Writing Genomes From Scratch

[SOFIA] Okay, so today we're doing something a little ambitious. Not one paper — a whole arc. The dream of writing genomes from scratch. And I mean actually sitting down with a blank page and composing DNA the way you'd compose software.

[DANIEL] Which, to be clear, nobody has fully done. Synthesizing a whole genome from nothing and booting it up — that's still mostly a handful of heroic projects. Mycoplasma, yeast chromosomes. Enormous teams, enormous budgets.

[SOFIA] Right, and the reason it stays hard is that there's this gap between what we can read and what we can write. We can sequence a bacterial genome for pocket change now. Writing one — designing it, physically building it, and getting it into a cell that'll run it — that's where everything bottlenecks.

[DANIEL] So think of it as three separate problems that all have to be solved at once. One, assembly — physically joining pieces of DNA into bigger pieces, scar-free, reliably. Two, delivery — getting large constructs into the organism you care about, which for most non-model bugs is genuinely painful. And three, function — do you even know what the genes you're writing actually do?

[SOFIA] And that third one is the sleeper problem. You can't write a genome if a big fraction of the parts are labeled "hypothetical protein, function unknown." It's like editing a manuscript in a language you half-speak.

[DANIEL] A few terms for anyone coming from outside the field. Conjugation — that's bacteria physically transferring DNA cell-to-cell through a bridge, as opposed to electroporation or chemical transformation where you force naked DNA across the membrane. And in bacteria it's transformation, never "transfection" — transfection is a eukaryote word.

[SOFIA] Then there's restriction-modification — the bacterial immune system that chops up incoming foreign DNA. It's a huge reason delivery fails. The cell sees your beautiful construct and just shreds it.

[DANIEL] And a couple of enzyme families worth naming up front, because they recur. Serine integrases, like Bxb1 — these drop DNA at a specific site, between a short "attP" and "attB" sequence, cleanly, in one direction. And transposases, which cut-and-paste DNA, historically more or less randomly.

[SOFIA] Great. So that's our toolkit vocabulary. Let's walk the arc. And I'd start it with the delivery problem, because there's this 2022 paper on Deinococcus radiodurans that I love.

[DANIEL] The organism that shrugs off radiation that would kill basically anything else.

[SOFIA] That one. And it's a nightmare to engineer — chemical transformation just fails for anything large. So what they did was use conjugation from E. coli to shove plasmids in, both the kind that can replicate in Deinococcus and the kind that can't. And then they went after those restriction-modification genes directly — swapped four of them out for resistance markers, one deletion at a time.

[DANIEL] Which is the elegant part. You're not just delivering cargo, you're disarming the defense system that makes delivery hard in the first place. Clearing the runway.

[SOFIA] And then — okay, this is the good stuff — they ran conjugation in reverse, conceptually. They put a nonreplicating plasmid in carrying about a kilobase of homology to a target, and they captured the entire 178-kilobase megaplasmid out of Deinococcus. Cloned it back into E. coli as a roughly 190-kb construct. Verified it on a MinION.

[DANIEL] A 178-kb capture at 62% GC. That's not a trivial amount of DNA, and high-GC is harder to handle. I'd want to know the efficiency honestly — the brief doesn't give me a number — but the capture working at all moves the ceiling on what you can shuttle between organisms.

[SOFIA] So that's the delivery pillar. Now jump to 2023 and the assembly-slash-editing side, because there's this method called ORBIT for E. coli.

[DANIEL] This one made me sit up. The old workhorse for editing E. coli is lambda Red recombineering — you feed in DNA with homology arms and it recombines. Great for small edits, terrible for big payloads. Efficiency falls off a cliff as the insert grows.

[SOFIA] ORBIT splits the job. You use lambda Red just to install a tiny landing pad — a Bxb1 attP site carried on an oligo. That's a small, easy edit. Then Bxb1 integrase does the heavy lifting, dropping in a kilobase payload at that site.

[DANIEL] And the number is the headline: a thousand-fold over lambda Red for that payload delivery, and it scales to libraries of thirty thousand members. The division of labor is the insight — use each enzyme for what it's actually good at.

[SOFIA] Which is such an engineering move. Don't make one tool do everything. Lambda Red places the address, Bxb1 delivers the package.

[DANIEL] And it connects straight to the integrase and recombineering lineage. It's not a clash with prior work, it's a clean build-on.

[SOFIA] Then 2024 just explodes. On pure assembly, there's the In- and Out-Cloning work — using SapI, which cuts and leaves three-nucleotide overhangs. Three nucleotides is a codon.

[DANIEL] That's the clever bit. Most Golden Gate assembly uses four-base overhangs that leave a little scar. Line your junctions up on codon boundaries with three-base overhangs and the scar disappears into the reading frame. Scarless transcription units.

[SOFIA] And "Out-Cloning" flips the usual MoClo setup — instead of being locked to prebuilt acceptor plasmids, you generate organism-specific acceptors on demand from modular parts. For any organism. That matters a lot if you're writing for non-model bugs.

[DANIEL] Then there's CAST — CRISPR-associated transposases. This is the transposase story getting an upgrade.

[SOFIA] Right, so classic transposases insert randomly. CASTs are RNA-guided — you point them with a guide RNA, and they insert large payloads at a programmable spot, and crucially without a double-strand break.

[DANIEL] Which is why it scales. No double-strand break means no reliance on the cell's repair machinery, which is where a lot of editing becomes lethal or messy at genome scale. They built genome-scale knockout, overexpression, and fusion libraries in bacteria. That's writing at scale, not one locus at a time.

[SOFIA] And then the function pillar — the "what do these genes even do" problem — gets its own 2024 answer. Coaux-Seq.

[DANIEL] This is the one I keep thinking about. They took roughly 3-kb shotgun fragments from eleven different bacteria, barcoded them, put them in an E. coli expression vector, and dropped them into twenty different auxotrophic knockout backgrounds — strains missing a gene they need to grow.

[SOFIA] So if a fragment rescues growth, that fragment carries a gene doing that job. Gain-of-function, read out by cheap barcode sequencing.

[DANIEL] First experimental validation for fifty-three protein functions. And one genuine surprise — a TauE-family protein working as a sulfate importer, when that family was only known as exporters. That's the kind of result that corrects the annotation you'd otherwise write into your synthetic genome by mistake.

[SOFIA] That's the whole point. You can't write what you've mislabeled. And then the last paper zooms way out — bacteria as delivery vectors into mammalian cells. Attenuated strains, lysis circuits, T4SS conjugation, outer membrane vesicles, nanoparticle coatings.

[DANIEL] Large cargo, mucosal access — places LNPs and viral vectors struggle. It's the delivery pillar pointed at a new destination.

[SOFIA] So the through-line: deliver big DNA, assemble it scarless, write it at scale without breaks, and actually know what the parts do. Four pillars converging on writing genomes you can trust.

[DANIEL] None of them alone writes a genome. Together they're starting to look like a workflow.

[SOFIA] That's the arc. Next up, we'll get into one of these in detail — st