Engineering Phages For Precision Medicine

[SOFIA] Okay, so viruses. We spend most of our lives trying to kill them, right? And today the whole segment is about the opposite — how a slice of the field has spent the last few years turning viruses into precision tools. Specifically bacteriophages, the viruses that infect bacteria.

[DANIEL] Which is a good moment for it, given where antibiotic resistance is heading.

[SOFIA] Right, that's the stakes. Carbapenem-resistant Klebsiella, drug-resistant E. coli — we're running out of small molecules. And phages have this property antibiotics don't: they're exquisitely specific. A given phage often infects one strain and leaves everything else alone.

[DANIEL] Which is the blessing and the curse. Specificity means you don't nuke the patient's microbiome. It also means you have to find the right phage for the right bug, every single time. That matching problem is what this whole arc is really about.

[SOFIA] So let's define the pieces for anyone coming from a different field. A phage lands on a bacterium and it has to recognize a specific molecule on the cell surface — the receptor. Could be a sugar, could be an outer-membrane protein. The part of the phage that does the recognizing is the receptor-binding protein, the RBP. Think of it like a key fitting a lock.

[DANIEL] And even if the key fits, the cell can still fight back. Bacteria carry immunity systems — restriction-modification, which chops up incoming DNA that isn't methylated the right way, and abortive infection, Abi, where the infected cell basically sacrifices itself to stop the phage spreading. So host range is this layered thing: get to the receptor, get past the defenses, then replicate.

[SOFIA] And historically all of that was — honestly kind of a mess to study. Everyone had their own phages, their own weird lab strains, nobody could compare results.

[DANIEL] Which is exactly why the first paper in the arc matters, even though it's not flashy. 2021, Harms's group, the BASEL collection. They isolated over a hundred and twenty E. coli phages on a host they'd engineered to remove its restriction-modification systems — a restriction-free K-12.

[SOFIA] That detail is doing a lot of work, by the way. If your isolation host is chewing up incoming DNA, you only recover phages that already dodge that defense. Strip the RM systems out and you catch a much broader, less biased set.

[DANIEL] Then they narrowed to 68 phages plus ten classical references, and here's the part I respect — they didn't just deposit them and walk away. They mapped the essential receptor for every single one. Against more than fifty single-gene mutants. And they quantified efficiency of plating against eleven immunity systems and seven strains, everything at three or more replicates.

[SOFIA] Which is the part I love, because it turns a phage zoo into an actual parts catalog. And they found things — LptD as a terminal receptor for seven small siphoviruses, which nobody had pinned down before. Plus swappable RBP loci matched to seven different receptors.

[DANIEL] "Swappable" being the word that sets up everything after. If the receptor-binding region sits in a defined, modular locus, you can imagine changing the key without rebuilding the whole phage.

[SOFIA] So that's the foundation — a rigorous, shared, receptor-mapped collection. Turning point one: the field gets a common reference set.

[DANIEL] The next paper zooms all the way in. 2023, Degroux and colleagues — the first atomic-resolution structure of a phage protein bound to its receptor. Phage T5's RBP, called pb5, in complex with the E. coli receptor FhuA, by cryo-EM.

[SOFIA] And the mechanism is gorgeous. The business end of pb5 is intrinsically disordered — floppy, no fixed shape — until it touches FhuA. Then it folds. Binding templates the structure.

[DANIEL] And the folding imposes a forty-five-degree kink, which they propose triggers the phage to eject its DNA into the cell. So recognition and injection aren't separate steps — the act of binding is the trigger.

[SOFIA] Which, if you're an engineer, is thrilling and terrifying. It means the RBP isn't just a passive address label. Its shape change is mechanically coupled to firing the genome. So when you swap receptor specificity, you might be messing with the trigger too.

[DANIEL] It's the mechanistic caution flag under everything the BASEL work made look modular. Supports the receptor-specificity concept, but adds, "careful, these parts are coupled."

[SOFIA] Okay, and here's where it starts scaling up. 2024, PhageHostLearn — predicting phage-host matches for Klebsiella at the strain level. They took receptor-binding proteins and the bacterial K-locus proteins, the capsule genes, ran them through ESM-2 protein language model embeddings, and trained XGBoost on top.

[DANIEL] Numbers, because they matter here. Cross-validated ROC AUC of 81.8 percent in silico, 79.3 on 28 carbapenem-resistant clinical isolates. And a top-5 candidate contained at least one matching phage 93.8 percent of the time.

[SOFIA] That last number is the clinically useful one. You don't need to nail the single best phage — you need to hand a clinician a short list that almost certainly contains a winner.

[DANIEL] I'll flag it's a modest test set — 28 isolates — so I'd want to see it hold on more strains and more phage families. But it's an honest external validation, not just cross-validation on training data. That distinction earns my trust.

[SOFIA] And notice the through-line: BASEL said receptors are the key, the structure said RBPs are the recognition module, and now a model is reading RBP sequence plus capsule to predict matches. Same logic, three scales.

[DANIEL] Same year, PhageMaP goes functional instead of predictive. They built barcoded genome-scale knockout libraries inside the phages themselves — T7 and Bas63 — using a Cas9-RecA approach, and mapped conditional gene essentiality across 44 hosts.

[SOFIA] Knocking out phage genes at scale. And they found modular defense inhibitors — anti-defense genes you can move from one phage genome into another and it still works. That's the "swappable" idea from BASEL, now applied to the counter-defense arsenal.

[DANIEL] Which is the piece that gets you past the immunity layer, not just the receptor layer.

[SOFIA] Then 2025 gives us two engineering payoffs. First, T3-T7 evolutionary steering. They retargeted these phages to bind a nanobody as an artificial receptor, and used that to ask how bacteria escape.

[DANIEL] And two findings fall out. Escape depends on how much receptor the host expresses — more receptor, different escape dynamics. And the capsid contributes to host range independently of the RBP. Which pushes back a little on the tidy RBP-is-everything story.

[SOFIA] I love that it complicates things. The key matters, but the body of the phage matters too.

[DANIEL] It's the field being honest with itself. Host range is multi-factorial.

[SOFIA] And the newest one — Meta-SIFT — this is the good stuff. They combined deep mutational scanning of the T7 RBP with metagenomic motif mining. Basically, use the DMS to learn which positions tolerate change, then go mine natural phage sequences from the environment for motifs that fit.

[DANIEL] Built 17,000 variants, 24.5 percent active. And critically, they got T7 infecting Shiga-toxin E. coli O121 — at high salt — which they couldn't reach with natural sequence diversity alone.

[SOFIA] That's the whole arc landing. A target you cannot hit by browsing what nature already made, reached by designing the RBP. From "here's a mapped collection" to "here's a phage we built to hit a pathogen on demand."

[DANIEL] With the honest caveat that a 24 percent hit rate means three-quarters of the designs fail. There's real work left on prediction.

[SOFIA] Which is where PhageHostLearn and Meta-SIFT are clearly heading toward each other. Predict the match, then design the RBP to make it. Daniel, take us out.

[DANIEL] Four years, from a shared parts catalog to atomic mechanism to models and designer phages. The field turned a virus into an engineerable tool — and it did the controls along the way. That's the part that holds up.