De novo design of DNA origami with a generative diffusion model Authors: Chien Truong-Quoc, Kyounghwa Jeon, Jinho Kim, Seo Hyun Kwon, Dongsik Seo, Chanseok Lee, Do-Nyun Kim Published: 2026 ABSTRACT Generative models that have advanced inverse design in protein engineering can be extended to DNA origami to explore broader design spaces enabling complex geometries and functions for emerging nanotechnologies. However, progress in generative DNA origami design has been limited by the lack of large, standardized structural datasets containing structural information. To address this challenge, we introduce a diffusion-based generative design framework trained on simulated equilibrium conformations obtained using a multiscale computational model. Given user-defined target geometries, our model produces physically plausible DNA origami designs through guided diffusion sampling and strand routing, with integrated structure prediction for quantitative evaluation. Among more than 100 generated candidates, selected structures are experimentally validated, demonstrating proper folding and functional behaviors such as auxetic transformation and modular assembly. Our results highlight the potential of generative modeling for complex DNA origami design, expanding the accessible design space and facilitating the creation of sophisticated, reconfigurable architectures. Introduction DNA origami provides a versatile framework for constructing nanostructures with programmable shape and properties through controllable DNA hybridization. Although designs with DNA duplexes arranged on regular lattices remain widely used, recent advances have led to a growing demand to utilize geometrically more complex, free-form structures targeted for emerging applications. Such structures enable intricate functional assemblies1–5, geometry-guided multivalent interfaces6–13, and reconfigurable nanosystems exhibiting dynamic mechanical behaviors14–20. Despite the proliferation of computer-aided design (CAD) tools21–29, most frameworks remain constrained by algorithmic regularity and periodic crossover placement, following the inherent geometry of double-stranded DNA. While these approaches are effective in forming regular and stable structures, they are limited in their capacity to generate arbitrary, free-form conformations and often require manual intervention30–32. Recently, diffusion-based generative AI models33 have demonstrated astonishing performance in the inverse design and generation of structures including complex protein structures34,35 and small molecules36,37. By learning implicit structural priors from data, these models offer a data-driven alternative to conventional algorithmic approaches. However, despite recent advances in machine learning-based predictive modeling38 and computational characterization of DNA nanostructures39,40, developing a generative AI model for designing DNA nanostructures remains challenging. This difficulty primarily arises from the lack of a sufficiently large and standardized dataset of their designs suitable for training generative models41 and highly limited structures determined either experimentally or computationally42–47. Here, we present a diffusion-based design framework, Generative SNUPI (Structured NUcleic acids Programming Interface), for the design of complex, free-form DNA origami structures. We leveraged our multiscale simulation, SNUPI48–51, to overcome data scarcity by generating equilibrium configurations of wireframe DNA origami designs, which served as a physically validated training set for learning structural priors. Line-based shapes extracted from a target geometry through manual sketching, CAD, or AI-assisted design can serve as inputs, thereby expanding the design space from simple shape tracing to biomimetic and nanomechanical components (Fig. 1a). Generative SNUPI translates these input geometries into atomic structural designs through diffusion sampling and DNA routing, followed by in silico evaluation of structural stability and stiffness (Fig. 1b). We experimentally validated selected structures and confirmed that our framework expands the accessible design space beyond conventional tools, allowing the design of structures that exhibit reconfigurable mechanics and modular organization (Fig. 1c). Our results demonstrate how generative modeling can be used to create physically realizable, geometrically complex DNA origami structures, opening new avenues in data-driven design for structured DNA assemblies.Fig. 1Design and synthesis flow of DNA origami using generative SNUPI.a Inputs can take the form of line-based target shapes created through various methods. Target shapes range from simple tracings to biomimetic and nanomechanical designs. Examples shown: Mona Lisa image (Image courtesy of the Centre de recherche et de restauration des musées de France (C2RMF), via Wikimedia Commons. Public domain.), assembly of the phiX174 procapsid (PDB ID: 1CD3), and reconfigurable kirigami metasheet with triangle cuts. Scale bar, 10 nm. b Generative SNUPI converts a random 3D point cloud into a DNA origami structure, which is then turned into an atomic design using a routing program. SNUPI-based simulation is incorporated to provide the prediction of structural features. Scale bars, 20 nm. c Annealing and analyzing monomer yield using agarose gel electrophoresis (L: ladder, S: scaffold, U: unpurified sample after folding, P: purified sample by ultrafiltration.) and