Self‐Folding Graphene‐Based Interface for Brain‐Like Modular 3D Tissue

Electrophysiology of 3D neuronal cultures is of rapidly growing importance for revealing cellular communications associated with neurodevelopment and neurological diseases in their brain‐like 3D environment. Despite that the brain also exhibits an inherent modular architecture that is essential for cortical processing, it remains challenging to interface a modular network consisting of multiple 3D neuronal tissues. Here, a self‐folding graphene‐based electrode array is proposed that enables to reconstruct modular 3D neuronal tissue and investigate firing dynamics among moduli. A graphene‐sandwiched parylene‐C film self‐folds into a cylindrical structure within which living cells can be encapsulated. Culture of encapsulated cells inside the folded graphene enables to spontaneously construct 3D cell aggregates and ensure firm contact between the graphene surface and encapsulated cells. As the inner graphene surface can be utilized as an electrode, the reliable cell–electrode contact allows for long‐term electrical recording from multiple 3D aggregates. Additionally, the modular network consisting of multiple 3D aggregates exhibits richer firing patterns than a conventional homogenous 2D network, which demonstrates that the approach enables measurements of firing dynamics in complex 3D neuronal networks. The deformable graphene electrode will be a powerful platform for investigating complex cellular communications in brain‐like 3D cultures.

The ORCID identification number(s) for the author(s) of this article can be found under https://doi.org/10.1002/adfm.202301836.
connectivity of neuronal networks and functions of the human brain. For instance, microfabrication allows for spatial control of cell distribution to construct in vitro modular networks. [1][2][3][4][5][6][7] The modularity facilitates the emergence of complex dynamics that are essential for cortical processing. [6][7][8][9] Furthermore, 3D neuronal cell cultures such as spheroids and organoids provide a native cellular environment in which cells are intimately connected and form a complex system. [10][11][12] Hence, modular 3D brain-like tissues are considered to promote realistic physiological behaviors and provide insight into their complexities. [13][14][15][16][17] In these in vitro brain models, one of the key challenges is a functional analysis to investigate the relationship between the architectures and functions. Among the available methods, cellular electrophysiology with microelectrode arrays (MEAs) promises to improve our understanding of the cellular interactions associated with the development of the nervous system. [18,19] Conventional planar MEAs exhibit a performance limitation due to their poor contact with 3D tissues. Therefore, deformable MEA substrates that utilize stress-based self-folding to wrap 3D tissues have recently attracted great attention and have achieved reliable spatiotemporal recordings from a 3D spheroid and assembloid. [20][21][22][23][24][25][26] To extend the application of 3D electrophysiology for investigation of modular neuronal networks, it is preferable to record firing dynamics among multiple interconnected 3D tissues. However, it is still challenging to individually interface multiple 3D tissues using self-folding devices due to the difficulty in aligning the tissues with electrodes. Given that an electrode itself selffolds into a curved 3D structure and encapsulates dissociated cells, the encapsulated cells should be able to aggregate inside the electrode and, with the curved 3D structure used as a scaffold, spontaneously form a 3D tissue. It would then be possible to easily couple the electrode with multiple 3D tissues. Consequently, an array of self-folding electrodes will take reliable recordings from a network composed of interconnected 3D neuronal modules. Thanks to the encapsulation and selfaggregation of neurons, neurons would be spatially organized into clusters prior to the circuit wiring. This process is similar to brain development where connections are built and reshaped

Introduction
In vitro formation of brain-mimetic architectures has gained great importance in neuroscience because they recapitulate after organization of neuronal clusters such as brain regions, cortical layers, and cortical columns. [25,26] The developmental process is crucial for normal brain function and consequent miswiring drives dysfunction in the adult brain. Thus, an array of self-folding electrodes will spatiotemporally recapitulate brain development and then will help us better understand how brain-mimetic architectures affect the network development in physiological and pathological conditions. In this work, to validate this MEA design concept, we developed a self-folding electrode array and used it to demonstrate how the 3D modular structure affects network function.
Among the various self-folding materials, graphene is an ideal candidate for self-folding electrodes because it has been widely utilized as bioelectrodes and cell scaffolds thanks to its excellent electroconductivity, mechanical flexibility, transparency, and biocompatibility. [25][26][27][28][29] Unlike metal films, transparent graphene-based bioelectrodes allow for simultaneous imaging of the morphology, growth, and calcium signals of encapsulated cells. To date, self-folding graphene-based films have been designed to assemble graphene in a variety of microscale 3D shapes, such as cubic, [30] gripper-shaped, [31][32][33][34] and cylindrical structures, [35,36] which are applicable to conformal encapsulation of microscale biological tissues. [36,37] For example, encapsulation of living cells within a graphene-based self-folding skin has been shown to enhance Raman signals to facilitate 3D optical molecular sensing, indicating that the confined structure ensures firm contact between the graphene-surface and encapsulated cells. [37] Furthermore, it has been proven that self-folding graphene can be used as a scaffold to form a 3D neuronal cellular construct. [36] Therefore, self-folding graphene makes it possible to construct 3D neuronal tissue while reliably interfacing with it. To implement self-folding graphene as multiple electrodes, it is essential to develop a self-folding method that can be incorporated into an MEA substrate. We have demonstrated a cell-friendly method to self-fold a bilayer of graphene and poly(chloro-p-xylylene) (parylene-C), which has been widely used as a reliable insulator in bioelectronics. [27,35,36] This method does not require physical stimuli such as solvent exchange or heating to maintain the folded structure, making it suitable for long-term cell culture. Although the graphene exists on the outer surface of the curved structure in the previous method, a method to fold graphene in the opposite direction would realize an MEA with self-folding graphene electrodes.
