Radical Numerics Tests Local Pathogen Screening With Omnii-1

Radical Numerics’ Oct. 20 research preview detailed a local pathogen-screening API built on Omnii-1. Evaluations showed gains in sequence detection, while performance on real-world field samples remains unestablished.

Omnii-1 is a genome language model. The API combines its predictions about biological function with conventional matching against databases of known sequences. The company designed that additional layer to flag sequences that have diverged from familiar references. The API is restricted, and local execution is intended for places with unreliable internet or where sequence data must remain on site.

For an edge test, Radical Numerics paired an Oxford Nanopore MinION sequencer with a MacBook Pro running an Apple M3 Max chip and 128 gigabytes of memory. The setup analyzed environmental metagenomic samples, then added sequences of concern to the resulting reads. Because those sequences were spiked into the reads, the test does not establish detection of naturally present pathogens in unaltered field samples.

On a labeled set of 47,754 protein sequences, including controls, recall for pathogenicity-related sequences was 95.8% at a 0.5 probability threshold, compared with 78.6% for DIAMOND alone. Precision was 63.8% with Omnii-1, versus 64.9% without it.

A separate simulation excluded the target virus’ entire genus from both the reference databases and Omnii-1’s training. The combined API recovered 44.6% of dengue read pairs and 22% of Bundibugyo pairs; the conventional workflow recovered less than 0.1%. In simulated metagenomic tests, the false-positive rate was no more than 0.5% across tested conditions, averaging 0.23% with one million background reads and 0.17% with 200,000. The averages came from simulated conditions and do not establish a general false-alarm rate in field use.

The protein test measured recall and precision, while the viral simulation measured read-pair recovery; those figures are separate results, not one overall accuracy score. The evaluations measure sequence screening and recovery, not whether a sequence was designed with AI or whether a flagged organism poses a real-world threat. They also do not establish the system’s performance in routine deployment. Bronwyn MacInnis, director of pathogen surveillance at the Broad Institute of MIT and Harvard, said, “detecting a signal is still a big leap.”