Three questions inside one score
Put a threat score between zero and one beside a genome and the first question is what the number measures. Radical Numerics’ new Omnii Defense model produces separate predictions about pathogenicity, possible AI origin and taxonomic similarity. The company combines predictions about unnaturalness and function in its final score. The tool is designed to operate with a laptop and USB sequencer without an internet connection. Access is currently limited to vetted external testers.[1]
The distinction that matters to me is between recognizing an unfamiliar sequence and recognizing a dangerous event. Sequence resemblance can offer clues about biological function. Naming a number a threat score does not turn that clue into a threat. Engineered origin, pathogenic potential and harm under actual conditions can become blurred when represented by one label. I would want to see which question the model answers before interpreting the combined score. Its separate outputs offer a starting point for that examination, rather than a conclusion.[1]
What withholding a virus can tell us
In the experiment described by the company, Bundibugyo virus is removed from training data. The exclusion is applied first to the species and then to its entire genus. That differs from looking up an identical record in a library. It asks whether the model can find a signal associated with pathogenicity in sequences from a withheld group. In Wired’s account, the company distinguishes Omnii from tools that match sequences against libraries of known organisms on this basis. The reported experiment targets recognition beyond an exact match.[1]
The design asks a useful question: does a signal survive when close relatives are removed from training? A positive signal still need not show that the model distinguishes engineered origin. One alternative is that it recognizes broader similarities among natural viruses. That could be useful, but it is a different capability. Withholding a species or genus cannot substitute for an experiment comparing natural and engineered origins. I read the result as a test of generalization to a withheld biological group, rather than counting the same evidence again as a test of provenance.[1]
The measurement that reaches the field
Bronwyn MacInnis, who works on pathogen surveillance at the Broad Institute, supports using technology for vigilance in her comments to Wired. She then marks the limit: detecting a signal remains far from establishing a pathogen’s real consequences or whether a threat was engineered. Radical Numerics says national laboratories, government agencies and independent evaluators are assessing performance. This is a stage of external evaluation, rather than a completed demonstration of field validity.[1]
I would build the next useful evaluation for a field researcher around that distinction. Define pathogenicity, origin and biological consequence as separate targets, then test a frozen model on externally selected examples for each. Unusual but natural sequences belong in the comparison. I would want to see false alarms alongside detected cases. Their burden affects how much material requires follow-up and how researchers spend their time; waiting for a perfect method also has a cost. Offline access does not remove that burden, although making a measurement where connectivity is absent is a separate benefit. My criterion is not how impressive the combined score looks, but whether it clearly helps a specified decision with a measured error rate.[1]