Here, we propose a self-folding graphene/parylene-C/ graphene trilayer film and demonstrate simultaneous electrophysiological measurements from a network composed of multiple 3D neuronal constructs. To fold graphene in the opposite direction, another graphene layer was stacked on the self-folding parylene-C/graphene bilayer, resulting in a graphene-sandwiched parylene-C film. We examined the folding direction with this configuration and the mechanisms underlying the sloped internal strain in the film using Raman spectroscopy. The fabricated trilayer film was applied to 3D tissue formation. Neurons encapsulated within the folded trilayer were visualized to examine the process of spontaneous tissue formation and confirm the long-term maintenance of tissue. To highlight the reliable recording from multiple 3D tissues, the self-folding film was incorporated into an MEA and electrically tested. The firing dynamics of interconnected 3D neuronal aggregates were characterized to demonstrate the capability of recording in modular 3D neuronal networks for long periods of time. We believe that the cell encapsulation approach with the selffolding graphene-based electrode array will shed light on the electrophysiological behavior in brain-like networks.

Self-Folding of Graphene/Parylene-C/Graphene Trilayer Film
We fabricated a self-folding film by lithographical patterning as described previously [35,36] (see Figure S1, Supporting Information for the details of the fabrication process). As illustrated in Figure 1a, the self-folding film consists of two graphene layers (an upper and lower graphene layer), a parylene-C layer grown by chemical vapor deposition, and a sacrificial Ca-alginate layer. The number of graphene sheets in the upper (N upper ) and lower (N lower ) layers was varied to regulate the folding direction. The multilayer film was patterned by reactive ion etching (RIE) with oxygen plasma through a photoresist mask. We developed several rectangular films varying in size from 300 × 600 to 1000 × 2000 µm 2 . The size range was appropriate for observing the folding process with an optical microscope.
We investigated the transition from 2D templates to 3D architectures by microscopic observations (Figure 1b-f). To examine the direction of folding, we fabricated rectangular films (300 × 600 µm 2 ) with a varied number of graphene sheets (N upper and N lower ) and observed the transformation by inclining the films under a confocal fluorescence microscope ( Figure S2, Supporting Information). Dissolution of Ca-alginate with an ethylenediaminetetraacetic acid (EDTA) solution triggered upward or downward self-folding of the film. The folded film formed a cylindrical structure, which is called a microroll ( Figure 1b). As described previously, [35] the graphene/parylene-C bilayer film (N upper = 1, N lower = 0) folded downward (mountain fold, Figure 1c). The dissolution of the Ca-alginate layer did not directly trigger the self-folding. The bilayer remained flat more than 2 min after adding EDTA and then immediately folded ( Figure S3, Supporting Information). In contrast, the parylene-C/graphene bilayer film (N upper = 0, N lower = 1) folded upward (valley fold). The bilayer was progressively released and folded upward from the substrate along the gradual dissolution of Ca-alginate. The response time from EDTA addition was shorter (<10 s) and a long-side rolling was more preferred in the case of parylene-C/graphene bilayer film (N upper = 0, N lower = 1; Figure S3, Supporting Information). Note that a parylene-C film without graphene (N upper = 0, N lower = 0) did not self-fold ( Figure 1f). Thus, in these bilayer films, the strain mismatch between graphene and parylene-C initiated the self-folding. The inverted stacking order also affected the curvature radius of microrolls (Figure 1d). The parylene-C/graphene bilayer film (N upper = 0, N lower = 1) created microrolls with smaller curvature radii than the graphene/parylene-C bilayer (N upper = 1, N lower = 0), which implies a steeper gradient of internal strain. As a result, the graphene/parylene-C/graphene trilayer film (N upper = 1, N lower = 1) folded upward owing to the in-plain strain gradient from the bottom to the top (Figure 1c). Although the curvature radii of trilayer films were larger than those of the bilayer ones, adding more graphene sheets to the lower layer (N upper = 1, N lower = 2) resulted in reduction of the curvature radii to an appropriate size for cell encapsulation (50-100 µm in diameter, Figure 1c,e). As shown in Figure 1f, when additional graphene was transferred to the upper layer (N upper = 2, N lower = 1), the folding direction did not reverse, supporting the idea that the lower graphene produced a larger strain mismatch than the upper graphene. Furthermore, folding characteristics of the trilayers, including the preference in folding orientation and the response time, were similar to the parylene-C/graphene bilayer film (N upper = 0, N lower = 1), suggesting that the lower graphene mainly contributed to the self-folding of the trilayers ( Figure S3, Supporting Information). Importantly, the upper graphene formed the inner side of the microrolls and could be utilized as a cellular interface with encapsulated tissues.

Characterization of Strain in Self-Folding Graphene
To reveal why the bilayers and trilayers self-fold, we investigated the strain state in graphene before and after the folding. Figure 2a, potential candidates for the driving force of folding are a compressive pre-strain in the parylene-C layer and a tensile pre-strain in graphene. Thus, we examined these pre-strains using the parylene-C/graphene bilayer film (N upper = 0, N lower = 1) and then considered a mechanism that would explain why the strain mismatch differs between the upper and lower graphene.

