Introducing Omnii-1: a leap toward general biological intelligence

Language models and the agents built on them can draw on what they have learned to tackle new tasks in different environments. We believe the ultimate measure of a general biological model is how broadly the model can be applied, and how much of what it learns can transfer to new problems.

With Omnii-1, we are beginning to demonstrate this kind of transfer in biology. Today, we are releasing eight applications built on the same foundation, spanning genome interpretation, protein engineering, disease detection, therapeutic design, and biological defense. Each begins with the same model, adapted through mid-training or post-training to bring its learned understanding of biological sequence and function to a different problem.

Our goal with Omnii-1 is to build the engine behind general biological intelligence: a model that can learn from the full breadth of biological data, transferring what it learns in one context to another, to help us reason across the complexity of living systems and ultimately understand, predict, and engineer biology.

• Omnii Genome. Advances regulatory prediction through generative multiomic modeling. Omnii-1 sets a new state of the art on regulatory-track and variant-effect benchmarks, ahead of specialized sequence-to-function models such as AlphaGenome and Borzoi. It also enables new capabilities, including single-CpG DNA methylation prediction and reconstruction of missing assays conditioned on measurements from observed modalities in unseen cellular contexts.
• Omnii Phage. Brings directed phage design to antibiotic-resistant infections. Omnii-1-designed phages were administered to a human patient with an antibiotic-resistant bacterial infection under a government-regulated compassionate use program. In host-prediction benchmarks, Omnii-1's top two recommendations included a biologically active phage for 92.1% of eligible, randomly held-out E. coli strains.
• Omnii Genetics. Translates genetic signals into testable mechanisms across rare and complex traits. Omnii-1 nominates candidate causal variants in four of five unsolved congenital myopathy cases, ranks a functional variant within its top four picks in 90% of evaluated GWAS credible sets, ahead of every tested method, and proposes cell-type-specific mechanisms across 153 Parkinson's disease GWAS signals.
• Omnii Diagnostics. Adapts to a new blood-based biomarker for enhanced early-stage cancer detection. Post-training enables Omnii-1 to predict cell-free 5-hydroxymethylcytosine (5hmC) coverage profiles, even though the entire 5hmC modality was absent from pretraining. Its features augment detection and tissue-of-origin prediction at 98% specificity in the evaluated cohort, with the largest gains at very low tumor fractions. It predicts effects of genetic variants on 5hmC at 0.854 Pearson correlation.
• Omnii Defense. Boosts screening capabilities against novel or engineered pathogens. Adding Omnii-1 to traditional pipelines can raise pathogenicity-related protein recall from 78.6% to 95.8%. With entire target genera excluded from training and reference databases, the combined system recovers 44.6% of simulated dengue read pairs and 22.0% of Bundibugyo read pairs, versus less than 0.1% for the conventional cascade.
• Omnii Rare Earths. Transfers biological learning to rare earth-binding protein design. Omnii-1 outperforms every public model tested on overall selectivity prediction across 105 element pairs in 609 proteins, despite no rare-earth binding data in pretraining. With Lawrence Livermore National Laboratory, we are using this capability to develop protein candidates for more efficient extraction and recycling.
• Omnii Antibody. Improves on the antibody affinity and expression design frontier. Two Omnii-1 designs bind more tightly than the assay's limit of detection (<0.2 nM). Its representations also achieve the highest point estimates for expression and potency prediction among the evaluated models, linking design to measurable function.
• Omnii Cancer Vaccine. Outperforms specialized immunogenicity predictors to help select tumor-specific vaccine targets. Omnii-1 scores antigen presentation and immune recognition across MHC classes I and II, reaching 0.750 allele-averaged AUROC for class I immunogenicity versus 0.558 for BigMHC-IM. This connects patient-specific target selection and mRNA vaccine design through the same biological foundation.

• Predict in context. Omnii-1 uses genome-aligned modalities, such as chromatin accessibility or DNA methylation, to inform its predictions and reconstruct missing measurements. The same DNA can yield different predictions depending on the supplied cell or tissue context.

• Align new modalities. Omnii-1's flexible architecture is designed to incorporate new modalities at different stages of training. As new biomarkers and molecular signals emerge, the model can transfer what it has learned across existing modalities to align these new measurements within a shared biological representation. For example, Omnii-1 learned to reconstruct patient-specific cell-free 5hmC profiles after post-training, even though the entire 5hmC modality was absent from its pretraining data.

