Omnii-1: AI-designed viruses administered to a patient, a claimed first by Radical Numerics

Radical Numerics introduces Omnii-1, a biological model trained on 28 modalities, from which AI-designed bacteriophages were administered to a patient.

Omnii-1: AI-Designed Viruses Reach a Human Patient as Radical Numerics Unveils Its Biology Model

Viruses designed with artificial intelligence have been administered to a patient suffering from an antibiotic-resistant bacterial infection. That is the clinical milestone claimed by Radical Numerics, which is simultaneously introducing Omnii-1, a foundation model designed to understand and engineer biological systems. Unlike systems specializing in proteins, gene expression, or medical genetics, Omnii-1 was pretrained across 28 biological modalities, with the aim of transferring knowledge between different areas of biology. The company presents eight applications spanning cancer research, antibody design, biological surveillance, and rare-earth separation. Its bacteriophage program represents the most significant clinical development, although the patient's treatment outcome remains undisclosed. AI-Designed Bacteriophages Administered to a Patient

Bacteriophages are viruses that infect bacteria. Their specificity creates both their therapeutic potential and a major practical challenge: a phage that attacks one bacterial strain may have no effect on another, even a closely related one. Treating antibiotic-resistant infections can therefore require extensive screening and, in some cases, adapting existing phage candidates to the patient's bacterial infection.

Radical Numerics says it used Omnii-1 to design bacteriophages for a patient with an antibiotic-resistant infection. The candidates underwent experimental evaluation before being administered through an authorized, physician-led compassionate-use pathway. According to the company, this represents the first human administration of bacteriophages designed with an AI model.

The announcement does not establish that the treatment worked. Radical Numerics has not disclosed detailed information about the patient, the progression of the infection, or the clinical outcome. Speaking to The New York Times, the company cited medical confidentiality and said it intends to submit its findings for peer-reviewed publication. The technical report available at launch has not yet undergone that process.

Separately, the company reports that Omnii-1 can prioritize phages likely to act against particular bacterial strains. In tests using held-out E. coli strains, its top two recommendations included an experimentally active phage for 92.1% of eligible strains. This is a laboratory selection result, not a patient recovery rate. Identifying a virus that can infect a bacterial target does not, by itself, demonstrate clinical efficacy or safety. One Foundation Model Across 28 Biological Modalities

Omnii-1 addresses a longstanding limitation in biological AI. Many systems are built for individual tasks, such as protein structure prediction, genetic variant interpretation, gene expression modeling, or molecular design. Yet those problems are connected through biological processes that specialized models do not necessarily represent together.

Radical Numerics trained Omnii-1 across 28 biological modalities, including DNA, RNA, proteins, chromatin accessibility, DNA methylation, histone modifications, splicing, and gene expression. Pretraining involved trillions of tokens of genetic and multiomic data, with information about the cell types and tissues involved. Additional training stages then adapt the shared model to specific applications.

This architecture is designed to capture relationships that cannot be understood from DNA sequence alone. The same genome can produce different patterns of gene activity in liver, muscle, and immune cells, depending on how genes are regulated. By combining genetic sequences with molecular measurements, Omnii-1 can reconstruct certain missing observations while accounting for biological context. Results Against AlphaGenome on Gene Regulation

The first application, Omnii Genome, focuses on predicting the molecular processes that regulate gene activity. Radical Numerics compares its system with Google DeepMind's AlphaGenome and Borzoi across benchmarks covering molecular signal reconstruction and the effects of genetic variation.

On RNA-seq gene expression prediction, Omnii-1 reaches a Pearson correlation of 0.830, compared with 0.800 for AlphaGenome. It also leads on chromatin accessibility, scoring 0.904 versus 0.860, and on splicing prediction, with 0.884 versus 0.863. DNA methylation prediction at individual CpG sites reaches a correlation of 0.879 in the reported evaluation.

However, not every measurement is directly comparable. Radical Numerics notes that some AlphaGenome results involve supervised evaluation, that the splicing benchmark uses different cohorts, and that equivalent scores are unavailable for several categories. The results document promising performance on specific tasks without proving that Omnii-1 is universally superior to AlphaGenome. Rare Diseases and Early Cancer Detection

Two additional applications use Omnii-1 to interpret genetic and molecular evidence associated with disease. Omnii Genetics aims to distinguish potentially causal mutations from other variants linked to a condition. Across nearly 1,800 experimentally grounded sets of candidate variants, the model ranks a functional variant among its top four choices in 90% of evaluated cases. Radical Numerics also reports candidate findings in four of five previously unsolved congenital myopathy cases, alongside proposed mechanisms for 153 genetic signals associated with Parkinson's disease.

Omnii Diagnostics focuses on molecular traces of cancer found in blood. In collaboration with ClearNote Health, Radical Numerics adapted the model to predict 5-hydroxymethylcytosine, or 5hmC, an epigenetic marker absent from the original pretraining data. This provides a test of whether knowledge acquired from other biological modalities can be transferred to a new type of biomarker.

In the evaluated cohort, incorporating Omnii-1 representations improved sensitivity for detecting