As illustrated in
We first evaluated the compressive pre-strain in the parylene-C layer. To characterize the intrinsic strain in graphene, Raman spectroscopy was performed. As shown in Figure 2b, the Raman spectrum of the graphene shows the characteristic D, G and 2D band at 1350, 1580, and 2680 cm −1 , respectively. The D band arises from defects of graphene and it was overlapped by the parylene-C-derived peak. Although it could not be observed because of the overlapping, there were no apparent changes in the peak height of D band after folding, suggesting that the film folded without any damage. Since the D' band at 1600 cm −1 was also overlapped by the parylene-C-derived peak, we could not examine the emergence of the D' peak caused by the breaking of lattice symmetry. Meanwhile, the G and 2D bands were clearly observed even after the parylene-C coating. The  shift of the Raman frequency of G (ω G ) and 2D (ω 2D ) modes determines the magnitude of in-plain strain at the recording point. [38,39] To evaluate the spatial distribution of strain, we recorded multipoint spectra at 1 µm intervals from 30 × 30 µm area and then made 2D plots of the ω G -ω 2D (Figure 2c  Raman spectra of flat graphene (black trace) and flat and folded parylene-C/graphene bilayer (blue and cyan trace). c) Optical images and Raman map for the G-mode (ω g ) and 2D-mode frequency (ω 2D ) obtained from the flat and folded film. The white dotted lines indicate the edge of the bilayers. Each Raman mapping was performed in the area indicated by the blue dashed boxes spanning 30 × 30 µm 2 . Scale bars indicate 100 µm. d) Correlation plots of the G-2D Raman frequencies (ω g , ω 2D ) obtained from a single Raman map with 900 measuring points. The plots were created for the folded (blue) and flat (light blue) parylene-C/graphene bilayer and flat graphene (gray). Stars indicate the averaged position of all plots. Note that a strain in graphene ε g can be estimated by vector decomposition in the G-2D space. e) Bar plots that represent the averaged value of estimated strain obtained from six independent Raman maps ( = 6 films). f) AFM topography images of flat graphene on a glass substrate and Ca-alginate layer. The lower green traces indicate the height cross sections along green dotted lines in images. g) Schematic illustration of possible mechanisms of self-folding for the graphene-sandwiched trilayer film. h) Bar plots of averaged ε g obtained from oppositely stacked bilayers (blue, N upper = 0, N lower = 1; red, N upper = 1, N lower = 0). i) Plots of ω 2D versus ω g measured from flat and folded graphene-sandwiched films. j) Estimated strain ε g of flat and folded graphenesandwiched films. n = 900 measuring points for the each Raman map, and n = 6 films for each bar plot. ***p < 0.001 and **p < 0.01; Welch's t-test. Data are shown as mean ±SD.
into a graphene sheet, the plots of ω G -ω 2D shift on a single line with a slope (Δω 2D /Δω G ) of 2.2. [38] As shown in Figure 2d,e, the 2D Raman analysis shows that the average position shifted after the parylene-C deposition, indicating a compressive strain (−0.1 ≤ ε g ≤ 0) caused by the parylene-C deposition. Interestingly, there was a slight change in the average position after the folding. These results suggest that the compressive pre-strain induced the self-folding, while the total strain within the graphene was retained. The origin of pre-strain might be a residual stress generated when heated monomers were condensed and polymerized on the substrate during the CVD process. [40,41] It is worth noting that the CVD process itself produced a small stress gradient in the parylene-C layer because the pristine parylene-C film (N upper = 0, N lower = 0) remained flat even after floating, as shown in Figure 1f. Thus, the mechanical heterogeneity in the bilayer leads to the strain gradient after the removal of the sacrificial layer. As described in the previous study, control experiments with the bilayer of graphene and other materials such as photoresist and hydrogel indicate that the bilayer fabrication process or the material properties of parylene-C were required to induce the self-folding ( Figure S4, Supporting Information). [35] Additionally, a finite element analysis supports the idea that the pre-strain in parylene-C induced the folding ( Figure S5, Supporting Information). In the simulations, an increased compressive pre-strain in the parylene-C layer promoted folding behavior, leading to a reduced curvature radius. However, the thicker parylene-C layer decreased the curvature radius, which disagrees with the experimental trend shown in Figure 1d. Thus, this simulation also implies the existence of pre-strain in graphene.
We then examined the pre-strain in graphene. As shown in Figure 2c,d, there was an apparent reduction in strain variation after the folding. Furthermore, atomic force microscopy shows that out-of-plane wrinkles with an interval of 0.5-5 um were formed in the graphene that had been transferred onto the Caalginate layer (Figure 2f and Figure S6, Supporting Information). These results suggest that the wrinkles caused the strain variation, and they disappeared after folding. Given that the disappearance of the wrinkles compensates for the expansion of graphene, the surface area of the bilayer film on the graphene side would be increased after it folds. It supports the existence of graphene on the outer surface of microroll while the total strain within the graphene was retained even after folding. Note that the wrinkles were not generated when graphene was transferred onto a glass substrate ( Figure S6, Supporting Information). Thus, the wrinkles were probably formed by the strain mismatch between the graphene and Ca-alginate layer, which might be caused by deswelling of Ca-alginate after the wet transfer of graphene during the film fabrication. Control experiments supported this idea. When a Cr thin layer was used as a sacrificial layer and then dissolved using Cr etchant, few films folded with a large curvature radius, which suggests that the Ca-alginate layer is required for the self-folding ( Figure S4, Supporting Information). There was also only a slight difference in the folding behavior between films composed of different Caalginate ( Figure S4, Supporting Information). Taken together, we concluded that the self-folding was driven by pre-strain both in the parylene-C layer and the graphene. Importantly, the driving force for self-folding does not require any continuous stimulation. In previously developed methods of self-folding graphene, a strain gradient in the self-folding film was artificially created and then the tunable driving force resulted in a controllable and reversible folding driven by stimuli such as solvent exchange and heating. [32][33][34] However, when the self-folding graphene is used as a cell scaffold, a long-term application of stimuli should affect cell growth. Despite the irreversibility, the self-folding of parylene-C/graphene bilayer is driven by intrinsic strain without any continuous stimuli and the bilayer consists of biocompatible materials. Therefore, this choice of film materials is essential for applications involving long-term cell culture.
