variant-workbench

Open-source genetic variant database: a GRCh38 warehouse of genes, studied variants and their evidence from gnomAD, ClinVar, GENCODE, GTEx, the GWAS Catalog and MaveDB. Research use only.

Research use only. This is not medical advice, not a diagnostic tool and not for clinical decisions. It is not a personal genome tool: it holds no individual genotypes and cannot read anyone's genome. Behavioral and cognitive GWAS hits in it are weak observational claims. The reports were written by an AI agent and are published as written.

This repository holds the code that builds the warehouse, the build plan, the QC notes and the written reports. The warehouse tables themselves are published as GitHub Release assets (see Release assets). The raw downloads (about 768 GB) are not redistributed; the scripts fetch them from the providers.

What this is, and is not

• It is a reproducible build: about 20 single-file Python scripts that turn six public sources into Parquet tables queried with DuckDB, plus the derived tables.
• It models public evidence about genes and about variants someone has already studied or measured: ClinVar labels, GWAS Catalog associations, GTEx expression and QTLs, MaveDB assay scores, gnomAD frequencies and gnomAD constraint. Each kind of evidence stays a separate claim.
• It is not a complete map of what the genome does. Most of GRCh38 falls outside gene exons, lead QTLs and GWAS loci; the build measures that fraction (warehouse/reports/dark_genome.md) instead of explaining every base.
• It contains no DNA sequence and no individual genotypes, only variant keys, annotations and aggregate statistics.
• It is not a polygenic-score tool, a linkage-disequilibrium resource or a colocalization analysis (see Limitations).

Sources and versions

Source	Version used	What was used	Terms (details in NOTICE.md)
gnomAD	v4.1	Sites VCFs for genomes and exomes, chr1-22, X and Y (48 VCFs with indexes); v4.1 constraint metrics	Primary data CC0; do not attempt to re-identify participants; "gnomAD" is a trademark of The Broad Institute, Inc.
ClinVar (NCBI)	Weekly variant_summary.txt.gz, file dated 2026-11-01	GRCh38 rows	NCBI places no restrictions; NCBI disclaimer applies
GENCODE	Release 49 (Ensembl 115)	Comprehensive annotation GTF	Open access; EMBL-EBI Terms of Use
NHGRI-EBI GWAS Catalog	releases/latest retrieved 2026-11-03; TSV dated 2026-10-29	Associations, ontology-annotated full download	EMBL-EBI Terms of Use
GTEx	v8 (RNA-SeQC v1.1.9)	Median TPM by tissue; single-tissue cis-eQTL and cis-sQTL results	Open-access files; acknowledge GTEx
MaveDB	Bulk dump as of 2026-06-24 (mavedb-dump.2026062418131, Zenodo DOI 10.5281/zenodo.20840937)	Score sets, scores, mapped variants	CC0, with per-score-set notes

All files were downloaded on 2026-11-03. Exact URLs are in scripts/fetch_sources.py. This tool includes data from the gnomAD v4.1 release. This project is not affiliated with or endorsed by gnomAD, the Broad Institute, NCBI, EMBL-EBI, GENCODE, GTEx or MaveDB.

Repository layout

plan.md                     the build plan the agent followed (rules, schemas, waves, limits)
scripts/                    the build: fetch, load, build, annotate, QC and report helpers
warehouse/
  reports/                  17 written reports (Markdown), as written by the build agent
    sql/                    the DuckDB SQL behind 16 of them (run with scripts/run_report_sql.py)
    figures/                cognitive_loci_map.png
  qc/                       QC notes per lane, the truth-set report, wave notes
  manifest/                 the raw-file manifest (files.tsv) and disk note, paths relative to ROOT
logs/fetch.status           the download's final status line (DONE)
docs/RELEASE_NOTES.md       the text of the GitHub Release notes
LICENSE                     GPL-3.0-or-later (code)
DATA-LICENSE.md             CC0 1.0 (this project's data contributions)
NOTICE.md                   source terms, required notices, exclusions
CITATIONS.md, CITATION.cff  how to cite this project and its sources

Not in git: raw/ (downloads), db/ (SQLite catalog), logs/*.log, and every warehouse data table (.parquet, .tsv). See .gitignore.

Warehouse layout

All tables are Parquet under warehouse/, GRCh38, queried with DuckDB. Column-level detail is in plan.md and in each lane's QC note in warehouse/qc/.

