American laboratory Radical Numerics announced that viruses designed by its artificial intelligence model had been administered to a real patient with a bacterial infection that did not respond to antibiotics. It described the case in a technical article on its website on October 20. According to the company, this is the first time a treatment of this type has made it from a computer to a human being. Citing privacy protections, the laboratory did not disclose the patient’s name, diagnosis, or treatment outcome. Moreover, the results have not yet been reviewed by other scientists in an academic journal, so independent verification is still to come.
The reason for seeking something like this is well known. Bacteria defend themselves, mutate, and exchange resistance genes, meaning that drugs that once worked are now failing. In 2021, approximately 4.7 million deaths worldwide were associated with antibiotic resistance. Estimates cited by Radical Numerics indicate that resistant bacteria will claim nearly forty million lives between 2025 and 2050.
One possible answer has existed in nature for billions of years and is called a bacteriophage, or phage for short. This virus attacks bacteria and does not harm human cells. It attaches itself to the surface of a bacterium, injects its genetic information into it, replicates inside, and ruptures the cell. Unlike antibiotics, which also wipe out beneficial bacteria in the gut, a phage is highly selective. That is both its strength and its weakness.
The existing process is laborious. Doctors collect a bacterial sample from a patient and then test collections of phages one by one in a laboratory to see whether any of them work against that particular strain. If none is found, they search in nature, wastewater, or soil. For someone running out of options, weeks can mean the difference between life and death. Radical Numerics therefore trained a model called Omnii-1, which first learned to read genomes from a broad mix of genetic data and then specifically from phage genomes. The idea is simple: instead of searching for a suitable virus, one can be designed directly.
How AI Changes a Genome
The model does not build an entire virus from scratch. It takes an existing phage and rewrites only a selected section of its genetic information that determines which bacterium the virus can attach to. The scientists worked with phage T7, a well-studied virus that attacks the gut bacterium E. coli. Its genome contains approximately forty thousand letters, more than seven times as many as the virus used in an earlier, similar experiment. Omnii-1 rewrote the gene for a protein on the virus’s tail fiber, which the phage uses to attach to the bacterium, while taking the surrounding genomic context into account so that the rest of the organism remained functional.
The results look convincing. When the team introduced eleven to fifteen changes into the designed proteins, nearly seventy percent of the proposed variants came to life and replicated, compared with less than five percent of randomly modified controls. Of the 6,694 variants whose effects on bacteria were measured, 2,161 were more potent than the original natural virus, including variants that differed from the original at twenty-eight positions. The random controls outperformed the natural phage only when they contained no more than five changes.
The model can also rank in advance the variants that are worth testing at all. For two bacterial strains, 97 and 96 of its one hundred highest-rated designs produced successful hits. For the third and most difficult strain, where only 4.5 percent of the entire library worked, 24 percent of its top one hundred candidates were effective, a fivefold improvement.
Its second capability is equally important for personalized treatment. Omnii-1 learned from public data about which phages attack which bacteria and then predicted outcomes for strains omitted from training. When the scientists deliberately withheld an entire related branch of thirty-seven E. coli strains, it was able to identify an effective phage among its first two recommendations for more than ninety percent of them. The team then tested it on Pseudomonas aeruginosa, a bacterium that afflicts patients with burns, non-healing wounds, and cystic fibrosis. Significantly less data is available for it: approximately 7,300 measured pairs, compared with 32,700 for E. coli. It turned out that when the model was also given E. coli data for training, its predictions for Pseudomonas became more accurate without the addition of a single new measurement involving that bacterium. This is good news for less-studied bacteria.
From the Screen into the Vein
For the patient, the process unfolded as follows. A bacterial strain was isolated from the infection, and a cocktail containing phage variants designed by the model was prepared to target it. Researchers first verified that it actually destroyed the bacterium in a test tube. Only then was the treatment administered to the patient under a compassionate-use program, a procedure that allows patients with serious illnesses and no other options to receive a treatment that has not yet been approved, under medical supervision and with authorization from regulators.
The company is now seeking hospitals, phage centers, and physicians willing to send it bacterial samples from their patients. The goal is to produce doses tailored to one specific person and their specific infection, shortening the treatment-design process from months of searching to a few days of work by the model followed by test-tube verification.
Radical Numerics was founded this year by people who previously worked on the Evo model at the Arc Institute, a project that recently produced the first complete virus genome designed by a computer. In June, the company emerged from stealth with fifty million dollars in its first funding round, led by Emergence Capital. Its CEO is Eric Nguyen, a Stanford graduate specializing in bioengineering and artificial intelligence.
Sources: radicalnumerics.ai, futurism.com, and independent.co.uk
