Main

In our search for dodecahedral silica-cage structures (Fig. 1), we began with the early stages in the self-assembly of silica structures directed by surfactant micelles14. Our synthesis system contained cetyltrimethylammonium bromide (CTAB) surfactant micelles and tetramethyl orthosilicate (TMOS) as a sol–gel silica precursor (see Methods). We added hydrophobic mesitylene (TMB) into the aqueous CTAB micelle solution, increasing the size and deformability of the micelles15. We selected TMOS as the silica source because of its fast hydrolysis rate in water, and we adjusted the initial reaction pH to about 8.5. When TMOS was added, its hydrolysis to silicic acid reduced the reaction pH to neutral16,17. The lowered pH accelerated silane condensation, leading to primary silica clusters18 with diameters of around 2 nm. The negatively charged silica clusters were attracted to the positively charged CTAB micelle surface, assembling into micelle-templated nanostructures14. This experimental design—in which fast hydrolysis and condensation of the silica precursor quickly terminated the reaction—allowed the early stages in the micelle-directed self-assembly of silica structures to be preserved17.

Fig. 1: Representations of a dodecahedron.
figure 1

Left, among the platonic solids, the dodecahedron best fills out its circumscribed sphere—that is, a sphere that passes through all of its vertices. Right, the inscribed sphere that passes through all of the dodecahedron’s facets is shown for comparison.

Full size image

In order to improve the dispersity of particles on transmission electron microscopy (TEM) grids, we added low-molar-mass, silane-modified monofunctional polyethylene glycol (PEG) into the solution one day before preparing TEM samples. This process covalently coated the accessible silica surface19, yielding PEGylated nanoparticles (Extended Data Fig. 1) that could be further purified and isolated from the synthesis solution. Using TEM, we observed particles of narrow size distribution, with an average diameter of around 12 nm (Fig. 2a, including inset), consistent with silica structures wrapped around TMB-swollen CTAB micelles15. The detailed particle structure was difficult to identify, however. Therefore, we plasma-etched TEM samples on carbon grids for five seconds before imaging in order to remove excess organic chemicals (such as PEG-silane) that would otherwise contribute to background noise. To further improve the signal-to-noise ratio, we acquired and averaged a series of images of the same sample area. Stripes and windows in zoomed-in images of individual particles became more clearly recognizable, suggesting the presence of cage-like structures (Fig. 2b, including insets).

Fig. 2: TEM and cryo-EM characterizations of silicages.
figure 2

a, Low-magnification TEM images of PEG-coated silicages on a carbon substrate. The inset is a zoomed-in image. b, Averaged TEM image obtained using 11 images of the same sample area of PEG-coated silicages; insets show representative individual structures at a higher magnification. The sample was plasma etched for five seconds before TEM characterization to reduce background noise. c, Comparison of silicages observed by TEM and cryo-EM with projections of simulated dodecahedral cages and models. d, Cryo-EM images of silicages without PEG coating. Scale bars in the insets of b–d represent 5 nm.

Full size image

Analysis of thousands of such single-particle TEM images revealed the prevalence of two cage projections, one with two-fold symmetry and one with five-fold symmetry (Fig. 2c)—too few to allow a successful 3D reconstruction. We therefore shifted our attention to cryo-electron microscopy (cryo-EM) characterization of the native reaction solution. We omitted the silica-surface PEGylation step, as the high PEG concentration substantially increased the sensitivity of the samples to radiation, resulting in difficulties in obtaining clear cryo-EM images.

Cryo-EM provided direct visualization of particles in solution with arbitrary orientation—that is, without disturbances resulting from drying of the samples on TEM substrates, including structure deflation. The background noise was much reduced because of the absence of a TEM substrate and of chemicals dried onto the substrate during sample preparation (Fig. 2c). Although particle aggregation was observed occasionally through cryo-EM (Fig. 2d), individual silica nanoparticles with cage-like structures could always be identified (Fig. 2c, d). No particle aggregation was observed in dry-state TEM of PEGylated particles, suggesting that particle aggregation observed by cryo-EM was a reversible process that could be overcome via insertion of PEG chains.