AFM. AFM images are aligned and averaged to produce a mean image. Each monomer image is then compared to the target shape to calculate RMSD for accuracy estimation (n = 200). Scale bars, 100 nm. Source data are provided as a Source Data file. Uncropped gel images with molecular weight markers are provided in the Source Data. Molecular graphics in (a) (phiX174 procapsid) and (b) were generated using UCSF Chimera66, developed by the Resource for Biocomputing, Visualization, and Informatics at the University of California, San Francisco, with support from NIH P41-GM103311. Design and synthesis flow of DNA origami using generative SNUPI. a Inputs can take the form of line-based target shapes created through various methods. Target shapes range from simple tracings to biomimetic and nanomechanical designs. Examples shown: Mona Lisa image (Image courtesy of the Centre de recherche et de restauration des musées de France (C2RMF), via Wikimedia Commons. Public domain.), assembly of the phiX174 procapsid (PDB ID: 1CD3), and reconfigurable kirigami metasheet with triangle cuts. Scale bar, 10 nm. b Generative SNUPI converts a random 3D point cloud into a DNA origami structure, which is then turned into an atomic design using a routing program. SNUPI-based simulation is incorporated to provide the prediction of structural features. Scale bars, 20 nm. c Annealing and analyzing monomer yield using agarose gel electrophoresis (L: ladder, S: scaffold, U: unpurified sample after folding, P: purified sample by ultrafiltration.) and AFM. AFM images are aligned and averaged to produce a mean image. Each monomer image is then compared to the target shape to calculate RMSD for accuracy estimation (n = 200). Scale bars, 100 nm. Source data are provided as a Source Data file. Uncropped gel images with molecular weight markers are provided in the Source Data. Molecular graphics in (a) (phiX174 procapsid) and (b) were generated using UCSF Chimera66, developed by the Resource for Biocomputing, Visualization, and Informatics at the University of California, San Francisco, with support from NIH P41-GM103311. Results Generative SNUPI design framework We adopt a denoising diffusion probabilistic model (DDPM)33 as the generative diffusion model for learning and generating DNA origami structures (Supplementary Note 1). This model operates on a base-pair-level representation of DNA origami structures, where each base pair is described by its 3D coordinates. We treat all base pairs within the DNA structure equally, akin to a point cloud, and develop a scalable graph Transformer for DNA origami structures with random graph construction (Supplementary Note 2). The training dataset consists of 450 wireframe structures with two-helix-bundle (2HB) edges (216 2D and 234 3D designs) whose base-pair coordinates were determined using SNUPI48. To prepare the training inputs, we added 3D Gaussian noise to them with up to 500 perturbation steps per structure (forward process in Fig. 2a and Supplementary Note 3). We incorporate SE(3) symmetry-based geometric invariance52–54 to directly predict base-pair coordinates from noisy inputs, minimizing the mean-squared error (MSE) between the predicted and ground-truth structures without requiring any alignment34 (Fig. 2b and Supplementary Note 2). This loss function ensures denoising paths match the data distribution, leading to constructible DNA origami structures.Fig. 2Core design steps of the Generative SNUPI.a From the training set, a forward diffusion process gradually transforms DNA origami structures at the base-pair scale into random point clouds (right to left). The reverse process (left to right) generates new structures by iteratively denoising the initial random noise XT into a plausible structure X0. b The denoiser consists of SE(3)-Transformer layers and a random graph neural network with long-range connectivity. c The Wasserstein Distance (WD) is used to directly bias sampling towards target shapes. d Conditional DDPM performs stably across a wide range of shapes and sizes of designed patterns (as indicated by the initial WD values, 1294.99 ± 458.52 nm, using gray circles). The generated DNA structures match well with their corresponding target shape (as indicated by the final WD values, 2.21 ± 1.32 nm, using blue circles). e Generated structures from DDPM are converted into atomic designs through a routing process. f The quality of generated designs. The RMSD values are calculated between target and generated configurations, and the average RMSF values are obtained by averaging base-pair RMSF values of each structure. Structures marked with blue circles are shown in the top row of (g), while those marked with orange circles appear in the bottom row. Corresponding samples between (f) and (g) are indicated by matching symbols. g A representative set of DNA origami designs generated using conditional DDPM. Values in parentheses represent RMSD and average RMSF. Scale bars, 20 nm. Source data are provided as a Source Data file. Molecular graphics were generated using UCSF Chimera66, developed by the Resource for Biocomputing, Visualization, and Informatics at the University of California, San Francisco, with support from NIH P41-GM103311. Core design steps of the Generative SNUPI. a From the training set, a forward diffusion process gradually transforms DNA origami structures at the base-pair scale into random point clouds (right to left). The reverse process (left to right) generates new structures by iteratively denoising the initial random noise XT into a plausible structure X0. b The denoiser consists of SE(3)-Transformer layers and a random graph neural network with long-range connectivity. c The Wasserstein Distance (WD) is used to directly bias sampling towards target shapes. d Conditional DDPM performs stably across a wide range of shapes and sizes of designed patterns (as indicated by the initial WD values, 1294.99 ± 458.52 nm, using gray circles). The generated DNA structures match well with their corresponding target shape (as indicated by the final WD values, 2.21 ± 1.32 nm, using blue circles). e Generated structures from DDPM are converted into atomic designs through a routing process. f The quality of generated designs. The RMSD values are calculated between target and generated configurations, and the average RMSF values are obtained by averaging base-pair RMSF values of each structure. Structures marked with blue circles are shown in the top row of (g), while those marked with orange circles appear in the bottom row. Corresponding samples between (f) and (g) are indicated by matching symbols. g A representative set of DNA origami designs generated using conditional DDPM. Values in parentheses represent RMSD and average RMSF. Scale bars, 20 nm. Source data are provided as a Source Data file. Molecular graphics were generated using UCSF Chimera66, developed by the Resource for Biocomputing, Visualization, and Informatics at the University of California, San Francisco, with support from NIH P41-GM103311. The model starts from random base-pair coordinates and progressively refines them over multiple denoising iterations while incorporating noise. Through successive steps, the design space narrows and the predictions converge toward realistic DNA origami geometries (reverse process in Fig. 2a and Supplementary Note 4). We impose the given geometry as a condition on the DDPM so that it can generate a DNA origami design following the target shape accurately (Fig. 2c), while the unconditional DDPM can generate wireframe designs without external guidance (Supplementary Fig. 22). We integrate the conditioning approach based on optimal transport55 which provides tools for identifying correspondences and defining similarities between objects in different domains. We accomplish this by adding classifier gradients based on Wasserstein Distance (WD) between the designed pattern and the generated structure (Supplementary Note 1). These conditions help the generated structures converge to target shapes in the final step of the reverse diffusion process (Fig. 2c). This conditional guidance also enhances the model performance by narrowing the design space around a defined structural objective, particularly in terms of routing success rate (Supplementary Note 5). To evaluate the capability of Generative SNUPI and the performance of conditional DDPM, 100 DNA origami structures (Supplementary Fig. 23) with diverse shapes and sizes were generated (Fig. 2d). The performance was assessed by measuring the WD between the target shapes and the generated structures at both the initial and final stages of the reverse diffusion process. The initial WD values exhibited substantial variation, ranging from 192.69 to 2178.54 nm, reflecting the broad diversity of target shapes, with structure sizes spanning from hundreds to ten thousand base pairs. In contrast, the final WD values showed a clear reduction, averaging 2.21 nm with a standard deviation of 1.32 nm (Fig. 2d and Supplementary Table 2). These results suggest that the framework successfully accommodates common DNA origami structures across a wide range of sizes, achieving high fidelity in shape generation even for structures consisting of approximately 15,000 base pairs (Supplementary Fig. 24). In addition, we assessed novelty and sampling variability using two complementary analyses to evaluate potential overfitting. For unconditional sampling of 400-500 base-pair structures, nearest-neighbor root mean square deviation (RMSD) evaluations within the generated samples and RMSD comparisons to the training set showed that generated structures are sufficiently distinct from the training data ( < 1.0% of samples below 0.5 nm RMSD; Supplementary Fig. 25). Furthermore, the conditional sampling with both target shape and the number of base pairs fixed confirmed non-deterministic sampling without collapse despite strong conditioning (Supplementary Fig. 26). These results demonstrate that Generative SNUPI learns a generative prior rather than simply reproducing samples from the training set and remains non-deterministic under target shape conditioning without mode collapse. Subsequently, we integrate DDPM with a routing program to enable a fully Generative SNUPI design framework that produces both structures and their corresponding encoded sequences (Fig. 2e and Supplementary Note 5). The routing process refines the generated structure by the reverse diffusion process into an atomic DNA origami model with a corresponding scaffold route and staple sequence set, providing functionality similar to that offered by previous design programs24,56–58. First, the generated structure is converted into a loop representation, identifying continuous strand paths. Next, a spanning tree is constructed to establish a simplified backbone connectivity while preserving the overall topology. The scaffold routing step