Programmable self-assembly of three-dimensional nanostructures from 10,000 unique components.

Nucleic acids (DNA and RNA) are widely used to construct nanometre-scale structures with ever increasing complexity, with possible application in fields such as structural biology, biophysics, synthetic biology and photonics. The nanostructures are formed through one-pot self-assembly, with early kilodalton-scale examples containing typically tens of unique DNA strands. The introduction of DNA origami, which uses many staple strands to fold one long scaffold strand into a desired structure, has provided access to megadalton-scale nanostructures that contain hundreds of unique DNA strands. Even larger DNA origami structures are possible, but manufacturing and manipulating an increasingly long scaffold strand remains a challenge. An alternative and more readily scalable approach involves the assembly of DNA bricks, which each consist of four short binding domains arranged so that the bricks can interlock. This approach does not require a scaffold; instead, the short DNA brick strands self-assemble according to specific inter-brick interactions. First-generation bricks used to create three-dimensional structures are 32 nucleotides long, consisting of four eight-nucleotide binding domains. Protocols have been designed to direct the assembly of hundreds of distinct bricks into well formed structures, but attempts to create larger structures have encountered practical challenges and had limited success. Here we show that DNA bricks with longer, 13-nucleotide binding domains make it possible to self-assemble 0.1-1-gigadalton, three-dimensional nanostructures from tens of thousands of unique components, including a 0.5-gigadalton cuboid containing about 30,000 unique bricks and a 1-gigadalton rotationally symmetric tetramer. We also assembled a cuboid that contains around 10,000 bricks and about 20,000 uniquely addressable, 13-base-pair 'voxels' that serves as a molecular canvas for three-dimensional sculpting. Complex, user-prescribed, three-dimensional cavities can be produced within this molecular canvas, enabling the creation of shapes such as letters, a helicoid and a teddy bear. We anticipate that with further optimization of structure design, strand synthesis and assembly procedure even larger structures could be accessible, which could be useful for applications such as positioning functional components.

Without altering the fundamental design principle of the original 32-nt DNA bricks, we empirically optimized domain dimensions to generate the new 52-nt DNA bricks that allow the self-assembly of 0.1 – 1 gigadalton structures from 104 bricks (Fig. 1a, b). We investigated structure formation yields by tuning the brick lengths to 52 nt (four 13-nt domains) or 74 nt (two 18-nt and two 19-nt domains), such that the inter-brick binding pattern remains perpendicular. For example, two neighboring 52-nt DNA bricks form a 13-basepair (bp) duplex that corresponds to a 90° inter-brick angle. Comparing 6 helices (H)×6H×8x base pairs (B), where×= {8, 13, 18.5}, we find that cuboids assembled from 52-nt bricks show substantially higher formation yields in both 72-hour thermal annealing (Supplementary Figs. 4, 6, 7, and 8) and isothermal annealing reactions (Supplementary Figs. 5, 11, Supplementary Table 2) than the other two brick structures. Direct comparison of 52-nt brick structures and 32-nt brick structures with similar overall dimensions also revealed that 52-nt brick structures assemble with higher yield and thermal stability (Supplementary Figs. 9 – 12).

Figure 1

Three dimensional nanostructures self-assembled from DNA bricks

a, Comparison of 3D DNA origami (~200 components, ~5 megadalton),[4] and DNA brick nanostructures assembled here (~30,000 unique components, ~500 megadalton). b, Detailed helical (top) and brick (bottom) models of incorporated 52-nt DNA brick strands. c, A ~150 MDa DNA brick cuboid (left) as a molecular canvas (middle) composed of ~20,000 13-bp voxels (right). Scale bar in a and c measures 100 nm. d, A 3D teddy bear rendering (left) can be approximated using the ~20,000 voxel canvas (middle) to form the cavity of a cuboid structure (right).

Given the importance of annealing conditions, we tested the influence of a number of factors including salinity, temperature ramps, and reaction times on the folding of a 20H×20H×260B structure at a 5 nM strand concentration to obtain an optimal protocol. This 67.6 MDa cuboid had the highest gel yields of ~6% after annealing in 20 mM MgCl2 isothermally at 51.4°C, or when using a 1.5°C narrow annealing ramp (52.5 °C to 51°C) over 5 – 7 days (Supplementary Figs. 16, 17).