Folder	Tables	Contents
intervals/	genes, transcripts, exons, cds, utrs	GENCODE v49 gene models, 1-based inclusive coordinates
constraint/	transcript_constraint, gene_constraint	gnomAD v4.1 constraint; one row per gene picked by MANE, then canonical, then longest CDS
clinvar/	alleles, allele_dedup	ClinVar GRCh38 records with VCF-style variant keys and review-star buckets; one row per variant key
gwas/	associations, traits, loci	All 38 catalog columns plus parsed fields; traits by EFO URI; associations clumped within 1 Mb per trait
gtex/	median_tpm, gene_expression, egenes, sgenes, eqtl_pairs, sqtl_pairs	Expression by tissue with tau; lead and significant eQTL and sQTL pairs
mavedb/	experiments, scores, mapped_variants, mapped_variants_scored, gene_coverage	Score-set inventory, scores, GRCh38-mapped variants, coverage per gene
crosswalk/	genes, unmatched	Gene identifier crosswalk (Ensembl, symbol, NCBI, HGNC, MANE) and join misses
marts/	gene_mart, gene_mart_protein_coding	One row per GENCODE gene: identity, constraint, expression, ClinVar, GWAS, MaveDB, QTL counts
variants/	tier_a, tier_a_genome, tier_a_exome, tier_a_keys.tsv, tier_a_unresolved.tsv, field_map.tsv	Tier A variants (ClinVar, GWAS with a resolved key, MaveDB mapped, lead QTLs) with gnomAD genome and exome frequencies in separate columns and gnomAD's VEP consequences
ledger/	evidence	Evidence ledger: one row per claim, with the evidence class on every row (clinvar, gwas, eqtl, sqtl, mavedb, constraint)

The variant_key is chrom:pos:ref:alt with a 1-based VCF position and uppercase alleles. Ensembl ids are joined without their version. Genome and exome allele counts are never added together.

Release assets

The warehouse tables are attached to the GitHub Release as 11 zip files, each under 2 GiB. Each zip keeps the warehouse/<folder>/ layout (members are <folder>/<file>), so unpacking them all into one warehouse folder under your ROOT gives the paths the scripts and the report SQL expect. Check the downloads against SHA256SUMS, and MANIFEST.tsv lists every table with bytes and rows. Both SHA256SUMS and MANIFEST.tsv are attached to the release.

sha256sum -c SHA256SUMS
for z in variant-workbench-*.zip; do unzip -o "$z" -d warehouse; done

PowerShell equivalent (run in the folder that holds the zips and SHA256SUMS; set $root to your ROOT, the folder that should contain warehouse):

# Compare each zip's SHA-256 with SHA256SUMS (prints False for any mismatch)
Get-Content SHA256SUMS | ForEach-Object {
  $hash, $name = $_ -split '\s+', 2
  $name = $name.TrimStart('*').Trim()
  (Get-FileHash -Algorithm SHA256 $name).Hash.ToLower() -eq $hash.ToLower()
}
# Members are <folder>/<file>, so extract into the warehouse folder
$root = 'E:\vw-data'
Get-ChildItem variant-workbench-*.zip | ForEach-Object {
  Expand-Archive -LiteralPath $_.FullName -DestinationPath "$root\warehouse" -Force
}

gtex/eqtl_pairs.parquet is in its own zip and is not split. All tables are Parquet except variants/tier_a_keys.tsv, tier_a_unresolved.tsv and field_map.tsv. The reports, QC notes and manifests are in this repository, not in the release.