then processes the spanning tree to define a continuous scaffold path, ensuring structural stability. To maintain regular base-pair steps within helices, a radius graph generation method is employed, enforcing a consistent base-pair bond length of 0.34 ± 0.05 nm. In the final sequence design step, nucleotide sequences are assigned to both scaffold and staple strands, with staple lengths ranging from 20 to 60 nucleotides, ensuring compatibility with experimental synthesis constraints. Importantly, while the routing procedure is deterministic given the generated geometry, the diffusion model produces a different base-pair-level structural geometry at each sampling. Consequently, repeated generations under the same target shape can induce distinct routed crossover and topology patterns (Supplementary Fig. 27). We export the final design into a CanDo59 representation and then convert it into an atomic PDB representation via a deterministic post-processing pipeline using tacoxDNA60 (CNDO → oxDNA → PDB). Atomic models generated through this process enable further validation through experimental realization and computational analysis. We also employed a SNUPI-based simulation model48–51 within the Generative SNUPI workflow to establish a unified framework that integrates design and simulation at the base-pair level. Unlike conventional workflows that require format conversion between design and simulation stages26, our approach seamlessly integrates these stages, maintaining consistency throughout the process and eliminating inefficiencies. Moreover, this SNUPI-based analytical model enables rapid prediction of structural features in full-sized DNA origami with makedly lower computational cost compared to conventional molecular dynamics-based methods61,62, making it a practical resource for iterative design validation and refinement. We calculated the RMSD, quantifying the structural deviation between the target and predicted conformations, and the average root mean square fluctuation (RMSF), indicating the dynamic flexibility, for 100 designs whose equilibrium shape and thermal fluctuations are predicted using the simulation model (Fig. 2f). A majority of the designs (72%) fall within one standard deviation of the mean RMSD (2.49 ± 1.29 nm) and the mean average RMSF (1.72 ± 0.15 nm), suggesting their high structural fidelity and mechanical robustness. Figure 2g shows three representative designs well matched with their target shapes (top) and three other designs distorted from their target shape with relatively high flexibility due to sparse edge connection (bottom). Figure 3 showcases the ability of Generative SNUPI to design geometrically diverse and complex nanostructures with high fidelity. For example, artwork-inspired designs highlight the model’s capacity for fine-grained encoding of free-form geometric features (Fig. 3a). Kirigami-/origami-inspired designs demonstrate a great potential of Generative SNUPI to be used to construct a structure with periodic unit cells that have unique mechanical properties and reconfiguration mechanisms (Fig. 3b). Unlike conventional design tools, 3D wireframe architectures with curved or tilted edges can be easily designed using Generative SNUPI (Fig. 3c). Topologically more intricate designs, such as 3D knots, can be readily designed as well (Fig. 3d). Together, Generative SNUPI offers an effective way of expanding the creative boundaries of structural DNA nanotechnology by enabling the generation of highly diverse, noncanonical structures beyond the reach of rule-based or manually curated design workflows.Fig. 3Design gallery of free-form DNA origami structures.Representative 24 designs created using Generative SNUPI (The complete set of 100 designs is provided in Supplementary Fig. 23). a Artwork-inspired designs. b Kirigami-/origami-inspired designs. c Three-dimensional wireframe designs with curved or tilted edges. d Knot-inspired designs. Scale bars, 20 nm. Molecular graphics were generated using UCSF Chimera66, developed by the Resource for Biocomputing, Visualization, and Informatics at the University of California, San Francisco, with support from NIH P41-GM103311. Design gallery of free-form DNA origami structures. Representative 24 designs created using Generative SNUPI (The complete set of 100 designs is provided in Supplementary Fig. 23). a Artwork-inspired designs. b Kirigami-/origami-inspired designs. c Three-dimensional wireframe designs with curved or tilted edges. d Knot-inspired designs. Scale bars, 20 nm. Molecular graphics were generated using UCSF Chimera66, developed by the Resource for Biocomputing, Visualization, and Informatics at the University of California, San Francisco, with support from NIH P41-GM103311. Experimental validation of Generative SNUPI designs A main advantage of the Generative SNUPI design framework is its ability to predict the structural flexibility of free-form designs, offering an opportunity to revise the design before experimental synthesis. To demonstrate, we designed two dog faces (Face 1 and Face 2) and examined their structural characteristics experimentally (Fig. 4a). According to the simulation result, the Face 1 design exhibited high RMSF values in some parts of the structure due to an insufficient connection between edges while the mean shape matched relatively well with the target. These local regions with high RMSF may exhibit collapsed shape due to structural instability. Indeed, AFM