Phage Engineering For Precision Therapeutics

[SOFIA] So here's a thing I love: some of the best tools in molecular biology aren't things we built from scratch. They're things we domesticated. Restriction enzymes, CRISPR — those came from bacteria fighting off viruses. And today I want to talk about the viruses themselves. Bacteriophages. Turning them into tools.

[DANIEL] Which is funny, because for decades phages were the tool. T7 RNA polymerase, phage display, the T4 ligase in your freezer right now. We just didn't think of the phage as something you'd engineer on purpose.

[SOFIA] Right, we used their parts. Now people want to program the whole particle. And the reason this matters beyond the bench is antibiotic resistance. Carbapenem-resistant Klebsiella, Shiga-toxin E. coli — we're running out of drugs, and phages kill bacteria very precisely. One phage, maybe one strain.

[DANIEL] Which is the double-edged part. That specificity is why they're safe — they won't wipe your gut microbiome the way a broad-spectrum antibiotic does — but it's also why phage therapy is hard. You need the right phage for the right bug, and there are a lot of bugs.

[SOFIA] So let's define the key piece for anyone outside this world. A phage recognizes its host through a receptor-binding protein — an RBP — usually sitting on the tail. That protein finds a specific molecule on the bacterial surface, the receptor. Could be a sugar, could be an outer membrane protein. Lock and key. Get the key wrong, the phage bounces off.

[DANIEL] And even if it binds, the host might kill it internally. Restriction-modification systems chew up incoming DNA that isn't methylated the right way. Abortive infection systems make the cell commit suicide before the phage finishes. So there are layers — surface recognition, then a gauntlet of immune systems inside.

[SOFIA] Which is exactly the mess the first paper in our arc walks into. 2021, the Harms group, the BASEL collection. And the reason I'd call this the starting gun — before you engineer anything, you need a clean, characterized parts bin.

[DANIEL] This is the paper I keep coming back to for how it was done. They isolated over 120 E. coli phages, but on a restriction-free host — K-12 with the RM systems deleted. That's a deliberate choice. You don't want your collection biased toward only the phages that could beat that one strain's defenses.

[SOFIA] Then they curated down to 68, plus ten classical reference phages, and here's the part I find genuinely useful — they mapped the receptor for every single one. Against more than fifty single-gene knockout mutants.

[DANIEL] Every one. That's not a screen you can hand-wave. And the immunity side was quantified, not just observed — efficiency of plating across eleven systems, six restriction-modification, five abortive infection, at three or more replicates. When someone gives me effect sizes and replicates instead of "we saw resistance," I trust it.

[SOFIA] And they found new biology falling out of the characterization. LptD as a terminal receptor for seven small siphoviruses — nobody had that. Plus swappable RBP loci matched to seven different receptors.

[DANIEL] That "swappable" word is the hinge for the whole arc. If the receptor-binding module is a discrete, movable part, you can imagine changing a phage's target the way you'd swap a promoter.

[SOFIA] Okay, so BASEL gives you the catalog. Next turning point, 2023, Degroux — they solve the first atomic-resolution structure of a phage RBP actually bound to its receptor. T5's RBP, called pb5, sitting in FhuA.

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

[SOFIA] And folding isn't the end of it — the paper proposes that binding imposes a forty-five-degree kink, and that kink is what triggers the phage to eject its DNA into the cell. So recognition and injection are mechanically coupled.

[DANIEL] Now, "proposes" is the honest word there. The disordered-to-ordered transition, that's in the structure. The kink cascade initiating ejection is a model built on it. I'd want functional mutants to nail causation. But as a hypothesis it reframes the RBP from a sticky patch into a machine.

[SOFIA] And for an engineer that's huge, because if binding does mechanical work, you can't just graft any binding domain on and expect injection. That tension shows up later, hold that thought.

[DANIEL] Meanwhile the field splits into two engineering problems. One: predict which phage hits which host so you don't have to test everything. Two: figure out which genes inside the phage matter once it's in.

[SOFIA] The prediction fork is the 2024 Klebsiella paper — PhageHostLearn. They take RBP sequences and the host's K-locus proteins — that's the capsule, the sugar coat Klebsiella hides behind — run them through ESM-2 protein language model embeddings, then XGBoost to classify matches.

[DANIEL] Eighty-two percent ROC AUC in cross-validation, seventy-nine on twenty-eight actual carbapenem-resistant clinical isolates. And the number that matters clinically — a matching phage in the top five candidates almost ninety-four percent of the time.

[SOFIA] Which is the difference between screening a thousand phages and screening five.

[DANIEL] For a patient, that's the difference between useful and too slow. I'll flag the sample — twenty-eight isolates is a real-world test, but it's small. I want to see it hold on hundreds.

[SOFIA] The other fork, same year — PhageMaP. This one I get excited about because it's genetics done at genome scale on the phage itself. They use a Cas9-RecA system to build barcoded knockout libraries in T7 and in Bas63 — which, note, is a BASEL phage. The collection paying off.

[DANIEL] Knock out every gene, see which are essential and under what conditions. And they found defense inhibitors — phage proteins that block host immune systems — that are modular. You can move an inhibitor from one phage genome into another and it still works.

[SOFIA] Transferable counter-defense parts. So now you've got a receptor module you can swap and immunity-evasion modules you can port. The toolbox is filling in.

[DANIEL] Which sets up the two 2025 papers, and this is where that pb5 tension comes back to bite.

[SOFIA] The evolutionary steering one — they retarget T3 and T7 to bind a nanobody as an artificial receptor. And the finding that surprised me: escape depended on how much receptor the host displayed, and the capsid — not just the tail fiber — contributes to host range independently.

[DANIEL] Which means retargeting isn't only an RBP problem. The head matters too. That complicates the clean "swap the key" story.

[SOFIA] And the last one, Meta-SIFT, tackles it head-on. Okay, this is the good stuff — they don't just shuffle natural RBPs. They use deep mutational scanning to weight a search through metagenomic sequence — all that environmental phage DNA — for motifs, build seventeen thousand T7 RBP variants, and twenty-four and a half percent are active.

[DANIEL] A quarter active in a designed library is a strong hit rate. And they got T7 infecting Shiga-toxin E. coli O121 at high salt — an activity you couldn't reach with natural sequences alone.

[SOFIA] Which is the whole arc in one result. BASEL cataloged the parts, the structure showed binding is mechanical, prediction narrowed the search, PhageMaP mapped the guts, and now we're reaching targets nature never evolved to hit.

[DANIEL] With the honest caveat that the capsid contribution means we don't fully control it yet. Plenty left to falsify.

[SOFIA] Which is where we'll pick it up after the break. Domesticating viruses — still very much in progress.