Skip to content


Repository files navigation

Open Target Genetics fine-mapping pipeline

Fine-mapping pipeline for Open Targets Genetics. In brief, the method is:

  1. Detect independent loci across the summary stat file using either (i) GCTA-cojo and a given plink file as an LD reference, (ii) distance based clumping. Method specified with --method argument.
  2. If --method conditional, for each independent locus condition on all other surrounding loci (configurable with cojo_wind).
  3. Perform approximate Bayes factor credible set analysis for each independent locus.

For FinnGen, we incorporate each new release by directly taking the SuSIE fine-mapping outputs from FinnGen to determine top loci. Each time this is done, we have to be careful to NOT to run GCTA fine-mapping on FinnGen sumstats, since their SuSIE fine-mapping is superior (and we don't have a good LD reference).


  • Spark v2.4.0
  • GCTA (>= v1.91.3) must be available in $PATH
  • conda
  • GNU parallel

Setup environment

git clone
cd ~/genetics-finemapping
. ~/.profile # Reload profile so that conda works
conda env create -n finemap --file environment.yaml

Configure pipeline

Many of the pipeline parameters must first be specified in the analysis config file: configs/analysis.config.yaml

Run a single study

A single study can be fine-mapped using the single study wrapper

# Edit config file (this needs selecting with --config_file arg)
nano configs/analysis.config.yaml

# View args
$ python finemapping/ --help
usage: [-h] --pq <file> --ld <file> --config_file
                               <file> --type <str> --study_id <str> --chrom
                               <str> [--phenotype_id <str>]
                               [--bio_feature <str>] --method
                               [conditional|distance] --pval_threshold <float>
                               --toploci <file> --credset <file> --log <file>
                               --tmpdir <file> [--delete_tmpdir]

optional arguments:
  -h, --help            show this help message and exit
  --pq <file>           Input: parquet file containing summary stats
  --ld <file>           Input: plink file to estimate LD from
  --config_file <file>  Input: analysis config file
  --type <str>          type to extract from pq
  --study_id <str>      study_id to extract from pq
  --chrom <str>         chrom to extract from pq
  --phenotype_id <str>  phenotype_id to extract from pq
  --bio_feature <str>   bio_feature to extract from pq
  --method [conditional|distance]
                        Which method to run, either with conditional analysis
                        (gcta-cojo) or distance based with conditional
  --pval_threshold <float>
                        P-value threshold to be considered "significant"
  --run_finemap         If True, then run FINEMAP
  --toploci <file>      Output: top loci json file
  --credset <file>      Output: credible set json file
  --finemap <file>      Output: finemap snp probabilities file
  --log <file>          Output: log file
  --tmpdir <file>       Output: temp dir
  --delete_tmpdir       Remove temp dir when complete

Note: The capability of running FINEMAP has been used but not extensively tested.

Running the pipeline

Step 1: Prepare input data

Prepare summary statistic files, including "significant" window extraction to reduce input size.

Prepare LD references in plink bed|bim|fam format, currently using UK Biobank downsampled to 10K individuals and lifted over to GRCh38.

Download to local machine.

mkdir -p $HOME/data/ukb_v3_downsampled10k
gsutil -m rsync gs://open-targets-ukbb/genotypes/ukb_v3_downsampled10k/ $HOME/data/ukb_v3_downsampled10k/

# To optimise the GCTA conditioning step, it helps to split the UKB reference panel
# into smaller, overlapping "sub-panels". This is because (I believe) GCTA loads
# in the index for the whole chromosome (*.bim file) to determine which SNPs match.
time python --path $HOME/data/ukb_v3_downsampled10k/ukb_v3_chr{chrom}.downsampled10k

To process only new data, we should only download "significant windows" from new studies. The pipeline runs fine-mapping for all available significant windows.

cd $HOME/genetics-finemapping
mkdir -p $HOME/genetics-finemapping/data/filtered/significant_window_2mb/gwas/
mkdir -p $HOME/genetics-finemapping/data/filtered/significant_window_2mb/molecular_trait/

# Identify new GWAS studies: list all studies, order by date
gsutil ls -l gs://genetics-portal-dev-sumstats/filtered/significant_window_2mb/gwas/*/_SUCCESS | sort -k 2 >
# Manually copy the list of new file paths into "gwas_to_download.txt"
# Copy the .parquet folder names, not the _SUCCESS file names

# Download only new files into the destination folder
cat gwas_to_download.txt | gsutil -m cp -r -I $HOME/genetics-finemapping/data/filtered/significant_window_2mb/gwas/