Asset	Contents (member: rows)	Bytes	SHA-256
variant-workbench-clinvar.zip	clinvar/allele_dedup.parquet: 4,639,253, clinvar/alleles.parquet: 4,639,270	562,360,623	9e4d5846cd3bb81c7b0219845505064f0078bf6fa19ce489e273c9105dc6308f
variant-workbench-constraint.zip	constraint/gene_constraint.parquet: 18,716, constraint/transcript_constraint.parquet: 211,523	32,237,879	12556c3108b9e638fbd18c900f8ac16b482d779325f42cd0681209cbd846019d
variant-workbench-crosswalk.zip	crosswalk/genes.parquet: 78,691, crosswalk/unmatched.parquet: 11,339	3,469,927	de765d0015af52009d9af0fcf17f8769b2b0bf9f2acd300dba757b306a899ac9
variant-workbench-gtex-eqtl_pairs.zip	gtex/eqtl_pairs.parquet: 71,478,479	1,969,181,641	994bce0b68f9f7d3a6b6931b9ce8f1b34f7b02efe006cf57a5f19ea52d70f3b2
variant-workbench-gtex.zip	gtex/egenes.parquet: 1,207,976, gtex/gene_expression.parquet: 56,200, gtex/median_tpm.parquet: 3,034,800, gtex/sgenes.parquet: 618,932, gtex/sqtl_pairs.parquet: 24,445,206	1,000,184,695	34cf2c3fa82e7676956adc8303c996cc900ecf94df21b4cf5f176446b59d0d9a
variant-workbench-gwas.zip	gwas/associations.parquet: 1,205,644, gwas/loci.parquet: 311,160, gwas/traits.parquet: 16,886	88,815,755	5ac842c98813b44d5a653e482ed4410ba791b41a661696f16b6a43bab492eca2
variant-workbench-intervals.zip	intervals/cds.parquet: 2,498,156, intervals/exons.parquet: 3,673,949, intervals/genes.parquet: 78,691, intervals/transcripts.parquet: 507,365, intervals/utrs.parquet: 770,688	38,852,112	6e3979d3adcc84178a47e1e9f63a3f0d39af462785656feba280719fdb981199
variant-workbench-ledger.zip	ledger/evidence.parquet: 10,866,762	381,884,271	0d6684151bbe512f81e83bb3ef1e6299cf3515ffa085fb7dd7e5dd4616d3e68a
variant-workbench-marts.zip	marts/gene_mart.parquet: 78,691, marts/gene_mart_protein_coding.parquet: 20,097	5,155,988	33e6bd8393f8102f39465f150bc16af6f231576b0f6118daec70baee88935161
variant-workbench-mavedb.zip	mavedb/experiments.parquet: 2,771, mavedb/gene_coverage.parquet: 1,135, mavedb/mapped_variants.parquet: 6,329,224, mavedb/mapped_variants_scored.parquet: 6,329,224, mavedb/scores.parquet: 7,777,824	907,432,269	1d090d66605877192b65479658e150e89793580a5ad9aad7911d334a6b5b85c9
variant-workbench-variants.zip	variants/field_map.tsv: 43, variants/tier_a.parquet: 5,928,738, variants/tier_a_exome.parquet: 5,928,738, variants/tier_a_genome.parquet: 5,928,738, variants/tier_a_keys.tsv: 5,928,738, variants/tier_a_unresolved.tsv: 416,792	952,078,910	e38123251628c451da31af29844f9f601ef83591398ea2c5f5354d5e2c2c4384

Total: 5,941,654,070 bytes (5.534 GiB) in 11 assets, each under 2 GiB. Row counts are rows in the table, not lines.

The release tables differ from the original build in two ways (see What was left out): the 14 predictor-score columns are removed from the Tier A tables (and their 14 rows from field_map.tsv), and six MaveDB score sets are removed from the MaveDB tables.

Rebuilding from the sources

Nothing here has to be rebuilt to use the release tables. A full rebuild downloads about 768 GB and takes a large machine.

Requirements

• Python 3.13 with the packages in requirements.txt (pip install -r requirements.txt).
• Disk: about 768 GB for the downloads (761.6 GB of it is the gnomAD VCFs), about 7 GB for the warehouse and about 1 GB for the SQLite catalog. fetch_sources.py refuses a download that would leave less than 80 GiB free, and load_gtex.py stops before writing the QTL pair tables if its size estimate would leave less than 120 GiB free. Plan on roughly 1 TB free.
• Memory: the heaviest steps (build_tier_a.py, annotate_gnomad.py, build_gene_mart.py, build_ledger.py) set DuckDB to memory_limit='64GB' and 8 threads; the report SQL uses 12-16 GB. The original build ran on a Windows 11 desktop with about 127 GB of RAM. On a smaller machine, lower those settings; DuckDB spills to warehouse/tmp/. Peak memory and run times were not measured.
• No bcftools, htslib or pysam needed: scripts/vcf_query.py and scripts/vcf_stream.py read the bgzip VCFs and tabix indexes in pure Python (with Biopython's BGZF reader).

