The GWAS Catalog, Explained: What It Is and How It Becomes Your DNA Health Report

The GWAS Catalog is the public, curated list of genome-wide association study results, run by NHGRI and EMBL-EBI, and every row in a Full Human health report comes from it. Here is what the Catalog is, which studies it accepts, how we reduce 1.2 million associations to 12,110 rows, what a row in your PDF maps to, and why two tools read one file differently.

What a GWAS is, in one paragraph#

A genome-wide association study (GWAS) genotypes a large group of people at hundreds of thousands to millions of positions and asks, position by position, whether one allele is more common in people with a trait (a disease, a blood measurement, a habit) than in people without it. Per variant it reports a p-value, for how surprising the difference is, and an effect size: an odds ratio for yes/no conditions, a beta for measured quantities. Because a GWAS tests about a million hypotheses at once, the field adopted a strict bar for a "genome-wide significant" hit, p < 5 × 10⁻⁸, proposed in 2008 by Pe'er and colleagues [6]. A GWAS finds statistical signals near a trait; it does not say which variant does the work, or why.

What the GWAS Catalog is#

The NHGRI-EBI GWAS Catalog is where those results are collected. The US National Human Genome Research Institute founded it in 2008; since 2010 it has been a joint project with EMBL-EBI, which has hosted it since 2015 (about page and EMBL news, 7 April 2015, both checked 14 October 2026). Curators read each qualifying paper and extract, per association, the variant (rsID), the risk allele, the p-value, the odds ratio or beta with its confidence interval, sample sizes and ancestries, and the trait, mapped to the Experimental Factor Ontology (EFO) so that "T2D", "type 2 diabetes" and "NIDDM" from three papers land on one term.

Its size on the release of 3 October 2026 (checked 14 October 2026 against the Catalog's statistics endpoint and REST metadata):

• 1,198,394 curated associations from 7,845 publications, plus 203,002 full summary-statistics datasets; the one-million mark was passed in late 2025 (release note, 10 November 2025).
• Variants on genome build GRCh38.p14 and dbSNP build 156; traits on EFO v3.94.0.
• New curated data about every two weeks (FAQ, A7).

Its 2025 database paper [5] counted 625,113 associations in July 2024; the Catalog has nearly doubled since.

What gets in, and what that means for you#

The Catalog's eligibility criteria (checked 14 October 2026) require a study to:

• be published in a peer-reviewed journal, in English, on humans (summary statistics from preprints are accepted but labelled as not curator-reviewed until the paper is published);
• test at least 100,000 variants before quality control, on an array or by sequencing, with a genome-wide backbone;
• not be a candidate-gene study. A paper that only looks at MTHFR, COMT or a handful of favourite genes is excluded, however often it is cited online.

Within an eligible study only "top associations" are extracted: the strongest variant per locus (within a gene or a 100 kb window, unless the authors show independent signals), and only if p < 1 × 10⁻⁵. Papers are found by a weekly PubMed search triaged with the machine-learning tool LitSuggest.

Two consequences for a reader of a consumer health report. The Catalog's bar, 1 × 10⁻⁵, is looser than genome-wide significance, so a tool built on it has to apply its own threshold. And the popular "methylation" and "detox" variants rest on candidate-gene studies, which the Catalog excludes: no disease row for MTHFR rs1801133 or rs1801131, COMT rs4680 or CYP1A2 rs762551 passes our filters, so none of them is scored in a Full Human report (rs1801133 reaches the nutrition catalog only through two folate-level studies that give no allele frequency, so it cannot be compared with the average person). That is a property of the evidence, not an oversight.

From 1.2 million rows to 12,110: how Full Human filters the Catalog#

Full Human reads the Catalog's "all associations, ontology-annotated" download and rebuilds two catalogs from it with a small Rust program. The health catalog keeps a row only when all five rules hold:

1. Genome-wide significance. The reported p-value is below 5 × 10⁻⁸.
2. A reported effect size. The "OR or BETA" field is filled, and the interval is not marked "NR", given as a z-score or expressed in measurement units. For the health catalog the effect must be an odds ratio with an interval; betas with units are dropped by this rule.
3. A plausible confidence interval. Two numbers in the right order, the upper less than three times the lower. This removes small studies with very wide error bars.
4. A single-base risk allele. The letter after the dash is A, C, G or T, or an insertion or deletion written I or D. Rows whose risk allele is "?" or longer than one base are dropped when the catalog is built, because no consumer genotype can be compared with them.
5. A trait mapped to a disease. The EFO term must be a MONDO disease term, or map to one through the EFO-to-MONDO mapping file from EBI's Ontology Lookup Service. MONDO is the Monarch Initiative's disease ontology, which "aims to harmonize disease definitions across the world" (mondo.monarchinitiative.org, checked 14 October 2026). "Height" fails this rule; "type 2 diabetes mellitus" passes.

