Our Science: How GenomeNarrator Reads Your DNA

This page explains how GenomeNarrator analyzes a 23andMe, AncestryDNA or VCF file, and how well it does it. Every figure was checked against the code or a recorded measurement, and the accuracy results are dated, including the weak ones. Reviewed 24 September 2026.

Key numbers: more than 350,000 ClinVar variant records indexed, rebuilt monthly; all 84 ACMG SF v3.3 genes screened on every upload; 133 gene-drug pairs, 55 at CPIC level A; allele frequencies from 807,162 people in gnomAD v4.1; 121 curated traits; 80 blood tests; ancestry calls correct for 98.0% of 2,347 reference genomes; medication calls agreeing with PharmCAT for 99.4% of 46,341 gene calls in 3,202 people.

Your file

Your file is read in your browser. Only the genotypes the analysis needs leave your device, and the service that scores them keeps no copy. The file itself is never uploaded.

The pipeline

Read and normalize. Every genotype is mapped to one genome build and to standard rsIDs. Unreadable lines and no-calls are discarded, never guessed.
Infer ancestry. Ancestry-informative markers are compared with reference frequencies for eight populations. This decides which population's allele frequencies, polygenic reference and founder-variant panels the later stages use.
Match against the references. Each genotype is looked up in ClinVar, the GWAS Catalog and our curated condition list.
Filter and gate. Duplicates are merged and linked variants are clumped. A single-gene finding needs a Pathogenic or Likely Pathogenic ClinVar record with at least two review stars (one for ACMG genes), a matching allele and a germline condition.
Classify. Candidates are checked against the ACMG/AMP criteria genotype data can support (8 of 28), and carrier or affected is decided from the gene's inheritance pattern, not the disease name.
Pharmacogenomics. 101 marker positions across 37 genes and 4 HLA alleles are assembled into star-allele diplotypes and metabolizer phenotypes, then into guidance drawn from CPIC guidelines.
Score and explain. Polygenic scores are ranked against your inferred population, an absolute risk is estimated where a baseline prevalence exists, and every finding keeps its evidence trail.

How we check our work

Ancestry: 98.0% correct on 2,347 people from 1000 Genomes (25 September 2026), and the right regional population for 90.7% of 2,248 (21 October 2026). The admixed Puerto Rican (76.9%) and Colombian (89.4%) sets score lowest.
Chromosome painting: 4 of 5 gated population pairs pass. Middle Eastern against European and Finnish against European are weakly separated: the correct side usually leads, but as a visible colour blend rather than a clean boundary.
Pharmacogenomics: 99.4% agreement with PharmCAT, the reference caller from PharmGKB, across 46,341 gene calls in 3,202 people, with no gene called abnormal that PharmCAT called normal; 87.1% agreement with CDC GeT-RM reference samples across 961 comparisons, lower for CYP3A4 (most of its alleles are not on consumer arrays).
Rare chip calls: a disease variant rarer than 0.001% in gnomAD, read from a consumer chip, is labelled an unconfirmed chip call and never ranked urgent, because in UK Biobank only 16% of such calls were confirmed by sequencing (Weedon et al., BMJ 2021).

Every one of these measurements, with its answer key, sample, date, grading method and exclusions, is in the Technical Validation Note (version 1.2, 22 October 2026). Polygenic scores have not been measured against health outcomes, and no result has yet been reproduced by an independent laboratory.

Genetics and your labs

Bloodwork results are cross-referenced against your genes through 85 cited gene-biomarker links across 64 genes: 15 rest on causal or functional evidence, 57 on strong association and 13 on moderate association. 47 state a published effect size, drawn as an expected-for-your-genotype band beside the lab range.

Known limitations

A consumer chip reads roughly 600,000 to 700,000 positions. ClinVar lists 211,399 variant IDs as pathogenic or likely pathogenic, and a typical chip types about 2,962 of them, around 1.4%. A clean chip result is weak evidence that you carry nothing.
Large deletions, duplications and copy-number changes do not show up on a SNP array.
Pharmacogenomic star alleles are inferred from tag SNPs, and CYP2D6 gene duplications and deletions can be missed.
Most genetic research over-represents people of European ancestry, so scores are less accurate for everyone else.
GenomeNarrator is a health-education tool, not an FDA-cleared diagnostic. Confirm any finding you would act on with a clinical laboratory.

Sources

ClinVar (rebuilt monthly; every report names the release it was scored against), gnomAD v4.1, the NHGRI-EBI GWAS Catalog, the PGS Catalog, CPIC guidelines, the ACMG SF v3.3 list, and AlphaMissense. 1000 Genomes, GeT-RM and Genome in a Bottle are used only to measure our accuracy.

Data sources and credits

The open datasets behind this product, the licence each one is used under, and the credit its providers ask for.