We manually identified around 19,000 single-particle images from cryo-EM micrographs, then clustered them and averaged the images in each cluster in order to improve the signal-to-noise ratio20. The averages showed different orientations of silica nanoparticles with cage-like structures (silicages; Extended Data Fig. 2a). We identified averages that were consistent with selected projections of a pentagonal dodecahedral cage (Fig. 2c and Extended Data Fig. 2b). The dodecahedral silicage (icosahedral point group, Ih; Fig. 1) is the simplest of a set of Voronoi polyhedra suggested to form the smallest structural units of multiple forms of mesoporous silica21. Although such highly symmetrical ultrasmall silica cages have, to our knowledge, never been isolated before, it had seemed likely that this should be possible.

Guided by this structural insight, we carried out single-particle 3D reconstructions of silicages using the ‘Hetero’22 model-based maximum likelihood machine-learning algorithm, in which a two-class reconstruction is computed to overcome challenges associated with structural heterogeneity (see Methods) and rotational icosahedral symmetry is imposed on both classes (Extended Data Fig. 2). One of the two-class reconstructions was a dodecahedral cage, as visualized by UCSF Chimera (ref. 23; Fig. 3a, b). We identified a low-intensity signal inside this reconstructed cage, consistent with the presence of TMB-swollen CTAB micelles within the silicage, whose electron density is lower than that of silica but higher than that of the surrounding ice. The other class of reconstruction did not provide an interpretable structure, probably owing to heterogeneity in the structure of the corresponding particles. We carried out such two-class reconstructions using different numbers of single-particle images (2,000, 7,000 and 10,000) and found consistent results. We further performed single-class reconstructions with the Hetero algorithm, using only images from the class that showed dodecahedral cages in two-class reconstructions. Equivalent two-class and single-class reconstructions were also performed with the widely used RELION 2.1 system24,25. Dodecahedral cage structures were obtained in all of these reconstructions (Extended Data Fig. 3 and Supplementary Videos 1, 2). The resolution of the reconstructions was approximately 2 nm (ref. 26; Extended Data Fig. 4). Silica in these cages is amorphous at the atomic level, which prevented atomic resolution of these reconstructions.

Fig. 3: Single-particle reconstruction of the dodecahedral silicage.
figure 3

a, b, Reconstruction of the dodecahedral silicage (a), and its three most unique projections along the two-fold, three-fold and five-fold symmetry axes (b). The average dimensions of silicages in b were estimated from the reconstructed dodecahedral silicage. c, Representative comparison of nine unique projections from the reconstruction and cryo-EM cluster averages with projections of a 3D dodecahedral cage model. Corresponding single cryo-EM images are shown at the bottom, highlighting the difference between the raw data and the reconstruction. The scale bar in c represents 10 nm. The visualizations in a and b were made using UCSF Chimera23.

Full size image

The Hetero reconstruction algorithm provided estimates of the projected orientation (that is, three Euler angles) for each experimental image, which we used to compute predicted projections. Nine predicted projections and corresponding experimental images were manually clustered, and averages were computed for each cluster (Fig. 3c). The similarity of the projections of the 3D reconstruction and the averaged experimental images supports the dodecahedral cage structure. Furthermore, we calculated the theoretical probabilities of finding each of the nine projections (Fig. 3c) on the basis of the assumption that the orientations of silicages in cryo-EM are random. We then compared the results with the probabilities observed by single-particle 3D reconstruction (Extended Data Fig. 5). The high consistency between the theoretical and experimental projection probabilities further supports the dodecahedral cage reconstruction.

