AI Automates Biology's Slowest Loops

[THEO] Okay, picture the classic image of a scientist. Someone hunched over a bench, pipette in hand, doing the same transfer four hundred times. For decades that's just been the job. The question we're chasing today is: what happens when you put an AI in that chair?

[DR. MARA] And to be clear up front — "AI at the bench" isn't one thing. It's a whole range, from a cheap robot that photographs Petri plates, all the way up to generative models that dream up proteins that have never existed. The interesting part is how those got connected.

[THEO] Right, so let's set the table for anyone whose PhD was in, say, geophysics. The classic loop in biological engineering is called the Design-Build-Test-Learn cycle. DBTL. You design a construct, you build it — clone it into a plasmid, transform it into your organism — you test whether it does the thing, and you learn from the result to design the next one.

[DR. MARA] And the painful truth is that cycle is slow. Each turn can take weeks. The build step and the test step are where you lose your life. So the whole subject we're tracing is really: can machines take over pieces of that loop — and eventually, can they collapse the loop entirely?

[THEO] A couple of terms we'll lean on. "Directed evolution" — that's the Nobel-winning approach where you mutate a protein, screen for the ones that work a bit better, and repeat. It's beautiful, but it's local search. You're feeling around the neighborhood of a sequence you already have.

[DR. MARA] Then there are foundation models. Same idea as the large language models everyone knows, except trained on biological sequence — DNA, or protein. You feed them enormous amounts of sequence and they learn the statistical grammar of what real biology looks like. A "protein language model" learns which amino acid tends to follow which, the way a text model learns words.

[THEO] And "zero-shot" — that means the model makes a useful prediction on something it was never explicitly trained to do. No fine-tuning, no examples. Just, here's a sequence, tell me if it'll fold or bind.

[DR. MARA] Hold those two ideas — the slow DBTL loop, and the model that can shortcut it — because the arc we're about to walk is the story of those two colliding.

[THEO] Let's start humble. 2023, Ohlsson and colleagues, a thing called SPIRO. It's a Raspberry Pi and a 3D-printed frame that photographs Petri plates on a schedule. That's it.

[DR. MARA] And I want to defend how important the humble version is. SPIRO's whole point is that a plant biologist with zero engineering background can build one and automate root-growth and germination phenotyping at scale. It even runs assays in the dark, which was impractical before — you'd have to open the incubator and ruin your dark condition every time you wanted a picture.

[THEO] It's automation as a camera that never sleeps. No AI designing anything yet. But it's the first move: get the machine to generate the data so the human isn't the bottleneck on observation.

[DR. MARA] The same year, BacterAI takes the next conceptual step. Now the machine isn't just watching — it's deciding what experiment to run next. The team reframed nutrient mapping as a reinforcement learning problem. The agent gets rewarded for removing ingredients from the growth medium.

[THEO] Oh, that's clever. So instead of testing every combination of nutrients — which is combinatorially insane — the agent is incentivized to strip things out and see what the bug can still live without.

[DR. MARA] Which drives its experiments right onto the "growth front" — the edge between survival and death, where the information is. It mapped auxotrophies, meaning which nutrients an organism can't make for itself, in days. That's the turning point from automated observation to automated experimental design.

[THEO] So SPIRO is the eye, BacterAI is the hand plus a bit of the brain deciding where to poke. And both of those lean on an older idea — Bayesian optimization, choosing your next experiment to be the most informative one. That's the thread BacterAI explicitly builds on.

[DR. MARA] Now jump to 2025, and the character of the field changes. The models stop just picking among real experiments and start generating candidates from scratch. GenomeOcean, from Zhou and colleagues, is a genome foundation model — four billion parameters, trained on metagenome co-assemblies.

[THEO] Metagenomes meaning you sequence everything in an environmental sample at once — soil, seawater — and you don't bother isolating individual organisms first.

[DR. MARA] Right, so it's trained on the messy real diversity of nature rather than a handful of lab strains. Two things stood out. One, it uses byte-pair encoding tokenization and generates protein-coding sequence about a hundred and fifty times faster than the Evo-7B model. And two, after a fine-tune they call bgcFM, it discovers biosynthetic gene clusters zero-shot.

[THEO] And BGCs — those are the stretches of genome that encode the machinery to make natural products. Antibiotics, a lot of our drugs come from BGCs that bacteria and fungi use to wage chemical war on each other.

[DR. MARA] Normally you'd hunt them with a tool like antiSMASH that pattern-matches against known clusters. GenomeOcean proposing novel ones is a different mode of discovery. Though I'd flag — the brief says it discovers them; what that means functionally still needs the wet-lab confirmation. The model proposing a cluster and the cluster actually making a molecule are two different claims.

[THEO] Fair. And then there's Germinal, which I love because it's so concrete. It designs antibodies.

[DR. MARA] It co-optimizes two things at once. AlphaFold-Multimer gives it structural confidence — does this designed antibody actually dock onto the target epitope — and an antibody language model, IgLM, scores whether the sequence looks like a real antibody nature might make. Push both together and it designed nanomolar binders against four different protein targets.

[THEO] Nanomolar meaning they bind tightly — that's a genuinely useful affinity, not a toy result. And they got there in something like 43 to 101 designs per target. Compare that to screening millions in a traditional campaign.

[DR. MARA] With a four-to-twenty-two percent success rate, which sounds modest until you remember de novo binder design against an arbitrary epitope was basically science fiction a few years ago.

[THEO] So that brings us to the two big-picture 2025 pieces — the enzyme discovery vision, and the LDBT paradigm. And here's where things get a little spicy, right?

[DR. MARA] This is the real argument of the whole arc. The enzyme-discovery vision says: a unified generative model encoding sequence, structure, and function together could reach chemistry that directed evolution never will — because directed evolution is stuck doing local search. And LDBT reorders the whole cycle. Not Design-Build-Test-Learn. Learn first — the model's zero-shot prediction comes before you design anything — then build and test cheaply with cell-free synthesis.

[THEO] Cell-free meaning you run the gene-expression machinery in a tube, no living cell to coddle. You can test at massive scale.

[DR. MARA] And the dream is single-round. One pass through the loop instead of twenty. But notice the tension the brief flags — this sits uneasily with machine-learning-guided directed evolution, which says, no, you still need the iterative feedback, the model guides the loop rather than replacing it. That's an unsettled fight.

[THEO] So the arc runs from "put a camera on the incubator" to "let the model skip the loop." SPIRO watched, BacterAI chose, GenomeOcean and Germinal generated, and LDBT wants to throw out the loop order entirely.

[DR. MARA] Where it lands depends on a boring, decisive question: how often do these designs actually work at the bench. The models have gotten bold. The validation is still catching up. That gap is the whole story going forward.

[THEO] Keep an eye on that gap with us. We'll be back after the break.