Live Nov 18 · Cohort 3Nov 18, 19:00 MSK — free hands-on AI agents live workshop · Live Cohort 3 starts Nov 23Join Nov 18 Live →Cohort 3 (Nov 23)
ReleaseInsight · On RadarWired→ original

Radical Numerics says its AI detected 22–33% of viral sequences withheld from training

Searching a genome database can identify a familiar pathogen but may miss an unknown one. Radical Numerics has introduced Omnii Defense: a model that analyzes genetic sequences on an offline laptop and assesses their potential danger. In its own test, the company reported detecting 22–33% of sequences from a virus excluded from training. Independent validation is still pending.

AI-processed from Wired; edited by Hamidun News
Radical Numerics says its AI detected 22–33% of viral sequences withheld from training
Source: Wired. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceWired ↗
Editor-in-Chief
Key metric #1The model looks for signs of danger without requiring an exact database match.
Key metric #2The proposed field setup pairs a laptop with a USB sequencer.
Key metric #3Access is currently limited to vetted external testers.

In a field laboratory, reading the genetic material in a sample is only part of the job. Researchers also need to determine whether any of the resulting sequences could be dangerous. Searching a reference database helps when the organism is already known. An unfamiliar sample is harder to assess: a lack of matches says nothing, on its own, about safety.

Radical Numerics proposes adding AI to this process. According to WIRED, the startup has introduced Omnii Defense—a model that runs on an offline laptop and analyzes sequences obtained using a USB sequencer. Its purpose is straightforward: help researchers identify which samples warrant further investigation, even when there is no exact entry in the database.

The main caveat is apparent from the outset. In the experiment described by the company, the model flagged only 22–33% of sequences from a virus deliberately excluded from training as potentially pathogenic. That is an interesting finding for further research, but a reliable threat detection tool remains a long way off.

What the model adds to database searches

Imagine a researcher has obtained a genetic sequence of unknown origin. A conventional search looks for similarities to organisms that have already been described. This approach is useful, but its capabilities depend on the contents of the reference database and how closely the sample resembles known entries.

According to its developers, Omnii Defense also uses patterns the model has learned from biological data. It makes several predictions: whether a sequence could belong to a pathogen, how likely it is to have been artificially created, and which taxonomic group it most closely resembles.

The system then assigns the genome an overall threat score, combining an assessment of how unusual it is with predictions of biological functions. It is important to distinguish the metric's name from what it actually means. A high score is a reason to examine a sample more closely. It does not, by itself, prove that the researcher is dealing with a dangerous organism or a product of deliberate engineering.

For a product developer, this is a familiar architecture: the model helps select cases for the next stage of review. It adds value if it picks up useful signals that existing tools miss without overwhelming specialists with false alarms.

What 22–33% actually means

Radical Numerics described an experiment involving Bundibugyo virus. First, information about its species was removed from the training data. The model then flagged 33% of the sequences presented to it as potentially pathogenic. When the entire corresponding genus was excluded from training, the figure was 22%.

These figures should not be described as the system's overall accuracy. They describe detection under a specific experimental setup. The reported results do not make clear how many harmless sequences the model incorrectly flagged as suspicious, how the detection threshold was chosen, or how well the findings generalize to other organisms.

The company also says that tools without AI produced almost no danger signal in this comparison. However, the available description is insufficient to conclude that the model outperforms all existing approaches. That would require details of the test, clear comparison conditions, and an independent assessment.

Bronwyn MacInnis, who leads pathogen surveillance at the Broad Institute, supports the idea of using technology for biodefense. However, she emphasizes the gap between detecting a signal and establishing that a pathogen poses a real danger or has an artificial origin.

Why this matters to teams building AI products

The most practical aspect of the announcement is local operation. If the claimed capabilities are confirmed, researchers could obtain preliminary assessments where connectivity is unreliable or unavailable. But a laptop running the model remains one part of the laboratory workflow: it does not replace high-quality data collection or subsequent verification of the findings.

For founders, the choice of problem is instructive: applying a large biological model to a specific operational constraint—on-site analysis without a persistent connection. When evaluating such a product, it is worth examining missed threats, false alarms, hardware requirements, and what users do after an alert.

For now, Omnii Defense is available only to vetted external testers. Radical Numerics says it is working with national laboratories, government bodies, and independent evaluators. The next meaningful milestone will be the results of those assessments. They will show whether the model helps researchers make better-informed decisions under real-world conditions.

Story

Key Questions Answered

What the model adds to database searches?

Imagine a researcher has obtained a genetic sequence of unknown origin. A conventional search looks for similarities to organisms that have already been described. This approach is useful, but its capabilities depend on the contents of the reference database and how closely the sample resembles known entries. According to its developers, Omnii Defense also uses patterns the model has learned from biological data.

What 22–33% actually means?

Radical Numerics described an experiment involving Bundibugyo virus. First, information about its species was removed from the training data. The model then flagged 33% of the sequences presented to it as potentially pathogenic. When the entire corresponding genus was excluded from training, the figure was 22%. These figures should not be described as the system's overall accuracy. They describe detection under a specific experimental setup.

Why this matters to teams building AI products?

The most practical aspect of the announcement is local operation. If the claimed capabilities are confirmed, researchers could obtain preliminary assessments where connectivity is unreliable or unavailable. But a laptop running the model remains one part of the laboratory workflow: it does not replace high-quality data collection or subsequent verification of the findings.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…