At this early point we can only speculate about the exact mechanism by which the observed silicage structure is formed, but there are clues in the details of the reconstruction. The silica occupying the vertices of the dodecahedral cage has a diameter of around 2.4 nm (Fig. 3b)—only slightly larger than the diameter of primary silica clusters, being approximately 2 nm (ref. 18; Extended Data Fig. 6). The interstitial spacing between two nearby vertices is estimated to be about 1.4 nm (this is the length of an edge, 3.8 nm, minus the diameter of the vertices, 2.4 nm; see Fig. 3b), which is much less than the diameter of such clusters. Bridges between vertices forming the edges of the dodecahedron are substantially thinner than the size of the primary clusters (Fig. 3a, b). This suggests that negatively charged primary silica clusters formed in solution may start to descend onto the positively charged micelle surface, attracted by Coulomb interactions. As more and more silica clusters assemble on the micelle surface, their repulsive interactions and possible interactions with other micelles may cause them to move to the vertices of a dodecahedron. Further silane condensation onto the surface of growing clusters may lead eventually to bridge formation, resulting in the final observed cage structure (Fig. 3). The origin of icosahedral symmetry in viruses has been associated with the energy minimization of two opposing interactions: repulsive interactions associated with bending rigidity, and attractive hydrophobic interactions27. In a similar way, in addition to electrostatic interactions, deformation of the micelle surface around the silica clusters may be another important contributor to the free energy in our system. This idea is supported by experiments showing that the cage structures do not form in the absence of TMB (Extended Data Fig. 7), which is expected to enhance the deformability of the micelle surface.

We have found that the self-assembly of ultrasmall cage structures directed by surfactant micelles can also occur when using other inorganic materials that have similarly sized features and similar surface chemistry to silica. In preliminary experiments, we replaced silica with one of two metals, gold or silver, or a transition metal oxide, vanadium oxide. Gold and silver structures were prepared by the reduction of metal precursors—gold chloride trihydrate and silver nitrate, respectively—in the presence of the micelles (Extended Data Fig. 8; see Methods). We used tetrakis(hydroxymethyl)phosphonium chloride (THPC) as both the reductant and the capping agent to stabilize primary gold and silver nanoparticles and to provide negative surface charges28. By contrast, primary vanadium oxide nanoparticles with a native negatively charged particle surface were prepared via sol–gel chemistry29, in a similar way to the synthesis of silicages (see Methods). Images of individual particles obtained by TEM revealed similar internal structures (Fig. 4). These nanoparticles did not appear to be dense but instead showed cage-like structures (compare Figs. 2 and 4), as corroborated by associated projection averages20 revealing cages with rotational symmetry (bottom insets in Fig. 4), similar to the prevalent projection in case of the silicage (vide supra). Therefore, micelle-self-assembly-directed cages like the dodecagonal structures described here may not be unique to amorphous silica, but may provide direct synthesis pathways to crystalline material cages (Extended Data Fig. 9).

Fig. 4: Cage-like structures with different inorganic compositions.
figure 4

a–c, Cage-like nanoparticles similar to silicages were obtained when silica was replaced by other materials, including gold (a), silver (b) or vanadium oxide (c). The insets show zoomed-in images of individual particles (top two rows) and averaged images20 (bottom row). The scale bars in all insets represent 2 nm.

Full size image

There are several ramifications of our silicage discovery. First, such cages are generally considered to be individual structural units from which larger-scale mesoporous silica is built in a bottom-up manner15,21. However, given that a dodecahedron cannot be used to generate a tessellation of 3D space, other silica cage structures must be required. This knowledge should motivate better understanding of the early formation pathways of surfactant-directed silica self-assembly, including a search for micelle-directed ultrasmall silicages with other structures, and from other materials (vide supra). Second, the chemical and practical value of this polyhedral structure may prove immense. Given the versatility of silica surface chemistry, and the ability to distinguish the inside and outside of the cage via micelle-directed synthesis17, one can readily conceive of cage derivatives of many kinds, which may exhibit unusual properties and be useful in applications ranging from catalysis to drug delivery. For example, given recent success in the clinical translation of ultrasmall fluorescent silica nanoparticles with similar particle sizes and surface properties to those described here30, one can envisage a range of new diagnostic and therapeutic probes with drugs hidden inside the cages.