# Identify new molecular trait studies: list studies, order by date
gsutil ls -l gs://genetics-portal-dev-sumstats/filtered/significant_window_2mb/molecular_trait/*/_SUCCESS | sort -k 2 >
# Manually copy the list of new file paths into "moltrait_to_download.txt"
# Copy the .parquet folder names, not the _SUCCESS file names

# Download only new files into the destination folder
cat moltrait_to_download.txt | gsutil -m cp -r -I $HOME/genetics-finemapping/data/filtered/significant_window_2mb/molecular_trait/

#gsutil -m rsync -r gs://genetics-portal-dev-sumstats/filtered/significant_window_2mb/gwas/ $HOME/genetics-finemapping/data/filtered/significant_window_2mb/gwas/
#gsutil -m rsync -r gs://genetics-portal-dev-sumstats/filtered/significant_window_2mb/molecular_trait/ $HOME/genetics-finemapping/data/filtered/significant_window_2mb/molecular_trait/
#gsutil -m rsync -r gs://genetics-portal-dev-sumstats/filtered/significant_window_2mb/molecular_trait_new/ $HOME/genetics-finemapping/data/filtered/significant_window_2mb/molecular_trait/

# Remove FinnGen GWAS (if any), since we don't run fine-mapping for them
rm -r $HOME/genetics-finemapping/data/filtered/significant_window_2mb/gwas/FINNGEN*

# Remove extra files that can screw up data loading later
find /home/js29/genetics-finemapping/data/filtered/significant_window_2mb -name "*_SUCCESS" | wc -l
find /home/js29/genetics-finemapping/data/filtered/significant_window_2mb -name "*_SUCCESS" -delete

Step 2: Prepare environment

# Activate environment
source activate finemap

# Set spark paths
export PYSPARK_SUBMIT_ARGS="--driver-memory 80g pyspark-shell"
#export SPARK_HOME=/home/ubuntu/software/spark-2.4.0-bin-hadoop2.7

Step 3: Make manifest file

The manifest file specifies all analyses to be run. The manifest is a JSON lines file with each line containing the following fields:

  "type": "gwas",
  "study_id": "NEALE2_50_raw",
  "phenotype_id": null,
  "bio_feature": null,
  "chrom": "6",
  "in_pq": "/home/ubuntu/data/sumstats/filtered/significant_window_2mb/gwas/NEALE2_50_raw.parquet",
  "in_ld": "/home/ubuntu/data/genotypes/ukb_v3_downsampled10k_plink/ukb_v3_chr{chrom}.downsampled10k",
  "out_top_loci": "/home/ubuntu/results/finemapping/output/study_id=NEALE2_50_raw/phenotype_id=None/bio_feature=None/chrom=6/top_loci.json.gz",
  "out_credset": "/home/ubuntu/results/finemapping/output/study_id=NEALE2_50_raw/phenotype_id=None/bio_feature=None/chrom=6/credible_set.json.gz",
  "out_log": "/home/ubuntu/results/finemapping/logs/study_id=NEALE2_50_raw/phenotype_id=None/bio_feature=None/chrom=6/logfile.txt",
  "tmpdir": "/home/ubuntu/results/finemapping/tmp/study_id=NEALE2_50_raw/phenotype_id=None/bio_feature=None/chrom=6",
  "method": "conditional",
  "pval_threshold": 5e-08

Note that the wildcard {chrom} can be used in in_ld field.

The manifest file can be automatically generated using:

cd ~/genetics-finemapping

# Edit the Args and Paths in ``

# Reads variants filtered for p value, and outputs a single json record in
# tmp/filtered_input for each study/chromosome combination with at least one
# significant variant. Takes a couple of minutes for 200 GWAS.
time python

# Creates manifest file, one job per study/chromosome. Output path `configs/manifest.json.gz`

Step 4: Run pipeline

mkdir logs
tmux   # So run continues if connection is lost

# Set number of cores based on machine size used, then run all commands
time bash $NCORES | tee logs/run_pipeline.out.txt

# Commands can be regenerated and run separately if needed
#python --quiet
#time zcat commands_todo.txt.gz | shuf | parallel -j $NCORES --bar --joblog logs/ | tee logs/run_pipeline.out2.txt 2>&1

# Exit tmux with Ctrl+b then d

The above command will run all analyses specified in the manifest using GNU parallel. It will create two files commands_todo.txt.gz and commands_done.txt.gz showing which analyses have not yet/already been done. The pipeline can be stopped at any time and restarted without repeating any completed analyses. You can safely regenerate the commands_*.txt.gz commands whilst the pipeline is running using python --quiet.

If you get this error: ModuleNotFoundError: No module named 'dask' then I've solved it just by deactivating conda and reactivating. This seems to happen especially when using tmux... I'm not sure why.

