Genome wide association study data: the field's evidence base as deliverable rows

A genome wide association study scans hundreds of thousands of variants across a population and reports the ones whose frequency tracks a trait. The reusable asset is not the paper - it is the structured record set underneath it: which variant, in which gene neighborhood, shifted the odds of which trait, by how much, in how many people.

Genome wide association study data therefore arrives as two linked layers. Study records describe each published scan: cohort composition and ancestry, sample sizes for discovery and replication, genotyping platform, design flags. Association records describe each finding: variant, mapped trait, p-value, odds ratio or beta with its unit, risk allele frequency. One row per variant-trait pairing, every row carrying a stable accession back to its study.

GWAS Catalog curates exactly that pair, jointly run by EMBL-EBI and NHGRI: 229,900 studies and 1,191,572 SNP-trait associations as of August 2026, reaching back to 2005 - the debut year of the genome-wide approach itself. The first accession, GCST000001, is the Science paper that put complement factor H on the macular-degeneration map: 96 European-ancestry cases against 50 controls, a p-value of 4e-08, an odds ratio of 4.6 per allele at rs380390-C. Twenty years later that record shape still holds, which is why two decades of findings can sit in one queryable table.

The difference between reading GWAS results and computing on them lives in four field groups.

Effect sizes typed as numbers. Every association carries orPerCopyNum or betaNum with betaUnit and betaDirection, plus standardError and riskFrequency - precisely the inputs polygenic risk score methods consume. P-values ship as separate mantissa and exponent, so no pipeline parses scientific-notation strings to sort significance.

Ancestry as structured data, not prose. Each study breaks its cohorts into ancestral groups with headcounts (ancestries, plus split initialSampleSize / replicationSampleSize). Published GWAS skew heavily European-ancestry; these fields turn that skew from an anecdote into a measurable column, and let replication planning check where an effect was actually found before a validation cohort gets budgeted.

Ontology-mapped traits. Author-reported trait strings are kept verbatim for provenance, while a parallel mapping assigns EFO and MONDO terms - celiac disease resolves to MONDO_0005130 regardless of which of forty phrasings a paper used. Aggregation belongs on the mapped side; counting on free text quietly splits one trait into many.

Design flags that screen studies before modeling. fullPvalueSet marks studies whose complete summary statistics exist (as opposed to top-hits-only reporting), beside booleans for imputed genotypes, pooled designs and gene-environment or gene-gene interaction testing. Filtering on those flags first avoids building a model on a publication that cannot support it.

A single association names a variant and a trait; a target decision needs genes, networks and tissue evidence attached. The chain runs in four joins, each anchored to a record Datadory delivers in the same keyed fashion:

1. Pull the trait's association rows from GWAS Catalog, carrying the EFO/MONDO mapping and ancestry annotations through every downstream step.
2. Resolve variants to genes with the Ensembl Genome Browser. Release 116 (June 2026) annotates roughly 60,000 human genes on GRCh38 under stable identifiers - ENSG00000141510 is TP53 this release and last - so locus-to-gene mapping survives symbol renames.
3. Intersect candidates with interaction networks. STRING Protein Interaction Database scores more than 20 billion protein-protein associations across 59.3 million proteins in 12,535 organisms, each pair carrying a combined confidence score assembled from seven evidence channels. Network position separates a plausible candidate from a lonely hit.
4. Filter by expression and localization. Human Protein Atlas v25.1 maps 27,883 antibodies against 17,407 proteins - tissue immunohistochemistry for 15,312 genes, subcellular localization for 13,603 across 49 organelle classes - so a target gets tissue-relevance and druggability screens before anyone orders an assay.

Where oncology applies, the disease layer closes the loop: cBioPortal carries alteration and clinical records for roughly 400,000 tumor samples across 539 studies, and the NCI Genomic Data Commons harmonizes over 1.3 million files across 50,571 cases from TCGA, TARGET and CPTAC. Run end to end, the chain converts a catalog row into a shortlist of genes with network position, tissue evidence and cohort support already attached.

Three tiers of evidence sit behind any published association, and they behave differently once delivered.

Curated top results - the whole catalog. All 1,191,572 associations are the strongest reported findings per study, complete with effect sizes and mapped traits. For evidence maps, PRS inputs and target triage, this is the working set, and it covers essentially the entire GWAS-era literature.