images confirmed structural distortion of the Face 1 design (Fig. 4a, top row; Supplementary Fig. 28) even though the monomer yield evaluated from gel electrophoresis was as high as 79.7%. As a result, high RMSF regions of the structure (near nose and chin) appeared blurry in the mean AFM image (Fig. 4a, inset and Supplementary Note 6). Furthermore, the distribution of RMSD between individual monomers and the target shape showed notable deviation from the intended structure (4.07 ± 0.48 nm). To reduce structural deviation, we enhanced the stiffness the design by adding edges to the flexible regions, resulting in the Face 2 design (Fig. 4a, bottom row). AFM images for the Face 2 design demonstrated markedly reduced structural distortion with a slightly increased monomer yield of 83.1% (Fig. 4a, bottom row; Supplementary Fig. 29). The RMSD distribution quantitatively confirmed an improved match with the target shape (3.45 ± 0.35 nm). In addition to local flexibility, extreme geometric features, such as vertices with sharp turns, can further increase structural deviations from the target shape due to electrostatic repulsion (Supplementary Note 7). Although performed manually in this study, this revision process could be automated by integrating Generative SNUPI with an optimization algorithm, given its predictive capability.Fig. 4Experimental validation of the generated designs.a Example of the design-revision-validation process using the predicted RMSF values from Generative SNUPI. b Experimental validation of the representative designs inspired by (i) Mona Lisa painting, (ii) Admiral Yi statue (in Republic of Korea), (iii) One Pillar Pagoda (in Vietnam), (iv) dog side face, (v) design inspired by OpenAI logo (OpenAI is not affiliated with or does not endorse this work), (vi) hexa-chiral metastructure, and (vii) rotating triangle metastructure. Gel electrophoresis results are for ultrafiltrated samples after annealing in 12 mM MgCl2. This experiment was repeated independently twice with similar results. Scale bars, 20 nm for atomic models and simulation results; 100 nm for AFM images. c RMSD distributions. RMSD was calculated from n = 200 structures for each sample type. How samples were selected is described in “Statistcs & Reproducibility” in Methods. Source data are provided as a Source Data file. Uncropped gel images with molecular weight markers are provided in the Source Data. Molecular graphics were generated using UCSF Chimera66, developed by the Resource for Biocomputing, Visualization, and Informatics at the University of California, San Francisco, with support from NIH P41-GM103311. Experimental validation of the generated designs. a Example of the design-revision-validation process using the predicted RMSF values from Generative SNUPI. b Experimental validation of the representative designs inspired by (i) Mona Lisa painting, (ii) Admiral Yi statue (in Republic of Korea), (iii) One Pillar Pagoda (in Vietnam), (iv) dog side face, (v) design inspired by OpenAI logo (OpenAI is not affiliated with or does not endorse this work), (vi) hexa-chiral metastructure, and (vii) rotating triangle metastructure. Gel electrophoresis results are for ultrafiltrated samples after annealing in 12 mM MgCl2. This experiment was repeated independently twice with similar results. Scale bars, 20 nm for atomic models and simulation results; 100 nm for AFM images. c RMSD distributions. RMSD was calculated from n = 200 structures for each sample type. How samples were selected is described in “Statistcs & Reproducibility” in Methods. Source data are provided as a Source Data file. Uncropped gel images with molecular weight markers are provided in the Source Data. Molecular graphics were generated using UCSF Chimera66, developed by the Resource for Biocomputing, Visualization, and Informatics at the University of California, San Francisco, with support from NIH P41-GM103311. We further validated Generative SNUPI by using seven distinct shapes from the design library (Fig. 4b). Gel electrophoresis confirmed that all structures achieved a high monomer yield exceeding 80%. The first four designs (simplified line drawings of the Mona Lisa, the Admiral Yi statue in the Republic of Korea, the One Pillar Pagoda in Vietnam, and a dog’s side face) were well matched with their target shapes as predicted by simulations and confirmed by experiments (Fig. 4b ⅰ-ⅳ; Fig. 4c; Supplementary Figs. 30-33). In contrast, the design inspired by the OpenAI logo showed distorted shapes compared to its target as expected from high levels of local fluctuations predicted (Fig. 4b ⅴ; Supplementary Fig. 34), resulting in high RMSD values of 5.64 ± 0.95 nm (Fig. 4c). While the mean AFM image more closely resembled the target shape, it was due to self-correcting effects during the averaging process arising from its multi-axial symmetry. The last two designs, hexa-chiral and rotating triangle inspired by kirigami-based auxetic metastuctures exhibiting a negative Poisson’s ratio, showed RMSD values similar to those of the OpenAI logo-like design (Fig. 4b ⅵ-ⅶ; Supplementary Figs. 35-36). Note that it was not because of structural distortion due to insufficient edge connections but because of the open-close mechanism inherent to these structures, activated by electrostatic repulsion between edges around flexible junctions. While we could suppress this mechanism by making junctions stiffer, flexible junctions would be useful