AI is revolutionizing protein design. Binders, enzymes, and even complex structures! But it hasn’t reached DNA nanotech yet.
The problem? We don’t have enough data. There isn’t a Protein Data Bank for DNA yet (some people are working on it!), with hundreds of thousands of experimentally validated structures.
Well, this paper takes a different approach: are simulations enough? Well, Generative SNUPI seems to do fine! Are these the first steps of generative AI in DNA nanotech?
Don’t keep this newsletter a secret: Forward it to a friend today!
Was this email forwarded to you? Subscribe here!
DNA Goes Generative
Researchers created Generative SNUPI, a generative AI model that designs wireframe DNA origami from just a target shape.
The DNA Origami Design Problem
DNA nanotech is my favorite way to build nanostructures.
Using the predictability of Watson-Crick base pairing, you can program DNA strands to self-assemble into nanoscale objects. And the applications are everywhere: vaccines, synthetic cells, and even DNA-based hard drives!
DNA nanotech is powerful. But protein design is starting to challenge DNA’s dominance.
The culprit? AI.
Well, machine learning and AI-based design tools, to be precise. These models have been designing binders and enzymes for a while. But now, they’re moving into structural design too, with virus-inspired nanocages and protein assemblies!
So, the obvious question is: Can AI do the same for DNA nanotech?
Today, most DNA nanostructures are still designed manually. Even complex structures like DNA origami begin with someone drawing the geometry in CAD software and deciding how the DNA should travel through it.
DNA origami uses one long single-stranded DNA molecule as a scaffold. Hundreds of smaller “staple” strands bind to different regions of that scaffold, folding it into a chosen nanoscale shape.
And yeah, you have to route this long scaffold along your structure yourself!
Researchers have developed tools to automate parts of this process, especially for wireframe structures. I actually worked on automated DNA origami design during my PhD!
But many of these systems struggle to create completely arbitrary shapes, and often still require a lot of manual work. The system I developed, for example, worked well for regular geometric shapes, but struggled with completely free-form shapes!
A generative AI model could change that.
Instead of building structures from predefined rules, it could learn the general principles behind successful DNA origami and generate new and more custom designs → just like in protein design!
So, why don’t we have an RFdiffusion for DNA origami?
The answer is simple: we don’t have the data. DNA nanotech doesn’t have a Protein Data Bank equivalent with hundreds of thousands of experimentally determined structures. And without data, you can’t train much!
No Data? No Problem
Well, the authors of today’s paper tackled the problem directly!
They introduce Generative SNUPI (Structured Nucleic acids Programming Interface), a diffusion-based AI framework for designing DNA origami nanostructures from free-form inputs.
And the data-shortage problem?
They solved it themselves. The team took 450 DNA origami structures with 2-helix-bundle (2HB) edges and simulated them. These simulated structures became the training data to teach Generative SNUPI how to design DNA origami!
And the results are pretty amazing.
The model generated DNA origami structures that could actually be produced! I was surprised by the high synthesis yield, even with minimal manual optimization. Sometimes DNA origami design requires a few tries; here, some structures reached > 80% yield on the first attempt!
But let’s see how Generative SNUPI works.
Under the Hood: Training Data and Generative Pipeline
Generative SNUPI is a denoising diffusion probabilistic model.
The basic idea is straightforward. During training, the model takes real DNA origami structures and gradually adds noise to their base-pair coordinates. Step by step, the nanostructure becomes a random cloud of points.
The neural network learns the reverse process: from noisy coordinates → predict where every base pair should be to recover a realistic DNA structure. After enough training, the model starts from random noise and reshapes it into a target DNA origami geometry!
Once the model is trained, the complete pipeline looks like this:
The user gives a 2D/3D target shape.
The model refines noise into a pair-based representation of that shape.
The DNA scaffold is routed through the structure.
Complementary DNA staples are generated.
Nucleotide sequences are assigned.
SNUPI simulations predict the final structure and its flexibility.
An automatic pipeline for DNA origami design! The AI generates the geometry, while conventional DNA algorithms handle routing and sequence design.
