Automating Biology's Experimental Bottleneck

[SOFIA] Okay, so here's a question that sounds simple and absolutely is not: how fast can science actually go? Not how fast can we think — how fast can the bench keep up with the thinking.

[DANIEL] And for most of the history of molecular biology, the answer is: the bench is the bottleneck. You can have a brilliant hypothesis Monday morning and still be waiting on plates and pipettes Thursday.

[SOFIA] Right! So today we're tracing one through-line — AI at the lab bench. And I want to be careful, because "AI in biology" has become this huge fuzzy cloud. What we're actually following is a specific arc: machines and models that don't just analyze data, they decide what experiment to run next, or they design the molecule you build.

[DANIEL] Let me set the vocabulary, because this crosses a few fields. The old loop everyone in synthetic biology lives by is DBTL — Design, Build, Test, Learn. You design a construct, you build it in cells, you test whether it works, you learn, you go around again. Each lap is slow and expensive.

[SOFIA] And every one of these papers is attacking that loop from a different angle. Some are speeding up the Test. Some are trying to make the Design so good you barely need to loop at all.

[DANIEL] The other term worth defining up front: a foundation model. Same idea as the language models everyone's used — train a big neural network on an enormous pile of sequence data, no specific task, and it picks up the statistical grammar. For us that pile is DNA or protein sequence instead of English.

[SOFIA] And the thing a PhD from, say, immunology needs to hear is — these models learn what "normal" biological sequence looks like well enough to generate new ones. That's the leap. Not reading biology, writing it.

[DANIEL] So where does the arc start? I'd argue it starts humble, and I like that it does. 2023, SPIRO.

[SOFIA] I love SPIRO so much, and it's the least glamorous thing on the list. It's a Raspberry Pi and some 3D-printed parts that photograph Petri plates on a schedule. That's it.

[DANIEL] A plate imager. For germination and root-growth phenotyping. And the reason it belongs at the root of this story is the problem it names out loud: you cannot feed data-hungry methods if a human has to walk to the incubator and eyeball every plate.

[SOFIA] And it did something genuinely new in the biology, not just the engineering — it could run assays in the dark. Roots care about light. If every time you measure you flood the plate with light, you've perturbed the thing you're measuring. SPIRO images in darkness, at scale, no engineering degree required.

[DANIEL] Open hardware, reproducible, cheap. From my seat that's the quiet virtue — automation that removes the human as the source of variance. The measurement becomes consistent before you ever point a model at it.

[SOFIA] So that's the floor: get the data flowing, cleanly, without a person in the loop. The next turn is where it gets fun — what if the machine doesn't just record the experiment, it chooses it?

[DANIEL] BacterAI. Same year, 2023. This one made me sit up.

[SOFIA] Okay, this is the good stuff. The problem they're attacking is bacterial metabolism — specifically, which nutrients does an organism actually need? For non-model bugs we often have no idea. The classic approach is leave-one-out: drop one ingredient from the medium, see if it still grows. But the combinations explode. You can't test them all.

[DANIEL] And their reframing is the clever bit. They cast it as a reinforcement learning problem — a Markov decision process. The agent gets rewarded for removing ingredients while keeping the cells alive. So it's constantly pushed toward the edge, the boundary between growth and no-growth.

[SOFIA] The growth front! Which is exactly where the information is. Testing a medium that obviously works teaches you nothing. Testing one that obviously fails teaches you nothing. The learning is all at the knife's edge.

[DANIEL] And it mapped auxotrophies — the nutrients a strain can't make for itself — in days. The agent did the picking, a liquid-handling setup did the doing. That's a closed loop with the model steering.

[SOFIA] And notice this builds on Bayesian optimization in bioprocessing — same philosophy, pick the next experiment to maximize what you learn, don't grid-search blindly.

[DANIEL] That's the honest ancestor. Active learning dressed for the wet lab. What BacterAI adds is the organism in the loop in near-real time.

[SOFIA] So now jump to 2025, and the whole arc tilts. Up to here, AI is choosing experiments. Now it starts designing molecules. GenomeOcean — a genome foundation model, four billion parameters, trained on metagenome co-assemblies.

[DANIEL] Metagenome being the key word. They didn't train on tidy reference genomes, they trained on bulk environmental DNA — all the uncultured stuff. And a tokenization choice, byte-pair encoding, lets it generate protein-coding sequence about 150 times faster than a comparable model.

[SOFIA] And the application that matters here — biosynthetic gene clusters. BGCs. Those are the stretches of genome that encode the assembly lines for natural products, antibiotics, all the interesting chemistry. Normally you hunt them with tools like antiSMASH, which pattern-match against things we already know.

[DANIEL] Which is the limitation — you find what resembles what you've seen. GenomeOcean, fine-tuned, claims zero-shot BGC discovery. Novel clusters, no labeled examples.

[SOFIA] Now, there's tension in the field over genomic foundation models — whether they're really learning biology or just very good autocomplete. I don't think that's settled, and I wouldn't oversell it. But the direction is clear.

[DANIEL] Hold the skepticism in one hand, agreed. "Novel" needs wet-lab confirmation that the cluster makes a real molecule. The generation is cheap; the validation still isn't.

[SOFIA] Which is the perfect bridge to Germinal, because that one actually ties the design back to structure. Antibody design — they want the CDRs, the loops that grab the target — against any protein epitope.

[DANIEL] And the move is co-optimization. They push on AlphaFold-Multimer's structural confidence and an antibody language model's sense of what a "natural" antibody looks like, at the same time. Reported nanomolar binders against four targets, in something like 43 to 101 designs.

[SOFIA] A hundred designs to a nanomolar binder is wild if it holds. Compare that to the old way — immunize, screen, hope.

[DANIEL] "If it holds" is doing work. Four targets, success rates from four to twenty-two percent — that's a real spread, and a small number of targets. I'd want to see it on epitopes people have historically failed on. But the structure-plus-language joint objective is a genuinely good idea.

[SOFIA] And that philosophy — predict first, build second — is exactly what the last two pieces name explicitly. The enzyme-discovery vision wants one unified generative model tying sequence, structure, and function together, so you could in principle encode any chemistry you want.

[DANIEL] And it's openly in tension with machine-learning-directed evolution. Directed evolution does local search — mutate, select, climb the nearby hill. The generative pitch is: skip the hill-climbing, jump straight to the sequence. Whether a model can really leap across sequence space like that is the open question of the whole field.

[SOFIA] Which brings us to the capstone — the LDBT paradigm. They literally reorder the old loop. Learn, Design, Build, Test. The model's zero-shot prediction comes first, and then cell-free synthesis — expressing proteins without living cells — lets you build and test at massive scale.

[DANIEL] The dream being a single-round campaign. No more laps.

[SOFIA] So that's the arc, start to finish — SPIRO gets the data flowing, BacterAI lets the machine pick the experiment, and the 2025 crew tries to design the answer before the experiment even runs.

[DANIEL] With the bottleneck quietly migrating. It used to be building and testing. Increasingly it's trusting — validating what the model generated.

[SOFIA] And that's a great problem to have. More on that next.