The magnitude of these pre-strains differs between the upper and lower graphene, leading to the self-folding of the trilayers (N upper = 1, N lower = 1) ( Figure 2g). As far as the prestrain in parylene-C is concerned, the compressive strain was larger in the lower graphene (N upper = 0, N lower = 1) than in the upper graphene (N upper = 1, N lower = 0), suggesting that the effect of pre-strain in parylene-C was larger in the lower graphene ( Figure 2h). This is because the lower graphene more tightly adhered to the deposited parylene-C layer, whereas the upper graphene was just transferred onto the parylene-C layer. Interestingly, when an underlying layer was changed from the Ca-alginate to a glass substrate, the compressive strain in the upper graphene was further reduced, suggesting that the compression requires the material properties of Ca-alginate ( Figure S7, Supporting Information). Furthermore, the size and the number of wrinkles might also be larger in the lower graphene because it was directly stacked on the Ca-alginate layer. Therefore, the larger strain mismatch in the lower graphene resulted in a valley fold of the trilayer film (N upper = 1, N lower = 1). In the case of the trilayer film, the Raman spectrum contains overlapped signals that were recorded from both the upper and lower graphene. Note that there were no clear differences in the peak height of the 2D band after folding, which suggests that the self-folding did not cause any exfoliation of either the upper or lower graphene ( Figure S7, Supporting Information). Since the upper graphene forms the inner face of a microroll, the upper graphene needs to be compressed through the folding process. As a result, unlike in the bilayer film, the value of (ω G , ω 2D ) moved parallel to the vector for compressive strain (Figure 2i,j).
To confirm that the stress-based self-folding maintains the availability of graphene for an electrode, we evaluated the conductivity of the folded graphene using two-point probes ( Figure S8, Supporting Information). The I-V curve for the upper graphene in the flat state showed linear behavior with a sheet resistance of 550 ohms sq −1 . Although the sheet resistance increased to 645 ohms sq −1 after folding, the electrical properties were retained. In contrast, the self-folding altered the electrical properties of the lower graphene, and the I-V curve became nonlinear ( Figure S8, Supporting Information). As previously described, folded graphene nanochannel structures show nonlinear I-V curves. [42] Additionally, it has recently been demonstrated that self-folding graphene microstructures also exhibit nonlinear behavior, which is consistent with our observations in the lower graphene. [34,35] Interestingly, even though the compressive strain was larger in the upper graphene, it was the lower graphene that exhibited the nonlinear behavior. Further study with gate voltages applied would better clarify the electrical modulation observed. Importantly, the selffolding process never led to the exfoliation of the graphene, and the linear behavior of upper graphene was retained even after folding, which indicates the availability of the upper graphene for an electrode.

Cell Encapsulation within the Self-Folding Graphene
Since the stable conductivity of graphene was confirmed, we next utilized the graphene-sandwiched parylene-C film (N upper = 1, N lower = 2) to encapsulate living cells (Figure 3). Aside from its high Young's modulus (≈1 TPa), graphene has unique surface characteristics and versatile chemical properties that allow the inner graphene surface to work as a scaffold for cell culture. Dissociated cells were seeded onto the film prior to its release from the substrate (Figure 3a). Geometries of the self-folding film and the detailed cell encapsulation method are summarized in Figures S9 and S10 (Supporting Information). To improve the cell attachment on the film surface, the film was coated by polyethyleneimine (PEI) and laminin. Additionally, to improve the culture condition inside the microrolls, pores with a diameter of 6 µm and spacing of 50 µm were incorporated into the microroll surface. [36] As previously described, the incorporation of pores had no effect on the microroll curvature. [36] The pores facilitate the diffusion of oxygen and nutrients while cells are inside the microroll. Furthermore, to efficiently observe the  Note that small pores with diameter of 6 µm were created on the film surface to improve the culture environment inside. b) Viability assay by staining neurons with dyes of calcein (live; green) and ethidium homodimer-1 (EthD-1; dead; red) before and after folding. There is no significant reduction in cell viability after folding. c) 3D reconstructed images and d) cross-sectional images of encapsulated neurons at 0, 5, and 14 DIV. Neurons were stained with anti-tau1 antibody (green) and Hoechst 33342 (blue). The microroll surface was visualized using the FITC-labeled self-folding film (red). e) Increase of ratio of tau1-positive area to the internal space of microroll with increasing culturing period. n = 24 z-plane images in eight films. f) 3D reconstructed images of aggregated neurons cultured on the flat and folded films at 5 and 14 DIV. Long-term trends in g) volume of aggregates and h) % viable cells compared between the flat and folded state. n = 7 films for volume analysis and n > 40 films for viability assay. Data are shown as mean ±SD.
