South Korean Researchers Used AI to Design DNA That Folds Itself into Tiny Shapes

Key points

• South Korean researchers published a study in Nature Communications showing an AI model, Generative SNUPI, can design DNA structures from user-drawn shapes.
• DNA origami, a technique for bending genetic material into precise shapes, has existed for roughly 20 years but has been slow to develop because the design process is expensive and labour-intensive.
• Generative SNUPI uses a diffusion model, the same category of AI that powers image generators like DALL-E, to translate a target shape into a workable DNA blueprint.
• The research teams are based at Seoul National University and Hanyang University in South Korea.
• Future versions of the tool may support flexible, moving structures needed for drug delivery and cancer treatment.

Imagine folding a piece of paper into an origami crane, except the paper is a strand of DNA and the finished crane is thousands of times smaller than a human hair. That is the basic idea behind DNA origami, and it has scientists excited about possibilities ranging from tiny medical robots to structures that can deliver drugs directly inside the body.

The catch has always been the design step. Figuring out exactly how to sequence, or arrange, the chemical letters of DNA so that the strand folds into the right shape requires specialist knowledge, a lot of manual tweaking, and significant time.

A new AI model called Generative SNUPI, short for Structured Nucleic Acids Programming Interface, aims to remove that bottleneck. Teams at Seoul National University and Hanyang University built the tool and described it in the journal Nature Communications. As first reported by IEEE Spectrum AI, the model can take a user-drawn outline, think a dog's face or the silhouette of the Mona Lisa, and output a DNA design that physically works.

The model uses what researchers call a diffusion model. Think of it like sketching over a shape with millions of tiny, precise dots until the dots themselves become the blueprint. The model knows the chemical rules DNA follows: specifically, which molecular building blocks are attracted to each other. It uses those rules to populate the target shape with a valid DNA sequence.

Once the design is done, lab scientists synthesise, or chemically manufacture, short DNA strands called staples. These staples pull a longer strand, called a scaffold, into the intended shape. Kyounghwa Jeon, a PhD candidate at Seoul National University who worked on the project, says the process is "very similar to stapling paper."

The team found that some shapes did not hold at first, not because of an error in the AI, but because the input shape itself was structurally unstable. They added a check before the design stage to catch those cases early.

What does this mean for medicine?

For now, it mainly affects researchers. The current version of Generative SNUPI produces rigid structures, and many real-world medical uses, like guiding a drug to a cancer cell, need structures that can flex and move. Do-Nyun Kim, an assistant professor at Seoul National University, says the team plans to extend the work to dynamic, reconfigurable designs in future research.

If that goal is reached, the practical outcomes could include more precise drug delivery systems and new tools for immunotherapy, treatments that train the immune system to fight disease. That is still years away. But today, the tool already makes a demanding, expert-only process accessible to a much wider group of scientists.

Rebecca Taylor, a professor of mechanical engineering at Carnegie Mellon University who was not involved in the study, put it plainly: "The entire field is sort of enabled and held back by its tools. When you make a new tool that enables a new capability, that's just such a big advance for the field."