How greatgenes works and how reliable it is
greatgenes compares the raw data file from a consumer DNA test with public scientific databases and writes down what those databases say about the variants in your file. This page explains which databases we use, how the report is put together, how much each part can be trusted, and what a DNA chip cannot see.
No guesswork. greatgenes does not use computational “pathogenicity predictors” (such as CADD or SpliceAI) and does not build its own risk models. Each line of the report repeats what a curated database (ClinVar, CPIC, ACMG SF), a published study (GWAS Catalog, PGS Catalog) or a reference resource records — with the source named and its reliability level shown. Built for consumer DNA chips: greatgenes is designed for raw data from 23andMe, AncestryDNA, MyHeritage and FamilyTreeDNA; it reads only the positions a chip measures and says which ones it could not read.
- Databases
- 13 public databases, each named with its licence
- Knowledge base built
- 2026-10-23
- ClinVar records
- 198,495 single-nucleotide variants, of which 195,720 are classed as pathogenic or likely pathogenic (release of 2026-10-04)
- Medicines
- 10 genes by CPIC tables (snapshot 2026-10-13)
- Polygenic scores
- 15 published scores from the PGS Catalog
- Traits from GWAS
- 54 traits and conditions from the GWAS Catalog
- Chip versions recognised
- 11
- Where the analysis runs
- in your browser; the file is not uploaded
Databases and versions
The knowledge base is built from these databases. The version or date is that of the copy used for the current knowledge base (built 2026-10-23). Licences and terms are on the page Data sources and licences.
| Source | What we take | Version / date |
|---|---|---|
| ClinVar | clinical significance of variants | release of 2026-10-04 |
| GWAS Catalog | variant-trait and variant-disease associations | downloaded 2026-10-12 |
| PGS Catalog | polygenic scores | 15 scores, each named by its PGS ID |
| SNPedia | variant descriptions and significance (magnitude, repute) | downloaded 2026-10-12 |
| Ensembl (1000 Genomes) | population allele frequencies (1000 Genomes) | GRCh37, downloaded 2026-10-20 |
| CPIC | pharmacogenetic guidelines: gene, drug, recommendation | snapshot 2026-10-13 |
| DPWG | pharmacogenetic guidelines of the Dutch working group | cited per guideline publication |
| PharmGKB / ClinPGx | gene-drug annotations, references in the dictionary and panel | clinical annotations of 2025-07-05 |
| ClinGen | gene-disease validity, actionability, ACMG SF list | file of 2026-10-13 |
| MedlinePlus | gene and condition descriptions, inheritance type | version 2025-06-02 |
| Orphanet | rare disease codes and names, inheritance type | export of 2026-06-23 |
| 1000 Genomes Project (IGSR) | allele frequencies of pharmacogenes in populations | phase 3, snapshot 2026-10-14 |
| gnomAD v4.1 | allele frequencies of pharmacogenes in populations | v4.1, snapshot 2026-10-14 |
Databases are revised and sometimes disagree. The report shows what the databases said in these versions; it does not decide which source is right, and we do not check that the information in the databases is correct.
How the report is organised
The report is built in your browser and goes from the most practical to the most general:
- What we read from your file — sex by DNA, a few likely visible traits and the number of positions read, so you can check that the file is yours.
- Summary — up to 8 main findings, each with what it means, a next step and a reliability word.
- Health-related findings, Carrier status, Lifestyle and nutrition, Traits.
- Medications (pharmacogenetics) — how your genes may affect the way your body handles some medicines.
- Sensitive results — three collapsed blocks that every report has, whatever the result: Alzheimer’s disease risk (APOE), Parkinson’s disease risk (LRRK2, GBA) and genes of the ACMG secondary-findings list. Some people prefer not to learn these results, so they are not in the Summary and open only if you choose; a block you open is printed, and in the saved file they stay collapsed.
- Polygenic scores — only the results that differ from the average are shown openly.
