Radical Numerics says a patient with an antibiotic-resistant E. coli infection received a cocktail of bacteria-killing phages, including phages with a segment of their genome rewritten by AI

Radical Numerics says a patient with an antibiotic-resistant E. coli infection received a cocktail of bacteria-killing phages, including phages with a segment of their genome rewritten by AI

On October 20, Radical Numerics introduced Omnii-1, a neural network trained on DNA, RNA, proteins and other data, and published a report on its work with phages. According to the company, the model proposed variants of the gene encoding the protein that phage T7 uses to attach to E. coli. The company describes the administration of these phages to a patient as the world's first clinical use of AI-designed phages.

Each phage infects only certain strains, which are variants within a bacterial species. Finding a suitable phage involves searching collections of naturally occurring phages. “You have to test thousands of candidates, mostly learning which ones do not work,” says Marisa Azad, an infectious disease physician at The Ottawa Hospital who collaborates with the company, in the press release. According to The New York Times, the patient's E. coli had resisted antibiotics and all available phages.

The company chose phage T7, which infects only E. coli. The New York Times compares it to a lunar lander: a roughly spherical protein shell containing DNA rests on six legs, which the phage uses to attach to a bacterium before injecting its DNA. The legs consist of the protein gp17, which binds to receptors on the bacterial surface and largely determines which strains the phage recognizes. Researchers had modified the parts of these legs responsible for recognition before using AI: in 2019, Timothy Lu's group at the Massachusetts Institute of Technology built libraries containing millions of variants of these regions in T3, a phage related to T7, and obtained phages with an altered range of bacterial hosts.

The developers of the Evo neural networks launched Radical Numerics in June. In August, Science published a study in which these neural networks generated complete genome variants of phage ΦX174 (5,4 thousand DNA letters). The T7 genome is more than seven times as long, and Omnii-1 rewrites one segment of it, the gp17 gene. According to the report, a short segment can be synthesized more quickly, and its variants can be tested in large batches.

The model fills in DNA at a selected location using the sequences on either side, much as one might fill in a missing word. According to the report, it produces functional variants much more often than random selection does: with 11–15 amino acid substitutions in gp17, phages could be detected after synthesis of the corresponding DNA for 69,9% of the model's variants and 4,8% of the random variants.

For this patient, the researchers say they trained the model on data from the strain causing the infection and asked it to design T7 phages with legs suited to those bacteria. According to the press release, the design process took weeks. The researchers say they then synthesized DNA for the variants, generated phages from it and tested their safety. A cocktail containing these phages killed the bacteria in a test tube and was administered to the patient through compassionate use, which allows a seriously ill patient to receive an unapproved treatment outside a clinical trial when no suitable treatment options are available. The company says the FDA authorized this use for a single patient.

“To protect the patient's privacy, we are not disclosing details of the clinical course or outcomes,” the report states. CEO Eric Nguyen hopes that, if the treatment succeeds, the time needed to create a phage for a patient can be reduced from weeks to days.

Harvard University microbiologist Michael Baym, who was not involved in the work, tells the newspaper, “If it works, this is incredible,” but adds, “My instinct is to be skeptical.” Uppsala University microbiologist Brandon Berryhill says laboratory measurements cannot establish how well these phages will treat an infection.