Phage Specificity Rewritten By Design

[THEO] Okay, picture this: You’ve got a microscopic predator, a bacteriophage, and it’s really good at hunting one specific type of bacteria. But what if you need it to hunt a *different* type, or maybe *several* different types, or even avoid one particular type it usually targets? It’s like trying to retrain a highly specialized hunting dog to go after foxes instead of rabbits, or maybe even *all* small furry things except your pet hamster.

[DR. MARA] That’s a fair analogy, Theo. In molecular terms, we're talking about engineering a phage's host range, specifically altering which bacterial strains it can infect. Phages are viruses that infect bacteria, and they do this by recognizing specific receptors on the bacterial cell surface. This recognition is mediated by what we call receptor-binding proteins, or RBPs, on the phage.

[THEO] Right, the RBP is the dog's nose, sniff-sniffing out the right prey. And if you want to change the prey, you need to change the nose. But these proteins are complex, right? You can't just swap out one part and expect it to work perfectly.

[DR. MARA] Precisely. Changing an RBP to alter host specificity isn't straightforward. Small changes can have cascading effects, sometimes improving infection on one strain while inadvertently reducing it on another, or broadening it more than desired. Traditionally, you'd make mutations, test them, and iterate, which is a slow and often inefficient process, especially when you have multiple, sometimes conflicting, objectives.

[THEO] So, slow, expensive, lots of trial and error. That sounds like a job for AI, then. What did this new work from Naia Novy and the team do?

[DR. MARA] They used a combination of deep mutational scanning and multi-objective machine learning to design T7 bacteriophage RBPs. T7 is a well-studied phage that infects *E. coli*. They started with a massive library of 26,838 variants of the T7 RBP, each with slightly different mutations.

[THEO] Twenty-six thousand variants! That's a huge number to test. How did they figure out what all those different "noses" were good at?

[DR. MARA] They tested each of those RBP variants for their ability to infect five different *E. coli* strains. This gave them a comprehensive dataset on how specific mutations affected infection across multiple hosts. Then, they fed all that data into a machine learning model. The "multi-objective" part means the model wasn't just optimizing for one thing, like maximum infection on one strain. It was simultaneously trying to achieve up to 26 different goals—like increasing infection on strain A, decreasing it on strain B, or broadening it to include strains C and D.

[THEO] So, instead of just training the dog to catch rabbits, they're training it to catch foxes *and* leave the hamster alone *and* maybe even ignore the squirrels, all at once? That’s clever. What was the big takeaway from what the AI designed?

[DR. MARA] The model successfully designed RBPs to meet all 26 of their specified objectives, including some that required opposite targeting specificities. And here's the interesting part: they found designs that were only three mutations apart in the RBP could switch from targeting one *E. coli* strain to preferentially infecting a different one.

[THEO] Just three mutations? That’s like tweaking a couple of tiny knobs on a complex machine to completely change its function. That seems incredibly efficient.

[DR. MARA] It is. What this indicates is that subtle changes can rapidly reroute host tropism. The ML model was able to navigate that complex fitness landscape to find these precise, minimal changes for significant functional shifts. It moves beyond trial-and-error by predicting which specific RBP changes will yield the desired infection profile, rather than just optimizing for a single, broad goal.

[THEO] So this means we could potentially engineer phages with much greater precision for things like targeted antimicrobials, or even as tools for delivering genetic material to specific bacteria in a mixed population?

[DR. MARA] That's one of the key implications. By understanding these mutational landscapes and being able to design RBPs with such fine-tuned specificity, we gain a more robust platform for phage therapy or for using phages as highly specific delivery vehicles in complex microbial communities. It's a step towards more predictable and efficient phage engineering.