to design a dynamically reconfigurable structure in response to external stimuli. Design of transformable structures and modular assemblies High design flexibility and geometric fidelity offered by Generative SNUPI enable us to embed diverse functionalities into DNA origami through controlled reconfiguration. To demonstrate, we first constructed two auxetic metastructures with distinct open-close mechanisms: a rotating triangle design in which rigid triangular units rotate about junctions, and a re-entrant design, in which re-entrant edge angles change during shape transition (Figs. 5a, 5c and Supplementary Fig. 37). Access to both open and closed configurations requires precise and facile realization of an intermediate configuration that facilitates smooth shape transformation, as supported by a simplified energy model (Supplementary Note 8). Generative SNUPI is particularly well-suited for this purpose, as it can generate designs with high geometric fidelity from user-provided line-based drawings, allowing us to specify complex intermediate patterns that serve as a foundation for mechanical reconfiguration.Fig. 5Design of transformable structures and modular assemblies.a–d Rotating triangle (a, b) and re-entrant structures (c, d) in open and closed states. Open states formed through electrostatic repulsion, while closed states were induced by adding connectors. For each design, schematics for transformation, gel electrophoresis (O: open, C: closed, This experiment was repeated independently twice with similar results.), and representative AFM images of each state are shown (a, c, Scale bars, 100 nm), along with area distributions calculated from AFM images (b, d, n = 200, How samples were selected is described in “Statistcs & Reproducibility” in Methods.). e Schematics of modular assembly using dog face and body components. Shared interfacial edges (visualized with curvature combs) enable seamless connectivity across various modular combinations. An atomic model of the interface illustrates geometric compatibility. f Designs of two distinct body modules and representative AFM images. g, Modular assemblies combining face and body modules. Gel electrophoresis of dimerized samples (10 nM each component) is shown with band intensities. This experiment was repeated independently twice with similar results. Representative AFM images of assembled configurations are provided. Scale bars, 100 nm. Source data are provided as a Source Data file. Uncropped gel images with molecular weight markers are provided in the Source Data. Design of transformable structures and modular assemblies. a–d Rotating triangle (a, b) and re-entrant structures (c, d) in open and closed states. Open states formed through electrostatic repulsion, while closed states were induced by adding connectors. For each design, schematics for transformation, gel electrophoresis (O: open, C: closed, This experiment was repeated independently twice with similar results.), and representative AFM images of each state are shown (a, c, Scale bars, 100 nm), along with area distributions calculated from AFM images (b, d, n = 200, How samples were selected is described in “Statistcs & Reproducibility” in Methods.). e Schematics of modular assembly using dog face and body components. Shared interfacial edges (visualized with curvature combs) enable seamless connectivity across various modular combinations. An atomic model of the interface illustrates geometric compatibility. f Designs of two distinct body modules and representative AFM images. g, Modular assemblies combining face and body modules. Gel electrophoresis of dimerized samples (10 nM each component) is shown with band intensities. This experiment was repeated independently twice with similar results. Representative AFM images of assembled configurations are provided. Scale bars, 100 nm. Source data are provided as a Source Data file. Uncropped gel images with molecular weight markers are provided in the Source Data. To promote open-to-closed shape transitions, we introduced short single-stranded gaps63 at the junctions after shape generation (Supplementary Note 9). These manually modified gaps reduce local stiffness, thereby increasing the junction flexibility and preventing the formation of multimers. We successfully induce the open-to-closed transformations for both designs by using site-specific connectors64, which outperformed glue-overhang pairing20 in both precision and yield (Supplementary Note 10 and Supplementary Figs. 38 and 39). Band migration shifts in gels and AFM images confirmed substantial changes in the enclosed area—an unique property of auxetic metastructures—with mean area reduction of 34.9% and 47.3% for rotating triangle and re-entrant designs, respectively (Fig. 5b and 5d). These results highlight the capability of Generative SNUPI to embed reconfiguration mechanisms into DNA origami. Furthermore, Generative SNUPI can be used to design DNA origami structures for their modular assembly that requires precise shape matching at connecting interfaces. This geometric consistency at these interfaces is important for stable dimer formation of free-form structures with minimal structural strain. To illustrate, we designed a modular dog-shaped structure composed of interchangeable face and body modules (Fig. 5e). Each module was drawn with a shared curved edge at the interface, and converted into a corresponding DNA origami design using Generative SNUPI. The