cultures, an array of rectangular films was formed, and it was mechanically supported by cross-shaped hinge patterns as illustrated in Figure S9 (Supporting Information). We cultured and encapsulated primary rat hippocampal neurons (see Experimental section for details of cell culture) within microrolls with the diameter of 50-80 µm to confirm the capability of forming 3D tissues. As shown in Figure 3b, there is no clear difference in the ratio of living cells between before and after folding, indicating that the process of cell encapsulation had no effect on cell viability. The encapsulated neurons adhered to the inner graphene surface of the microrolls even after folding due to graphene's cytocompatibility and the surface treatments of PEI and laminin. The encapsulated neurons grew and increased in volume inside the microroll due to the maturation of neurons and proliferation of glial cells (Figure 3c-e). They occupied 70% of internal space at 5 d in vitro (DIV) and then the ratio of occupation exceeded 80% at 14 DIV. As a result, the encapsulated neurons formed a cylindrical aggregate, as opposed to a spherical aggregate on a flat film ( Figure 3f). As shown in Figure 3g, the volume increase of the cylindrical aggregate was slower because the confined structure caused contact inhibition of cell proliferation. Although the higher cell proliferation on the flat film produced a larger aggregate with a higher ratio of living cells by 14 DIV, excessive proliferation caused cell death, which resulted in the reduction both in the tissue volume and viability ratio after 14 DIV. Although the higher cell proliferation on the flat film produced a larger aggregate with a higher ratio of living cells by 14 DIV, excessive proliferation caused cell death, which resulted in the reduction both in the tissue volume and viability ratio after 14 DIV (Figure 3h). On the other hand, the aggregates in the microroll gradually increased their volume even after 14 DIV without a significant reduction of cell viability. It is worth noting that the pores on the surface of the microroll enhanced permeability of oxygen and nutrients and then allowed the long-term culture of aggregates. While the culture environment inside the microroll was sufficient for the long-term maintenance of the culture, further investigation of cell growth using different pore patterns will allow earlier aggregate formation. The cell encapsulation method can be applied to other kinds of electrogenic cells, such as neurons, cardiomyocytes, and C2C12 myoblasts, which also grew and formed into a functional 3D cellular tissue inside the microrolls ( Figure S11, Supporting Information). These results confirm that the porous microroll architecture allows the formation of a cylindrical aggregate that can be cultured for a long period of time. Furthermore, the most important point here is that the tissue formation is spontaneous, resulting in the automatic creation of coupling between the tissue and folded graphene surface.
The porous microroll architecture also allows the encapsulated neurons to interact with their surroundings via axons. As described in the previous study, [36] the pore diameter of 6 µm was the largest one that allows axonal elongation while retaining neurons inside. To observe the axon elongation, we encapsulated neurons within a free-standing self-folding film and then transferred the neuron-laden microrolls with a diameter of 80 µm onto a PEI and laminin-coated glass-bottom culture dish (Figure 4a). Time-lapse images of a neuron-laden microroll in Figure 4b indicate that their axons extended to the surroundings through the edge of the microroll and the porous wall. As shown in Figure 4c, MAP2-positive cell bodies were sustainably located inside the microroll. On the other hand, they extended their axons (tau1-positive) to the surroundings. Therefore, with the 3D aggregate architecture retained, encapsulated neurons were able to connect with surrounding neurons, particularly with the other aggregate (Figure 4d). Immunocytochemical images in Figure 4e visualize structural connections between the two aggregates via axons. Axons connected the two aggregates at the bottom surface of the culture dish (z = 0 µm). Additionally, spontaneous Ca 2+ activities in the two aggregates were synchronized, indicating the establishment of functional connections between them (Figure 4f). Thus, 3D neuronal aggregates in microrolls can be utilized to create interconnected neuronal modules. Furthermore, these results indicate that the optical transparency of the microrolls facilitates structural and functional investigation of the modular 3D neuronal network using imaging techniques.

Electrophysiological Measurements of Encapsulated Cells
As a proof of concept for electrophysiological measurements, we developed a self-folding graphene-based microelectrode array (sf-gMEA) in which the self-folding graphene-sandwiched parylene-C films (N upper = 1, N lower = 2) were implemented as 64-channel electrodes. To distinguish the spatiotemporal firing patterns of the modular 3D network from that of the conventional homogenous network, an array pattern of 8 × 8 was developed. In order to form sparse connections among the aggregates, the electrode spacing was set at 1 mm, which is larger than in conventional MEAs. The schematic illustrations in Figure 5a show the configuration of the sf-gMEA (see Figure S12, Supporting Information for details of the fabrication process). The Au/Cr (gold/chromium) interconnect lines were extended from rectangular 200 × 400 µm 2 electrodes to connection pads, through which the recorded extracellular potentials were transferred from the electrodes to an amplifier. The Au/Cr interconnect lines were passivated by a 1 µm thick parylene-C layer. The transferred signals were recorded and digitalized using amplifiers with a stimulator. Since the Au/Cr interconnect lines were covered with a thick parylene-C layer, the dissolution of the sacrificial Ca-alginate layer did not release the interconnect lines but released the electrodes from the substrate. The controlled release allowed for cell encapsulation independently within each electrode (Figure 5b).
