How a MarkerView report is built, and what it can and cannot tell you

What goes in, and what we do to it

The only genetic input is the raw data file you download from AncestryDNA or 23andMe. That file is a genotyping array read-out: a fixed set of predefined positions, a few hundred thousand of them, with the letters found at each one. It is not a sequence of your genome. Our converter turns it into a standard VCF, the format annotation tools expect.

• The layout is read from the file, not from the brand. Three tab-separated layouts are handled: 23andMe's four-column and five-column phased exports, and AncestryDNA's five-column export. The column count comes from the file's own header row, or from the most common width across a sample of rows; rows that do not match are skipped, never reinterpreted. No-call tokens and the vendor's insertion and deletion markers carry no usable base, so only A, C, G and T are carried forward.
• Positions come from an rsID lookup, not from your file. Consumer arrays report letters against the older GRCh37 build and carry no reference or alternate allele, so the vendor's chromosome and position columns are ignored entirely. Each rsID is resolved instead against the ANNOVAR avsnp150 table — a snapshot of dbSNP build 150 on GRCh38 — which supplies chromosome, position and both alleles directly. There is no liftover step, only single-letter changes resolve, and a marker whose rsID does not resolve is not analysed.
• Strand is handled by zygosity. At A/T and C/G sites the array's letters cannot say which strand the probe read. A heterozygous call means the same thing either way, so it is kept; a homozygous call is resolved on the forward strand both vendors declare their data uses. Positions where you match the reference are not carried forward, because annotation only needs the differences.

The result then gets a tidying pass — splitting multi-allelic rows, removing duplicates, sorting. That is housekeeping, not the population-level quality control a research study runs: those checks compare thousands of people to each other, and there is one person here. Which files we accept → · How to download yours →

Where the evidence comes from

Findings are ranked by the strength of the evidence behind them, and the report says which kind each one is.

• ClinVar — the NIH archive of variant–condition classifications submitted by laboratories and expert panels. It is the first gate on clinical significance.
• A curated list of common trait markers. ClinVar's categories are right for rare disease alleles and wrong for common trait variants, which it often files as benign or uncertain — a well-replicated coronary marker in the LPA gene is classed benign, and was being discarded for that reason. About seventy markers are therefore admitted by name at both gates, each needing a large, well-replicated effect and a clear interpretation.
• Polygenic scores, for two conditions only. Coronary artery disease and type 2 diabetes are the only panels that can be scored, and the list is an allowlist. A percentile is not a measurement and not a prediction — it is where your combined pattern sits against a population model built from published European-reference allele frequencies. A panel is dropped entirely when your file covers fewer than 70% of its markers, and the report says a measured result outranks it every time.
• Pharmacogenomics from published guidelines. Gene–medicine findings carry the body that issued the guidance and the PubMed ID of the guideline paper — printed as, for example, CPIC guidance · PMID 28198005. Where a finding has no guideline behind it the report says so, and it never asserts a metabolizer status, because single markers are not a full diplotype.
• Drug information from official labeling. Monograph text comes from the prescribing information published on DailyMed by the National Library of Medicine, indexed so the passages for the medicines you actually listed can be retrieved. Interaction rules are a small curated table, each naming the labeling section behind it.
• Supplements, from your measurements rather than your genes. Suggestions are driven by out-of-range lab values and the medicines you listed, never by genotype. Anything that interacts with something you take is dropped rather than printed with a caution, and no dose is ever given.

How the plain-English wording is written

The science is deterministic code. An AI writing model writes sentences, and only sentences. Classification, filtering and every safety rule are ordinary program logic. The model cannot add a finding, remove one, or change how one is classified — the findings are fixed before it sees anything.

Two model calls run for a typical report: one looking for intersections between your DNA, your labs and your medicines, one writing a single sentence per marker for the associations table. Both are given the structured findings only. Where medicines are involved, the relevant passages of the labeling for those specific medicines are retrieved and supplied as context, so the wording is grounded in a document rather than in recall. Anything you write in the free-text health box is passed in fenced and labelled as data, never as instructions.

Every sentence is then re-checked in code before it can print. A sentence is rejected, never rewritten, if it carries directive medication or dose language, an invented statistic, medical advice, an rsID or a gene name, or more than one sentence. A rejected sentence falls back to text the program wrote itself, so a bad generation costs one cell rather than the report. The model is not the source of any number you read.

What this report cannot establish

Consumer arrays are reliable at common variants and unreliable at rare ones. When one laboratory re-tested variants that had appeared in raw direct-to-consumer data, 40% of them proved to be false positives, and some flagged as raising risk were classified as benign on confirmation (Tandy-Connor et al., Genetics in Medicine, 2018 — reference 1).

So MarkerView never headlines a rare alarming call. A pathogenic-looking call is trusted only when it sits on a marker we have deliberately curated — a pharmacogenomic star allele, a well-replicated common marker, or one of two founder variants the chips were designed to call. Everything else goes to a separate appendix, described as unconfirmed and kept away from the action items. In fact no genetic finding of any kind becomes a page-one action item: that list is drawn from medication interactions and measured results. More on that decision →

• Coverage is not completeness. Every count in the report is a count of markers, never a share of your genome or of a gene.
• Absence of a finding never rules a condition out. The array reads predefined positions and does not sequence genes.
• Some things are simply not readable here — which parent a variant came from, how many copies of a gene you carry, insertions and deletions, and repeat expansions.
• Nothing here is a diagnosis, a prescription or a treatment plan, no clinician reviews your individual report, and no dose is ever given. Do not start, stop or change any medicine on the strength of it.
• The path for anything that matters is confirmatory clinical testing ordered by your own clinician, alongside the screening your guidelines already recommend for you.