neck edge served as the interfacial boundary during dimerization of face and body modules, and the same curved line was used to define this edge in all input drawings. We constructed two face modules (Fig. 4a and Fig. 4b iv) and two body modules (Fig. 5f and Supplementary Figs. 40 and 41). The original designs were slightly modified to induce vertical dimer assembly while avoiding horizontal assembly, which would induce structural strain due to inner–outer edge mismatch (Supplementary Note 11). All four distinct modular combinations were successfully realized with relatively high dimerization yield exceeding 65% (Fig. 5g, Supplementary Figs. 42–45, Supplementary Note 12 and Supplementary Table 6). A smeared band above the dimer band indicates aggregation of the structures, consistent with AFM imaging. Together, these results support the feasibility of expanding the modular framework by introducing additional components designed with compatible interfacial geometry. Discussion Generative SNUPI is a diffusion-based generative design framework that extends the capabilities of DNA origami design beyond conventional rule-based methods. To address the lack of large, standardized training data in DNA nanostructure design, we constructed a robust training dataset using SNUPI to simulate equilibrium conformations of wireframe structures with 2HB edges. This platform enables the generation of physically realizable, geometrically complex free-form structures while allowing users to assess structural and mechanical properties during the design process, reducing the design-build-test cycle through rapid refinement prior to experimental validation. Our framework expands the design space to support free-form, transformable, and modular structures. To demonstrate this capability, we designed over 100 structures with diverse structural characteristics and experimentally validated selected cases, including artistic drawings, auxetic lattice patterns, and modular assemblies. These demonstrations highlight the framework’s capacity to produce structures with noncanonical geometries, programmable mechanical transformations, and scalable component-based architectures, which are challenging to achieve using previous design methods. While Generative SNUPI establishes a foundation for generative modeling of sophisticated DNA architectures, it is currently limited to producing wireframe-based structures with 2HB edges. However, the flexibility of the framework allows further extension to 6HB-edge systems, multilayer lattices, and sheet-like geometries as sufficient training datasets become available. While a substantial amount of 6HB wireframe structural data can be obtained from related studies25,58 combined with SNUPI simulations, multilayer lattice designs remain scarce due to the limited number of reports. Incorporating data augmentation strategies could help address this limitation by expanding the training dataset and enabling the generation of a broader range of structural variations. Integration with deep learning–based analysis tools such as DeepSNUPI41 is expected to further accelerate the design and validation of new structures by enabling rapid and accurate predictions of structural and functional behaviors. This extensibility will be valuable as the field progresses toward structurally and functionally diverse DNA assemblies. Further, Generative SNUPI streamlines the design process for complex structures that traditionally required extensive expert intervention. Its ability to produce editable, nucleotide-level outputs makes it compatible with downstream design and fabrication pipelines, facilitating integration with broader applications. Overall, this work highlights a shift in DNA nanostructure design toward data-driven generation, moving beyond rule-based methods through generative modeling. By enabling intuitive, automated, and scalable design of free-form DNA architectures, Generative SNUPI can expand the accessible design space and accelerate the development of functional nanostructures in fields ranging from molecular robotics to biomedical engineering. Methods Datasets We collected 450 wireframe DNA origami structures sourced from ATHENA24 (Supplementary Notes 3). To obtain the configurations in equilibrium, we employed the multiscale computational model for structured DNA assemblies, SNUPI48–51. The solution shape was iteratively determined using thirty incremental steps through the Newton–Raphson method, until the tolerance values for the internal force and total energy were reduced below 10−3 and 10−6, respectively. The data collection was performed using MATLAB (R2022b), Python (3.9), PyTorch (1.13.1), and SNUPI (2.01). Generative diffusion model for DNA origami structure Technical details of the DDPM are described in Supplementary Notes 1 and 4. All computational analyses were performed using MATLAB (R2022b), Python (3.9), PyTorch (1.13.1), PyTorch Geometric (2.2.0) and DGL (1.1.2). Simulation of DNA origami designs from Generative SNUPI To verify the quality of each DNA-origami structure obtained using Generative SNUPI in the design gallery, we first calculated an equilibrium configuration with SNUPI-based simulations, which served as the reference for RMSD analysis. Starting from this relaxed state, we carried out one-microsecond stochastic dynamics simulations for each design. Simulations were performed with a 1 ps integration timestep over one million steps, with atomic