Where the data go: ROOT and VW_ROOT

Every script reads and writes under one folder, ROOT. ROOT is the VW_ROOT environment variable when it is set; otherwise it is the folder above scripts/, so by default the data go inside your clone. Inside ROOT the scripts use raw/, db/, logs/ and warehouse/.

# Windows PowerShell: keep the code in the clone, put the data on a big drive
$env:VW_ROOT = 'E:\vw-data'

# Linux or macOS
export VW_ROOT=/data/vw

The scripts import each other from scripts/, so run them as python scripts/<name>.py from anywhere; they also work under python -I.

Windows notes

The project was built and run on Windows 11 only.

• fetch_sources.py calls curl.exe, which ships with Windows 10 and later. On Linux or macOS, change "curl.exe" to "curl" in that script (untested).
• The pure-Python VCF readers should run on Linux and macOS but were not tested there.
• plot_cognitive_loci.py uses Arial from C:\Windows\Fonts and falls back to Pillow's default font elsewhere.
• Disk-space checks use ROOT's drive on Windows and the filesystem that holds ROOT elsewhere.

Build order

This is the order in plan.md (Step 0 to Step 4). Each step reads the outputs of the steps before it.

# Step 0: download (about 768 GB) into raw/; "core" and "gnomad" run the two halves
python scripts/fetch_sources.py all          # writes logs/fetch.log and logs/fetch.status (must end DONE)

# Wave 0: catalog, manifest, truth set
python scripts/load_core.py                  # db/variants.sqlite (its gene table is used by later steps)
python scripts/build_manifest.py             # warehouse/manifest/files.tsv and disk.md; SQLite source_file
python scripts/build_truth_set.py            # warehouse/qc/truth_set.tsv (18 cases looked up in the sources)

# Wave 1: one loader per source (independent of each other)
python scripts/load_intervals.py             # warehouse/intervals/
python scripts/load_constraint.py            # warehouse/constraint/
python scripts/load_clinvar.py               # warehouse/clinvar/
python scripts/load_gwas.py                  # warehouse/gwas/
python scripts/load_gtex.py                  # warehouse/gtex/
python scripts/load_mavedb.py                # warehouse/mavedb/ (inventory and mapped variants)
python scripts/load_mavedb.py --scores-only  # warehouse/mavedb/ scores joined to mapped variants
python scripts/build_crosswalk.py            # warehouse/crosswalk/ (after intervals, constraint, ClinVar)
python scripts/truth_set.py                  # warehouse/qc/truth_set_report.md

# Wave 2: integration
python scripts/build_gene_mart.py            # warehouse/marts/
python scripts/build_tier_a.py               # warehouse/variants/tier_a_keys.tsv, tier_a_unresolved.tsv
python scripts/annotate_gnomad.py            # warehouse/variants/tier_a*.parquet (one pass over the 48 VCFs)
python scripts/build_ledger.py               # warehouse/ledger/evidence.parquet and reports/dark_genome.md
python scripts/truth_set.py                  # rerun; all 18 cases passed in the original build

# Wave 3: reports (optional)
python scripts/run_report_sql.py <report>    # see "Running the report SQL"
python scripts/plot_cognitive_loci.py        # warehouse/reports/figures/cognitive_loci_map.png

Each loader writes a QC note to warehouse/qc/. Rebuilding overwrites the QC notes, dark_genome.md and the figure in your ROOT.

A rebuild from the sources reproduces the original tables, including the predictor columns and the six MaveDB score sets this release leaves out. If you rebuild, the terms in NOTICE.md for those data apply to you.

Regenerating the predictor columns locally

The release removes CADD, REVEL, SpliceAI, Pangolin, phyloP, SIFT and PolyPhen (genome and exome forms, 14 columns) from the Tier A tables. They come from the gnomAD v4.1 sites VCFs, and scripts/annotate_gnomad.py extracts them again:

1. Download the gnomAD VCFs and indexes into ROOT: python scripts/fetch_sources.py gnomad (about 761.6 GB).
2. Put warehouse/variants/tier_a_keys.tsv in place: it is in variant-workbench-variants.zip, or rebuild it with build_tier_a.py (which needs the ClinVar, GWAS, MaveDB and GTEx tables).
3. Run python scripts/annotate_gnomad.py. It writes tier_a_genome.parquet, tier_a_exome.parquet and tier_a.parquet with every extracted field, predictors included, and field_map.tsv. It overwrites the release copies of those files in your ROOT.

