SLALOM

SLALOM (suspicious loci analysis of meta-analysis summary statistics) is a summary statistics-based QC method that identifies suspicious loci for meta-analysis fine-mapping by detecting association statistics outliers based on local LD structure. SLALOM only takes GWAS summary statistics and ancestry-matched external LD reference (e.g., gnomAD) as input and predicts whether each locus shows a suspicious pattern that called into question fine-mapping accuracy. The outlier detection was built upon the simplified version of the DENTIST method.

Analysis and figure generation code for Kanai, M. et al. (2022) is available here. Fine-mapping pipeline is available here.

What's new: Hail-free

SLALOM is now a pip-installable package with a slalom command-line tool and needs no Hail or Spark at runtime. The gnomAD LD BlockMatrix is read directly from cloud storage in pure Python via ldcov, and every annotation that used to read a Hail .ht (allele alignment, CUPs, VEP consequence, gnomAD frequency) now reads a Parquet reference table instead. See CHANGELOG.md for the full list of changes from the original single-file slalom.py.

Installation

pip install slalom-qc

The distribution is published on PyPI as slalom-qc; the import package and CLI are both slalom (import slalom, slalom --help).

Python ≥ 3.9. Reading LD or reference tables from Google Cloud Storage uses gcsfs (installed automatically); reading a BlockMatrix from AWS S3 (e.g. Pan-UKB) needs the s3 extra: pip install "slalom-qc[s3]".

Usage

slalom \
    --snp example/example.snp \
    --out example/example.slalom.txt \
    --out-summary example/example.summary.txt \
    --reference-genome GRCh38 \
    --align-alleles \
    --annotate-consequence \
    --annotate-cups \
    --annotate-gnomad-freq \
    --export-r \
    --lead-variant-choice prob \
    --weighted-average-r afr=n_afr amr=n_amr eas=n_eas fin=n_fin nfe=n_nfe \
    --dentist-s \
    --abf \
    --summary \
    --case-control

Run slalom --help for the full flag list. The tool can also be used as a library:

from slalom import SlalomConfig, run_slalom

run_slalom(SlalomConfig(snp="example/example.snp", out="out.txt",
                        reference_genome="GRCh38", dentist_s=True))

Reading requester-pays / private buckets

gnomAD LD variant indices and the SLALOM reference tables live in requester-pays buckets. Pass fsspec options as JSON via --storage-options, e.g. to bill your own project:

slalom ... --storage-options '{"requester_pays": true, "project": "YOUR_GCP_PROJECT"}'

Reference data

All reference paths default to hosted copies and are overridable on the command line.

Downloadable bundles (no requester-pays setup)

The requester-pays reference tables are also published as tar.gz bundles. Download them once on a machine that has gcloud and a billing project, extract, and point SLALOM at the local copies with the override flags; then runs need no requester-pays billing.

{build} is b37 (GRCh37) or b38 (GRCh38); ldcov also publishes Pan-UKB variant-index bundles (panukb.{build}.variant_index.tar.gz).

BUILD=b38   # or b37
gcloud storage cp \
    gs://finucane-requester-pays/slalom/bundles/slalom.${BUILD}.parquet.tar.gz \
    gs://ldcov-requester-pays/bundles/gnomad_v2.${BUILD}.variant_index.tar.gz \
    . --billing-project YOUR_PROJECT
tar -xzf slalom.${BUILD}.parquet.tar.gz                          # -> sites + CUP directories
mkdir -p ld_index && tar -xzf gnomad_v2.${BUILD}.variant_index.tar.gz -C ld_index

# then point SLALOM at the local paths (GRCh38 shown):
slalom ... --reference-genome GRCh38 \
    --gnomad-sites-parquet gnomad.genomes.r3.1.2.sites.most_severe.b38.parquet \
    --cup-parquet FASTA_BED.ALL_GRCh38.cups.parquet \
    --ld-variant-index-dir ld_index

This avoids requester-pays reads, but is not a fully offline mode: the gnomAD LD BlockMatrix is still read at runtime from the public gs://gcp-public-data--gnomad bucket (free, but GCS network access is required; the matrices are far too large to download).

The Parquet variant indices are provided by ldcov (see its README for pre-computed bundles and the gnomAD/Pan-UKB populations available). The gnomAD sites and CUP Parquet tables are built once from the original Hail Tables with the helpers under scripts/ (Hail is only needed for this one-time step; install it with pip install "slalom-qc[convert]"):

# gnomAD sites: written via Spark as a *partitioned Parquet directory* (the genome-wide HT
# has ~760M variants). pyarrow.dataset reads the directory transparently.
python scripts/make_gnomad_sites_parquet.py \
    --ht gs://.../gnomad.genomes.r3.1.2.sites.most_severe.ht \
    --reference-genome GRCh38 --coalesce 256 \
    --out gs://YOUR_BUCKET/.../gnomad.genomes.r3.1.2.sites.most_severe.b38.parquet

python scripts/make_cup_parquet.py \
    --cup-ht    gs://.../FASTA_BED.ALL_GRCh38.novel_CUPs.ht \
    --reject-ht gs://.../FASTA_BED.ALL_GRCh38.reject_2.ht \
    --out FASTA_BED.ALL_GRCh38.cups.parquet

Custom LD reference

To use your own LD instead of gnomAD, point SLALOM at a Hail BlockMatrix and its Parquet variant index (e.g. one built with ldcov's scripts/make_bm_variant_index.py):

slalom ... \
    --ld-reference custom \
    --custom-ld-path my_ld.bm \
    --custom-ld-variant-index-path my_ld.variant_index.parquet \
    --custom-ld-label my_panel

Input format

(Required) per-locus .snp file

Required minimum columns are as follows:

• chromosome: chromosome either in GRCh37 (1, 2, 3...) or in GRCh38 (chr1, chr2, chr3, ...). Users can specify a reference genome by --reference-genome GRCh38.
• position: position
• allele1: reference allele in a specified reference genome. If users are unsure about reference/alternative alleles, set --align-alleles to make it consistent with gnomAD.
• allele2: alternative allele in a specified reference genome. This allele is assumed to be an effect allele regardless of --align-alleles.
• beta: effect size
• se: standard error
• p: P-value

Other input column specifications are as follows:

• If --weighted-average-r is specified, sample size columns supplied by this argument are also required, such as n_afr, n_eas, n_nfe, ... (--weighted-average-r requires --export-r).
• If a total sample size n_samples (and n_cases for --case-control) exist in the input, additional output columns min_neff_r2 and max_neff_r2 will be added.
• Any other input columns will remain in an output except for those overwritten by SLALOM.
• --summary reports DENTIST-S outliers and max-PIP statistics, so it requires --dentist-s and --abf.
• LD variants are matched to the reference panel in exact ref/alt orientation. --align-alleles additionally allows matching variants stored in the panel with ref/alt swapped (sign-flipping their r), consistent with its role of reconciling allele orientation against gnomAD.

To make SLALOM-compatible per-locus .snp files from a genome-wide summary statistics, you can also use make_finemap_inputs.py from our fine-mapping pipeline.

Legacy WDL pipeline

The legacy/wdl/ pipeline predates this packaging and targets the standalone Hail/Spark script. It is kept for reference only; the recommended entry point is now the slalom CLI above.

Citation

Kanai, M. et al. Meta-analysis fine-mapping is often miscalibrated at single-variant resolution. Cell Genomics 2, 100210 (2022)

Contact

Masahiro Kanai (mkanai@broadinstitute.org)