AI Automates Biology's Design-Build-Test Cycle

[THEO] Okay, picture this. You walk into a lab, and there's nobody there. But the science is still happening. A little robot's photographing seedlings in the dark, a language model's spitting out enzyme sequences that don't exist in nature yet, and somewhere a computer is designing an antibody against a target it's never seen. That's roughly where we are, and I want to trace how we got here.

[DR. MARA] It's a good frame, as long as we're honest that "AI at the bench" means very different things at different points in this arc. Some of it is automation — a machine doing a tedious task faster than a grad student. Some of it is generative modeling — proposing molecules that evolution never tried. Those are not the same claim, and the confusion between them is where a lot of hype lives.

[THEO] Right, so let's define the pieces for anyone who doesn't live in this world. The through-line here is the design-build-test-learn cycle. That's the loop biology engineering runs on. You design something — a gene, an enzyme, a strain. You build it. You test it. You learn from what broke, and you go around again.

[DR. MARA] And historically that loop is slow and painful. Each turn can take weeks. Directed evolution — which won a Nobel — is essentially brute-forcing that loop. You make random variants, you screen thousands, you keep the winners, you repeat. It works, but it's local search. You can only climb the hill you're standing on. You rarely make a big jump to a better solution across the valley.

[THEO] That "local search" limitation is going to matter for the whole story, so hold that thought. The dream is: what if a model could just look across the valley and tell you where the better hill is?

[DR. MARA] The other term worth defining up front is a foundation model. In language, it's a model trained on enormous text that learns the statistics of how words go together. In biology, you train the same architecture on DNA or protein sequences instead of English. It learns the grammar of what real proteins look like. That's the tool that keeps showing up in the recent papers.

[THEO] So where do the roots actually start? For me it's the least glamorous paper in the stack, and I love it for that. 2023, SPIRO — the Smart Plate Imaging Robot. Ohlsson and colleagues. It's a Raspberry Pi and a 3D-printed frame that photographs Petri plates on a schedule.

[DR. MARA] It's not AI in the generative sense at all. And that's exactly why it belongs at the base of this arc. Before you can have any machine learning worth the name, you need data — clean, consistent, at scale. SPIRO automates plant phenotyping: germination, root growth, measured plate after plate, image after image, by a device that doesn't get bored or introduce variability.

[THEO] And the detail I keep coming back to — it can image in the dark. If you're doing a seedling assay in the dark, every time you open the incubator and flip the lights on to take a photo, you've changed the experiment. A robot sitting in there shooting on its own schedule just... doesn't have that problem.

[DR. MARA] And it was built for biologists without engineering expertise. Open hardware, low cost. That democratization piece is the quiet turning point. You don't get a data-hungry field unless ordinary labs can generate the data. SPIRO is the "test" arm of that loop getting cheap and automated.

[THEO] So that's the build-and-test machinery getting industrialized. Then the design side explodes. Jump to 2025 and suddenly there's a cluster of papers all reaching for generative models. Where do you want to start, Mara?

[DR. MARA] Start with the base editor work, because it's the cleanest head-to-head. The TadA ortholog diversification study. TadA is a deaminase — the enzyme at the heart of base editors, the CRISPR tools that chemically convert one DNA letter to another without cutting the double helix. They took 95 natural TadA orthologs, shuffled them, and trained a generative model on that pool.

[THEO] And here's the payoff — in a single round, the model found compact cytosine editors and efficient adenosine editors that beat directed evolution starting from the same pool. One round. That's the valley-crossing dream I mentioned, actually landing.

[DR. MARA] With a caveat I want on the record. The paper is listed as in tension with generative protein sequence design more broadly. Which tells you the field itself isn't settled on how far these methods generalize. It worked beautifully here, from a curated natural pool. Whether that transfers to an enzyme with no such family to train on is a separate question.

[THEO] That tension shows up again in GenomeOcean, right? Zhou and colleagues. A four-billion-parameter genome foundation model, trained on metagenomes — DNA pulled straight from environmental samples, not tidy single organisms.

[DR. MARA] And two things stand out. It generates protein-coding sequences about 150 times faster than a comparable model, Evo-7B, partly through smarter tokenization — how you chop DNA into the units the model reads. And it can discover biosynthetic gene clusters zero-shot after fine-tuning. BGCs are the stretches of genome that encode the machinery for natural products — antibiotics, a lot of our drug scaffolds come from them.

[THEO] Zero-shot meaning it finds new ones without being explicitly trained on labeled examples of each. That's the leap. The classic tool there is antiSMASH, which scans for known signatures. A foundation model potentially sees clusters that don't match anything we've catalogued.

[DR. MARA] Though, again — it's flagged as in tension with genomic foundation models as a category. Bigger and faster is not automatically better biology. The open question is whether the sequences it generates actually fold and function, or just look statistically plausible.

[THEO] Which is the perfect segue to Germinal, because that one closes the loop between sequence and structure. It designs antibodies — specifically the CDRs, the loops that actually grip the target — against any protein epitope you hand it.

[DR. MARA] And the method is the interesting part. It co-optimizes two signals at once: AlphaFold-Multimer's structural confidence — does this thing fold and dock the way I want — and an antibody language model's sense of whether the sequence looks like a real antibody. Structure and naturalness, pulled together. They report nanomolar binders against four targets in something like 43 to 101 designs.

[THEO] Which, if you've ever done antibody discovery the old way — screening enormous libraries — testing dozens of designs instead of millions is a completely different sport.

[DR. MARA] It is. And it points at the paper that ties the whole arc into a thesis: the LDBT proposal. They reorder the loop. Instead of design-build-test-learn, it's learn-design-build-test. The model's zero-shot prediction comes first, before you design anything, and cell-free synthesis — making proteins in a tube without living cells — lets you build and test at massive scale.

[THEO] The ambition being a single-round campaign. You go around the loop once instead of twenty times.

[DR. MARA] That's the vision. And I'd separate vision from demonstration carefully. The TadA and Germinal papers are real single-round wins in narrow domains. LDBT and the enzyme-discovery vision paper are arguing that this becomes the default. The tension links in this arc are the field arguing with itself about whether that's true.

[THEO] From a robot photographing seedlings in the dark to models that skip most of the loop — in two years. Where it's heading is basically: how much of that loop can you actually delete without lying to yourself.

[DR. MARA] Well put. And the honest answer is: we don't know yet. Which is why it's worth watching.

[THEO] That's the arc. We'll be back with the mailbag after this.