Sequencings

Method

How this works, and where it stops

Everything here should be checkable by somebody who knows the field. This page exists so that it is.

Reading the file

Four vendor formats are supported: 23andMe and AncestryDNA as tab-separated text, MyHeritage and Family Tree DNA as quoted CSV. AncestryDNA splits the two alleles across separate columns and numbers the sex chromosomes 23, 24, 25 and 26 rather than naming them, both of which are handled. Files are read in eight-megabyte slices so that the same code path handles a 20 MB chip export and a multi-gigabyte whole-genome VCF without the tab ever holding more than one slice.

Markers are matched by rsid first and by chromosome and GRCh37 position second. The fallback matters more than it sounds: Y-chromosome markers are frequently written with vendor-internal identifiers rather than rsids, and a tool that only looks up rsids silently misses most of them.

Strand orientation, which is where tools go wrong

DNA is double stranded. A position that reads C and T on one strand reads G and A on the other, and both are the same fact written two ways. Vendors report some positions on the forward strand and some on the reverse, and the file does not say which. So a report that looks up a marker and asks “is it T?” is, for some files, silently asking the wrong question and answering it with confidence.

Every marker in this panel carries its forward-strand allele pair, taken from Ensembl GRCh37. Each observed genotype is resolved against that pair: used as it is if the alleles match, flipped if their complements match, and refused if neither does. The report tells you when a genotype was flipped.

One class cannot be solved this way. If a variant’s two alleles are complements of each other, which is to say A and T, or C and G, then flipping produces the same pair and no test can tell the strands apart. Those markers are declined by computation rather than by anyone remembering to exclude them, and the report shows them as declined with the reason. Two markers in the current panel fall out this way.

What is deliberately excluded

No rare pathogenic variants. Not BRCA founder mutations, not CFTR, not any of them. Ambry Genetics took 49 people’s raw consumer files to a clinical laboratory and found 40 per cent of flagged variants were false positives, with 8 further real variants misread by interpretation services.

That failure is not evenly spread. It concentrates in rare variants, where a genotyping array has very little signal and a single mis-clustered probe produces a confident wrong answer. So the panel has a frequency floor: nothing below roughly one per cent minor allele frequency in every major population, and nothing imputed.

Imputation is the usual way to claim more coverage than a chip has. It works well for common variants and poorly for the ones that carry clinical weight: 0 to 1 per cent of very rare variants are recovered reliably. More analysis does not make thin data thick, and this site does not pretend otherwise.

The coverage audit

For each of 25 genes we hold its GRCh37 coordinates and the number of variants ClinVar classifies as pathogenic or likely pathogenic. The report then counts how many positions inside those coordinates your own file actually contains. The ratio is the honest answer to whether raw data covers that gene, and for every gene on the list the answer is no.

ClinVar counts retrieved 2026-09-22 through NCBI E-utilities, using the query GENE[gene] AND ("pathogenic"[Clinical significance] OR "likely pathogenic"[Clinical significance]).

Haplogroups

Both lines are called by walking a published phylogeny from the root. At each fork, the defining mutations of every candidate branch are looked up in your file and scored as supporting, contradicting or untestable, and the walk descends into the best-supported branch until no branch has support.

The report then states where it stopped and why, which matters more than the label. A chip carries a few thousand Y positions and a few thousand mitochondrial ones, chosen years ago, and below a certain depth the defining mutations for a branch simply are not on it. Most tools present wherever they halted as the answer. The same haplogroup name from a file that tested four markers and one that tested forty are not the same claim, so both numbers are shown.

Scale, for context

A consumer test reads about 650,000 positions out of roughly 3,100,000,000. A 30x whole-genome sequence reads every position about thirty times and typically finds 4,800,000 variants, at 99.98 per cent at 30x. If you want materially more than a chip file can give, the answer is more data, not more interpretation.

Sources

Corrections

If you think something here is wrong, please say so. Interpretations get revised, sources get superseded, and the cheapest way to find an error is to make it easy for the person who spotted it to report it. The contact form has a topic for exactly this.