- Checked, no signs found — what was looked at and found nothing, including 17 well-known variants listed by name and whether your chip reads them.
- About your data, Sources and Not shown — the rest of the matches, which you can download as a full table.
Each finding names its sources. Short quotes from databases are marked as quotes; the rest is retold in our own words.
Finding priority and reliability
In short: a finding ranks higher when a well-reviewed clinical database calls the variant pathogenic, and its reliability word (high, medium, low) shows how well the sources support the link between the variant and the description. Reliability is not the probability of a disease.
How ClinVar review status is used
ClinVar records how thoroughly each classification was reviewed and shows it as stars (ClinVar: review status). greatgenes turns the review status into a weight:
| ClinVar review status | ClinVar stars | Weight in greatgenes |
|---|---|---|
| practice guideline | four (★★★★) | 4 |
| reviewed by expert panel | three (★★★) | 4 |
| criteria provided, multiple submitters, no conflicts | two (★★) | 3 |
| criteria provided, single submitter | one (★) | 2 |
| criteria provided, conflicting classifications | one (★) | 1 |
| no assertion criteria provided | none | 1 |
Records with conflicting classifications are not used as pathogenic. Pathogenic records without assertion criteria (no stars) do not raise a finding on their own; the “Not shown” section counts those left in the full table.
Priority
- Highest priority: ClinVar calls the variant pathogenic or likely pathogenic with a weight of at least 2 — that is, with assertion criteria and without conflicts.
- Next: the sources point to a risk, with SNPedia magnitude of at least 3 or a reliability score of at least 3, and no reliable source says the genotype is normal.
- Then: hand-checked panel entries, ranked by the strength of their evidence, and medicine-related findings.
- Everything else goes to the full table under “Not shown”.
Where our hand-checked dictionary has an entry for the variant and genotype, its category and priority are used.
Reliability score
The score is the weight of the strongest source, plus 0.5 for each further source that agrees, plus 0.5 if another variant in the same gene points the same way. That other variant counts only if it lies more than 100 kb away: closer variants are often inherited together and do not confirm each other. A score of 3.5 or more is high, 2 or more is medium, anything lower is low.
If the chip reports a pathogenic or likely pathogenic variant that is rare (below 0.1 % in the population) or has no known frequency, the finding is not one of the main findings in the Summary, whatever the gene: chips often misread such variants. Its reliability is shown as low — needs confirmation, the finding starts with a note that a clinical confirmation test is needed before any conclusion, and the Summary gives only their number. If the chip reports two copies of such a variant, the report also says that this is a likely chip error.
Medicines (pharmacogenetics)
In short: for 10 genes the report works out the likely gene variants and the metabolizer type using CPIC tables, and retells the CPIC recommendation in plain words. Never change a medicine or dose on your own; show the result to your doctor or pharmacist.
Genes: CYP2B6, CYP2C19, CYP2C9, CYP2D6, CYP3A5, DPYD, G6PD, NUDT15, SLCO1B1, TPMT. The rules are cautious on purpose, so as not to claim more than the chip shows:
- A chip does not tell which copy of a chromosome each variant sits on, so all compatible pairs of gene variants are considered. A metabolizer type is named only if it reaches 95 % of the probability in every population group in the CPIC frequency tables (your ancestry is not known); otherwise the possible types are listed.
- Common gene variants with altered function that a chip cannot read (insertions, deletions, or positions missing from the file, such as an insertion in CYP3A5 that is common in people of African ancestry) are included in this calculation with their population frequencies: a copy read as normal may be one of them. If this alone lowers the share below 95 % but not below 90 %, the type is given as “likely …”; if lower, the possible types are listed.
- A gene variant found in fewer than 0.1 % of people (or of unknown frequency) is reported only as “possible, confirm with a clinical test”, without a recommendation. Between 0.1 % and 1 %, the recommendation is given with a note that the result needs confirmation.
- Two copies of a variant that occurs only in very rare gene variants are treated as a likely chip error, and the gene is not named.