The photograph and phase-contrast images of the fabricated sf-gMEA in Figure 5c,d demonstrate that the electrode folded upward along the hinge to form a cylindrical structure with a diameter of 80 µm and encapsulated primary rat hippocampal neurons within its folded structure. To improve the culture environment inside the folded electrode, 6 µm pores were created on the electrode surface, the same as in the experiments on aggregate formation. As shown in Figure 5e and Figure S13 (Supporting Information), the encapsulated neurons were sustainably located for the long-term culture and grew inside the folded electrode for 70 DIV. While the pores enhanced axonal elongation, neurons extended their axons mainly from the end of the roll-shape to the outside because the substrate underneath the sacrificial layer was not treated for cell adhesion. In addition to the encapsulated neurons, surrounding neurons migrated and grew on the outer surface of the folded electrode over 14 DIV. Scanning electron microscopy (SEM) images in Figure S13 (Supporting Information) show that neurons adhered to the outer surface of the electrode, while axons elongated through the pores. Meanwhile, time-lapse phase-contrast and immunochemical images in Figure 5f,g reveal that they extended their axons to the outside of the folded electrode and formed structural connections with different neuronal aggregates. As previously reported, the growth speed of axons was 3 µm h −1 when one isolated aggregate was cultured. [36] This suggests that it would take more than 10 d to connect with adjacent aggregates. Nevertheless, as shown in Figure 5f, the connection was formed at 7 DIV, which implies that growth factors secreted from the adjacent aggregates promoted and guided the growth of axons. The connections were sparse due to the restricted outlet for axon extension from the folded electrode. The dense connection in the aggregates and the sparse connection among aggregates created network modularity on the sf-gMEA. It is also worth noting that axons extended to an adjacent aggregate at 7 DIV when neurons had formed an aggregate, as indicated in Figures 3c-e and 5e. The network development inside individual aggregates probably occurred earlier than the formation of interconnections between different aggregates. The developmental process is similar to that in the human brain, where neurons in modules such as cortical microcolumn form intramodule connections prior to the formation of inter-module connections. [26] Therefore, growing the modular 3D culture recapitulates one of the characteristics of brain development.
To verify the structural advantages of the folded electrodes for recordings from 3D neuronal tissue, we compared flat and  folded electrodes. As shown in Figure S14 (Supporting Information), the 3D aggregates are formed on flat electrodes in a way similar to how the culture is formed on the flat graphenesandwiched film (Figure 3h). Imaging analysis was performed  reconstructed images of neuronal aggregates that were cultured on the flat and folded electrode. Neurons were stained with anti-tau1 antibody (green) and Hoechst 33342 (blue) at 14 and 28 DIV. The electrode surface was visualized using the FITC-labeled self-folding electrode (red). c) Comparison of cell-electrode contact region between flat and folded electrode along the culturing period. n = 24 z-plane images in eight films. d) Raw signals of spontaneous neuronal activities recorded from flat (black) and folded (violet) electrodes. e) Representative spike waveforms recorded from the folded electrode. All spikes obtained from a single electrode were sorted into spikes of individual neurons using a spike-sorting method. For each cluster, waveforms of 100 spikes were superimposed; the averaged waveform is indicated by the thick line. Comparison of f) spike amplitude and g) firing frequency (spike number per second), recorded from electrodes in flat and folded states at 14 DIV. n > 100 electrodes. Long-term trends in h) the number of active electrodes and i) firing frequency. The plotted value was obtained from data obtained with a 5 min recording of spontaneous activities. Note that the folded electrode sustainably exhibits more frequent spiking activities than the flat electrode. n = 8 (folded) and 6 (flat) MEAs. Data are shown as mean ±SD. ***p < 0.001; Welch's t-test. 28 DIV. On the other hand, encapsulated neurons were sustainably located inside the folded electrode, suggesting a better cell-electrode contact. As indicated in Figure 6c and Figure S16 (Supporting Information), the cell-electrode contact region in the flat state drastically decreased at 21 DIV, while the folded electrode could maintain the contact region. The better contact likely results in higher seal resistance at the cell-electrode interface and thus allows stable electrical recordings.
We next performed electrophysiological measurements. The folded electrodes were used to record spiking activities and local field potentials from aggregates of electrogenic cells such as neurons and cardiomyocytes ( Figure 6 and Figure S17, Supporting Information). We recorded and compared spontaneous spiking activities from aggregates that were grown on flat and folded electrodes (Figure 6d). The extracellular spikes recorded from a folded electrode are shown in Figure 6e. The spikes were clustered by individual neurons using a spike sorting method. It is known that the typical shape of an extracellular spike recorded from a cell body or an axon is biphasic or triphasic, respectively. [43,44] The averaged waveform in each cluster exhibited the two types of spikes, which indicates that the electrode can record a single spike. Meanwhile, complicated spike shapes that contained an additional bump just after the main spike were observed, suggesting that an electrode recorded from multiple axon branches extending from a single cell. [44] As indicated in Figure 6f,g, there were negligible differences in the averaged amplitude of the spikes, whereas the firing frequency (number of spikes per second) recorded on folded electrodes (2.4 Hz) was six-fold higher than that recorded on flat electrodes (0.4 Hz) at 14 DIV. The increased firing frequency was maintained for over 70 DIV thanks to the capability of sustaining a long-term culture inside the folded electrode (Figure 6i). Additionally, the number of active electrodes in the folded state was sustained for over 70 DIV, whereas the flat electrodes lost contact with neuronal aggregates during the culture, which led to a decreased number of active electrodes after 14 DIV (Figure 6h). Therefore, we mainly attribute the increased firing frequency to the larger number of neurons sitting on the electrode surface or in close proximity to it. On the other hand, previous reports have demonstrated that extracellular spikes are amplified by increasing the sealing resistivity with covering structures, such as a microgripper, [20] microchannel, [44][45][46] or glial-cell sheet. [47] Given that the confined structure inside the folded electrode improves cell-electrode contact and the sealing resistivity, it may enhance the spike amplitude. Although the ratio of large spikes (>50 mV) recorded from folded electrodes (2.9%) was higher than that recorded from flat electrodes (1.3%), the spike amplification mainly increased the number of small spikes (<30 mV) that were comparable to the noise level ( Figure S18, Supporting Information). As a result, since the large number of small spikes determined the averaged amplitude, the averaged amplitude was not altered (Figure 6f). Thus, the spike amplification may also contribute to the increased firing frequency. These results confirm that the cell encapsulation approach provides a structurally reliable electrical interface with 3D neuronal tissues.
