atlageneScience
Evidence first, with its limits in plain view.
A report is a reading of what one file measured, against what public evidence supports. Both halves have limits, and the report should show them.
Four rules every result follows.
Coverage, not confidence
Genotyping arrays measure a fraction of known variants. Each category reports the share of its panel the file typed, so a thin result is visibly thin.
Unmeasured is not normal
A position the file never measured is left out of the result. It is not scored and it is never presented as a reassuring finding.
Safety flags per variant
Whether a finding is safety-critical is decided for each variant, not for a whole category, and a flagged finding always gets the full write-up, whatever the report settings.
Reviewed definitions only
A variant definition is drafted with AI, challenged by a separate AI review, then countersigned by a person before any report can use it. Every change is kept in its history.
What an array cannot tell you.
Genotyping arrays read single positions. Most deletions, copy-number changes and repeat expansions are invisible to them. For genes where those are the common cause of disease, a result without a variant is not a clear screen, and the wording has to say so.
Arrays can also report rare variants that are not there: the rarer the variant, the more likely the call is wrong. A rare pathogenic finding from an array needs confirmation by a validated clinical test before anyone acts on it.
Results are informational. They support, and never replace, the judgement of the clinicians who review them. Routine clinical use of any result needs validation with the laboratory’s own assay.
We do not currently report polygenic risk scores as part of the laboratory pilot offering described here. Laboratory reports do not include them.
Sources.
Definitions draw on ClinVar for clinical significance, the GWAS Catalog for trait associations and PharmGKB for medication response. Allele frequencies come from gnomAD.
AI is used upstream, to draft and challenge variant definitions, and a person countersigns each one. When a report is generated, its text comes from those reviewed definitions and fixed templates. No AI model writes report text.
Read: what an array can and can’t tell youBring us your assay.
Tell us what you run, who reads the report and what you want an evaluation to answer. We reply with a scoped plan.