Google DeepMind AlphaMissense predictions (CC BY 4.0). AlphaMissense pathogenicity predictions are from Google DeepMind and are used under the Creative Commons Attribution 4.0 International licence (CC BY 4.0).
Cheng J, et al. Accurate proteome-wide missense variant effect prediction with AlphaMissense. Science. 2023;381:eadg7492. PubMed
NCBI ClinVar (US government data, no restrictions). ClinVar, National Center for Biotechnology Information (NCBI), is a data source for the variant classifications in this product. NCBI asks that its disclaimer and copyright notice be evident to users: see https://www.ncbi.nlm.nih.gov/home/about/policies/.
Landrum MJ, et al. ClinVar: improving access to variant interpretations and supporting evidence. Nucleic Acids Res. 2018;46:D1062-D1067. PubMed
NCBI dbSNP (US government data, no restrictions). dbSNP, National Center for Biotechnology Information (NCBI), is a data source for variant identifiers and merged-rsID records. NCBI asks that its disclaimer and copyright notice be evident to users: see https://www.ncbi.nlm.nih.gov/home/about/policies/.
Sherry ST, et al. dbSNP: the NCBI database of genetic variation. Nucleic Acids Res. 2001;29:308-311. PubMed
gnomAD (Broad Institute) (CC0 1.0). Allele frequencies and gene constraint come from the Genome Aggregation Database (gnomAD), Broad Institute, version 4.1.
Chen S, et al. A genomic mutational constraint map using variation in 76,156 human genomes. Nature. 2024;625:92-100. PubMed
NIST Genome in a Bottle stratifications (US government data, acknowledgement requested). The Genome in a Bottle genomic stratifications are from the National Institute of Standards and Technology (NIST). The file we use was modified: it is a subset covering low-mappability and segmental-duplication regions.
GTEx Portal (GTEx Portal terms). The Genotype-Tissue Expression (GTEx) Project was supported by the Common Fund of the Office of the Director of the National Institutes of Health, and by NCI, NHGRI, NHLBI, NIDA, NIMH, and NINDS. The data used for the analyses described in this manuscript were obtained from the GTEx Portal on 08/16/2026 (GTEx Analysis Release V10, dbGaP accession phs000424.v10.p2). These data are from GTEx v10 and may not be the latest release available from the GTEx Portal. GTEx has not endorsed this product.
GTEx Consortium. The GTEx Consortium atlas of genetic regulatory effects across human tissues. Science. 2020;369:1318-1330. PubMed
NHGRI-EBI GWAS Catalog (EMBL-EBI terms of use). Genome-wide association results are from the NHGRI-EBI GWAS Catalog.
Sollis E, et al. The NHGRI-EBI GWAS Catalog: knowledgebase and deposition resource. Nucleic Acids Res. 2023;51:D977-D985. PubMed
PGS Catalog scoring files (default EBI terms) (EMBL-EBI terms; each score carries its own licence). Polygenic score weights are from the PGS Catalog, EMBL-EBI. Only scores whose own licence allows commercial use are included.
Lambert SA, et al. The Polygenic Score Catalog as an open database for reproducibility and systematic evaluation. Nat Genet. 2021;53:420-425. PubMed
1000 Genomes Project Phase 3 (IGSR) (IGSR data reuse terms). Ancestry reference haplotypes and allele frequencies come from the 1000 Genomes Project, distributed by the International Genome Sample Resource (IGSR).
The 1000 Genomes Project Consortium. A global reference for human genetic variation. Nature. 2015;526:68-74. PubMed
Human Genome Diversity Project (Bergstrom et al. 2020) via IGSR (Fort Lauderdale data sharing principles). Sub-population reference frequencies come from the Human Genome Diversity Project whole-genome data, distributed through IGSR.
Bergstrom A, et al. Insights into human genetic variation and population history from 929 diverse genomes. Science. 2020;367:eaay5012. PubMed
PhyloTree Build 17 Forensic Update (genepi/phylotree-fu-rcrs) (MIT). Mitochondrial haplogroup definitions come from PhyloTree Build 17 as packaged in the phylotree-fu-rcrs repository. Copyright (c) 2022 Institute of Genetic Epidemiology. Used under the MIT licence.
van Oven M, Kayser M. Updated comprehensive phylogenetic tree of global human mitochondrial DNA variation. Hum Mutat. 2009;30:E386-E394. Dur A, Huber N, Parson W. Fine-tuning phylogenetic alignment and haplogrouping of mtDNA sequences. Int J Mol Sci. 2021;22. PubMed
Reactome (CC0 1.0). Pathway membership is from Reactome.
Milacic M, et al. The Reactome Pathway Knowledgebase 2024. Nucleic Acids Res. 2024;52:D672-D678. PubMed
Ensembl (REST API and data) (Ensembl terms, no restrictions). Genome coordinates and variant lookups use the Ensembl REST API and data, EMBL-EBI and the Wellcome Sanger Institute.
Yates A, et al. The Ensembl REST API: Ensembl data for any language. Bioinformatics. 2015;31:143-145. PubMed
CDC/NCHS NHANES public-use files (US government data). Population lab-value percentiles are computed from the National Health and Nutrition Examination Survey (NHANES), US Centers for Disease Control and Prevention, National Center for Health Statistics. We use weighted summary percentiles only, never individual records.
Natural Earth (Public domain). Map boundaries and place names are from Natural Earth.

See pricing, start your genomic analysis, or read our GWAS guide and pharmacogenomics guide.