Importantly, cell encapsulation within the self-folding electrode enables us to construct 3D cellular tissue while maintaining the cell-electrode coupling. In previous electrophysiological measurements of 3D cell spheroids, the 3D cell spheroids had to be manually positioned on electrodes. [21][22][23][24] On the other hand, the cell encapsulation approach provides an electrical interface with a 3D cellular tissue without the need to align cells with the electrode. Since the cell-electrode coupling can be achieved in multiple electrodes with a batch release of the sacrificial layer, the individually addressable electrodes allow the simultaneous measurement of multiple 3D cellular tissues. In particular, the simultaneous recording can characterize the firing dynamics of interconnected 3D neuronal tissues.

Multisite Recording of Interconnected 3D Neuronal Tissues
We employed simultaneous and multisite recording to characterize the network behavior of interconnected 3D neuronal tissues, which are called modular 3D cultures. Raw signals of spontaneous activity and spike raster plots are shown in Figure 7a and Figure S19 (Supporting Information). At 7 DIV, neurons began firing action potentials randomly, and there was no clear evidence of collective activity. A synchronized activity, called a network burst (NB), emerged after 9 DIV, and it dominated the activity throughout the rest of development. The NB patterns shifted with increased frequency and a shortened duration during the culturing period. In particular, the culture at 9 and 14 DIV exhibited variable burst patterns mixed with single-spike activity, and then the pattern variety decreased after 21 DIV. Importantly, although axons extended from an aggregate to an adjacent aggregate at 7 DIV (Figure 5f), NBs emerged after 9 DIV, suggesting that it took 2 d to establish the excitatory synaptic connections required for their emergence. Other experiments also confirmed the excitatory synaptic connections ( Figure S20, Supporting Information). Pulse injection to an electrode evoked spike activities in other electrodes at 19 DIV. Note that spatial patterns of evoked activities depended on the stimulation site, which suggests the existence of several small networks in the culture. In addition to electrical recordings, calcium imaging at 14 DIV confirmed that spontaneous activity was synchronized between two aggregates encapsulated within folded electrodes. The spontaneous activity was dramatically suppressed by pharmacological treatments with receptor blockers of glutamatergic synapses (50 × 10 −6 m CNQX and AP-5) at 23 DIV. This reveals that the spontaneous activities were mainly mediated by excitatory synaptic inputs. [48] As described in previous reports, typical burst behavior emerges with synaptic development at the early stage of maturation (7)(8)(9)(10)(11)(12)(13)(14) in homogenous 2D cultures, [49,50] which is consistent with the generation of NBs in our observation. This suggests that the time course of synapse development may not be affected by the network dimensionality and topology.
To further investigate whether the topology affects the network behavior, we quantified the characteristics of the NBs and compared them with homogenous 2D cultures (Figure 7b-f and Figure S21, Supporting Information). Homogenous 2D cultures were constructed by coating the whole surface of the device with PEI and laminin solutions. Additionally, modular 3D cultures were grown on the flat electrode as a control condition. The method of creating these different cultures is shown in Figure S14 (Supporting Information). An NB was www.afm-journal.de www.advancedsciencenews.com

(12 of 17)
defined as a period when high-frequency firing occurred in a large number of electrodes. As described in Figure 7c, modular 3D and homogenous 2D cultures exhibited different activity patterns. Activity patterns in homogenous 2D cultures were fully synchronized at 12 DIV (Figure 7c and Figure S21, Supporting Information), whereas random spikes outside the NBs remained at least 10% in activities of modular 3D cultures. This implies that local circuits generating random spikes were formed within each aggregate (intra-cluster connection), while different aggregates were weakly interconnected (inter-cluster connection). The weak inter-cluster connection could change the NB properties. Modular 3D cultures exhibited half-length NBs more frequently compared to homogenous 2D cultures (Figure 7c and Figure S21, Supporting Information). These NB properties are similar to those reported in previous work with a grid-patterned network in which connection paths were restricted, [1,51] where the shorter burst duration was attributed to reduced re-entrant pathways of excitatory and inhibitory inputs. Simultaneous intracellular recordings of inhibitory and excitatory postsynaptic currents from encapsulated neurons will help us reveal how the synaptic inputs affect NB properties. In our observations, the sparse inter-cluster connection probably reduced the re-entrant excitation from different aggregates. Figure 7. Simultaneous multisite recordings of activity patterns in interconnected 3D neuronal cell culture. a) Examples of raw signals and spike raster plots of spontaneous neural activities recorded from folded electrodes at 7, 9, 14, and 21 DIV. The raster plots show spike trains in all electrodes. b) Schematic illustrations of modular 3D culture formed by cell encapsulation within folded electrodes and homogenous 2D culture (Homo. 2D) grown on flat electrodes. c) Spike raster plots obtained from a modular 3D (violet) and a homogeneous 2D (black) cultures. Note that the modular 3D culture exhibits richer patterns of network bursts (NBs) compared to the homogenous 2D culture. d) Probability density functions of fraction of NBparticipating electrodes. The fraction of NB-participating electrodes is more broadly distributed in the modular 3D culture than in the homogenous 2D culture. e) Comparison of fraction of NB-participating electrodes between homogenous (black) and modular cultures grown on flat (gray) and folded (violet) electrodes. f) Developmental changes in fraction of NB-participating electrodes in the modular 3D and the homogenous 2D cultures. n = 8 (modular 3D/folded), 6 (modular 3D/flat), and 3 (homogenous 2D/flat) MEAs. Data are shown as means ±SD. *p < 0.05, ***p < 0.001; Student's t-test.
