AI built a working phage genome. It still cannot tell you which phage to use

Stanford researchers used AI to design the DNA of a complete bacteriophage, and when that synthetic DNA was built in the laboratory it produced a functioning virus, according to a new study in Science ^[4]. That is a genuine demonstration that a model can learn enough about a biological system to specify something that works in a flask, and it is not the same as the result an antimicrobial resistance program actually has to buy: knowing in advance which phage will infect which bacterium ^[3][12].

The distinction matters because the source material is explicit about where the demonstrated capability stops. The Stanford work started with PhiX174, a phage that infects E. coli and one of the smallest, simplest and best-studied phages known ^[5]. Phages with therapeutic potential are five to 50 times larger, carry double-stranded DNA genomes, and can contain hundreds of genes, along with machinery for recognizing particular bacteria, replicating inside them and defeating the defense systems bacteria have evolved ^[6]. A medicine-grade phage has to find and infect the right bacteria, work in the conditions inside a patient, and ideally stay useful as bacteria evolve resistance ^[7]. Building a viable minimal genome does not address any of those three.

What AI is doing usefully today in this field is narrower and already deployed. Tools such as AlphaFold predict three-dimensional protein structure from amino acid sequence, which is most valuable for the large fraction of phage proteins whose function is unknown: a gene that read as an anonymous stretch of DNA now yields a shape, and a shape yields a hypothesis ^[8]. A researcher at the Becky Mayer Centre for Phage Research at the University of Leicester, writing in phys.org, puts the limit plainly: AI does not replace experiments, it supplies new ideas to test ^[9].

That group's argument is about training data, and it is the part operators should read twice. The centre isolates, sequences and studies phages against E. coli, Klebsiella, Pseudomonas and other bacteria that are hard to treat ^[10], and reports that detailed characterization keeps turning up unexpected biology: different phages recognize different parts of the bacterial cell, use different proteins to overcome defenses, and behave differently when combined with antibiotics or placed in conditions closer to those inside the human body ^[11]. There are thought to be more phages on Earth than any other biological entity ^[2], so the library is enormous, but the useful signal is a link between sequence, protein structure and measured behavior, and the proposal is to have AI connect those three, including identifying the genetic features that determine host range ^[12]. That is stated as a question, not a finding ^[13].

So the screening bottleneck is still wet. Anyone budgeting a phage program should assume the cost of isolating, sequencing and phenotyping candidates against clinical isolates, because the predictive layer that would let you shortlist in silico is the thing being asked for rather than delivered ^[12][13].

Watch for the datasets: whether anyone publishes host-range measurements at a scale and consistency that supports supervised prediction, whether the Stanford design approach transfers from a single-stranded minimal genome to large double-stranded phages ^[5][6], and whether any predicted match survives testing alongside antibiotics and in body-like conditions ^[11].