• Ask biological questions and design for function. Omnii-1 can be prompted to reconstruct missing measurements, score variants, or propose DNA with a desired profile or biological function.

Read more: Omnii Genome.

Less than 2% of the human genome codes for protein. Much of the biology connecting a sequence to a trait lies in how genes are regulated: the same DNA can behave differently in the liver, muscle or an immune cell. Experimental measurements help reveal that context, but data availability across assays and cell types is uneven.

Omnii-1 connects these measurements in one model. Alongside DNA sequence, it can take in measurements from other modalities in a biosample and reconstruct missing ones. This makes it possible to use the information already available for a cell or tissue to predict assays that were never run, while preserving differences between biological contexts.

DNA methylation is one example: predictions at individual CpG sites agree with measurements at a Pearson r of 0.879 across our evaluation set. The “Multimodal reconstruction” animation follows the same ORM2 sequence in liver and skeletal muscle, recovering high expression in one and almost none in the other.

Read more: Omnii Phage.

Antibiotic-resistant infections can leave patients with few treatment options. Phages, viruses that infect bacteria, offer a highly targeted approach, but that specificity makes finding a suitable treatment difficult. A phage active against one bacterial strain may fail against another, and screening collections for the bacterial strain causing a patient's infection takes time.

Omnii-1 brings phage genome design and host-interaction prediction into the same model. It can propose changes within an existing genome while accounting for the surrounding sequence, and help prioritize which phages to test against a new bacterial strain. The aim is to make experimental search more directed while preserving the functions a phage needs to remain viable.

On randomly held-out E. coli strains, its top two recommendations included an experimentally active phage for 92.1% of eligible strains. Omnii-1-designed phages were administered to a human patient with an antibiotic-resistant bacterial infection under a government-regulated compassionate use program.

Read more: Omnii Genetics.

Genome-wide association studies identify regions of the genome linked to disease, but because nearby variants are inherited together, each signal leaves many plausible candidates. Experiments are still needed to determine which variants change molecular function, and choosing which to test first is a central bottleneck in turning genetic associations into biological understanding.

Omnii-1 compares reference and alternate sequences using both DNA sequence scores and predicted molecular tracks in a chosen tissue or cell type. On credible sets with experimental ground truth, Omnii-1 ranked a functional variant within its top four picks in 90% of nearly 1,800 sets, ahead of every other method tested. Applied prospectively to 153 Parkinson's disease GWAS signals, where the functional variant is unknown, Omnii-1 nominated a candidate variant at each signal and proposed a testable mechanism, including a predicted weakening of the NUCKS1 promoter consistent with independent expression and experimental data.

Read more: Omnii Diagnostics.

Early-stage cancers can leave faint traces in blood, mixed into a much larger background of cell-free DNA from normal cells. Detecting those signals also requires high specificity: when screening people without symptoms, false positives can lead to unnecessary imaging, cost and anxiety.

We post-trained Omnii-1 to predict cell-free 5-hydroxymethylcytosine (5hmC), an epigenetic mark collected by ClearNote Health. Although the model saw no 5hmC during pretraining, it learned to reconstruct patient-specific 5hmC profiles from inherited genetic variation and cell-free whole-genome sequencing. This tests whether knowledge learned across biological modalities can transfer to a new modality with a practical diagnostic use.

Across 250 held-out quantitative trait loci (QTLs), evaluated in 32 held-out patients, predicted alternate-allele dosage effects on 5hmC agree with measured effects at a Pearson correlation of 0.854, compared with −0.057 for matched whole-genome sequencing coverage. The reconstructions recover genetic associations with this new modality that are not apparent in sequencing coverage alone. In the evaluated cohort, classifiers incorporating Omnii-1 features showed higher sensitivity for early-stage cancer detection at 98% specificity than classifiers using 5hmC alone or 5hmC plus WGS. The largest differences in detection sensitivity were observed in patients with very low tumor fractions. Adding these features also improved tissue-of-origin discrimination. Prospective screening studies are still needed to establish clinical benefit.

Read more: Omnii Defense.

Speed and fidelity are key to strengthening biological defenses; for example, in identifying and understanding risks from concerning sequences before DNA is synthesized and as fast as possible when samples are taken. Today, searches against references of known sequences are central to many workflows, but offer less help when a sequence has no close match on file. Yet a protein can retain its function even as its sequence changes enough to become difficult to recognize.