The regenerated scores are under their providers' terms (CADD, REVEL and SpliceAI are free for non-commercial use only; see NOTICE.md). Do not redistribute them without checking those terms.

Running the report SQL

The report SQL files are published exactly as the build wrote them, so they hard-code the original project folder D:/variant-workbench. Do not edit them; run them through the helper, which substitutes your ROOT at run time:

python scripts/run_report_sql.py study_redundancy             # prints each result table
python scripts/run_report_sql.py ancestry_audit --max-rows 0  # print every row
python scripts/run_report_sql.py splicing --dry-run           # show the rewritten SQL only

Expect differences against the release tables: conflicts.sql, predictor_disagreement.sql and splicing.sql read the predictor columns and fail until you regenerate them, and counts that involve MaveDB change slightly because six score sets are excluded. The report text itself was written by the agent from these results; no script regenerates it.

What was left out, and why

• Third-party predictor scores (14 columns). CADD, REVEL and SpliceAI scores are offered for non-commercial use only, and none of their terms grants redistribution in an unrestricted public release. The terms for Pangolin, phyloP, SIFT and PolyPhen-2 values as distributed by gnomAD were not found, so they are left out as unclear. Regenerate them locally (above).
• Six MaveDB score sets. urn:mavedb:00000044-a-1, -a-2, -b-1 and -b-2 are SARS-CoV-2 receptor-binding-domain scans (viral, not human) whose usage note is a BSD 3-Clause licence text. urn:mavedb:00001255-a-1 and -a-2 (AIRE) say the data "must be pre-printed prior to usage", and no preprint is listed. They are removed from the published MaveDB tables (experiments 2,777 to 2,771 rows, scores 8,205,740 to 7,777,824, mapped_variants and mapped_variants_scored 6,350,898 to 6,329,224 each, gene_coverage 1,137 to 1,135). The evidence ledger, the gene marts, the crosswalk and the Tier A tables never held rows from these sets (checked: 0 ledger rows reference them), so they are unchanged. The 28 Fort Lauderdale score sets (PMID 31395865) and the RHO score set urn:mavedb:00001275-a-1 are kept, and their publications are cited in CITATIONS.md.
• raw/ and db/. The raw downloads are too large to host, and they include the GTEx sample-level TPM file (17,382 donor-sample columns) and MaveDB metadata with researcher names. The SQLite catalog is a derived convenience copy; load_core.py and build_manifest.py rebuild it.
• Local details. Personal paths, hardware model names and process IDs were redacted from plan.md and warehouse/qc/wave0.md, and the manifest paths were made relative to ROOT. The substance of the plan is unchanged.

Limitations

From plan.md ("What this disk can and cannot support") and the build's QC:

• Not every base is explained. Most of GRCh38 falls outside gene exons, lead QTLs and GWAS loci. dark_genome.md publishes the fraction.
• No linkage disequilibrium and no real polygenic score. gnomAD sites VCFs carry no individual genotypes, and the GWAS Catalog stores reported significant associations, not genome-wide summary statistics. Do not build a score from lead SNPs.
• No personal genome report. There are no individual genotypes anywhere in the project.
• No statistical colocalization. Overlap between a GWAS mapped gene and an eGene in a relevant tissue is a hypothesis; colocalization needs full summary statistics.
• Not a complete functional map. MaveDB covers a small slice of proteins (in the original build, assays for 134 of 20,097 protein-coding genes by symbol).
• GWAS rows carry no alleles. The GWAS Catalog gives no ref/alt, so variant_key is null in gwas/associations; GWAS variants enter Tier A only through an rsID match to ClinVar. 416,791 GWAS rsIDs with a position stayed unresolved (tier_a_unresolved.tsv) because the genome-wide gnomAD rsID scan was deferred (416,792 rows in tier_a_unresolved.tsv; one rsID appears twice). GWAS-versus-frequency results cover only a small share of GWAS variants.
• Ancestry. Most catalog discoveries come from European-ancestry samples. A frequency in the discovery group is not a frequency for everyone. Reports give coverage and missingness and do not rank groups.
• Annotation versions differ. Gene models are GENCODE v49, but the vep_* consequence columns come from gnomAD's own VEP v105 run on GENCODE 39 transcripts.
• Sources change. ClinVar classifications are revised weekly; this is a snapshot.
• One claim was stopped. constraint_versus_function.md does not estimate a MaveDB effect size, because the span join credited scores to overlapped genes rather than assayed genes (warehouse/qc/issues_constraint_versus_function.md). plan.md asked for a warehouse/qc/issues.md; the build did not write one.
• Other gaps. The plan's bcftools acceptance check was replaced by the pure-Python reader because bcftools has no win-64 conda build. The build also found an extra GENCODE file in raw/gencode/ (pc_translations) that no script fetched or used.

