Multimodal genome language model trained across 28 biological modalities that takes assay tracks as input to reconstruct missing measurements.
No providers recorded yet. Browse all providers
A cell's regulatory state is measured one assay at a time, and no cell type has them all: a tissue may have RNA-seq and an ATAC-seq peak set but no methylation, while another has methylation and nothing else. Sequence-to-function models such as AlphaGenome and Borzoi handle this by treating assay tracks purely as training targets, learning to predict a fixed panel of biosamples from DNA alone. At prediction time they see only sequence, so a new modality or an unseen cellular context is out of reach. Omnii-1 inverts that relationship: tracks are inputs as well as outputs, so the measurements a lab already has for a sample can condition predictions of the assays it never ran.
Omnii-1 is the multimodal genome language model released by Radical Numerics on October 20, 2026, and the successor to the company's June 2026 Omnii-preview, which learned from DNA sequence and conservation tracks only. Where a genome language model can score how unusual a variant is, it cannot say what that variant does in a particular tissue. Omnii-1 was pretrained natively across 28 biological modalities — DNA, RNA, proteins, chromatin accessibility, transcription factor binding, histone modifications, DNA methylation, splicing and gene expression — conditioned on cell type and tissue, which puts both questions inside one model.
The lab was founded by members of the team behind Evo and Evo 2, and Omnii-1 continues that lineage of long-context genomic sequence modeling while adding genome-aligned experimental readouts as a first-class input.
Radical Numerics reports training on trillions of tokens of genetic and multiomic data with a multi-scale architecture reaching a context window of millions of base pairs; the parameter count and the full pretraining recipe are undisclosed. Every modality is aligned to the genome's coordinate system, which is what permits joint modeling of sequence and signal. On the AlphaGenome zero-shot eQTL causality panel — 21,948 variant–gene–tissue triples — Omnii-1 reaches 0.810 tissue-weighted AUROC against 0.705 for an ensemble of four Borzoi replicates and AlphaGenome's reported 0.713, and its margin widens with distance to the transcription start site. It reaches 0.814 on the AlphaGenome splicing-QTL panel against 0.721 for Pangolin, and 0.898 and 0.802 AUROC on the African- and European-ancestry caQTL panels against 0.794 and 0.710 for Borzoi and 0.792 and 0.680 for ChromBPNet. All of these figures are self-reported in company blog posts; no preprint or peer-reviewed paper accompanies the release, and the AlphaGenome numbers are quoted from its Nature publication because Radical Numerics states it cannot run that model commercially.
The eight launch applications span genome interpretation and design. Omnii Genetics ranked a functional variant in its top four picks for 90% of nearly 1,800 GWAS credible sets with experimental ground truth and nominated candidate mechanisms across 153 Parkinson's disease signals. Omnii Phage designed phage variants whose cocktail cleared a patient's antibiotic-resistant isolate in vitro and was then administered under a physician-led compassionate-use pathway — the first clinical administration of AI-designed phages. Omnii Diagnostics augments a cell-free 5hmC cancer test, Omnii Defense runs behind a restricted biosurveillance API and on edge hardware, and the Rare Earths, Antibody and Cancer Vaccine adaptations cover metal-binding selectivity, single-domain antibody design and MHC class I/II immunogenicity scoring.
Omnii-1's substantive contribution is architectural rather than numerical: making experimental assays a conditioning input turns a genome model into something that can answer questions about cell types and modalities it was never trained to predict, which is the limitation that has kept sequence-to-function models tied to their training panels. The caveats are equally concrete. No weights, code, training data or paper have been released; access is by request as a research preview, with the biodefense interface restricted to vetted users. Every benchmark figure is the company's own, and its headline clinical result is a single patient whose outcome is undisclosed. Independent replication will determine how much of the claimed generalization holds.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.