Radical Numerics post-trains Omnii for personalized mRNA cancer-vaccine design

Radical Numerics has post-trained its Omnii genome model to design personalized mRNA cancer vaccines, ranking neoantigens and writing mRNA cassettes from tumor mutations and MHC alleles, but has not yet presented human, animal, or manufacturing results. The company, founded by Evo co-creators Eric Nguyen, Michael Poli, Stefano Massaroli, and Armin W. Thomas, announced the pipeline on September 8th via X, offering early access to researchers and vaccine developers. The work remains computational, with validation pending.

Radical Numerics post-trains Omnii for personalized mRNA cancer-vaccine design The Evo founders say one genome model can rank neoantigens and write an mRNA cassette; Omnii's evidence remains computational. By RuntimeWire Staff /author/runtimewire-staff · Published Primary source: Radical Numerics https://x.com/RadicalNumerics/status/2097355738456998316 Why it matters Radical Numerics is testing whether one biological foundation model can replace a chain of specialized vaccine-design tools. Success will depend on experimental and clinical validation, not benchmark gains. Eric Nguyen https://erictnguyen.com/?ref=runtimewire , Michael Poli https://zymrael.github.io/?ref=runtimewire , Stefano Massaroli and Armin W. Thomas said in a thread on X https://x.com/RadicalNumerics/status/2097355738456998316?ref=runtimewire on September 8th that Radical Numerics https://www.radicalnumerics.ai/?ref=runtimewire has post-trained its Omnii genome model to turn tumor mutations into a personalized mRNA cancer-vaccine design. The intended pipeline takes a patient's tumor variants and MHC allele sequences, predicts which mutated peptides the tumor will display and which ones could provoke a T-cell response, ranks the candidates, and writes an optimized mRNA cassette encoding the selected targets. Omnii is available through an early-access program https://www.radicalnumerics.ai/waitlist?track=health&utm campaign=omnii-cancer-vaccine&ref=runtimewire for researchers and vaccine developers. Radical Numerics has not presented human, animal or manufacturing results from a vaccine designed by the model. That distinction sets the boundary around Tuesday's announcement. Nguyen and his co-founders have produced a research pipeline and a set of computational benchmarks. They have not produced a clinically validated treatment. The work is still a consequential test of the thesis behind Radical Numerics: a model trained broadly on biological sequences can learn enough transferable biology to handle jobs that previously required separate, specialized predictors. Nguyen came to that thesis through an unusual route. Before completing a Stanford PhD in bioengineering and AI, he studied civil engineering at UC Berkeley and Stanford, then computer science at Cornell. His research moved from structures made of concrete and steel to the architectures needed for million-token genomic models. Poli, a Stanford computer science researcher and former founding scientist at Liquid AI, worked on Hyena and HyenaDNA, model architectures built to process sequences far longer than conventional attention systems could comfortably handle. The four co-founders later worked on Evo and Evo 2, generative genomics models capable of reading and generating DNA. When Radical Numerics raised a $50 million seed round in June /article/radical-numerics-50m-seed-general-biological-intelligence , the founders pitched Omnii as the next step: one multimodal model spanning DNA, RNA, proteins, epigenomics and protein structure. Cancer-vaccine design gives that broad research program a sharply defined medical problem. Picking the few mutations that matter A personalized cancer vaccine starts with mutations found in a patient's tumor. Those mutations can produce neoantigens, peptide fragments absent from healthy tissue that may give the immune system a target. Most candidates will fail somewhere between sequencing and immune response. A useful target needs to clear two difficult predictions. The tumor must process and present the peptide through one of the patient's MHC molecules, making it visible on the cell surface. A T cell must then recognize the peptide-MHC complex and mount a response. Presentation without immunogenicity gives the immune system a target it ignores. In a technical post https://www.radicalnumerics.ai/blog/omnii-cancer-vaccines?ref=runtimewire authored by Alexander Fields, Poli and Nguyen, Radical Numerics said it post-trained Omnii for presentation and immunogenicity prediction across MHC class I and class II. Radical Numerics reported a 0.939 AUROC on its class I presentation evaluation, compared with 0.933 for BigMHC-EL. On class I immunogenicity, Omnii scored 0.750, compared with 0.558 for BigMHC-IM and 0.514 for PRIME2.1. The class II results followed the same pattern in Radical Numerics' evaluation. Omnii scored 0.945 for presentation against 0.935 for