The design rule (rule 11)

plan.md rule 11 governs the whole project. In plain words:

• The warehouse describes public biology. It is not a tool for ranking people, and it must not be used to score anyone.
• A behavioral or cognitive GWAS association is not a reading of a person's nature. Those hits stay in the catalog only as weak observational claims, shaped by who was studied, by environment and by social confounding.
• Population-frequency work reports coverage and missingness. It does not rank groups or infer disease risk for a population from a lead SNP.

Reading the reports

The 17 reports in warehouse/reports/ are AI-generated research notes, published exactly as written. During publication, 87 of their numbers (spread across all 17 reports and the figure) were spot-checked against the warehouse and all matched; the rest were not re-derived. Read them with these points in mind:

• common_disease.md lists cognitive, behavioral and social traits (intelligence, cognitive function, income, risk-taking and others) next to diseases, under a "Common-disease" heading, and does not repeat rule 11's caveat. Those rows are weak observational associations, not disease findings and not statements about people.
• figures/cognitive_loci_map.png maps catalog lead SNPs for educational attainment, intelligence and cognitive function. A highlighted region is where a study reported a lead variant, not a segment that causes a trait. The discovery cohorts were mostly of European ancestry, the associations are confounded by environment and social factors, and they do not predict anything about an individual or a group.
• predictor_disagreement.md uses phrases such as "the one to trust first". That is a statement about the warehouse's evidence, not guidance for interpreting anyone's variant.
• clinvar_versus_frequency.md and conflicts.md discuss common alleles with pathogenic labels (for example HBB, HFE, CFTR and G6PD). Carriers of one copy of a recessive pathogenic allele are usually unaffected.
• One external citation in rare_disease.md (gnomAD LOEUF guidance dates) could be only partly verified.
• The numbers describe the original build, which still included the predictor columns and the six excluded MaveDB score sets.

Disclaimers

• Research use only. Not medical advice, not a diagnostic tool, and not for clinical decisions about you or anyone else. If you have a health question, ask a qualified clinician.
• Not a personal genome tool. It holds no individual genotypes and is not designed to interpret a personal genome.
• Weak evidence stays weak. Behavioral and cognitive GWAS hits are weak observational claims. Nothing here ranks people or groups.
• No warranty. The code is provided under GPL-3.0 without warranty (see LICENSE), and the data as is. Verify anything you rely on against the original sources.

How this was made: AI assistance

This project was built by AI agents.

• A Grok agent ran the build on 2026-11-03 and 2026-11-04, following plan.md: it wrote the scripts, downloaded the sources, built the warehouse, ran the QC and truth set, and wrote the 17 reports and the figure.
• Claude (Anthropic) prepared the publication: a licence, privacy and ethics audit; the portable paths (VW_ROOT) and run_report_sql.py; the redactions; and this README, the licence files, NOTICE.md and the citations.

The owner chose the description and the licences; the exclusions follow the audit's licence and privacy findings. The code and reports have not been independently audited by a human expert. Treat them accordingly.

Licences

• Code (scripts/, warehouse/reports/sql/): GNU General Public License version 3 or, at your option, any later version (LICENSE).
• This project's data contributions (schemas, derived keys, aggregates, reports, QC notes and documentation): CC0 1.0 Universal (DATA-LICENSE.md).
• Source data keep their providers' terms (NOTICE.md). In particular, users of the gnomAD-derived values agree not to attempt to re-identify gnomAD participants.

variant-workbench
Copyright (C) 2026 netmaster296

This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.

This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
GNU General Public License for more details.

You should have received a copy of the GNU General Public License
along with this program.  If not, see <https://www.gnu.org/licenses/>.

Citing

See CITATIONS.md and CITATION.cff. Please cite the data sources as well: gnomAD, ClinVar, GENCODE, the GWAS Catalog, GTEx (with its acknowledgement) and MaveDB, plus the original publications of the MaveDB score sets you use.