Full summary statistics - flagged, not universal. Studies whose complete per-variant result sets exist carry the fullPvalueSet flag. Fine-mapping and colocalization want those whole distributions rather than lead SNPs, so screening on the flag before promising a fine-mapping deliverable is the difference between a plan and a hope.

Individual-level genotypes - a different product entirely. The raw reads behind published scans live in the sequencing archives: NCBI SRA, the largest raw-read store anywhere at tens of petabases, and the European Nucleotide Archive, whose INSDC submissions reach back to the early 1980s. Human-subject genotypes additionally route through controlled-access authorization, so consent review sits between a reanalysis plan and its data regardless of who delivers the rows. Expression context that helps interpret a hit is a separate archive again - NCBI GEO holds roughly 250,000 functional-genomics Series with 7.9 million-plus Samples, continuous since 2000.

Teams that conflate the three tiers usually discover the gap after compute is scheduled. Match artifact to tier first: curated associations for scale, flagged full statistics for fine-mapping, archived reads only when individual-level models genuinely require them.

Target identification and prioritization. Reported genes arrive resolved to Ensembl and Entrez identifiers on every significant association, so pipelines rank targets by the number and strength of independent associations rather than by citation counts.

Polygenic risk score construction. Effect sizes, standard errors and risk allele frequencies are exactly what PRS methods ingest; the full-statistics flag routes modelers toward studies that can support the model instead of top-hit summaries that cannot.

Disease-area evidence maps. Ontology-mapped traits collapse thousands of author phrasings into countable categories, turning "what has been associated with type 2 diabetes and how strongly" into a grouped query instead of a literature review.

Ancestry-aware replication planning. Cohort composition columns show which populations an effect was found - and replicated - in, before a validation cohort is budgeted for a population the evidence never covered.

Market and momentum reads. Counting large multi-cohort studies by therapeutic area turns scientific momentum into a series you can chart; investors and quant researchers use it as a funding signal, competitive-intel teams as a pipeline-heat map, market researchers as a sizing denominator.

The persona-level versions of these workflows live on the pairing pages - biotechnology data for data scientists for the modeling angles, biotechnology data for investors & quants for the signal reads.

Four disciplines separate defensible GWAS analysis from a citation with numbers attached.

First, pick your layer deliberately. Study records describe experiments; association rows describe findings. Most analytics want the second, cohort-quality screens need the first, and both hang off the same GCST accession - quoting a denominator without saying which layer quietly breaks any coverage claim.

Second, harmonize effects before pooling magnitudes. Case-control traits report odds ratios per allele copy; continuous traits report betas with their own units and directions. Pooling them unnormalized produces averages of things that are not the same quantity.

Third, aggregate on mapped traits, keep author text for provenance. The free-text trait preserves what the paper said; the EFO/MONDO terms make it countable. Rollups computed on the wrong column fragment one disease into dozens of near-names.

Fourth, respect the ancestry annotation as data. European-ancestry dominance is a fact about published genetics, not a defect in the corpus - the ancestry fields exist precisely so claims get scoped to populations the evidence supports, and so representation gaps become measurable instead of assumed.

Every delivery keys on accession and locus, so this month's pull joins cleanly to last month's and to whatever variant or target lists you already hold. When a curation pass adds fields or revises records, the change arrives flagged rather than discovered mid-model, and each delivery travels with the field dictionary attached - definitions and worked examples per column, verified against live records.

Name the trait list you care about when you request a sample and the sample comes back shaped to it: association rows for your indication, the study headers behind them, the ancestry breakdowns that scope your claims. The first join attempt then happens against evidence rather than hope. Get a sample and the schema in the sample is the schema you ship against.

This cluster sits inside a 24-record biotechnology pool spanning sequence archives, protein and structure resources, chemistry, registries and regulatory records. The biotechnology data hub lists every dataset with coverage and delivery options in one scorecard, and the best biotechnology datasets ranking shows where the GWAS Catalog's 9-out-of-10 score lands industry-wide.

Product pages worth opening next: the GWAS Catalog dataset page carries the full field dictionary and captured sample rows; STRING adds the network layer; Human Protein Atlas supplies expression evidence; ClinicalTrials.gov follows a prioritized target into interventional testing across 599,549 registered studies.