A new screening tool dramatically reduces false alerts on DNA‑order requests.
Researchers built a certified screener that combines three public‑order signals—a k‑mer Jaccard score, a five‑LLM judge trimmed‑mean, and an embedding‑centroid cosine similarity—and calibrates the fusion with Conformal Risk Control. In ten leave‑one‑family‑out tests on UniProt‑listed toxins, the model missed no hazardous sequences and produced false flags on only one of ten folds. The certified miss‑rate ceiling sits between 24% and 49% depending on shift, but empirical miss rate is zero; the only remaining limit is the size of the calibration set.
The importance lies in exposing the weakest link of DNA‑synthesis screening: data, not algorithms. Existing providers rely on static hazard lists that fail when a toxin comes from an unseen taxonomic family, leading to a 100% false‑flag rate. By certifying risk under distribution shift, the new method shows that robust screening is possible without exhaustive databases, provided enough calibrated examples are gathered. This could push vendors toward larger, publicly shared calibration corpora.
If the community expands the calibration pool to the full UniProt KW‑0800 corpus, the approach could meet procurement‑grade miss‑rate targets (α = 10⁻³). Until then, the next step is to collect and share enough labeled sequences to shrink the certified bound, turning a proof‑of‑concept into an industry‑standard safeguard.