Current DNA synthesis screening tools fail completely when a hazardous sequence comes from a taxonomic family absent from their reference database.
Researchers demonstrated that the standard screening approach, which matches synthesis orders against curated hazard lists, collapses to a 100% false-flag rate under that condition. In practice, the system flags every benign sequence as dangerous because the discrimination signal is too weak to set a meaningful threshold. To address this, the team built CRC-Screen, combining three signals: k-mer Jaccard similarity to known toxins, scores from a panel of five large language models, and cosine similarity to clustered sequence embeddings. A logistic aggregator fuses all three, and Conformal Risk Control, a statistical framework with certified guarantees, calibrates the final threshold. Tested across ten leave-one-family-out folds, the system achieved zero empirical misses on every fold and zero false flags on nine of ten.
The significance is not academic. DNA synthesis is commercially accessible, and screening is the primary technical barrier between a synthesis order and a dangerous sequence. A screening system that silently fails on novel or understudied families is not a safety net; it is a formality. The certified miss-rate bound CRC-Screen provides is an honest accounting: it includes a term for the distribution gap between calibration data and real-world test families, rather than claiming performance only on sequences the system has already seen.
The hard limit is data, not cleverness. The researchers calculate that reaching procurement-grade accuracy, a miss rate of 1 in 1,000, requires roughly 18 times the calibration data they used. The full reviewed UniProt KW-0800 corpus is apparently large enough to supply it, which means the bottleneck is assembly and standardization, not biology.