Methods

Chemicals and materials

All chemicals were used as received. Cetyltrimethylammonium bromide (CTAB), ammonia (2 M in ethanol), mesitylene (1,3,5 trimethylbenzene; TMB), tetramethyl orthosilicate (TMOS), gold chloride trihydrate (HAuCl4·3H2O), silver nitrate (AgNO3), tetrakis(hydroxymethyl)phosphonium chloride (THPC), dimethyl sulfoxide (DMSO), acetic acid and ethanol were purchased from Sigma-Aldrich. Vanadium oxytriisopropoxide was purchased from Alfa Aesar. Anhydrous potassium carbonate (K2CO3) was purchased from Mallinckrodt. Anhydrous ethanol was purchased from Koptec. Silane-modified monofunctional polyethylene glycol (PEG-silane) with a molar mass of around 500 g mol−1 (six to nine ethylene glycol units) was purchased from Gelest. Carbon-film-coated copper grids for TEM, and C-Flat holey carbon grids for cryo-EM, were purchased from Electron Microscopy Sciences.

Synthesis, TEM and cryo-EM characterization of silicages

Silicages were synthesized in aqueous solution through surfactant-directed silica condensation. 125 mg of CTAB was first dissolved in 10 ml of ammonium hydroxide solution (0.002 M). Then, 100 µl TMB was added to expand the CTAB micelle size; the water/CTAB/TMB molar ratio was about 1,620/1/2. The solution was stirred at 600 r.p.m. at 30 °C overnight, and then 100 µl TMOS was added. The reaction was then left again at 30 °C overnight under stirring at 600 r.p.m.

To prepare cryo-EM samples, we applied 5 µl of the native reaction solution to a glow-discharged CF-4/2-2C Protochips C-Flat holey carbon grid, which was then blotted using filter paper and plunged into a liquid mixture of 37% ethane and 63% propane at −194 °C using an Electron Microscopy Science plunge freezer. Cryo-EM images were acquired on a FEI Tecnai F20-ST TEM operated at an acceleration voltage of 200 kV using a Gatan Orius charge-coupled-device (CCD) camera. All cryo-EM images used for reconstruction were acquired at the same magnification, with a pixel size of 0.16 nm, and nominal defocus was kept between 1 µm and 2 µm.

To prepare dry-state TEM samples, we added 100 µl of PEG-silane into the reaction solution. The reaction solution was left at 30 °C overnight under stirring at 600 r.p.m. to surface-modify silicages covalently with PEGs, in order to improve their dispersity on TEM grids. Afterwards, 30 µl of the reaction solution was dropped onto a copper grid coated with a continuous carbon film, and blotted using filter paper. TEM images were acquired using a FEI Tecnai T12 Spirit microscope operated at an acceleration voltage of 120 kV. In order to improve the signal-to-noise ratio in recorded images, TEM sample grids were plasma etched for 5 seconds before TEM characterization, and a series of images was acquired of the same sample area, which were then averaged.

In order to quench individual primary silica clusters formed at the very early stages of cage formation, we added 100 µl of PEG-silane into the reaction solution about three minutes after the addition of TMOS. The remainder of the procedures—including particle synthesis, dry-state TEM sample preparation, and TEM characterization—were as described above.

Particle purification

In order to remove CTAB and TMB from the cages, after adding PEG-silane and stirring at 30 °C for a day (see above), we heat-treated the solution at 80 °C overnight to further enhance the covalent attachment of PEG-silane to the silica surface of the silicages. The PEGylated nanocages were dialysed (with a molecular weight cut-off, MWCO, of 10 kDa) in a mixture of acetic acid, ethanol and water (volume ratio 7/500/500) for three days, and then in deionized water for another three days17. In both cases the dialysis solutions were changed once per day. The dry-state TEM sample preparation and TEM characterization methods were as described above.

Synthesis of particles without TMB

The synthesis and TEM characterization methods used for particles without TMB were the same as those for particles with TMB, as described above, except that the TMB addition step was omitted.

