Awesome
Functional Annotation and Association Testing Pipeline
Pipeline to run variant effect prediction and rare-variant association tests for continuous phenotypes on UK Biobank exome sequencing (or other) data. It uses Snakemake to parallelize large parts of the analysis across chromosomes and phenotypes. Conda on a Linux system is the main requirement for getting started.
This pipeline was originally run on different virtual machines in local mode. A snakemake profile for slurm is provided at ./slurm.
Getting Started
Installing basic dependencies
After cloning the repository, input files need to be configured and the basic installation script ./install.sh needs to be run.
# download plink 1.9 and plink 2.0 , install snakemake in a separate local environment:
./install.sh
Input file configuration
Paths to phenotypes and covariates are defined in conf/config.yaml.
Phenotypes and covariates are expected to be given in separate tab-delimeted files with single header lines that correspond to the names of variables. The first column in both files is expected to be called iid and contain the individual identifiers of participants. The files can be gzipped. These files are expected to be already pre-processed and filtered. Typically, this involves quantile transformation for the phenotypes, and the removal of related individuals. The specific entries that need to be modified are covariates, phenotypes, and covariate_column_names.
Paths to the genotypes in PLINK (.bim, .bam, .bed) format are defined in a separate file, defined with samplesheet_plink in config.yaml. We provide an example file at conf/plinkfiles.tsv.
Note: The pipeline can't handle separate FIDs and IIDs, currently it is assumed that all FID == IID in your plink files (as is the case for UK Biobank data).
Some steps require a reference genome and gene annotation. We use Ensembl gene annotations. Make sure all chromosome names are matched to your other data (e.g remove the "chr"-prefix where necessary). Adjust the following entries in the config file: reference_fa and gene_annotation_gtf.
We used Ensembl 97 annotations available at ftp://ftp.ensembl.org/pub/release-97/gtf/homo_sapiens/. If you want to use later versions, make sure to also adjust ensembl_version in config.yaml as well.
Another input to the pipeline which we can't provide ourselves are the SpliceAI variant effect predictions for SNVs. These can be downloaded from Illumina directly here. Specifically, we used spliceai_scores.masked.snv.hg38.vcf.gz (v1.3). Make sure to adjust splice_ai_scores_snv_vcf in config.yaml.
Running rules
Basic setup rules are defined in ./Snakemake, whereas other rules are located in separate Snakemake modules within ./snake/. It is recommended to execute rules using ./run.sh. This script should be adjusted according to your cluster needs. By default, everything is run locally. To check if you have correctly configured your config file and initialize the work directory, run:
./run.sh setup_all
Users should adjust ./run.sh if they wish for it to submit jobs to their cluster. Some of the rules in ./snake/deepripe.smk require GPU support to perform predictions for RBP-binding.
The pipeline implements extensive variant annotation and association tests using functionally informed gene-based variant collapsing and kernel-based tests. As of March 2021, both the pipeline and the software it depends on to perform association tests (seak) are under active development. Forking is recommended.
Variant effect prediction
The rules needed to run variant effect prediction are evep_missense_proc_all for missense variants, and process_ensembl_vep_output_highimpact_all for protein loss of function variants (pLOF). The rule splice_ai_filter_and_overlap_with_genotypes_all overlaps all SNVs to published SpliceAI predictions. Finally, run_deepripe_vep_all runs DeepRiPe variant effect predictions for the binding of RBPs.
Association tests
To run a basic analysis using gene-based variant collapsing score tests for missense and pLOF variants, use the rule assoc_baseline_scoretest_all. This will also export (tiny!) HDF5-files containing the pre-processed burdens for all individuals which have complete covariates to work/gene_scores/indicator01/missense/all/ and work/gene_scores/indicator01/pLOF/all/.
To perform kernel-based tets with the sLRT, the follwing rules can be used: To run gene-based variant collapsing and kernel-based tests for for missense variants use rule assoc_missense_localcollapsing_all. To run gene-based variant collapsing and kernel-based tests for for splice-variants use rule assoc_spliceai_linw_all. To run kernel-based tests incorporating variant effect predictions for RBP-binding use rule assoc_deepripe_multiple_cholesky_all.
Citation
This code has been used in Identifying interpretable gene-biomarker associations with functionally informed kernel-based tests in 190,000 exomes by Remo Monti, Pia Rautenstrauch, Mahsa Ghanbari, Alva Rani James, Matthias Kirchler, Uwe Ohler, Stefan Konigorski & Christoph Lippert, Nature Communications, 2022.
Summary statistics for that study are available on the GWAS catalog, accession number 36088354.