NetMHCIIpan-4.3-EL, and 0.778 for immunogenicity against 0.753 for ImmuScope-IM and 0.613 for TLimmuno2. Those comparisons are company-run evaluations on public data. Radical Numerics acknowledges that immunogenicity datasets are small, heterogeneous and unevenly distributed across MHC alleles. The class I immunogenicity test included 1,290 positive and 4,701 negative examples, while the class II test included 511 positive and 4,567 negative examples. Results were averaged across three data splits. Radical Numerics also simulated a case in which a tumor produced 100 plausible candidates, 6% of which were immunogenic, and a vaccine could include five. Omnii selected an expected 1.3 immunogenic peptides, compared with 0.7 for BigMHC-IM and 0.6 for PRIME2.1. That exercise shows how improved ranking might affect a constrained vaccine design. It does not show whether a vaccine would generate an immune response in a patient. A general model looks for its wedge Radical Numerics describes the cancer-vaccine capability as a "surprising emergent capability." Omnii was pretrained largely through next-token prediction on biological sequences, rather than built specifically as a neoantigen-selection model. Focused post-training was then used to adapt it to the vaccine workflow. That sequence matters to the founders' business. Specialized tools already predict MHC binding, presentation or immunogenicity, and some AI-assisted vaccine programs are much further into clinical testing. Evaxion https://evaxion.ai/scientific-posters/aacr-2026-evx-01/?ref=runtimewire , for example, has evaluated a personalized peptide vaccine containing AI-selected neoantigens in a Phase 2 melanoma study. Radical Numerics is betting that a shared representation across several biological modalities will eventually outperform pipelines assembled from disconnected tools. Omnii's current benchmarks support part of that claim, particularly on immunogenicity prediction. Clinical and experimental evidence will determine whether the general model carries useful information that public benchmarks fail to measure. The timing gives Radical Numerics a credible opening. On August 19th, Merck and Moderna reported https://www.merck.com/news/merck-and-moderna-announce-phase-3-interpath-001-trial-of-intismeran-autogene-plus-keytruda-met-endpoints-of-recurrence-free-survival-rfs-and-distant-metastasis-free-survival-dmfs-in-patient/?ref=runtimewire that their 1,137-patient Phase 3 study of intismeran autogene plus Keytruda met its recurrence-free-survival and distant-metastasis-free-survival endpoints in patients with resected melanoma. The companies described it as the first positive Phase 3 readout for an individualized neoantigen therapy. Intismeran encodes up to 34 patient-specific neoantigens in an mRNA construct. That result establishes clinical momentum for the treatment category. It does not validate Omnii, and melanoma is only one disease setting. It does sharpen the commercial question Radical Numerics is pursuing: if individualized mRNA cancer treatments work, who supplies the algorithm that chooses what goes into each patient's cassette? The $50 million data hunt Radical Numerics launched publicly on June 15th with a $50 million seed round https://www.radicalnumerics.ai/blog/radical-numerics-seed?ref=runtimewire led by Emergence Capital https://www.emcap.com/thoughts/biology-just-entered-its-read-write-era?ref=runtimewire , alongside Obvious Ventures https://obvious.com/portfolio/?pillar=human-health&ref=runtimewire , Triatomic Capital, First Spark Ventures https://www.firstsparkventures.com/?ref=runtimewire and Factory. The valuation was not disclosed. At the time, Radical Numerics said it was building a data center with Nvidia Blackwell systems and pursuing human-health and biodefense applications under the same research program. The cancer-vaccine preview converts some of that infrastructure spending into a product-shaped proposition: researchers provide tumor and immune-system inputs, and Omnii returns ranked targets and an mRNA sequence design. The practical constraint is data. Radical Numerics says real-world performance will likely depend on stronger post-training data than public benchmarks can provide. Its request for collaborators with immunology datasets, experimental systems and vaccine-validation capacity is therefore central to the strategy. Those partnerships could give Omnii the feedback loop required to move from predicting labels in a dataset to selecting targets that survive laboratory and clinical testing. Nguyen and his co-founders are taking a broad foundation-model wager into a field that punishes approximation. A coding model can survive an imperfect suggestion. A vaccine-design system must choose a tiny number of targets from hundreds of plausible mutations, then prove those choices in biology. Omnii's early results earn the founders a serious experiment. The next evidence has to come from cells, animals and eventually patients.