AI Revolutionizes Biology's Design Build Test Cycle

[THEO] Okay, picture this. You walk into a lab, and nobody's hunched over a bench pipetting at midnight. Instead there's a little Raspberry Pi camera clicking away in the dark, a software agent deciding which experiment to run next, and a generative model that's already dreamed up the protein you were going to spend two years evolving. That's the arc we're tracing today — AI showing up at the lab bench, and what it's actually doing once it gets there.

[DR. MARA] And I want to be careful about the word "AI" from the start, because it's doing a lot of work in that sentence. The papers we're walking through don't all mean the same thing by it. One is literally a cheap robot with a camera. Another is a four-billion-parameter language model trained on genomes. Those are radically different animals.

[THEO] Right, so maybe that's the first thing to lay out — what's the actual problem all of this is attacking?

[DR. MARA] The core problem is throughput and the cost of a guess. Biology moves in cycles. You design something — a strain, an enzyme, a growth medium — you build it, you test it, you learn, and you go around again. In synthetic biology they call it the Design-Build-Test-Learn cycle, DBTL. The trouble is each lap is slow and expensive, and the search space is astronomically large. A protein a few hundred residues long has more possible sequences than there are atoms in the universe. You cannot brute-force that.

[THEO] So historically, how did people deal with a space that big?

[DR. MARA] The workhorse has been directed evolution. You take a protein, make a library of random variants, screen for the ones that do what you want a little better, and iterate. It won a Nobel Prize, it works. But it's local search — you're feeling your way uphill from where you already are. You tend to find the nearest peak, not a better one across the valley.

[THEO] That's the mountaineering-in-fog version, right? You can only step where you can feel the ground rising, so you summit the little hill you're standing on and never see the Everest two kilometers over.

[DR. MARA] That's a fair picture. And the question underneath this entire arc is: can computation let us teleport across the valley instead of crawling uphill? Hold onto that, because it's where the newest papers get ambitious — and where they start arguing with each other.

[THEO] So before we get to the dreaming-up-proteins stuff, I actually love that the roots here are humble. 2023, SPIRO.

[DR. MARA] SPIRO is the un-glamorous end of this story and I'm glad it's in the arc. Ohlsson and colleagues built an open-hardware plate imager — 3D-printed parts, a Raspberry Pi, a camera that photographs Petri plates on a schedule. The point was phenotyping at scale: watching seeds germinate, roots grow, without a human standing there. And crucially, it works in the dark.

[THEO] Which matters because the moment you flip the lights on to take a photo, you've changed the experiment.

[DR. MARA] For a root, yes. Light cues development. So dark-condition assays were genuinely impractical before — you'd be perturbing the thing you're measuring every time you looked at it. SPIRO just quietly removes the human from the loop. No engineering degree required. That's the "AI at the bench" story in its most literal, mechanical form: automate the looking.

[THEO] And then BacterAI, same year, automates the deciding.

[DR. MARA] This one I find genuinely clever. The problem: you want to know what nutrients a bacterium needs — its auxotrophies, the things it can't make itself. The naive approach is combinatorial. Dozens of ingredients, test every combination, you're there for the heat death of the universe.

[THEO] So what did they do differently?

[DR. MARA] They reframed it as a reinforcement learning problem — a Markov decision process. The agent gets rewarded for removing ingredients from the medium. So it's constantly pushing toward the edge: the minimal recipe that still allows growth. It drives the experiments right onto the growth front, where the information is, instead of wasting runs deep in territory it already understands.

[THEO] Oh, that's nice. It's not sampling randomly, it's pressing against the boundary where it's uncertain. And they mapped metabolism in days instead of—

[DR. MARA] Days, with no prior knowledge fed in. That's the headline. And notice the through-line forming: SPIRO automates observation, BacterAI automates the choice of next experiment. Both are closing the loop a human used to sit in. The models here aren't huge — the intelligence is in how you pose the question.

[THEO] And that's the hinge, right? Because the 2025 papers change what "the model" even is.

[DR. MARA] They do. This is where genomic foundation models enter. GenomeOcean, from Zhou and colleagues — four billion parameters, trained not on curated genomes but on metagenome co-assembly. Raw environmental DNA, everything mixed together.

[THEO] Which is the right diet if you want to find weird new stuff, because most microbes have never been cultured.

[DR. MARA] Exactly. And they used byte-pair encoding — BPE tokenization — to chunk the sequence efficiently, which is part of how they hit roughly 150 times faster inference than Evo-7B. The payoff: zero-shot discovery of biosynthetic gene clusters, BGCs, after a fine-tuning step they call bgcFM.

[THEO] BGCs being the stretches of genome that encode the machinery for making natural products — antibiotics, the molecules we actually want.

[DR. MARA] Right. Normally you'd hunt for those with a tool like antiSMASH, which pattern-matches against known cluster types. "Zero-shot" means the model proposes candidates it was never explicitly taught to recognize. Now — I'd flag that the paper sits in some tension with the broader genomic-foundation-model program. There's real debate about whether these models have learned biology or learned statistics of sequence. The brief marks that tension and I won't paper over it.

[THEO] And then Germinal is the one that made me sit up, because it's actually designing a binder.

[DR. MARA] Germinal co-optimizes two things at once: AlphaFold-Multimer's structural confidence that your designed antibody loop will actually fold against the target, and an antibody language model — IgLM — scoring whether the sequence looks like a real antibody nature might make. Those pull in different directions, and balancing them is the trick. They designed nanomolar CDRs — the binding loops — against four different protein targets, with success rates in the four to twenty-two percent range, in roughly forty-three to a hundred designs.

[THEO] Which, if you've ever done antibody discovery the old way, those numbers are wild. You're not screening millions, you're testing dozens.

[DR. MARA] That's the teleport-across-the-valley promise made concrete. And it feeds the two most speculative pieces in the arc — the enzyme-discovery vision and the LDBT paradigm.

[THEO] LDBT — they literally reordered the acronym.

[DR. MARA] Deliberately. Learn-Design-Build-Test instead of Design-Build-Test-Learn. The claim is that if a model can make good zero-shot predictions before you design anything, and you pair it with cell-free synthesis for massively parallel testing, you could collapse those iterative loops toward a single round.

[THEO] And the enzyme vision is the same bet aimed at chemistry — one generative model holding sequence, structure, and function together, so you could in principle encode any reaction you want.

[DR. MARA] Which is explicitly positioned against machine-learning-guided directed evolution — the camp that says keep the iterative search, just steer it smarter. That's the live argument. One side wants to replace the loop; the other wants to make the loop intelligent. Both are 2025. Neither has won.

[THEO] So from a Raspberry Pi watching roots in the dark to a model claiming it can skip evolution entirely — in two years.

[DR. MARA] And I'd keep my eye on the humble end. The grand generative claims still have to survive contact with a plate and a pipette. Automating the looking and the deciding is what makes testing those claims fast enough to matter.

[THEO] Which is a good place to leave it. We'll be back after the break with the mailbag — bring your questions about whether your pipetting job is safe.