Our impact
Clinical validity
From our DNA testing platforms and multi-ancestry PRS validation to new discoveries, this is the rigorous science powering our health reports.
23andMe offers two distinct DNA testing platforms to detect genetic variants. Detailed coverage and limitations for each are documented in the package insert and sample report below.
Method
Microarray-based genotyping that tests for the presence or absence of specific, pre-selected variants at hundreds of thousands of positions across the genome.
Analytical performance
Method
Next Generation Sequencing (NGS) that reads the exome, which includes the protein-coding regions of the genome. Although the exome covers roughly 1-2% of total DNA, it includes the vast majority of known disease-causing variants.
Analytical performance
11M+research participants
Built on the world's largest and most diverse genetics research cohort
Research does not stop at discovery. New findings become new or updated insights and improved health tools for your patients.
Step 1
Participate
11M+ members consent to research and share genetic and health data.
Step 2
Research
Scientists run GWAS, discovering new associations and building PRS models.
Step 3
Discovery
Findings are published and translated into validated reports.
Step 4
Insights returned
Updated results reach patients in the account they already have.
23andMe PRS models are developed using a rigorous pipeline designed to maximize risk discrimination while ensuring equitable performance across diverse ancestries.
The largest and most diverse genomic dataset used in PRS development.
Ancestry diversity
11M+ research participants across six ancestry cohorts, more than 3.5M of them of non-European ancestry.
Phenotype depth
1,000+ disease and condition cohorts pulling from self-reported phenotypic data points and linked health records.
Variants are identified within each ancestry cohort, then candidate models are trained and the best performer is selected per ancestry group.
Ancestry-specific GWAS, then meta-analysis
Genome-wide association studies are run separately within each ancestry cohort with sufficient sample size, then combined through meta-analysis to identify the variants most strongly associated with each condition across the included populations.
Candidate variant sets, not a single model
Candidate sets of variants are identified through different permutations of criteria such as p-value threshold and minimum distance between variants. Candidate models are trained for every set in a multi-ancestry training set.
Best model selected per ancestry group
Candidate models are assessed in ancestry-specific validation cohorts that were not part of the GWAS or model training, pairing a best-performing model with each ancestry group. Final metrics come from held-out test sets and are reported in the white paper documentation.
Ancestry treated as a continuum
Raw scores are recalibrated using genetic principal components rather than discrete category labels, so a patient's qualitative result reflects risk relative to a hypothetical population with identical ancestry, not a continental average.
Every PRS model is evaluated in held-out test sets across ancestry cohorts before release.
How performance is measured
Model performance is assessed using AUC, risk stratification odds ratios, and calibration in ancestry-specific held-out test sets. Select models have been independently validated in external cohorts: the UK Biobank for cancer conditions, and the ARIC (Atherosclerosis Risk in Communities) study for multiple cardiovascular conditions.
Performance across ancestry groups
Multi-ancestry development reduces but does not eliminate performance differences across ancestry groups. If a PRS model has not yet met performance standards for a particular ancestry group, a genetic result will not be available for some patients.
Validation spotlights
Area Under the Curve (AUC) describes how well the score separates people who develop the condition from those who do not. A 0.5 represents random chance and a 1.0 indicates perfect classification. PRS for common complex diseases typically yield moderate discrimination, with AUCs ranging from 0.60 to 0.70. Odds ratios (OR) compare the extremes, showing how much more likely the highest-scoring group is to develop the condition than the lowest.
External validation spotlight
European ancestry
Key finding: Performance in the UK Biobank closely mirrors performance in 23andMe cohorts, confirming the model generalizes to an independent European cohort.
View White Paper 23-25External validation spotlight
European ancestry
African American ancestry
Key finding: Risk stratification replicates across both ancestries in external validation, confirming that the model identifies meaningful separation between highest and lowest risk groups regardless of ancestry.
View White Paper 23-21A patient's raw score is calibrated using their birth sex and genetic ancestry, then translated into two outputs.
Qualitative relative risk result
The patient's polygenic score is compared against a reference value computed for a person with the same birth sex and genetic ancestry, so the result is always relative to a matched group rather than a continent-level population average. When the odds ratio typically reaches 1.5 or higher, the report describes an increased likelihood; below that threshold it reports a typical likelihood.
Absolute lifetime risk estimate
The raw polygenic score is calibrated using the patient's sex and genetic ancestry, then converted to an absolute risk, estimating an individual's overall likelihood of developing the condition by a specific age.
Report excerpts shown are from an example breast cancer polygenic risk score report.
A selection of clinically relevant papers from the 23andMe Research Institute, spanning genetic discovery to health outcomes.
2026 · Genetics in Medicine Open
Among 1,076 23andMe participants with actionable variants (BRCA1/2, APOB, LDLR, TTR, others), 46% shared results with a clinician, 86% of those received a medical recommendation, and 87.5% of recommendations were followed.
Read paper2026 · Science
In 3.3 million 23andMe participants with both genetic and lung cancer data, the inherited EGFR T790M variant conferred 25-fold higher odds of lung cancer, the largest known single-variant genetic risk factor for the disease. Prevalence is roughly 1 in 15,000 in the US but reaches 1 in 2,000 in parts of the southeastern US, traced to a founder event in Southern Appalachia about 200 years ago.
Read paper2022 · Frontiers in Genetics
In over 1.5 million participants, a 23andMe T2D polygenic score added meaningful risk stratification beyond age, BMI, and family history, with findings replicating in a Hispanic/Latino cohort.
Read paper2026 · Nature
A GWAS of 27,885 semaglutide and tirzepatide users identified variants in the GLP1R and GIPR genes associated with weight loss efficacy and/or nausea and vomiting.
Read paper2024 · Blood Advances
In 4.2M 23andMe participants spanning five genetic ancestry groups, sickle cell trait conferred an increased risk of VTE, independent of race or ancestry.
Read paperTesting is performed in laboratories certified under the Clinical Laboratory Improvement Amendments (CLIA) and accredited by the College of American Pathologists (CAP).
A completed HIPAA Security Rule analysis aligns controls for systems storing genetic and health-related information with the same federal standards hospitals and clinics use.
ISO/IEC certifications cover information security and related controls. Current certifications, audit scope, and compliance documentation are available through our Trust Center.
The FDA has granted 23andMe 9 authorizations and clearances for direct-to-consumer reports, following rigorous review of accuracy, reliability, clinical validity, and result communication.
The 23andMe Health Digest newsletter
Each month, our team of clinical and scientific genetic experts at the 23andMe Research Institute distills new genomic research, educational resources, and health product offerings into relevant insights for the healthcare community.
Monthly · Free · No spam