Silicage surface area and yield of production

We assessed the specific surface area of the silicages through a combination of nitrogen sorption measurements and theoretical estimations. After the synthesis and purification of PEGylated silicages, particles were first upconcentrated using a spin filter (Vivaspin 20, MWCO 10 kDa) and dried at 60 °C. Particles were then calcined at 550 °C for 6 hours in air. The production yield was then estimated by dividing the remaining weight after calcination, that is, the weight of inorganic silica, by the theoretical weight of silica, that is, the weight calculated on the basis of the amount of silica source added into the synthesis. Nitrogen adsorption and desorption isotherms were acquired using a Micromeritics ASAP 2020 (Extended Data Fig. 1c), yielding a specific surface area of 570 m2 g−1 by the Brunauer−Emmett−Teller (BET) method. For comparison, using the dodecahedral cage model with the dimensions from the reconstruction shown in Fig. 3, we estimated the theoretical surface area of silicages to be around 790 m2 g−1. The lower experimental value is consistent with the expected losses of surface area during sample calcination.

Synthesis and TEM of metal cage-like structures

The gold and silver cage-like structures were prepared by the reduction of metal precursors—HAuCl4·3H2O and AgNO3, respectively—in the presence of micelles with the same water/CTAB/TMB ratio as for the silicage work. In a typical batch, 50 mg of CTAB was dissolved in 4 ml of water at 30 °C, then 40 µl of TMB and 200 µl of ethanol were added to the mixture. After stirring the reaction at 30 °C overnight at 600 r.p.m., 16 µl of either HAuCl4·3H2O (25 mM) or AgNO3 (25 mM) was added, followed after 5 min by 8 µl of THPC (68 mM). After another 5 min, 6 µl of potassium carbonate (0.2 M) was added.

Dry-state TEM samples of gold and silver cage-like structures were prepared after one day and 6 hours of reaction, respectively, owing to the different reaction rates as described in further synthesis Methods sections below. In both cases, the samples were prepared by drying 8 µl of the native reaction mixture diluted three times in ethanol on a TEM grid in air overnight. In order to remove the thick CTAB layer before imaging, the grid was immersed in ethanol for 2 min and then dried in air. TEM images of metal cage-like structures were acquired using a FEI Tecnai T12 Spirit microscope operated at an acceleration voltage of 120 kV.

Synthesis and TEM of metal oxide structures

Vanadium oxide cage-like structures were prepared on the basis of sol–gel chemistry in a similar way to the silicages, using vanadium oxytriisopropoxide as the precursor. In a typical batch, 50 mg of CTAB was dissolved in 4 ml of water at 30 °C, and then 40 µl of TMB were added. After stirring the reaction at 30 °C overnight at 600 r.p.m., 50 µl of vanadium oxytriisopropoxide diluted in 100 µl of DMSO were added.

Dry-state TEM samples for vanadium oxide cage-like structures were prepared after one day of reaction by drying on a TEM grid 8 µl of the native reaction mixture, diluted ten times in water. At such a dilution, the amount of CTAB was low enough that the TEM samples did not require any plasma cleaning or soaking in ethanol before imaging. The TEM images of vanadium oxide cages were acquired using a FEI Tecnai T12 Spirit microscope operated at an acceleration voltage of 120 kV.

Particle reconstruction

We used the ‘Hetero’ model-based maximum likelihood machine-learning algorithm22 for particle reconstruction. This algorithm can simultaneously estimate: (1) a reconstruction for each type of particle shown in the images; (2) the type of particle shown in each image; and (3) the projection orientation for each image. Such a joint estimation is a central feature of the algorithm and is a natural approach for processing data from complicated mixtures. The estimates in (2) and (3), which are based on 3D structure, are independent of the clustering of 2D images, which is based on pixel values (see, for example, Extended Data Fig. 2). In addition to the Hetero algorithm, we applied the widely used RELION 2.1 system24 to compute equivalent two-class and single-class reconstructions. The images were corrected for the contrast transfer function (CTF) by phase flipping.