Then, within each trait, only one lead variant per region is kept: a variant within 250,000 bases of a kept, more strongly associated variant is dropped as the same signal. The written catalog also keeps each row's position, strength, risk allele frequency and the alleles the whole catalog names for the variant.

From the 2026-06-01 release this gives 12,110 lead associations, 409 catalog traits, 1,020 studies and 8,291 distinct variants. The largest blocks by mapped trait are prostate carcinoma (846 rows), type 2 diabetes (639), schizophrenia (507), atrial fibrillation (468) and gout (461). The conditions are grouped into 17 categories through the ontology; the complete health report has a section for every category with at least one condition for the file's sex (14 to 16 sections; Genetic and Metabolic has none at present), and categories are not sold as separate reports.

The nutrition catalog is built differently, because rule 5 would reject every vitamin and mineral measurement. Its filter keeps the rows of a fixed list of nutrition traits instead (vitamin and mineral levels, trace elements, food sensitivities and allergies, alcohol and coffee consumption) and applies the same p-value, allele and one-lead-per-region rules. It accepts odds ratios only for celiac disease, peanut allergy and wheat allergy, and betas with a stated direction only for measured levels: 3,428 lead associations, 27 traits, 81 studies and 2,084 variants. The nutrition report covers 13 vitamins, 7 macrominerals, 6 trace elements and 6 food sensitivities; for vitamins B1, B2, B3, B7 and K, sulfur, manganese and lactose intolerance the catalog has no scorable study, and those sections say so.

The pipeline is documented on the methodology page, which prints the release in use. The generator currently runs on the 2026-06-01 release; EBI's newest is 2026-10-03, and a scheduled workflow brings each new release in after review.

What a row in your PDF maps to#

Open the sample report at any condition and you see a table with seven columns, each a Catalog field or derived from one:

Take rs429358, one of the two variants that define the APOE ε2/ε3/ε4 alleles; it appears in 41 rows of the health catalog, 22 of them under Alzheimer's disease. One of those, from PubMed study 34122051, reads: risk allele C, odds ratio 3.245, interval [2.941–3.579], risk allele frequency 0.226, mapped trait "Alzheimer disease" (MONDO:0004975); among the Alzheimer studies that give a frequency, the narrowest interval is that of PubMed 39129223 (odds ratio 1.80 [1.78–1.81], frequency 0.16). The report has no Alzheimer's disease section at present, because it has no condition text for it, so this row stays in the catalog and never reaches a PDF. rs429358 does print under abdominal aortic aneurysm, myocardial infarction and metabolic dysfunction-associated steatotic liver disease, and the mechanism is the same for each. Whatever your file says at rs429358, the variant counts: with CC you hold two copies of C, with CT one, with TT none, and the report compares your copies with twice the frequency of the study it picks among that condition's studies of the variant that give a frequency, the one with the narrowest interval (for myocardial infarction, PubMed 33532862, odds ratio 1.11, frequency 0.14, so the average person carries about 0.28 copies of C). The copies above or below average, times the log odds ratio of each variant, add up across the condition's variants your file measures (for measured levels in the nutrition report, the same copies times the betas in standard deviations, or by direction only when units differ); the sum, taken back to an odds ratio, is your odds relative to the average person, and 1.20 or more prints as higher odds, 0.83 or less as lower, between them as no meaningful difference. When fewer than a third of the condition's known variants are measured, the condition is not assessable, and when fewer than a third of its known variants can be compared with the average person, it is not comparable. The guide to reading the health report walks through this on an actual page.

Three details follow from the Catalog's conventions. Since 2021 the Catalog no longer inverts odds ratios below 1 (file headers, checked 14 October 2026), so 2,082 of our rows describe an allele that lowers the odds. Where the paper did not name the risk allele, the Catalog stores "?"; the generator compares that letter with your genotype, so those rows are dropped when the catalog is built. And a condition can only be judged if your file contains its rsIDs: a consumer array reads a few hundred thousand positions, so a typical file holds roughly a quarter of the catalog's lead variants, and most conditions with many variants print as not assessable.

Why two tools read the same file differently#

Give the same 23andMe file to two services and you get two different lists. The reasons are choices, not errors.

