The Science
How we select variants, what the evidence requires, and where the honest limits are.
How this works
Your 23andMe or AncestryDNA raw data file is a record of hundreds of thousands of positions in your genome where people commonly differ. Elpis reads that file entirely in your browser — using a parser written in Rust and compiled to WebAssembly — extracts the specific positions we study, and maps them to published research findings.
The output is a set of interpreted results in plain English, linked to the research behind them. No raw genotypes leave your browser. The interpretation is done locally, and that is a technical constraint, not a promise we make and hope you believe.
How we choose what to include
Every variant in our panel must clear three requirements:
It has to be on the chip.
We only report variants directly genotyped on arrays underlying 23andMe v5 and AncestryDNA data. We do not impute. If the call isn't in your file, we say so.
It has to have peer-reviewed, replicated evidence.
A single preliminary study is not enough. We require published research with meaningful sample sizes, and we weight confidence labels accordingly — strong, moderate, or limited.
It has to be responsible to report.
Some variants carry psychological weight requiring clinical context to interpret safely. We exclude those — not because we want to withhold data, but because giving someone a result without the right support does more harm than good.
What we permanently exclude and why
✕
APOE (rs429358)
The strongest common genetic signal for Alzheimer's disease risk. Excluded on both technical and ethical grounds — this result requires a genetic counselor, not a browser report.
✕
FTO (rs9939609)
A well-replicated body weight variant. In an athletic context, reporting it creates stigma without clinical benefit. Replaced with cleaner metabolic variants.
✕
MAOA (rs6323)
The so-called "warrior gene." The framing has been thoroughly discredited and has caused real harm. We do not report it under any circumstances.
✕
5-HTTLPR (SLC6A4)
A variable number tandem repeat — not a SNP. It cannot be reliably called from array data. Excluded until better calling methodology is available.
Confidence levels
Strong · Replicated
Confirmed in multiple large studies across different populations. High confidence the finding is real.
Moderate
Published, peer-reviewed evidence exists, but replication is limited or effect sizes are modest.
Limited
Preliminary evidence only. Reported because the published signal exists, but interpret with significant caution.
The ancestry problem
Most large-scale GWAS studies have been conducted primarily in European-ancestry populations. This is a real problem, and it would be dishonest to pretend otherwise. Allele frequencies, effect sizes, and associations may be less accurate for people of non-European ancestry.
We address this the best we can with the data that exists. Every report notes this limitation explicitly. The bias in genomics was not created by users. Fixing it requires them.
Citation table coming soon. A full table linking each variant to its source studies, effect sizes, sample sizes, and population context is in development.