AI Designs Biology Beyond Evolution

[SOFIA] So here's a question I've been chewing on. When people say "AI is coming for the lab," what do they actually mean? Because there's the version where a language model writes your grant, and then there's the version I actually care about — where AI is standing at the bench, deciding what molecule to make next.

[DANIEL] And those are very different claims with very different evidence behind them.

[SOFIA] Right! So today we're tracing that second one. How AI moved from a helper on the side to something that's genuinely proposing biology — enzymes, antibodies, genome editors. And it's a shorter arc than you'd think.

[DANIEL] It's worth being precise about what problem this is solving, though. Because the classic way you improve a protein — an enzyme, a binder — is directed evolution.

[SOFIA] The Frances Arnold Nobel approach.

[DANIEL] Exactly. You take a starting sequence, you mutate it, you screen thousands of variants, you keep the winners, and you do it again. Round after round. It works beautifully, but it's local search — you're feeling around the immediate neighborhood of what you started with. You can't jump across the sequence landscape to something genuinely different.

[SOFIA] And it's slow, and it's expensive, and every round is more bench work. Which is where automation comes in — and honestly that's the root of this whole story. Before you get fancy AI, you need machines that can actually watch biology happen at scale. So the oldest paper in our arc is from 2023, and it's almost charmingly humble compared to the rest.

[DANIEL] SPIRO. The Smart Plate Imaging Robot.

[SOFIA] A 3D-printed, Raspberry-Pi-based imager for petri plates. That's it. It rotates your plates, takes photos on a schedule, and lets you phenotype germination and root growth without a grad student babysitting a camera at 2 a.m.

[DANIEL] And I want to flag why that's not trivial. It's open hardware, built for biologists with no engineering background. The clever bit is it works in the dark — so you can actually image germination assays that were previously impractical because turning the lights on to photograph them ruins the experiment.

[SOFIA] Which is such a small, real problem. But that's the point, right? The bench is full of these tasks that are tedious and quantitative and perfect for a machine. SPIRO is the "collect the data" end of the loop.

[DANIEL] And it earns its place in this arc because everything downstream depends on data at scale. You can't train a useful model on twelve plates someone measured with a ruler.

[SOFIA] Okay so hold that thought — data collection — because the next move is data generation. Fast forward to 2025 and suddenly the whole field jumps to foundation models for genomes. GenomeOcean, from Zhou and colleagues.

[DANIEL] This one impressed me. It's a four-billion-parameter model trained on metagenome co-assemblies — so, DNA pulled straight from environmental samples, not tidy reference genomes.

[SOFIA] Which matters because that's where the weird chemistry lives. The uncultured stuff.

[DANIEL] The reported numbers: it generates protein-coding sequences about 150 times faster than Evo-7B, and after fine-tuning into a version they call bgcFM, it discovers biosynthetic gene clusters zero-shot.

[SOFIA] And for anyone outside this world — biosynthetic gene clusters, BGCs, are stretches of genome where bacteria encode the machinery to make complex molecules. Antibiotics, natural products. They cluster the genes together, conveniently for us. Normally you hunt them with a tool called antiSMASH that scans for known signatures.

[DANIEL] Which means antiSMASH finds things that look like what we already know. "Zero-shot" here means the model proposes clusters it was never explicitly told to look for. That's the leap.

[SOFIA] Okay, this is the good stuff, because now the model isn't just reading biology, it's writing it. And the brief notes GenomeOcean sits in some tension with the broader genomic foundation model crowd —

[DANIEL] Right, and I read that as a healthy disagreement about how you should train these things. Do you build one giant general model, or does metagenome-specific training and their tokenization choice buy you something real? GenomeOcean's argument is the specialized recipe wins on speed and on finding novelty. That's a testable claim, and I'd want to see it hold up against more than their own benchmarks.

[SOFIA] Fair. But notice the through-line — we went from "watch a plate" to "generate genome." The next three papers all ask the same question: can AI replace an evolution round entirely?

[DANIEL] And this is where I get more careful, because "replace directed evolution" is exactly the kind of claim that needs controls.

[SOFIA] So there's a vision paper — the AI-driven enzyme discovery piece — that lays out the dream explicitly. One unified generative model that encodes sequence, structure, and function together, so you could in principle genetically encode any chemistry you want. Reach parts of the landscape directed evolution's local search can never crawl to.

[DANIEL] That's a manifesto, though. It's a vision. And notably it's in tension with the machine-learning-guided directed evolution camp — the people who say, don't throw out evolution, just steer it with ML.

[SOFIA] Which is the honest fault line in the whole field right now. Replace the loop, or guide the loop?

[DANIEL] And the TadA paper is the one that actually puts a number on it.

[SOFIA] Yes! This is the one I'd point a skeptic to. TadA is a deaminase — it's the enzyme at the heart of base editors, the CRISPR tools that chemically convert one DNA letter to another without cutting the strand.

[DANIEL] They took a pool of 95 natural TadA orthologs, shuffled them, and trained a generative model on that diversity. Then — single round — they pulled out both compact cytosine editors and efficient adenosine editors. And critically, they compared against directed evolution starting from the same pool.

[SOFIA] Same starting material. Head to head.

[DANIEL] And the generative approach beat it in one round, with lower off-target activity reported. That's the comparison I wanted. It's still one system, one enzyme family — I wouldn't generalize to "AI beats evolution everywhere." But as a controlled demonstration, it holds.

[SOFIA] And Germinal is the antibody version of the same idea. They co-optimize AlphaFold-Multimer's structural confidence with an antibody language model's sense of what a "natural" antibody looks like, and design nanomolar binders against four different protein targets.

[DANIEL] With a reported success rate of 4 to 22 percent across 43 to 101 designs. Which — for de novo binder design against arbitrary epitopes — is genuinely workable. Low, but workable, and honestly reported as a range.

[SOFIA] And then LDBT ties the bow. The classic engineering cycle is Design-Build-Test-Learn. LDBT reorders it — Learn first. You let a model make zero-shot predictions before you design anything, then you use cell-free synthesis to build and test at massive scale.

[DANIEL] The aspiration being a single-round campaign. No iterating.

[SOFIA] Which is the whole arc in one acronym, right? SPIRO watched the plate. GenomeOcean generated the sequence. TadA and Germinal proved the generated thing works. And LDBT says: put the AI at the front of the loop instead of the end.

[DANIEL] I'd just keep the asterisk visible. The vision papers promise "any chemistry." The data papers show one enzyme family, four targets, single-digit-to-twenty-percent hit rates. That gap is where the next three years of work lives.

[SOFIA] And that's a great place to be — the demos are real, the promises are bigger, and now somebody has to close it. We'll be watching. After the break, one organism that refuses to be transformed no matter what we throw at it.