Step 5: Process the results

rm -r /home/js29/genetics-finemapping/tmp/*

# Combine the results of all the individual analyses
# This step can be slow/inefficient due to Hadoop many small files problem
# You are likely to get out of memory errors if you don't increase the java
# heap space available to Spark with PYSPARK_SUBMIT_ARGS.
export PYSPARK_SUBMIT_ARGS="--driver-memory 80g pyspark-shell"
time python
# Make a note as to what this finemapping run contained. E.g.:
echo "Run with new GTEx sQTL dataset, and updated GWAS catalog studies. Re-ran the 8 studies that had flipped betas previously (IBD and lipids)." > results/README.txt

# Copy the results to GCS
version_date=`date +%y%m%d`
bash $version_date

Number of top_loci raw: 1,623,534 Number of top_loci after dups removed: 1,541,938 Number of top_loci in previous version: 689,726

Number of credset rows raw: 44,180,640 Number of credset rows after dups removed: 40,910,064

Step 6: Merge with previous fine-mapping results

Steps like the below are needed if we are adding on to previous fine-mapping results, rather than recomputing everything. We assume that studies for which fine-mapping has been run are not present in the previous fine-mapped results that we are merging onto, otherwise we may get duplicates.

# Copy down previous fine-mapping results into temp folder
mkdir -p finemapping_to_merge/220113_merged
gsutil -m rsync -r gs://genetics-portal-dev-staging/finemapping/220113_merged finemapping_to_merge/220113_merged

mkdir -p finemapping_to_merge/$version_date/
cp -r results/* finemapping_to_merge/$version_date/

# Merge all old finemapping results with new
mkdir -p finemapping_merged
time python --prev_results finemapping_to_merge/220113_merged --new_results finemapping_to_merge/$version_date/ --output finemapping_merged | tee finemapping_merged/merge_results.log

# If adding new FinnGen results, then skip this step
echo "Merged fine-mapping results from 220113_merged + 220224" > finemapping_merged/README.txt
gsutil -m rsync -r $HOME/genetics-finemapping/finemapping_merged/ gs://genetics-portal-dev-staging/finemapping/${version_date}_merged

# If there are all-new FinnGen results, then pass the --remove_previous_finngen flag.
mkdir -p finemapping_merged_w_finngen
time python --prev_results finemapping_merged --new_results finngen/results/ --output finemapping_merged_w_finngen --remove_previous_finngen | tee finemapping_merged_w_finngen/merge_finngen.log

echo "Merged fine-mapping results from 220113_merged + 220224 + FinnGen R6. Removed 3 studies with bad data: GCST007236, GCST007799, GCST007800." > finemapping_merged_w_finngen/README.txt
gsutil -m rsync -r $HOME/genetics-finemapping/finemapping_merged_w_finngen/ gs://genetics-portal-dev-staging/finemapping/${version_date}_merged

Other notes

I did a test run on two different VM instances where I fine-mapped 15 GWAS. One VM had a balanced persistent disk (200 Gb), one had an SSD (200 Gb). Otherwise they both were N2-standard-8 configurations. The result was that the SSD version took about 4% longer than the standard disk. I did not try with a local SSD, but I suspect that the disk makes no difference, since the pipeline is CPU-bound.

Useful commands
# Parse time taken for each run
grep "Time taken" logs/study_id=*/phenotype_id=*/bio_feature=*/chrom=*/logfile.txt
ls -rt logs/study_id=*/phenotype_id=*/bio_feature=*/chrom=*/logfile.txt | xargs grep "Time taken"

# List all
ls logs/study_id=*/phenotype_id=*/bio_feature=*/chrom=*/logfile.txt
  • Currently fails for sex chromosomes
    • Need to replace X with 23 in plink file or when specifying gcta command
    • Need to impute sex in plink file for X for cojo to work
  • Manifest NAs must be represented with "None"
  • P-value threshold is specified in Set to 5e-8 for GWAS, and (0.05 / num_tests) for mol trait


# The below steps were used when we found duplicate top_loci, which was due
# to duplicated lines in the eQTL catalogue ingest. This has since been fixed,
# so the below should not be needed.
# Concatenate together all top_loci and credset files
time find output -name "top_loci.json.gz" | while read -r file; do zcat -f "$file"; done | gzip > top_loci.concat.json.gz &
time find output -name "credible_set.json.gz" | while read -r file; do zcat -f "$file"; done | gzip > credible_set.concat.json.gz

# Remove duplicates
# This should only be necessary because when we last ingested eQTL catalogue
# I failed to remove duplicate rows first.
time zcat top_loci.concat.json.gz | sort | uniq | gzip > top_loci.dedup.json.gz &
time zcat credible_set.concat.json.gz | sort | uniq | gzip > credible_set.dedup.json.gz

time python