DEV Community

Cover image for Omnii-1 offline AI pathogen screening vs DIAMOND
Shamuddin
Shamuddin

Posted on Originally published at aitecharchive.com

Omnii-1 offline AI pathogen screening vs DIAMOND

Verdict: on November 11, 2026, Radical Numerics announced Omnii Defense, and its Omnii-1 offline AI pathogen screening beats the alignment tools it was tested against, with a reported mean recall gain of 52 percent on mutated viral metagenomic reads. For sequences that have drifted from anything in a reference database, Omnii-1 is the stronger tool in the company's own tests, and it suits teams that need results on site. It does not replace alignment tools yet: the numbers are company-run, and access is limited to vetted partners.

TL;DR

  • Omnii-1 is now offered as a restricted biodefense API and an offline build that runs on a laptop with a USB gene sequencer (Radical Numerics).
  • The company reports 98.5 percent recall on protein paraphrases against 83.3 percent for DIAMOND (primary source).
  • With a whole virus genus removed from its data, it still recovered 44.6 percent of dengue read pairs, where the standard workflow recovered under 0.1 percent (primary source).
  • All evaluations are company-run, not independently replicated; nothing here is medical advice.
  • Last verified: November 12, 2026

What did Radical Numerics actually announce?

Radical Numerics of San Francisco placed the Omnii-1 genome model behind a restricted-access biodefense API and packaged a version for offline edge deployment, per its announcement page and Wired. It succeeds a June 15, 2026 research preview. The model is trained on DNA, RNA, proteins and epigenetic signals; the defense product assigns each genome a threat score between zero and one, combining how unnatural a sequence looks with functional predictions, and estimates whether it appears pathogenic or AI-generated and which taxonomic group it resembles.

How does Omnii-1 compare with DIAMOND and the standard metagenomic workflow?

Omnii-1 recalls more divergent sequences than either baseline, at similar precision in the protein tests and a low false positive rate in the read tests. Figures below are from the company's evaluation write-up; the Paraphrase benchmark comes from a peer-reviewed Science paper by Wittmann et al., but Omnii-1's results on it are not.

Test Baseline Omnii-1
Viable pathogen paraphrases (16,307) DIAMOND vs UniRef100: 83.3% recall 98.5% recall at 0.5 threshold
Expanded set, incl. unviable and non-pathogen proteins DIAMOND: 78.6% recall, 71.1% F1 95.8% recall, 76.6% F1
Mutated viral reads, arbitrary mutations SecureBio workflow (Nucleaze + Bowtie 2) About 31% mean recall gain
Mutated viral reads, synonymous mutations SecureBio workflow About 52% mean recall gain

Precision in the expanded protein test was 63.8 percent for Omnii-1 against 64.9 percent for DIAMOND, per the same source. Read tests used about 150 base pair reads, 1,000 reads per variant from 24 mutated viral species at 60 to 100 percent sequence identity, spiked into wastewater-like backgrounds of up to 1 million reads. The false positive rate stayed at or below 0.5 percent in every condition, averaging about 0.23 percent at 1 million background reads (primary source).

The 52 percent headline therefore applies to synonymous mutations; for arbitrary mutations the gain is smaller.

Can it flag a virus it has never seen?

In the company's tests it can — the most notable result. The team retrained Omnii-1 checkpoints with Bundibugyo virus and dengue removed from both training data and reference databases, then tested reads from the missing species. With the entire genus removed, the standard workflow recovered fewer than 0.1 percent of read pairs, while Omnii-1's human-virus predictions recovered 44.6 percent for dengue and 22.0 percent for Bundibugyo (primary source). Radical's biodefense head Christine Parthemore likened it to Omnii "operating in 1976", before the genus was characterised (Wired).

The result matters because alignment tools match known patterns: an engineered or newly emerged sequence can keep its biological function while diverging from known sequences — what the company calls an asymmetric recall gap. The US Defense Threat Reduction Agency describes current approaches as slow, dependent on prior knowledge of engineered sequences, and struggling to keep pace with modern engineering techniques (quoted on the primary page).

How does Omnii-1 offline AI pathogen screening work on a laptop?

The tested setup pairs an Oxford Nanopore MinION sequencer with a MacBook Pro (Apple M3 Max, 128 GB RAM). Data streams through basecalling, bioinformatics and Omnii-1 at once; a local interface shows flagged reads, functional classifications and taxonomic summaries, with a locally running chat model to interpret them (Radical Numerics).

The design targets places where connectivity is unreliable or sequence data must stay on site. Hardware note: 128 GB of RAM is a high-end laptop, not a typical field machine.

Why does the detection gap matter in a real outbreak?

Assays only find what they were built to find. The company cites WHO's Disease Outbreak News for the 2026 Ebola outbreak in the Democratic Republic of the Congo and Uganda, where initial clinical testing returned negative before further investigation identified Bundibugyo virus, outside the assay's scope. This is a story about detection coverage, not clinical guidance.

The scale of the unknown is large: of 31,236 species the NCBI taxonomy assigns across RNA-virus families, 239 are reported to infect humans, and that set has grown by roughly three species a year since the 1950s, according to the company's analysis. SecureBio targets three days from sample collection to actionable early-warning results (SecureBio), which on-site screening could help meet.

How does it fit with the new US DNA synthesis screening policy?

The two are complementary layers. On November 5, 2026 the White House released the National Policy for Nucleic Acid Synthesis Screening: recipients of federal life sciences funding must buy synthetic nucleic acids and benchtop synthesis equipment only from certified suppliers that verify customer identity and screen every order of 50 or more nucleotides. Requirements roll out over 340 days, replacing a voluntary framework from 2010, as Science reported.

Supplier screening acts at the point of ordering; Omnii-1 acts at the point of sample collection. Science also relayed expert criticism, including from Kevin Esvelt and Gema Gronvall: orders under 50 base pairs fall outside the rule, existing benchtop machines are not covered, and identity verification is underspecified. A detection model does not close those gaps, but adds a second line of sight.

Who can use it, and what are the limits?

Access is restricted: the API is limited to vetted partners, the edge build to vetted testers. Radical Numerics says it is working with national laboratories, government agencies and independent evaluators; there is no public pricing or availability date (Wired).

Three limits stand out: results are company-run without independent peer review, the held-out tests used simulated reads rather than field samples, A threat score is a screening signal that needs expert follow-up. Earlier company work, including AI-designed bacteriophages used under special FDA permission in an October 2026 patient treatment (The New York Times), shows capability, not validation of this product.

For related coverage of AI in the lab, see our pieces on AI wet-lab benchmarking, science benchmark comparisons, neural-network physics simulation and AI mapping of bacterial structures.

FAQ

Q: What is Omnii-1?
A: Omnii-1 is a genome language model from Radical Numerics, trained on DNA, RNA, proteins and epigenetic signals, offered for biodefense through a restricted API and an offline build (Radical Numerics).

Q: Does Omnii-1 beat DIAMOND at spotting AI-redesigned proteins?
A: In the company's tests, yes: 98.5 percent recall versus 83.3 percent for DIAMOND on 16,307 paraphrases (primary source).

Q: Can Omnii-1 run without an internet connection?
A: Yes — tested on a MacBook Pro (Apple M3 Max, 128 GB RAM) with an Oxford Nanopore MinION sequencer (Radical Numerics).

Q: Can the public use Omnii Defense?
A: No — the API is limited to vetted partners, the edge build to vetted testers, with no public pricing or availability date (Wired).

Q: Is this medical or diagnostic guidance?
A: No. This article covers detection and biosecurity technology only; it is not medical advice.

Top comments (0)