coordinates recorded every 1000 steps. The resulting trajectories yielded well-converged RMSF for every base-pair in the structure, providing an efficient means of evaluating the mechanical stability of the generative designs. Synthesis of DNA origami structures DNA folding was performed using a mixture containing 5 nM of single-stranded scaffold DNA (M13mp18 7249-nt length, Guild Biosciences) and 100 nM of each staple strand (Standard HAP purification, Bionics) in 1× TAE buffer solution (40 mM Tris-acetate and 1 mM EDTA, T9650, Sigma-Aldrich) supplemented with 12 mM MgCl2 (M1028, Sigma-Aldrich). The solution was thermally annealed in a thermal cycler (T100, Bio-Rad), starting with incubation at 80 °C for 15 min to reduce undesired hybridization, followed by gradual cooling from 80 °C to 65 °C at −1 °C/4 min, and subsequently from 65 °C to 25 °C at −1 °C/h. After annealing, samples for purification were stored at 4 °C prior to purification, while samples intended for use without purification were stored at −20 °C until further experiments. Oligonucleotide sequences for each structure are provided in Supplementary Data 1. Sample purification DNA origami structures were purified by ultrafiltration using a 50 kDa MWCO centrifugal filter (UFC505096, Amicon Ultra-0.5, Millipore). The filter was first pre-rinsed with 500 µl of 12 mM MgCl2 by centrifugation at 5000 × g for 8 min at 20 °C. Subsequently, 100 µl of the sample was loaded into the filter and diluted with 400 µl of 12 mM MgCl2, followed by centrifugation under the same conditions. Buffer exchange was repeated four times with 500 µl of 12 mM MgCl2 to remove excess staple strands. The purified sample was then recovered by inverting the filter into a clean microcentrifuge tube and centrifuging at 10,000 × g for 3 min at 20 °C. Final concentrations were determined by absorbance at 260 nm using NanoDrop One (Thermo Scientific). Gel electrophoresis 1.5% agarose gels (C-9100-1, Bioneer) were prepared in a running buffer containing 0.5× TBE (45 mM Tris-borate and 1 mM EDTA, T6400, Sigma-Aldrich) supplemented with 12 mM MgCl2 and 0.5 ug/ml EtBr (17896, Thermofisher Scientific). Electrophoresis was conducted for 90 min at 7.5 V/cm in an ice-water cooled chamber using a Wide Mini-Sub Cell GT horizontal electrophoresis system (Bio-Rad). Gel images were scanned with the GelDoc XR+ system and quantified using Image Lab v 6.1 software to obtain the relative front (normalized to 1.5 kb band in ladder) and band intensities. A 1 kb DNA ladder (N3232S, New England Biolabs) was included as a reference in all runs. Corresponding uncropped gel images with molecular weight markers are provided in the Source Data. AFM imaging procedure The samples were diluted in folding buffer (1× TAE, 10 mM MgCl2) to a final concentration of 1.0–1.5 ng/µl. A 20 µl aliquot was deposited onto freshly cleaved mica (Ted-Pella Inc.) and incubated for 5 min at room temperature, followed by rinsing with deionized water and drying with an air blower. Imaging was performed in the dry-state using an NX10 AFM (Park Systems) equipped with a PPP-NCHR probe. Scans were collected at 1 Hz over multiple regions with non-contact mode. The acquired images were processed using XEI 4.1.0 software. Individual structures were cropped using custom Python code based on OpenCV. The mean AFM image and RMSD were calculated by aligning each structure to the target pattern using the iterative closest point (ICP) algorithm (Supplementary Note 6). Modular assembly For modular assembly, each monomer was folded with 5 nM of scaffold and 100 nM of staples in 1× TAE/12 mM MgCl2 buffer with cohesive overhangs. The purification was performed via ultrafiltration as described above, except that 400 µl of samples and 100 μl of 12 mM MgCl2 loaded in the first loading and exchanging buffer nine times with 500 µl of 12 mM MgCl2. Face and corresponding body modules were mixed at equimolar concentrations from 5 to 15 nM in 12 mM MgCl2 and annealed from 37 °C to 20 °C at −1 °C/30 min ( ~ 9 hours). The dimer yield was calculated as described in Supplementary Note 11. Statistics and reproducibility No statistical method was used to predetermine sample size. Gel electrophoresis experiments were repeated independently twice with similar results, and representative images are shown. AFM imaging was performed independently at least three times with similar results, and representative images are shown. The number of samples (n = 200) used for RMSD and area analysis was chosen to ensure sufficient sampling of structural variability. For sample selection, contour lines were extracted from AFM images as described in Supplementary Note 6. The contour length of each imaged structure was estimated and compared with the corresponding target shape. Since slender 2HB structures can undergo deformation during random deposition on mica, only well-deposited monomers with contour lengths exceeding 80% of the target contour length were selected for analysis. The experiments were not randomized, and the investigators were not blinded to allocation during experiments and outcome assessment. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Description of Additional Supplementary Files Supplementary Data 1 Reporting Summary Transparent Peer Review file Supplementary Information Description of Additional Supplementary Files Supplementary Data 1 Reporting Summary Transparent Peer Review file Source data Source Data Source Data