Omnii-1 adds a complementary kind of evidence: predictions of biological function when the sequence does not match sequences seen in the past. The Defense API combines those predictions with conventional analysis, spanning whole proteins and short fragments from mixed biological samples. This gives reviewers another basis for prioritizing unfamiliar sequences.

In our early tests, on a labeled set of 47,754 protein sequences including controls, adding Omnii-1 lifted recall of pathogenicity-related sequences from 78.6% with DIAMOND alone to 95.8%, with similar precision: 64.9% without and 63.8% with Omnii-1. In a separate evaluation of simulated viral read pairs, we removed each target's entire genus from both the reference databases and Omnii-1 training: the conventional cascade recovered fewer than 0.1% of read pairs, while the combined API recovered 44.6% for dengue and 22.0% for Bundibugyo. Building on our earlier biodefense preview, these results show how learned biological function can broaden coverage, including beyond the viral genera represented in training, while retaining reference-based evidence and giving researchers more candidate reads to investigate. Additional work extends this type of capability to the edge, for example to give operational insights in austere or contested environments.

Read more: Omnii Rare Earths.

Rare earth elements are essential to technologies such as permanent magnets, yet their similar chemistry makes separation difficult. Conventional refining can require many repeated extraction stages, consuming chemicals and energy. Selective proteins offer a possible route to gentler separation, provided they can distinguish between closely related metals.

In work with Lawrence Livermore National Laboratory, Omnii-1 predicts which metals a protein preferentially binds. Small sequence changes outside a binding site can affect selectivity without an obvious structural change. Learning those relationships could help identify promising candidates for laboratory testing and eventually guide the design of more selective binders.

On measurements from 609 natural proteins, Omnii-1 reached a median Spearman correlation of 0.340 across 105 element pairs, compared with 0.318 for ESM-C 6B, the strongest public comparator in this evaluation. Omnii-1 saw no rare-earth binding data during pretraining. The results demonstrate transfer to selectivity prediction; experimental validation of generated binders remains the next step toward a practical separation process.

Read more: Omnii Antibody.

Antibodies are widely used to treat diseases by binding to specific molecular targets. Designing new antibodies can expand the range of diseases and biological targets that can be treated. In addition to strong binding to the target antigen, successful designs must also express efficiently, remain stable and resist aggregation. Improving one property can compromise another, making antibody development a search for useful combinations of traits across an enormous space of possible sequences.

We adapted Omnii-1 to single-domain antibody design through an iterative experimental workflow where Omnii-1 proposes variants of promising measured sequences and predicts their properties to prioritize the next batch for testing. New laboratory measurements then guide subsequent designs and model improvement, connecting generation, prediction and experimental learning in one cycle.

Across three experimental rounds, Omnii-1's designs improved the measured tradeoff between binding affinity and expression beyond the starting cohort. Two designs from the final round bound more tightly than the assay could resolve, so their affinities are reported as bounds. Predictors built on Omnii-1's representations also rank experimentally measured antibody properties: across 327 designs they reach Spearman correlations of 0.359 for expression and 0.273 for binding potency, the highest point estimates among the evaluated models and baselines.

Read more: Omnii Cancer Vaccine.

A personalized cancer vaccine must teach the immune system to recognize a patient's own tumor. Hundreds of mutations may offer possible targets, but only a small set can go into a vaccine. A useful target must be displayed by the patient's MHC molecules and provoke a T-cell response. Predicting both is central to choosing which candidates to pursue.

We post-trained Omnii-1 to score these two steps, antigen presentation and immunogenicity, across both MHC classes. The model reads candidate peptides alongside the patient's MHC sequences, bringing the same biological foundation to several decisions in vaccine design. These scores help prioritize tumor-specific targets for assembly into an mRNA sequence.

On class I immunogenicity, Omnii-1 reached an allele-averaged AUROC of 0.750, compared with 0.558 for BigMHC-IM.

Help us build the next generation of biological intelligence.

Omnii-1 is a general-purpose foundation model designed to understand and engineer biology across a wide range of applications. We're looking for partners with ambitious biological problems, unique datasets, and experimental capabilities to push the frontier together.

Whether you're a research laboratory, clinical institution, biotechnology company, or government organization, we'd love to explore what we can build together.

Or write to us at health@radicalnumerics.ai and defense@radicalnumerics.ai.

We are hiring → join us