Scalability is demonstrated by assembling the cuboids 10H×10H×156B, 14H×14H×208B, 20H×20H×260B, 30H×30H×260B, 36H×36H×312B, 40H×40H×338B, 46H×46H×390B, which range in size from 10.1 to 536 MDa and were annealed isothermally in one-pot reactions with 20 mM MgCl2 (Fig. 2a, in grey). Using an 8H×8H×104B 4.3 MDa origami structure as benchmark (Fig. 2a, b, in blue; Supplementary Figs. 19 – 22), gel electrophoresis analysis indicates 1 – 20% formation yields that depend on the size of the structure and the strand concentration (Fig. 2a, Extended data fig. 1). Each structure has an optimal formation temperature range that tends to narrow as the complexity of the structure increases (Supplementary Fig. 18), suggesting that increased sequence diversity and larger number of components may limit effective nucleation and growth to a smaller window of reaction conditions. Transmission Electron Microscopy (TEM) revealed complete structures with expected dimensions and morphologies using purified samples (Fig. 2, Supplementary Figs. 23 – 46), along with some defective structures (Supplementary Fig. 36) that may reflect incomplete assembly or post-assembly damage during gel purification or TEM sample preparation.

Figure 2

Self-assembly of brick cuboids

Cylindrical models of DNA brick cuboids (gray) and an M13-scaffolded DNA origami cuboid (blue). a, Comparison of gel yields (top numbers), TEM images of helical end view (middle), and lateral projection (bottom) of DNA cuboids. b, Model (top) and TEM image (bottom) of 536 MDa brick cuboids and 4.3 MDa origami cuboids. 1.05 gigadalton cuboid model (c), selected TEM helical (top) and lateral (bottom) images (d), and wide-field TEM images (e). f, 3D DNA-PAINT super-resolution images of the 152 MDa canvas structure: a wide-field view (top) and different projections of a single representative cuboid (bottom). Color bars indicate height along the z-axis. All scale bars measure 100 nm.

Extended Data Figure 1

Gel electrophoresis analysis of DNA brick cuboids

Structures were assembled isothermally for 5 – 7 days at the temperatures indicated above the lane. The label below a band of interest indicates the gel yield.

The 46H×46H×390B cuboid, with a size of 536.4 MDa that is over 100 times that of an M13 scaffolded DNA origami[4], is the largest assembled structure composed of entirely unique components (Fig. 2b, Supplementary Fig. 43). It measures over 100 nm in each dimension, contains over 30,000 unique components (33,511 strands) with ~1.7 million nucleotides, and forms with over 1% gel yield. Due to the symmetry present in DNA brick structures, discrete multimer structures can be created by connecting strands across different symmetric planes[17] (Supplementary Figs. 47 – 58). We applied a side-to-side tetramer design to assemble a 1 gigadalton tessellation structure, which measures 72H×72H×312B and contains four identical 262.8 MDa monomeric units (see Supplementary Figs. 59–61 for design details). This was implemented by utilizing the C4 symmetry[17] present in the plane perpendicular to the DNA helical axis, with strands designed to connect one face of the structure, parallel to the helical axis, to an adjacent face of the same orientation to give a rotationally symmetric tetramer (Fig. 2c–e, Supplementary Figs. 62 – 64). This gigadalton structure also forms through a simple one-pot isothermal annealing reaction with ~1% gel yield, with TEM confirming that its morphology is as designed (Extended data fig. 1h, Supplementary Figs. 62 – 64, Supplementary Figs. 62, 63). A defect seen in the center of some particles seems likely due to the putative strain accumulated at the center of the tetramer.

The high component complexity of these cuboids also enables them to be used as programmable “molecular canvases” for complex shape patterning. As a demonstration, we selected the 30H×30H×260B cuboid, which is assembled from 9,700 unique bricks, measures 152 megadalton, and offers 18,000 voxels at a resolution of 13 bp per voxel (Fig. 1c, d). TEM imaging of this cuboid showed that 90% of the particles exhibited expected morphology with no severe distortions (Supplementary Fig. 83), and 3D DNA-PAINT super-resolution imaging[11,18] further confirmed the expected dimensions for the particles in solution and revealed that all eight corners of most structures are intact (Fig. 2f model, Supplementary Figs. 65, 66).

