Category: Forensic laboratory tooling | Status: Working prototype, in validation
An automated quality-control screener for DNA analysis that reads instrument exports and reviews plates for contamination, dirty blanks and cross-batch profile discrepancies — the whole-plate QC review that analysts currently perform by hand.
What it does
The tool ingests capillary-electrophoresis export data and produces a prioritised worklist of the sample pairs most worth a human look. Its primary detector compares profiles for the same sample across plates and ranks discrepancies by allele gain — a profile that picks up alleles between runs is a mixture emerging, which is exactly the signal contamination produces. Where cross-plate comparison is not available, a fallback rarity-weighted detector ranks the most suspicious pairs within a single plate.
Recall-first, by design
The single most important design decision is the error trade-off. In contamination screening, missing a real contamination is far worse than raising a false alarm. The tool is therefore tuned to favour recall over precision, and its output is explicitly framed as screening and triage items feeding the laboratory's existing disposition workflow — never as determinations. An analyst dispositions the flags exactly as they already disposition QC findings.
Why it earns its place
Manual whole-plate review is exhausting and unforgiving: a large amount of comparison work where the overwhelming majority of pairs are unremarkable and attention degrades over time. Turning that into a ranked worklist means the most suspicious pairs are examined while the analyst is freshest. Any reliable prioritisation is a real gain here.
Current status
The engine works, and on a labelled baseline plate set it reproduced known contaminations — the known-answer test passed, with a small, manageable number of deterministic flags.
Two things stand between this and validation. An importer is needed for a newer wide-format export (mechanical work, no tuning decisions involved). More significantly, expert confirmation of the complete ground-truth label set is underway. Right now, the underlying algorithm is validated against one baseline and not yet ready for widespread deployment.