- The number of gene copies (important for CYP2D6) and insertions or deletions cannot be read from a chip. For CYP2D6 a recommendation is given only when both copies read on the chip have no function; otherwise the type is given “by chip positions, without copy number”.
- CYP2D6 is not assessed at all when the file lacks the position of its most common non-working variant in people of European ancestry. This is the case for these chip versions: 23andMe v2; 23andMe v5; AncestryDNA v1 / FamilyTreeDNA (OmniExpress); AncestryDNA, array c5.0.0; AncestryDNA, array c5.1.0; Illumina GSA v5.1.0 (FamilyTreeDNA, 23andMe). For them the report says that CYP2D6 was not assessed.
- DPYD: a chip reads only some of the gene variants that matter for capecitabine and fluorouracil, so a “normal” or “likely normal” DPYD result from a chip does not replace a clinical test. The US Food and Drug Administration safety labeling update (published 2026-02-10) says the Boxed Warning of both medicines now advises DPYD testing before treatment starts, unless immediate treatment is necessary.
- G6PD lies on the X chromosome, so the reading depends on sex: if sex cannot be determined and the data do not allow a safe reading, the gene is not assessed.
How often the named types are right: we compared the report with the consensus genotypes of the CDC GeT-RM reference programme for 147 samples that are also in 1000 Genomes, using files with only the positions of each chip version. For a 23andMe v5 file the report named a metabolizer type or gene function for 66 % of the gene results, and 94.7 % of those were right; for AncestryDNA, array c5.1.0, 51 % and 93.2 %. With every position of the gene tables read, the figures are 80 % and 96.3 %. The rest were reported as not assessed or not determined (“likely …” counts as named). Most wrong types come from variants a chip cannot read (insertions, deletions, gene copies) or from positions missing on the chip; the others are samples where the reference data disagree with each other.
The metabolizer types and recommendations in this section follow CPIC only; recommendations of the Dutch Pharmacogenetics Working Group (DPWG) are not yet included. DPWG is cited in a few single-variant rows of the report (for example, UGT1A1 and irinotecan). Active substance names are used, not brand names.
Polygenic scores and traits
In short: a polygenic score adds up many small effects. It compares you with people of European ancestry; it is not the probability that you will get a disease. A score is used only if your chip covers at least 50 % of it, and the table shows openly only results at least 1 standard deviation from the average.
Published scores (PGS Catalog)
- 15 published scores were chosen because consumer chips cover them well; of these, 8 have high reliability (evaluated in an independent sample, with coverage measured from real allele frequencies), the rest medium. We use the published weights as they are: no weights of our own and no imputation.
- Coverage is measured as the share of the score’s variance that your chip’s positions explain (Choi, Mak, O’Reilly 2020). Below 50 % the score is marked “not assessed” and no percentile is given.
- The percentile is calculated against the genomes of the European participants of 1000 Genomes: the average and spread of the score are taken from their real genotypes, on exactly the variants found in your file, so that variants inherited together are not counted as independent. Scores are less accurate for people of other ancestries (Martin et al. 2019).
Traits from the GWAS Catalog
- 54 traits and conditions. For each, associations are taken with p ≤ 5 × 10⁻⁸, from studies whose first-stage sample is at least 80 % European, one estimate per study, combined by meta-analysis.
- Within 500 kb only the variant with the strongest association is kept, so that neighbouring variants inherited together are not counted twice.
- Psychiatric and cognitive traits are not included.
Sex-specific results
2 scores and 4 traits apply to one sex only (for example, breast cancer and prostate cancer). Sex is determined from the X and Y chromosome data in the file. If it is known, results for the other sex are not shown. If it cannot be determined, results for both sexes are shown, each marked “for women” or “for men”, with one sentence above the table — use the ones that apply to you.
What is shown openly
Rows where the result is about average, or where the score did not pass the coverage threshold, are folded into “Details” under the table, with a one-line count. The full table, sorted by how far the result is from the average, is always there.