To facilitate user-friendly design of large 3D brick structures containing the order of 104 components, we developed a software tool called Nanobricks. First, the user draws, imports, or programs (e.g. via mathematical scripting) a 3D shape by placing “voxels” that represent DNA strand domains. The software then converts the shape into associated DNA brick strands. Finally, the software outputs sequences by generating new or applying an existing set of sequences to the strands (Fig. 3a). The software contains features to add, remove, or modify on the voxel or strand level for each of three steps (Fig. 3a, Supplementary Figs. 67 – 74), and can output file formats compatible with other commonly used DNA structure design and analysis tools[19] (see Supplementary Section S8.4).

Figure 3

Cavity shapes formed from a 30H×30H×260B molecular canvas

a, Design software for complex DNA brick structures. Desired shapes can be designed by editing voxels through a 3D interface (top), translated automatically to strands (middle), and assigned sequences (bottom). (b, c) Cavity shapes can be generated by selecting or excluding (right) voxels to approximate 3D rendering files (b) or to satisfy mathematical equations (c). (d -n) Diverse cavity shapes. For each design, the top diagram depicts a 3D model of the designed shape. Expected projections (top or left) and averaged TEM images (bottom or right) are also shown. Individual particles used in averaged images are depicted in Supplementary Table 3 and Supplementary Figs. 77 – 82. Scale bars measure 100 nm.

We used Nanobricks to design 13 distinct complex cavity shapes from the 30H×30H×260B canvas (Fig. 3, Extended Data Fig. 2, Supplementary Figs. 75 – 83, and Supplementary Table 3). The shapes were designed using shape importing, mathematical scripting, manual designing or a combination of these methods. Nanobrick’s user-friendly 3D visualization and editing interface allowed for easy manipulation of the 18,000 voxels of the molecular canvas (Supplementary Figs. 67 – 74). To determine the minimal feature size, we patterned the surface of a hollow cuboid with varying pore sizes and found that a minimum of 4 helices between separated design features was needed for the structure to form completely (Extended data fig. 2b). Implementing these restrictions, we used the software to convert several open-source 3D designs to voxelized approximations (Figs. 3b, e, f, and Supplementary Fig. 76). Scripting capabilities allowed for design of complex mathematical cavities, including a helicoid, Möbius strip, hyperboloid, and cone, by identifying whether voxels were located within a given mathematical formula (Figs. 3c, g–i and Supplementary Section S9.4). Manual designs include a structure featuring the projections of “G”, “E”, and “B” along three axes (fig. 3d), one exhibiting the word “LOVE” in one single projection (Figs. 3e), one containing two interconnected loop cavities (Fig. 3j), one with a cavity that threads through itself (Fig. 3k), and other complex shapes (Fig. 3l, m, n).

Extended Data Figure 2

Characterization of 30H×30H×260B cavity shapes

a, Schematic depicting the 30H×30H×260B molecular canvas in gray compared with a DNA origami-sized structure in blue. b, For each structure, the top panels show the 3D models of the designed structure. The bottom left panels shows expected TEM projections. The bottom right panels shows TEM averages from at least six particles. c, The structures were folded with 5 nM/strand by annealing isothermally or using a narrow ramp from 52.5 to 51°C. Products were analyzed on a 0.5% agarose gel in the presence of 10 mM MgCl2. The percentage number listed below a target band indicates the gel yield. Lane labels correspond to those in Fig. 3 and in (b).

No “protector strands”[9] were used within the cavities (Supplementary Fig. 75). Surprisingly, these structures showed strong tolerance to the presence of a large number of exposed “sticky” single-stranded ends inside the cavities, and assembled at yield between 1.5 and 5.1% (Extended data fig. 2c). TEM characterization of the different shapes further showed that approximately 75% of the structures were intact and displayed the expected internal cavities (Supplementary Fig. 83).

Complex structural features were also analyzed in detail by using electron tomography (Fig. 4). We first performed 3D reconstruction on a 30H×30H×260B cuboid with sixteen-parallel 2H×2H×260B crossing channels (Fig. 4a, b). The reconstructions and the 3D visualization using mesh surface representation revealed the 3D channels network in the cuboid. The global topology of the reconstructed density is in agreement with the expected architecture of the object and showed typical shape artifacts at the extreme top and bottom of the particles in the direction of the electron beam due to the missing wedge.[20] We then performed electron tomography on four distinct cavity structures: teddy bear, bunny, helicoid, and “GEB” (Fig. 4a–c, e, Supplementary Figs. 86 – 93, and Supplementary Movies). Tilt-series images were collected for each of the three projection views to validate the fine 3D features. Thin features containing only few voxels, such as the teddy bear’s snout and limbs (red arrows in Fig. 4e) or the bunny’s ears (Supplementary Figs. 91 – 93), were confirmed through reconstructions.

