Closing The Lab Bench Loop

[SOFIA] Okay, so today I want to talk about the lab bench itself. Not one organism, not one tool — the actual physical act of doing biology experiments, and how AI is creeping into every single step of it.

[DANIEL] Which is a broader claim than it sounds, because "AI in biology" usually means someone trained a model on a database and made a prediction. That's the easy version.

[SOFIA] Right, the version where the AI never touches a pipette.

[DANIEL] Exactly. The interesting story is what happens when you close the loop — when the model doesn't just predict, it decides what experiment to run next, and something in the physical world actually runs it.

[SOFIA] So let me set the stakes for anyone who doesn't live in this world. The way we've done biology forever is a cycle. In synthetic biology we literally have a name for it — Design, Build, Test, Learn. DBTL. You design a construct, you build it, you test it in cells, you learn from the mess, you go around again.

[DANIEL] And each turn of that wheel is slow. Weeks. Growing cells, running assays, sequencing. The bottleneck isn't ideas — it's throughput and time. You can only test what you can physically make and measure.

[SOFIA] And that's the thread through everything today. Every one of these papers is attacking a different piece of that cycle. Some are attacking the "test" — the physical measurement. Some are attacking the "design" — what should we even try. And a couple are trying to reorder the whole thing.

[DANIEL] So maybe define the players for someone coming from, say, immunology or ecology, who doesn't sit in a synbio lab.

[SOFIA] Good idea. Foundation model — big neural network trained on a mountain of unlabeled data, protein sequences or genomes, that learns the statistical grammar of biology and can then do new tasks. Directed evolution — the old-school way to make better enzymes, you mutate, you screen, you keep the winners, repeat. It works, but it's local search — you only find what's near where you started.

[DANIEL] And cell-free systems — you take the transcription-translation machinery out of the cell, run it in a tube, so you can test a construct without keeping anything alive. That matters later.

[SOFIA] So let's go chronologically, because there's a real arc here. Start in 2023, and honestly start humble — with a Raspberry Pi. This is the SPIRO paper, Ohlsson and colleagues.

[DANIEL] The Smart Plate Imaging Robot.

[SOFIA] It's a 3D-printed imager for Petri plates. You put your plates in, it photographs them on a schedule, and it tracks things like seed germination and root growth over time. That's it. No language model, no billion parameters.

[DANIEL] And I love that it's the starting point, because it's attacking the least glamorous bottleneck — the "test." Someone standing at a bench photographing plates by hand. SPIRO automates the observation, at scale, and crucially it can image in the dark, which you can't do standing there with the lights on.

[SOFIA] Which for anyone doing circadian or light-response work in plants — that's an assay you basically couldn't run before.

[DANIEL] And it's open hardware, built by biologists without engineering training. That's the quiet point. You don't need AI to start automating the bench. You need to get the data off the human.

[SOFIA] Same year, same instinct, one level up — BacterAI. And this one, okay, this is the good stuff. The problem: you want to know what nutrients a microbe actually needs to grow. Its metabolism, mapped. Classically you'd guess, or you'd already need a genome and a model.

[DANIEL] And the combinatorics are brutal. Dozens of ingredients, every possible combination — you can't test them all.

[SOFIA] So they reframe it as a reinforcement learning problem. A Markov decision process. The agent gets rewarded for removing ingredients from the medium and still seeing growth. So it keeps pushing right up to the edge — the boundary where the microbe just barely survives.

[DANIEL] The growth front.

[SOFIA] The growth front! And by hugging that boundary it learns the auxotrophies — what the organism can't make for itself — in days, with no prior knowledge.

[DANIEL] And here's what earns my respect. The agent isn't predicting from a database someone else built. It's choosing the next round of experiments, a robot runs them, the results feed back in. That's the closed loop we were talking about. The AI is steering the actual bench.

[SOFIA] And it builds directly on Bayesian optimization ideas from bioprocess engineering — the whole framework of "don't sample randomly, sample where you'll learn the most."

[DANIEL] Which is the honest engine under a lot of this. Experimental design as an optimization problem.

[SOFIA] Okay, so jump to 2025, and the ambition explodes. GenomeOcean, Zhou and colleagues. A four-billion-parameter genome foundation model — but trained on metagenome co-assemblies. Real environmental DNA, mixed communities, not tidy reference genomes.

[DANIEL] Which matters, because most of life isn't in your reference database. If you train on metagenomes you're learning the actual diversity out there.

[SOFIA] And they use BPE tokenization — byte-pair encoding, chunking the DNA into learned subunits rather than single bases — and they report generating protein-coding sequences a hundred and fifty times faster than a comparable model, Evo-7B.

[DANIEL] And zero-shot discovery of biosynthetic gene clusters. BGCs — stretches of genome that encode the machinery to make natural products, antibiotics, that kind of chemistry. Normally you hunt those with a tool like antiSMASH, rule-based. GenomeOcean, fine-tuned into "bgcFM," finds novel ones without being shown examples first.

[SOFIA] So now the AI isn't just steering experiments — it's proposing the biology itself. New clusters that might make new molecules.

[DANIEL] Though I'll flag — it's in tension with the broader genomic foundation model literature. Not everyone agrees these models generalize the way the headline suggests. Generating a plausible sequence is not the same as it folding and functioning.

[SOFIA] Which is exactly the bridge to Germinal — antibody design, 2025. They co-optimize two things: AlphaFold-Multimer's structural confidence, does the complex fold and bind, and an antibody language model, IgLM, for naturalness — does this look like a real antibody your body might actually make.

[DANIEL] And they design nanomolar-affinity CDRs — the binding loops — against four different protein targets, at a success rate of four to twenty-two percent, in something like 43 to 101 designs.

[SOFIA] Which for de novo binder design is genuinely good. You're testing dozens, not millions.

[DANIEL] And the two-model trick is the clever part. Structure alone will happily design something that binds beautifully and is completely un-manufacturable or immunogenic. The naturalness term pulls it back toward biology.

[SOFIA] So that brings us to the paper that names the whole arc — the LDBT paradigm. They take Design-Build-Test-Learn and reorder it. Learn first. L-D-B-T.

[DANIEL] Meaning the machine learning zero-shot prediction comes before you design anything physical. And then cell-free synthesis at megascale handles the build-and-test in a tube.

[SOFIA] The dream being a single-round campaign. You predict, you build once, you test massively in parallel, you're done. No years of going around the wheel.

[DANIEL] And there's the enzyme-discovery vision paper sitting right alongside it — unified generative models that encode sequence, structure and function together, to reach chemistry that directed evolution's local search can't.

[SOFIA] Which is the same philosophical fight from GenomeOcean, just louder. It's explicitly in tension with machine-learning-guided directed evolution — the camp that says, no, keep the iterative loop, just make each step smarter.

[DANIEL] And I don't think that's settled. The generative "design it all up front" vision is beautiful. But every result today that I actually trust — BacterAI, Germinal — kept a real experiment in the loop.

[SOFIA] So that's the through-line. We went from a Raspberry Pi photographing plates to models proposing whole enzymes — but the bench never left. The open question is just how many times you still have to go around.

[DANIEL] Learn first, sure. Test always.

[SOFIA] On that note — we'll be right back after this.