AI Revolutionizes Biological Experimentation

[SOFIA] So there's this phrase floating around biology right now — "AI at the lab bench" — and I want to actually take it apart today, because it means about five different things depending on who you ask.

[DANIEL] Hm. And most of the time nobody tells you which one they mean.

[SOFIA] Right! Sometimes it's a robot literally taking pictures of your Petri plates. Sometimes it's a giant language model dreaming up a protein that's never existed. Those are wildly different animals, but they're part of one story — how machines went from watching the experiment to proposing the experiment.

[DANIEL] That's the arc worth tracing. Because the temptation is to jump straight to the flashy generative stuff and skip why any of it was hard in the first place.

[SOFIA] Okay, so let's ground people. The classic loop in engineering biology is Design-Build-Test-Learn. DBTL. You design a construct, you build it — clone it, transform it into your organism — you test whether it does the thing, and you learn enough to design a better one next round.

[DANIEL] And the painful word in there is "round." Each cycle can take weeks. In directed evolution — which is the older, brute-force version of this — you make a huge library of variants, you screen or select for the ones that work a little better, then you mutate those and go again. It's local search. You climb the hill you happen to be standing on.

[SOFIA] Which is powerful! People won Nobel Prizes for it. But you only ever explore the neighborhood of sequences you started with. If the really good enzyme lives across a valley, directed evolution basically never walks there.

[DANIEL] So two bottlenecks. One is throughput — how fast can you build and test. The other is search — how much of sequence space can you even reach. Hold those two, because every paper in this arc is attacking one or the other.

[SOFIA] And a few terms for anyone visiting from another field. A foundation model is a big neural network trained on a mountain of unlabeled data — for us, DNA or protein sequences — so it soaks up the statistical "grammar" of biology. A protein language model treats amino acids like text and learns which sequences look "natural." Generative just means it can spit out new sequences, not only score existing ones.

[DANIEL] And "zero-shot," since it comes up a lot — it means the model makes a useful prediction on a task it was never explicitly trained for. No labeled examples for that specific job.

[SOFIA] Great. So let's start at the bench, literally, because the roots here aren't glamorous AI at all. 2023, Ohlsson and colleagues — SPIRO, the Smart Plate Imaging Robot.

[DANIEL] This one I have a soft spot for. It's a 3D-printed rig, a Raspberry Pi, a camera, and it just rotates and photographs Petri plates on a schedule. That's it.

[SOFIA] And I love that it's aimed at biologists with zero engineering background. You don't need a machine shop. It automates germination and root-growth phenotyping at scale — and here's the clever part — it can do assays in the dark, which you basically couldn't do by hand because opening the incubator and turning on the light ruins the experiment.

[DANIEL] Which is the honest version of "AI at the bench" circa a couple years ago. The intelligence is downstream — you're generating clean, consistent image data so that some model can actually learn from it. Garbage in, garbage out. SPIRO is about the "in."

[SOFIA] It's the Test half of DBTL getting automated and standardized. No model designs anything yet. But you can't have machine learning at scale without machine-readable data at scale, and that's the quiet contribution.

[DANIEL] So that's the floor. Now the ceiling shoots up in 2025, and it does it on several fronts at once.

[SOFIA] Okay, this is the good stuff. Let's take the search problem first, because there's this vision piece — AI-driven enzyme discovery — arguing for one unified generative model that encodes sequence, structure, and function together. The pitch: if you can genetically encode function, you could in principle access chemistry that directed evolution just can't reach by hill-climbing.

[DANIEL] And it's explicitly in tension with the machine-learning-guided directed evolution camp. That camp says: don't throw out the iterative loop, just use ML to pick smarter mutations each round. The unified-model camp says: stop iterating, jump straight to the good region.

[SOFIA] And that same year you get a concrete version of that fight — the LDBT paradigm. They literally reorder the letters. Instead of Design-Build-Test-Learn, it's Learn-Design-Build-Test. Learning comes first.

[DANIEL] Which is a real conceptual move, not just branding. You put a zero-shot ML prediction before you design anything, so the model narrows the space up front. Then you pair it with cell-free synthesis — expressing proteins without living cells, in a tube — so the Build-Test step can go massively parallel. The dream is a single-round campaign. One pass instead of ten.

[SOFIA] Cell-free is doing a lot of work there, right? No transformation, no growth, no waiting for colonies. You mix the machinery, you get protein, you test.

[DANIEL] It is a dream, though. I want to see the hit rates before I retire directed evolution. But as a hypothesis about where the field goes, it's coherent.

[SOFIA] And then two papers actually deliver receipts. GenomeOcean — Zhou and colleagues — a four-billion-parameter genome foundation model trained on metagenome co-assemblies. So, DNA pulled straight from environmental communities, not tidy lab genomes.

[DANIEL] The number that matters to me: 150 times faster inference than Evo-7B, a comparable model. Speed is not a footnote when you're generating sequences. And they fine-tuned it into "bgcFM" to discover biosynthetic gene clusters zero-shot.

[SOFIA] BGCs — biosynthetic gene clusters — are the stretches of genome that encode the assembly lines for natural products. Antibiotics, a lot of drugs come from these. Finding novel ones usually leans on tools like antiSMASH that pattern-match against what we already know.

[DANIEL] And "zero-shot BGC discovery" means the model flags candidate clusters without being handed labeled examples of every class. If that holds up, you're finding chemistry you didn't already have a template for. Though it's noted this sits in tension with other genomic foundation model results — the field hasn't agreed on how much these models actually generalize.

[SOFIA] Then the one that really lands for me on the tools side — the TadA ortholog work. TadA is the deaminase inside adenine base editors, the CRISPR tools that chemically convert one base to another without cutting both strands.

[DANIEL] They took 95 natural TadA orthologs, DNA-shuffled them, and trained a generative model on that pool. One round. And they pulled out compact cytosine and efficient adenosine editors — with lower off-target activity — that beat directed evolution starting from the exact same pool.

[SOFIA] Same starting material, head to head, and the generative approach won in a single round. That's the LDBT thesis actually demonstrated on something you'd put in a cell.

[DANIEL] It's the cleanest test in the arc, because they controlled the input. It's in tension with some generative-design work too, so it's not settled — but "one round beats iterative from the same pool" is a claim I can evaluate, and it survived.

[SOFIA] So look at the through-line. SPIRO automated the watching. Then the models moved to proposing — first as a vision, then as a reordered workflow, and now as real editors and real gene clusters.

[DANIEL] The open question is the one that runs through all of it: how far can these models jump before they need the bench to check them. Search got cheaper. Testing didn't disappear.

[SOFIA] Learn first — then go build. That's where the field's pointing. We'll leave it there for today.