Watershed-SV

Watershed-SV extends Watershed to model the impact of rare SVs (DUP, DEL, DUP-CNV, DEL-CNV, INV, INS) on nearby gene expressions outlier. For running Watershed model, please refer to Watershed GitHub. This repository contains:

1. the pipeline and associated scripts used for generating structural variations (rare and common) annotations with respect to nearby genes.
2. the scripts for generating expression outliers.
3. the approach to merge annotations AND the expression outliers, and finally format data into desired format for evaluate_watershed.R and predict_watershed.R.

Instructions for collecting annotations.

We provide a default pipeline for collecting annotations to run Watershed-SV. This is the set of annotations we used to generate the benchmark on the paper figure 4. You are welcome to come up with annotations on your own but the performance is not guaranteed. Benchmarks are required to verify actual performance of a custom model.

Installation

Watershed-SV annotation pipeline is available on conda: conda install -c dnachun -c conda-forge -c bioconda watershed-sv=0.1.22 conda install -c bioconda ensembl-vep We however recommend installation with pixi for most straightforward application. pixi global install -c dnachun -c conda-forge -c bioconda watershed-sv=0.1.22 pixi global install -c bioconda ensembl-vep Now, we also need to download annotations for running Watershed-SV, and VEP which is internally called. To download annotations: `` To download VEP cache: we recommend downloading: wget https://ftp.ensembl.org/pub/current_variation/indexed_vep_cache/homo_sapiens_vep_113_GRCh38.tar.gz To replicate the environment for collecting annotations, see: WatershedSV.yml.

Parameter breakdown:

1. -p | --pipeline: which pipeline to use, select from population, smallset. If data is of sufficient size, ie > 100, select population, allowing for option --filter-ethnicity, --filter rare. Otherwise, select smallset.
2. -v | --input-vcf: input vcf file, it has to have at least 1 sample column. We only consider SVTYPEs: DUP, DEL, DUP_CNV, DEL_CNV, CNV, INS, INV.
3. -f | --filters: if variant record in vcf have these filters, keep for further analysis.
4. -k | --flank: how much flanking up and downstream of genes to consider. usually use 100000, 10000.
5. -r | --rareness: rareness, if --filter-rare == True, then this is the MAF threshold to set to filter for rare variants, 0.01 recommended or lower.
6. -l | --liftover-bed: if you have a crossmap/liftover SV coordinate you want to use, ie, if VCF is in older build, you lifted over coordinates to HG38, then provide the bed file in addition to original VCF to convert coordinates.
7. -o | --outdir: output directory name for annotations
8. -b | --genome-bound-file: a file depicting the chromosome/contig name, start and end coordinates.
9. -g | --gencode-genes: gencode transcript model file.
10. -c | --vep-cache-dir: vep_cache_dir for running vep annotations. we recommend setting up vep offline to run our pipeline smoothly.
11. -a | --metadata: metadata file for filtering ethnicity. In our case, training data is GTEx, we used GTEx metadata file from dbGaP.
12. -e | --filter-ethnicity: filter by ethnicity? GTEx relic, True to only train on EUR individuals.
13. -i | --filter-rare: filter rare variants if using population model

Evaluating Watershed-SV model against WGS-only model

Using Watershed-SV model learned on a reference dataset to prioritize gene-SV pairs in an independent test dataset.