Antibiotic resistance is quietly killing millions of people every year. Standard drugs fail. Superbugs win. But a startup named Radical Numerics just changed the equation by administering the first-ever artificial intelligence-designed virus to a human patient under a compassionate-use framework.
This isn't science fiction anymore. It's happening in clinical settings, and it forces us to rethink how we fight drug-resistant infections.
The Problem With Traditional Antibiotics
Bacteria evolve fast. Really fast. For decades, we threw broad-spectrum antibiotics at every infection we could find, selection pressures did their work, and now we face multi-drug resistant strains that laugh off standard pharmacology.
Traditional drug discovery moves at a snail's pace. Researchers screen chemical libraries, run years of preclinical trials, and hope molecules work inside a human body. By the time a new antibiotic hits the market, pathogens have often already developed resistance to it.
We need a living, evolving countermeasure. We need viruses that hunt bacteria.
Enter Phages and the Omnii-1 Model
Bacteriophages are viruses that specifically target and destroy bacteria without harming human cells. They've been studied for a century, mostly in Eastern Europe, but using them clinically has always been messy. Finding or isolating the exact phage that matches a specific bacterial strain takes weeks. Time that severe infection patients simply don't have.
That is where Radical Numerics steps in. Founded by researchers behind the Evo and Evo 2 genome models from the Arc Institute, the startup built an AI system named Omnii-1.
Instead of writing entire viral genomes from scratch—which yielded a punishingly low success rate in earlier research—Omnii-1 takes a surgical approach. It targets specific components of existing phages, such as the tail fibers responsible for latching onto bacterial surfaces.
The numbers released by the company show real progress. When testing designs with 11 to 15 amino acid changes from the wild type, nearly 70% of the AI-designed variants were recovered as viable phages. Compare that to a meager 4.8% success rate for random control groups. Out of thousands of tested variants, thousands actually outperformed the original virus against specific resistant strains.
Inside the Compassionate Use Milestone
The recent milestone involves an individual suffering from a stubborn, antibiotic-resistant infection. Standard treatments failed. Doctors turned to an experimental phage cocktail containing variants designed by Omnii-1.
In laboratory tests prior to administration, the custom AI-engineered cocktail wiped out the patient's isolate entirely. That successful lab result cleared the path for a physician-led compassionate use authorization.
Transparency here matters. Radical Numerics withheld granular details about the patient's clinical outcome to protect their privacy. This restraint is standard for single-patient compassionate use cases. Critics will point out that a single successful lab test and administration doesn't replace a randomized controlled trial. They are right. One data point does not equal a proven universal therapy.
What Happens Next in Biological AI
The hype around artificial intelligence in biotech usually outpaces reality. Every week brings another press release claiming a breakthrough.
Yet, engineering functional biology using large-scale language and diffusion models represents a genuine shift. We are moving from reading genomes to writing them with intent.
Antibiotic resistance won't vanish because of one AI-designed virus. But as these biological models get sharper, we gain a massive tactical advantage against pathogens that treat traditional medicine as an easy target. The wet lab and the neural network are officially working together.
Take a hard look at your clinical pipelines. Drug discovery is no longer just chemistry. It's code.