• Source. Promethease reads SNPedia, a wiki that summarises papers of every design, candidate-gene studies included, and that has not been actively maintained for several years according to third-party reviews. A Catalog-based report inherits the Catalog's eligibility rules instead; each source has entries the other lacks. Our Promethease review goes through that trade-off.
• Thresholds. One tool keeps every Catalog row down to 1 × 10⁻⁵, another cuts at 5 × 10⁻⁸, a third requires a confidence interval.
• Chip coverage. 23andMe, AncestryDNA and MyHeritage use different arrays and array versions; a variant the Catalog knows may not be on your chip at all.
• Strand and build. The Catalog takes risk alleles "exactly as reported in the original publication" and "cannot systematically correct" for strand flips between genome builds (FAQ, E7, checked 14 October 2026). Consumer files generally report the forward strand, so for an A/T or C/G variant a paper on the other strand can make "risk allele A" look like your "T". Any tool that matches letters, ours included, can misread such a row.
• Aggregation. A product of odds ratios, a sum of log-odds weighted by allele count, a polygenic score with LD adjustment: three defensible methods, three different totals from the same rows.

Limitations#

• Ancestry. As of mid-2021, 86.3% of GWAS samples in the Catalog came from people of European descent [8]. Effects transfer imperfectly to other populations; a report for someone of East Asian, African or South Asian ancestry rests on thinner evidence.
• Associations are not causes. A lead variant marks a region. The causal variant is often a neighbour in linkage disequilibrium (LD) with it, and neither the Catalog nor a consumer report says which gene matters or why.
• LD and double counting. Different studies pick different lead SNPs for the same signal, and counting each of them counts one signal several times. We keep one lead variant per region of 250,000 bases for each trait, but do not model finer linkage between variants.
• Winner's curse. The first study to report a variant tends to overestimate its effect, especially when underpowered [7]. We take each variant's most precisely estimated study of that condition, among those that give the allele frequency, rather than an average, and do not correct for this bias.
• No baseline risk. An odds ratio of 1.3 says nothing about how common a condition is. "Higher odds" is relative to the average person of the study population, not an absolute probability.

An association is a statistical signal in a study population, not a cause, and it does not tell you whether you have or will develop the condition. If a row worries you, bring the PubMed link to a doctor or genetic counselor.

References#

1. GWAS Catalog, "About". EMBL-EBI. ebi.ac.uk/gwas/docs/about.
2. GWAS Catalog, "Eligibility criteria". ebi.ac.uk/gwas/docs/methods/criteria.
3. GWAS Catalog, "File headers". ebi.ac.uk/gwas/docs/fileheaders.
4. GWAS Catalog, REST API metadata, release 2026-10-03. ebi.ac.uk/gwas/rest/api/v2/metadata.
5. Cerezo M. et al. "The NHGRI-EBI GWAS Catalog: standards for reusability, sustainability and diversity." Nucleic Acids Research 53(D1): D998–D1005, 2025. doi:10.1093/nar/gkae1070.
6. Pe'er I., Yelensky R., Altshuler D., Daly M. J. "Estimation of the multiple testing burden for genomewide association studies of nearly all common variants." Genetic Epidemiology 32(4): 381–385, 2008. doi:10.1002/gepi.20303.
7. Xiao R., Boehnke M. "Quantifying and correcting for the winner's curse in genetic association studies." Genetic Epidemiology 33(5): 453–462, 2009. doi:10.1002/gepi.20398.
8. Fatumo S. et al. "A roadmap to increase diversity in genomic studies." Nature Medicine 28: 243–250, 2022. doi:10.1038/s41591-021-01672-4.

EBI pages were checked on 14 October 2026.

FAQ#

Is the GWAS Catalog free to use?

Yes. It is a public resource funded by NHGRI and EMBL-EBI, searchable at ebi.ac.uk/gwas, with full downloads on EBI's FTP site. You can look up any rsID from your raw file there yourself, which is also how to check a row of your report; what a report adds is the matching, filtering and aggregation.

Why is MTHFR not scored in my report?

The methylation claims around MTHFR rs1801133 come from candidate-gene studies, which the Catalog excludes. No disease row for rs1801133, rs1801131, COMT rs4680 or CYP1A2 rs762551 passes our filters, so none of them is scored in a Full Human health report. rs1801133's only rows in our catalogs are two genome-wide significant folate measurements in the nutrition catalog (PubMed 37479695 and 39414775, allele A), which give no allele frequency, so the nutrition report's vitamin B9 section lists the variant when you carry A, marked not scored, and it does not count towards the tendency. The methodology page lists these exclusions and the reasons.

Does Full Human use the newest Catalog release?

Not always the very newest. The methodology page prints the release currently in use. At the time of writing the generator runs on the 2026-06-01 release while EBI's newest is 2026-10-03; a scheduled workflow rebuilds our catalogs from each release and brings it in after review, so the lag is normally a few weeks.

What is the difference between p < 1 × 10⁻⁵ and p < 5 × 10⁻⁸?

Both are cut-offs for how unlikely a result is by chance alone. The Catalog records associations down to 1 × 10⁻⁵ so that suggestive hits stay available to researchers. Genome-wide significance, 5 × 10⁻⁸, corrects for the roughly one million independent tests in a GWAS. Full Human keeps only rows below the stricter value.