iterators_1.0.14 digest_0.6.29
# [9] lifecycle_1.0.1 tibble_3.1.7 gtable_0.3.0 lattice_0.20-45
#[13] doRNG_1.8.2 pkgconfig_2.0.3 rlang_1.0.2 Matrix_1.4-1
#[17] foreach_1.5.2 DBI_1.1.2 cli_3.3.0 flock_0.7
#[21] parallel_4.1.3 bigassertr_0.1.5 generics_0.1.2 vctrs_0.4.1
#[25] getopt_1.20.3 grid_4.1.3 tidyselect_1.1.2 cowplot_1.1.1
#[29] glue_1.6.2 R6_2.5.1 fansi_1.0.3 ggplot2_3.3.6
#[33] purrr_0.3.4 scales_1.2.0 codetools_0.2-18 ellipsis_0.3.2
#[37] assertthat_0.2.1 colorspace_2.0-3 utf8_1.2.2 stringi_1.7.6
#[41] munsell_0.5.0 doParallel_1.0.17 crayon_1.5.1
Usage: prs.r [options]
Options:
-m MODEL, --model=MODEL
model to use, either automatic or grid. Default to automatic
-i INPUT, --input=INPUT
.bed file (with companion .bim and .fam files) or (indexed) .bgen file [required]
-s SUMMARY, --summary=SUMMARY
GWAS summary stats or pre-calculated .tsv (with header) containing beta scores [required]
--summarycols=SUMMARYCOLS
.json file defining columns to use
-p PHENOTYPE, --phenotype=PHENOTYPE
.tsv file with phenotype (and also covariates, if any)
--heritability=HERITABILITY
heritability, if known
-o OUTPUT, --output=OUTPUT
output prefix [required]
--threads=THREADS
computing threads [1]
--correlation=CORRELATION
the correlation matrix provided as a pre-computed .rds object
-h, --help
Show this help message and exit
#run with test data
Rscript prs.r --input test/public-data3.bed --summary test/public-data3-sumstats.txt --summarycols test/public-data3.json --threads 8 --output test/output/public-data3Available models are "grid" and "automatic". Default to "automatic". Detailed informations on each model are available in the ldpred2 tutorial introduction and the paper.
The pipeline accepts either a unique .bed file (with matching .bim/.fam) or a unique (indexed) .bgen file. Further informations on these file formats are available here. Examples are provided in the test folder.
Companion scripts are released in the scripts folder to generate a single input .bed/.bim/.fam or .bgen if not available already (from chromosome-specific .bed/.bgen files). For .bgen files, use this script. Positional arguments are a directory containing (indexed) .bgen files, the output, merged, .rds object and the number of computational threads. The script also generates companion .bk, .bgen and .bgi files - the (indexed) .bgen files is actually empty but can be given as input to the pipeline that will load the informative .rds/.bk files instead. A sbatch script that run the above mentioned R script is also available - adjust the parameters accordingly to the input files. For .bed files, those can be merged using plink2. A dedicated sbatch script to be used as template is available in the script folder.
The pipeline can be run on .bgen files in parallel on a cluster - one node for each chromosome using this script. .bgen files must be provided in file-of-file-names format (.fofn) and the other parameters in the scipt should be set accordingly to the experimental setting. The sbatch script generates one output folder for each chromosome. Results for each chromosome can be summed up using this simple bash script that produces a all.prs.tsv table in the output folder (the same used above). Once the 2 scripts have been modified according to the experimental setting, the bash script can be used to run the sbatch array and sum the downstream results.
A standard summary statics file or a file with pre-computed scores can be given as input. Summary statistics file format from GWAS catalog is described here. PGS scoring file format from PGS catalog is described here. An example summary statistic file is available in the test folder.
While the other inputs to the pipeline are standard files, the input .json file is used to handle summary statistics and matrices with pre-computed scores generated with different methods - so that the pipeline actually knows which column refers to which value. The input .json file follows standard json specification, where, for each "key":"value" pair, users must specify:
Table containing the phenotype as a third column (can be either categorical or continuous) and possible covariates afterwards. First and second columns can be anything (family and individual ID, for instance). An example is provided in the test folder.
Value of the heritability, if known. It will be calculated otherwise.
If a correlation matrix for matching summary statistics and .bed/.bim/.fam files is available (because stored by a previous run of the pipeline), skip the computation of the correlation matrix
The pipeline stores the correlation matrix and the ld as R objects (.rds format) - named with the output prefix followed by .corr.rds and .ld.rds respectively. The correlation matrix can be provided as input to skip the computation of the correlation matrix itself - which is time consuming.
When using the automatic model without a table of phenotypes as input, the pipeline returns beta scores and predictions from the automatic model - named with the output prefix followed by .auto.beta_scores.tsv and .auto.prs.tsv respectively
When using the automatic model with a table of phenotypes as input, the pipeline further tests the generated predictions on the provided phenotype. Performances are stored as summary files - named with the output prefix followed by .auto.summary.tsv and .auto.summary.rds respectively
Grid model can only be run with a table of phenotypes as input.
The pipeline returns beta scores and predictions from the grid model - named with the output prefix followed by .auto.grid_scores.tsv and grid.prs.tsv respectively - as well as performances of tested models as a tab-separated value file - named with the output prefix followed by .grid.allsummary.tsv - and those of the best model as an additional R object - named with the output prefix followed by .grid.bestsummary.rds.
A sbatch script is provided to run the pipeline using an entire computational node. Parameters in the sbatch file should be adjusted coherently.