Figure 4

Electron tomography analysis and computational 3D reconstruction of DNA brick structures

a, 3D model of a cuboid containing parallel channels, with extracted slices from the tomogram (right and bottom). b, 3D model of the cuboid in a showing the positions of two orthogonal slices (left), and the corresponding 3D mesh-rendered view of their tomographic reconstructions (right). (c -e), 3D model (left), expected shape projections (middle), and slices extracted from tomograms (right) for the teddy bear (c), helicoid (d), and “GEB” (e) structures. Red arrows point to thin but visible features. Numbers in images correspond to slice position extracted from each tomogram (see Supplementary Figs. 84 – 93 and Supplementary Movies for more details). All scale bars measure 50 nm.

To quantitatively examine the incorporation of each of the 104 bricks into the structure, we applied a DNA sequencing-based analysis[21] on the teddy bear structure. The assembled structure was gel purified and heat denatured. The resulting DNA strands were ligated with sequencing primers, amplified, sequenced, and compared with a sample of unreacted strands[21](see Supplementary Sections S11.1 and S11.2). Strands with a sequencing read number below a specific threshold are designated as low abundance. By applying this threshold-based analysis to all strands of the molecular canvas, we can extract information about the abundance of each strand in the formed product and thus the average voxel composition of the formed teddy-bear cavity structure (Fig. 5, Supplementary Figs. 104 – 112). Such analysis revealed that the majority (>98%) of the strands forming the teddy bear structure were present in high abundance according to our sequencing analysis (Supplementary Fig. 105), and only a small number of sparsely distributed voxels exhibited undesired low abundance (red voxels in Fig. 5a and Supplementary Figs. 105 – 106). Projections of the low abundance strands data along the different axes matched the expected projections of the design (Fig. 5b, Supplementary Figs. 111, 112). By normalizing the data, we observed a “hot spot” of low-abundance strands at the back of the teddy bear, which is consistent with some broken particles observed by TEM (Supplementary Fig. 112). This structural defect could potentially be caused by the presence of only a few crossovers at this tenuous spot.

Figure 5

DNA sequencing analysis of the teddy bear cavity structure

a, 3D model (left) and 3D representation of sequencing results (right) of the teddy bear design. Gray and red colors correspond to intended (in cavity) and unintended (in structure) low abundance species, respectively. Opacity of voxels corresponds to the number of strands for which a criterion applies: completely opaque – two, partially transparent – one. Voxels formed by two well incorporated strands are not depicted. b, Schematic 2D representations (top) and respective 2D plots of the fractions of low abundance strands along a given axis (bottom).

The successful construction of large and complex structures appears the result of the 52-nt bricks being able to mitigate the slow assembly kinetics that inevitably arises with the decreased component concentration encountered when assembling large DNA structures from a massive number of distinct components. While the detailed mechanism of brick structure formation remains to be explored, our results are consistent with the hypothesis that assembly involves delayed nucleation followed by fast growth.[9,28] In our case, we find that domain lengthening from 8-nt to 13-nt results in structures forming more rapidly. We also note that binding heterogeneity has been found to circumvent the emergence of multiple dominant competing nuclei[28], implying that component heterogeneity is further enhanced in our 52-nt brick design because the range of accessible binding energies becomes wider with longer domains due to the larger sequence space.

The 0.5 GDa structure we have constructed contains 33,511 unique components, 1,684,336 nt of sequence, and bridges two orders of magnitude in length in all three dimensions in a space filling fashion: from a feature resolution of 2.8×2.8×4.4 nm3 to assembled structures with an 100 nm length in each of the three dimensions. Although the present work focuses on constructing compact, spacing filling structures with 104 unique components packed in 1 attolitre (i.e. 100×100×100 nm3 = 10−21m3) space, it should also be feasible to use variations of the bricks to construct wire-frame or porous structures[10,11,13,14] with similar component complexity. Considering that the 10-fold leap in component complexity afforded by DNA origami opened the door for using DNA nanostructures in fields as diverse as single molecule biophysics,[22] structural biology,[23] synthetic biology,[24] nanofabrication,[25,26] and photonics[27], we anticipate that the 100-fold increase in complexity afforded by our DNA brick method will also enable new uses for DNA nanostructures (e.g. as scaffolds for patterning complex inorganic nanostructures[25] or for 3D positioning of diverse functional moieties[26,27]). In fact, even large DNA brick assemblies might be possible, considering that the high cost of purchasing a large number of synthetic DNA strands has restricted our testing to ~30,000 distinct bricks and that low-cost methods for synthesizing DNA strands (e.g. chip-synthesized DNA followed by parallel enzymatic amplification[29]) are available. Further scaling-up of assembly size could also be achieved by hierarchical methods via sticky end association or shape complementarity.[12,30]

Methods

Condensed descriptions of methods are described below. See the Methods section in the Supplementary Information for specific details.