What a DNA chip cannot see
In short: a consumer chip reads between 574,516 and 992,615 chosen positions, depending on the chip version, and makes mistakes on rare variants. If a variant is not in the report, it does not mean you do not have it.
- Rare variants are often wrong. In a study comparing chip results with sequencing in UK Biobank, for variants with a frequency below 0.001 % only 16 % of 4,757 chip calls were confirmed; for pathogenic BRCA1 and BRCA2 variants the positive predictive value was 4.2 % (Weedon et al., BMJ 2021).
- Raw data from direct-to-consumer tests need confirmation. A clinical laboratory did not confirm 40 % of the variants from such raw data that were sent to it for testing (Tandy-Connor et al., Genet Med 2018). That is why two copies of a rare pathogenic variant are flagged as a likely chip error.
- No imputation. greatgenes uses only the positions your chip actually measured; it does not guess unmeasured positions from reference panels. As a result, many modern polygenic scores, built for imputed genomes, are not used.
- Not read by a chip: most insertions and deletions, the number of gene copies, structural changes, and whole genes. BRCA1 and BRCA2 are not assessed by chip, and the report says so.
- Checked positions only. “Checked, no signs found” covers the pathogenic and likely pathogenic ClinVar variants at positions your chip reads, including those in the 81 genes of the ACMG SF v3.3 list. It always says “no signs in your data”, never “you do not have the disease”.
Where your file is processed
In short: the analysis runs in your browser tab. Neither the file, nor any part of it, nor its name, nor the report is sent to the server.
The browser downloads the analysis engine and the part of the knowledge base for your chip model, then reads the file and builds the report on your device. To get these, it sends the server only service requests:
- the answer to a small computing task (a check that the request comes from a browser);
- a session number in a cookie;
- the chip model name, which is the name of the knowledge base part being downloaded — the same for everyone with that chip;
- sometimes a request for the whole knowledge base instead of the part for your chip: when the chip model is not recognised or the file has positions outside the chip’s usual set — the server then learns that, but not which positions;
- short signals about the analysis steps: which step finished and how (“ok” or the kind of error), with no file data;
- the ordinary details any browser sends, and your connection address, as with any website. The greatgenes server program keeps the address in memory only to limit the number of analyses and does not write it to its log.
A message from the feedback form reaches the server only if you send one. You can check all of this yourself in the browser’s Network tab: How to check that your file stays with you. The full rules are in the Privacy policy.
What the report is not
greatgenes is an information service. The report records what public databases say about the variants in your file; it is not a diagnosis and not medical advice. Discuss any health finding with a doctor or a clinical geneticist, and do not make medical decisions before a clinical test confirms the result. Never start, stop or change a medicine on your own.
References
- ClinVar: review status and star ratings (NCBI).
- CPIC guidelines (Clinical Pharmacogenetics Implementation Consortium).
- Lambert SA et al. The Polygenic Score Catalog. Nat Genet 2021. doi:10.1038/s41588-021-00783-5
- Sollis E et al. The NHGRI-EBI GWAS Catalog. Nucleic Acids Res 2023. doi:10.1093/nar/gkac1010
- Choi SW, Mak TS, O’Reilly PF. Tutorial: a guide to performing polygenic risk score analyses. Nat Protoc 2020. doi:10.1038/s41596-020-0353-1
- Martin AR et al. Clinical use of current polygenic risk scores may exacerbate health disparities. Nat Genet 2019. doi:10.1038/s41588-019-0379-x
- Weedon MN et al. Use of SNP chips to detect rare pathogenic variants: retrospective, population based diagnostic evaluation. BMJ 2021. doi:10.1136/bmj.n214
- Tandy-Connor S et al. False-positive results released by direct-to-consumer genetic tests highlight the importance of clinical confirmation testing for appropriate patient care. Genet Med 2018. PMID 29565420