The team designed over 100 structures to explore Generative SNUPI’s potential. Examples included:
Artistic drawings
Biomimetic structures
Knot-like topologies
Many of these structures would be difficult or impossible to design using conventional design tools, because they have arbitrary curves, tilted edges, or weird connectivity!
Experimental Validation: Simulations, Gels, and AFM
Okay, but how do these structures perform in the real world?
Well, the authors designed and synthesized 3 types of structures:
Basic structures: simulations and wet lab
They started with basic wireframe structures. Generative SNUPI natively simulates structures and estimates their flexibility, so you can see how it correlates to folding quality! In general, floppier structures were less similar to their intended shapes in AFM images. But yields estimated from agarose electrophoresis were high, with peaks close to 90% and no structures below 79%! Generative SNUPI learnt how to make DNA structures that fold well.
Transformable structures
DNA origami nanostructures are easy to move. The team designed two structures that can change their shape, depending on the connectors. Generative SNUPI designed an intermediate structure, which was manually modified to be controllably contracted or expanded. These transformable structures can be hard to design manually!
Modular assemblies
Finally, the team tested modular assembly. They generated 2 versions of a dog head and 2 versions of a dog body. The interface between the heads and the body was kept constant, so that they could be swapped. The results were good, with > 65% combination efficiency for each of the 4 possible modular assemblies! Generative SNUPI managed to keep the interface constant in all 4 structures, even if the connections were manually added.
Is AI Coming for DNA Origami?
I had a similar idea years ago; I’m happy to see someone actually doing it!
Generative SNUPI is a step forward in bringing generative AI to DNA nanotech. These new models have made protein design so much easier; can they help expand DNA nanotech as well?
I was impressed by the yield of these structures! DNA origami design can be fickle, and it’s interesting to see that the model seems to learn how to fold structures with high yield.
In short, Generative SNUPI is new in 3 ways:
It’s generative, not rule-based: structures are created by sampling from a learned distribution, instead of assembling predefined motifs.
It operates directly at the base-pair level.
It integrates mechanical simulation into the design loop, evaluating flexibility and stability before synthesis.
Still, there are some limitations:
It’s trained only on 450 wireframe structures, a small dataset.
It’s limited to 2HB structures, and it doesn’t work for classic multilayer origami or other wireframe structures.
Sequence optimization and scaffold routing are still rule-based
The software isn’t easy to install and use. I wasn’t able to actually test it.
Only a small number of structures were validated experimentally.
But still, it’s a good step towards better DNA origami design software! With better software, we could get new designs, maybe with interesting material or light-interacting properties!
But for now, go here and read the paper!
If you made it this far, thank you! What do you think of DNA nanostructures and AI? Do you think we need better design tools? Reply and let me know!
P.S: Know someone interested in RNA nanotech and SynBio? Share it with them!
What did you think of today's newsletter?
More Room:
Cryo-EM and Cyto-Skeletons: Yeah, it’s a bad pun, but it’s a cool paper! The authors uncover the structural mechanism by which tubulin cofactors assemble αβ-tubulin, the building block of microtubules. Using cryo-EM, they show how these proteins form a cage-like complex that regulates tubulin assembly and degradation, providing new insights into how cells control microtubule formation. We still know so little about so many things!
DNA Growing… Metal? There’s so much cool work that interfaces DNA with metals. I was also working on it! In this paper, the authors develop DNA-templated gold–platinum nanoclusters with strong enzyme-like catalytic activity that outperforms natural horseradish peroxidase. By tuning the DNA template sequence, they optimize catalytic performance and demonstrate applications in miRNA biosensing and the degradation of environmental pollutants, highlighting the versatility of DNA-guided nanozymes.
Phage-Bacteria Relationships: Did you miss some CRISPR-Cas research? Oh well, the relationship between phages and bacteria just got more complicated. In this article, the authors identify Cas12p, a phage-associated CRISPR enzyme that uses the bacterial protein TrxA as an activator to enhance DNA degradation. Structural and biochemical analyses reveal how TrxA enables Cas12p activity, uncovering a new phage–bacteria interaction where host factors are repurposed to strengthen CRISPR-based defence mechanisms.