Viral Engineering For NonModel Organisms

[SOFIA] So here's a thing I love about biology — some of our best tools started out trying to kill us. Viruses, transposons, the little selfish bits of DNA that hijack cells for their own agenda. And over the last thirty years, the move has been: okay, if you're that good at breaking into cells and rewriting genomes, let's put you to work.

[DANIEL] Right. The whole subject is domestication. Take an agent that evolved to invade, and turn it into a delivery system you can point at a target on purpose.

[SOFIA] Which matters right now for two reasons that are honestly kind of urgent. One is antibiotic resistance — we're running out of drugs, and phages, viruses that infect bacteria, are back on the table as therapy. And two is that most of the microbes we actually care about, in soil, in your gut, in the ocean, we can't engineer them. They're non-model organisms. No easy way to get DNA in.

[DANIEL] So let's define the pieces for someone coming from, say, human genetics. A phage is a virus that infects bacteria. It finds its host by having a receptor-binding protein — an RBP — on its tail that latches onto a specific molecule on the bacterial surface. That lock-and-key is what makes phages exquisitely specific. One phage, often one strain of bacteria.

[SOFIA] And a transposon is a mobile piece of DNA — a "jumping gene" — that can cut itself out and paste itself into a genome. Conjugation is bacterial mating, one cell passing a plasmid to another through direct contact. All of these are natural gene-transfer tricks. The arc of this whole field is us learning to drive them.

[DANIEL] So where do we start the story?

[SOFIA] 1997. Pseudomonas fluorescens, a soil bacterium. Someone wants to knock out genes across the whole chromosome to find out which ones do what. And the tool is a mini-Tn5 transposon delivered on a suicide vector — pUT — by electroporation. You zap the cells, the DNA goes in, the transposon jumps into the genome, and then the delivery vector just… disappears.

[DANIEL] The self-curing part is the elegant bit. You don't want the delivery plasmid hanging around and confusing your genetics. And the numbers held up — 7.7 times ten to the fifth clean single-insertion mutants per picomole of DNA. Single insertion matters. If you get two transposons in one cell you can't tell which knockout caused the phenotype.

[SOFIA] So that's the founding move: a mobile element as a mutagenesis engine. But it's brute force, right? You're scattering insertions randomly and screening.

[DANIEL] Random, and confined to one organism you can already electroporate. The next turning point is about reach — can you move engineered DNA through a whole community, not one strain in a cuvette.

[SOFIA] That's the 2022 pMATING work. They took RP4 — a famously promiscuous conjugation system, meaning it'll mate with an enormous range of hosts — and they minimized it. Stripped it down to a synthetic core that still does the mating machinery. And then they let it loose in soil microbiomes to propagate an engineered gene cell to cell.

[DANIEL] And they found something I didn't expect. The barrier to spreading genes through the community wasn't host range. The minimized machinery could transfer broadly, even across kingdoms in principle. The thing actually stopping delivery was contact-killing — a type six secretion system, T6SS, in a Klebsiella strain, literally stabbing the donor cells before they could mate.

[SOFIA] Which is such a great result, because everybody assumed the limit was "can this DNA replicate in a weird host." No — the limit was bacterial combat. You've got to survive the neighborhood before you can edit it.

[DANIEL] It reframes the engineering problem. Delivery isn't just molecular compatibility, it's ecology. Now — up to here the arc is bacterial gene transfer. The next three papers pivot hard into phages, and specifically into that lock-and-key targeting.

[SOFIA] And this is where it gets, okay, this is the good stuff. Because if you want phages as precision tools — therapy, or delivery — you have to understand the receptor binding at the atomic level. 2023, the T5 phage. They solved a cryo-EM structure of the receptor-binding protein, RBP-pb5, docked onto its receptor, FhuA, on E. coli.

[DANIEL] First atomic-resolution picture of a phage protein bound to its receptor. And the mechanism is lovely. The business end of the RBP is intrinsically disordered — floppy, no fixed shape — until it touches FhuA. Then it folds. Binding templates the structure.

[SOFIA] Disorder-to-order. The protein is basically waiting for the right handshake before it commits to a shape.

[DANIEL] And they propose that folding imposes a 45-degree kink that gets relayed down the tail to trigger DNA ejection. That's the trigger — how the phage knows it's arrived and starts injecting its genome. Now, "propose" is doing work there. The structure is solid; the kink cascade is a model built on it.

[SOFIA] Fair. But now you know the physical basis of specificity, the obvious engineer's question is: can I predict, or redesign, which host a phage hits? And that's the 2024 Klebsiella paper — PhageHostLearn.

[DANIEL] Machine learning. They take the receptor-binding proteins and the bacterial surface proteins — the K-locus, which builds the capsule the phage grabs — run them through ESM-2, a protein language model that embeds sequences into vectors, and train an XGBoost classifier to predict which phage hits which strain.

[SOFIA] Strain level. Not "this phage likes Klebsiella," but this phage versus that specific clinical isolate. And the performance — 81.8% ROC AUC in cross-validation, and it held up on real carbapenem-resistant clinical isolates, 79% there.

[DANIEL] The number I actually care about for therapy: in almost 94% of cases, at least one phage that works is sitting in the top five predictions. That's a triage tool. A patient comes in with a resistant infection, you shortlist candidate phages in silico instead of screening hundreds at the bench.

[SOFIA] It's the payoff of the 2023 structure, in a way — specificity is legible enough now to predict from sequence.

[DANIEL] Then two 2024-25 papers go from predicting to manipulating. PhageMaP first — genome-scale knockout libraries in phages T7 and Bas63, using a Cas9-RecA scheme to barcode and delete genes across the whole phage genome.

[SOFIA] Which is the 1997 transposon dream, but for phages and way more systematic. Map which genes are essential, and when. And they found modular defense inhibitors — phage proteins that block bacterial immune systems — that you can transfer between phage genomes. Plug-and-play counters.

[DANIEL] Portable parts. That's the engineering vocabulary finally applying to viruses. And the last one, T3 and T7 retargeting — they engineered the phages to bind a nanobody instead of their natural receptor.

[SOFIA] A totally synthetic target. And two things fell out: evolutionary escape depended on how much receptor the host displayed, and the capsid itself — not just the tail fiber — contributes to host range. Which nobody was really crediting.

[DANIEL] So the arc lands here: from randomly scattering transposons in one bacterium, to a structural and predictive and now editable understanding of how phages choose their hosts. The tool is becoming programmable.

[SOFIA] And where it's heading — designed phages you can point at a resistant infection, or use to deliver genes into microbes we could never touch before. The virus as a chassis. That's the story. Daniel, take us out.

[DANIEL] We'll be back after this with a single new paper to pull apart — controls and all. Stay with us.