The MRC IEU OpenGWAS data infrastructure

Key results

1. ShownThe database holds thousands of complete GWAS summary datasets, keeping every analysed variant regardless of p-value, across disease, molecular and brain imaging traits, mostly from European-ancestry samples.

   14582 datasets with 125.8 billion associations, from 196 consortium publications plus biobank-wide analyses (e.g. UK Biobank, Biobank Japan). Phenotypes: 4126 binary, 3371 protein, 3217 other continuous, 3143 brain-imaging and 725 metabolite. Extracted independent top hits total 116918, and 7109 datasets have at least one. · Abstract; Summary of the database contents

2. ClaimedEvery dataset passes one harmonisation pipeline, so alleles, strands and rsIDs are consistent across studies, and each dataset gets a published quality-control report.

   gwas2vcf validates columns, orients the non-effect allele to GRCh37/hg19 (switching effect signs and frequencies when needed), left-aligns and trims variants with vgraph and annotates dbSNP build 153. Reports cover LD score regression, genomic inflation, strictly clumped hits, allele frequencies vs 1000 Genomes, and Z-scores from effect/SE vs from p-values. · Data harmonisation and quality control; Data processing pipeline; Per-study reports; LD score regression procedure

3. SuggestedAn automatic effect-allele check catches mis-annotated studies, which would otherwise reverse inferred causal directions in analyses such as MR.

   Reported effect allele frequencies of common variants (frequency 0.1-0.3) are compared with a population reference, and an inverse relationship flags a mislabelled allele or frequency; several studies across many batches were flagged and needed follow-up, but no counts are given. · Effect allele checks

4. ClaimedA REST API with R and Python clients supports positional and p-value queries, LD-proxy substitution, on-the-fly clumping, LD matrices and phenome-wide lookups of a variant across all traits.

   Proxies come from pre-computed variant pairs with LD R > 0.4 within 10Mb windows, returning the highest-LD proxy with phase-aligned alleles. Default clumping runs PLINK 1.90 on one of five 1000 Genomes phase 3 super-population panels (LD R < 0.001, 10Mb, p < 5e−8). Clients are ieugwasr and ieugwaspy. · Online access to the database; Linkage disequilibrium reference panel; LD proxy lookups

5. ClaimedFor heavy workloads the harmonised data can be downloaded as GWAS-VCF files with dedicated query packages, and the gwasglue R package passes API or file data straight into MR, colocalisation and fine-mapping tools.

   gwasvcf (R) and pygwasvcf (Python) query GWAS-VCF files for parallel HPC jobs; served data sit in an Elasticsearch cluster on Oracle Cloud Infrastructure, chosen for query speed and scalability, with study metadata and embargo permissions in a Neo4j graph. No query-speed benchmarks are reported. · Accessing the data for high performance…; Connecting the queried data to analytical tools; Metadata; GWAS summary data