Design and formation of structures

Structures were designed using our Nanobricks software. Depicted 2D strand diagrams were generated from associating caDNAno files.[19] Structures were annealed in 0.5×TE buffer (5 mM Tris, 1 mM EDTA, pH 8) containing 20 mM MgCl2 using either an isothermal hold[31] or a narrow annealing ramp. See Supplementary Methods and Supplementary Table 1 for the detailed annealing conditions and optimal temperatures. See supplementary information for sequences used for each structure.

Agarose gel electrophoresis

Samples were analyzed using 0.3–2% agarose gel electrophoresis and stained using SYBR Safe loading dye. Gels were visualized using the Typhoon FLA 9000 gel imager and quantified using ImageJ[32] or TotalLabQuant v12.2 (Cleaver Scientific, ltd).

Transmission electron microscopy imaging

Samples were deposited on glow-discharged formvar/carbon coated grids from Electron Microscopy Sciences. Samples were stained for 60 seconds with 2% uranyl formate solution containing 25 mM NaOH and imaged using a JEOL JEM-1400 TEM operated at 80 kV.

Electron tomography and image processing

Samples were deposited on glow-discharged, carbon-coated 300 mesh copper grids and stained using 1% uranyl acetate solution. The grids were then transferred into a JEOL 2200FS FEG transmission microscope using the JEOL high tilt holder. Series of tilted images were collected at a magnification of 50,000 folds by using a 4k × 4k slow-scan CCD camera (Gatan, inc.) with defocus values of −3 µm and −5 µm. The acquisition was performed semi-automatically using the Serial EM software package. Samples were tilted between −60° and 60° with 2° increment steps. For a detailed description of the alignment and reconstruction procedure see the Supporting Information.

3D DNA-PAINT super-resolution setup

Fluorescence imaging was performed using an inverted Nikon Eclipse Ti-E microscope (Nikon Instruments, Melville, NY) with the Perfect Focus System, applying an objective-type TIRF configuration with an oil-immersion objective (CFI Apo TIRF 100×, NA 1.49, Oil). 3D images were acquired using a cylindrical lens (FL = 1m) in the detection path.

Super-resolution DNA-PAINT images were reconstructed using spot-finding and 2D-Gaussian fitting algorithms programmed in LabVIEW.[18] A previously published calibration function[33] was used for 3D calibration. Drift correction was performed on the DNA structures, as previously described.[34]

Z-calibration was additionally corrected for refractive-index-mismatch by measuring a reference structure with given height, resulting in a correction factor of 1.3.[11] ViSP[35] was used to visualize single-particle localizations in three dimensions. After exporting from ViSP, images and corresponding color bars were contrast-adjusted using Fiji.[36] See the Supplementary Methods for additional details on sample preparation and image analysis.

Sequencing sample preparation and analysis

Sequencing analysis was prepared following a modified version of the barcode extension for analysis and reconstruction of structures (BEARS) protocol.[21] Samples were ligated to an adaptor sequence on the 5’ end using T4 RNA ligase 1 (New England Biolabs) and purified using polyacrylamide gel electrophoresis and electroelution. Subsequently, the 3’ end of the strands was ligated to a previously tested adaptor sequence[21] containing an integrated barcode. Then samples were amplified using Q5 polymerase.

Multiple samples with different barcodes were pooled and sequenced with an Illumina MiSeq machine according to the manufacturer’s instructions by using the MiSeq V2 paired end 50 kit (Illumina Inc., San Diego, CA). A modified library denaturation and loading protocol for lower concentration libraries was used.[37]

Data availability

The authors declare that the main data supporting the findings of this study are available within the paper and its Supplementary Information files. Sequences used to form the large structures are provided as well. Structure designs and software are posted at https://yin.hms.harvard.edu/bricks/try/#. All other data supporting the findings of this study are available from the corresponding authors on request.