Further details of metal- and metal-oxide-based cages

In contrast to the sol–gel reaction that produces the silicage, synthesis of the gold and silver cage-like structures relies on reduction reactions. To this end, we used THPC because it reacts in water at basic pH to form trimethoxyphosphine, which can be both the reductant and the capping agent for the metal nanoparticles. THPC has been widely used to synthesize ultrasmall (less than 3 nm) and negatively charged phosphine-stabilized gold nanoparticles28. These nanoparticles are often used as seeds for the subsequent growth of continuous gold shells on the surface of aminated silica nanoparticles thanks to their high affinity and binding efficiency with amine groups31,32. Alcohol was added to the reaction mixture in order to mimic the conditions of the silicage synthesis, where methanol is formed upon hydrolysis of TMOS. Early-stage preliminary experiments showed that the resulting metal structures improved in size dispersity when using ethanol at a slightly higher concentration than that of the released methanol in the silicage synthesis. Gold and silver cage-like structure syntheses were performed at a much lower concentration (silver or gold concentration = 93.7 µM) than that of the silicages (silica precursor concentration = 65.9 mM). Attempts at synthesizing gold and silver cages at higher concentrations resulted in much larger nanoparticles with no apparent internal structure.

Gold-based synthesis

The addition of gold precursor to the reaction initially resulted in the formation of a pale yellow precipitate that turned into a clear (that is, non-turbid), darker-orange solution within a couple of minutes under stirring at 30 °C (see Extended Data Fig. 8 for a survey of the absorption characteristics at each step of the synthesis). Given that neither the precipitate nor the darker-orange colouration was observed in the absence of CTAB, we attribute these observations to some interaction between the gold chloride anions and the ammonium groups of the CTAB. After the addition of THPC, the solution turned colourless within a couple of minutes, indicating that gold(iii) had been reduced to gold(i). The subsequent transformation of THPC into trimethoxyphosphine, following an increase in pH through the addition of potassium carbonate, happened within the first hour of reaction (see also the description of silver-based synthesis below). However, the reduction from gold(i) to gold(0) was rather slow, with the first hint of colouration appearing after 8 hours of reaction. After one day of reaction, the solution exhibited a brown colouration. This colouration is the signature of gold nanoparticles that are too small or not dense enough to exhibit a strong surface plasmon resonance, as shown by the absorption profile in Extended Data Fig. 8, which shows only a faint feature around 510 nm.

Silver-based synthesis

The addition of silver precursor to the reaction did not initially translate into any visible effects, either in the presence of CTAB/TMB or after adding THPC. Nevertheless, after adding potassium carbonate to the silver-based synthesis, the solution started to turn pale yellow within the first hour of reaction and resulted in an intense yellow colouration after 6 hours, at which point we prepared the TEM samples. This yellow colouration is typical of small silver nanoparticles with a surface plasmon resonance centred around 420 nm (Extended Data Fig. 8).

Vanadium-oxide-based synthesis

The vanadium oxide cage-like nanoparticles were prepared under the same conditions as the silicages; however, the pH was not adjusted with ammonia owing to the fast hydrolysis and condensation rate of the vanadium oxide precursor. Unlike in the synthesis of metal cage-like nanoparticles, no alcohol was added here, because the hydrolysis of the vanadium precursor—vanadium oxytriisopropoxide—produces alcohol at concentrations similar to those used in the silicage synthesis. The addition of this precursor to the TMB micelles resulted in the immediate formation of a red precipitate. Under stirring, the precipitate dispersed homogeneously in solution, which remained turbid, and turned orange after one day of reaction at 30 °C.

Data availability

Figures that have associated raw data are Figs. 2–4 and Extended Data Figs. 1–4, 5c and 6–9. There are no restrictions on data availability. The data sets generated and analysed during this study are available from the corresponding author on reasonable request.

Code availability

The custom code and algorithm used for the 3D reconstruction and related analysis are available from the corresponding author on reasonable request.