The most prominent feature was segregation behavior in the modular 3D culture (Figure 7c). Figure 7d shows the probability density functions of the number of NB-participating electrodes. The comparison shows that the fraction of electrodeparticipating NBs was more broadly distributed from 0.3 to 1 for the modular 3D culture than for the homogenous 2D culture, whose values were clustered at ≈1. The mean value of 0.40 for the modular 3D culture was significantly lower than the 0.95 for the homogenous 2D culture, while the coefficient variance was approximately six times higher at 14 DIV (Figure 7e and Figure S21, Supporting Information). The small and variablesized NBs reflected segregated networks in the modular 3D culture. It has been demonstrated that a modular culture promotes dynamical richness and that weak inter-cluster connectivity in the modular network is required for its emergence. [6] In our system, in concert with the weak intercluster connections, dense intracluster connections within each aggregate could contribute to the dynamical richness. Despite the sparse inter-cluster connections, a modular 3D culture grown on flat electrodes exhibited less variable and larger NBs compared with the folded state ( Figure 7e and Figure S21, Supporting Information). This indicates that the cell encapsulation facilitated the richness of activity with the formation of segregated networks. As previously reported, the 3D culture environment enhances synapse formation and synchronous activities compared with 2D cultures. [52][53][54] The structural support of the 3D culture by folded electrodes likely contributed to a dense inter-cluster connectivity mediated by enhanced synapse formation. Thus, we attribute the dynamical richness to the unique balance between inter-and intra-cluster connectivity resulting from the spatially distributed 3D aggregates. Note that the noninvasive recording with the sf-gMEA allowed long-term monitoring of network behavior (Figure 7f). The capability of long-term monitoring showed that the fraction of NB-participating electrodes increased from 14 to 28 DIV, indicating that the segregated networks were gradually integrated during the culture. This shows that the modular 3D culture on the sf-gMEA enables us to investigate the developmental profile in terms of the integration and segregation balance. [55] The sf-gMEA was designed to allow simultaneous recordings of interconnected 3D neuronal tissues that exhibit rich firing patterns. Although other designs of deformable MEAs have demonstrated electrical measurements with high spatiotemporal resolution in 3D, they were applied to a few 3D spheroids or 2D cell culture. [20][21][22][23][24] This work shows an extended application of deformable MEAs for investigation of network dynamics among multiple 3D aggregates. Our in vitro neuronal model mimics two important architectures of the human brain. One is the 3D morphology of the culture. The dynamics of 3D neuronal cultures differs from those of 2D cultures and better recapitulate what is observed in vivo. [52][53][54] Additionally, neuronal cultures have already contributed to our understanding of neurological disorders, such as microcephaly and Alzheimer's disease. [56,57] To create the 3D architecture, the curved shape of the sf-gMEA provides a scaffold of neuronal aggregates and suppresses excessive proliferation of glial cells by contact inhibition. The structural advantages of the 3D electrode lead to the maintenance of 3D neuronal aggregates, which allows the long-term recording of firing dynamics from 3D neuronal aggregates. The other structural feature is the modularity. The brain exhibits an inherent modular architecture that facilitates high robustness, adaptiveness, and resilience. [58] As described above, electrophysiological experiments showed that neurons growing on the sf-gMEA form a modular network with dynamical richness. Thus, the sf-gMEA provides a neuronal interface platform to investigate firing dynamics in an in vivo like 3D neuronal network. Further studies with different dimensions, such as electrode size, spacing, and array patterns, will help us understand the relationship between modularity and firing dynamics. On the other hand, a limitation of the current design of the sf-gMEA is the low recording resolution of intra-cluster activities due to its limited number of channels. Precise recordings from each aggregate would help us to estimate intra-cluster connectivity. Moreover, they would also improve the estimation of inter-cluster connectivity, leading to a more in-depth investigation of the balance between intra-and inter-cluster connectivity. Although the number of channels is limited in the conventional MEA system, increasing the channels by incorporating self-folding film into a high-density MEA will enable precise recordings from multiple aggregates. Several studies have evaluated intra-cluster connectivity by optical imaging of firing patterns. [6,14] The electrode materials of the sf-gMEA are optically transparent and relevant to imaging analysis. Thus, both the incorporation of multiple electrodes into each microroll and the combination with optical measurements will allow further investigation of complex dynamics in the modular 3D network.
Apart from in vitro electrophysiology, the proposed selffolding graphene-based film can be utilized for other biological applications. The film consists of highly biocompatible, transparent, mechanically flexible, and chemically inert materials. These traits can be leveraged for functional elements such as biosensors, cell-scaffolds, and implantable devices. For instance, modifications of graphene with biomacromolecules have been extensively studied for biosensing of proteins and DNA, [59,60] while cell encapsulation within the factionalized graphene could enhance their biosensing performance due to the close proximity between the graphene surface and cells. In in vivo applications, the spatially restricted cell aggregation within the folded structure enables us to construct transplantable cellular tissues with defined architectures. In particular, when neuronal constructs are transplanted into the brain, the porosity on the surface of the folded film allows the establishment of functional connections between encapsulated neurons and the surrounding neurons as described in Figure 3. Furthermore, the inner and outer graphene can potentially be used as electrodes to record and stimulate neuronal activity to investigate the formation of synaptic graft-host connections. These wide-ranging examples highlight the broad applicability of this platform.

Conclusion
We have demonstrated that a self-folding graphene/parylene-C/graphene trilayer film can be implemented as a foldable electrode to investigate the firing dynamics of modular 3D networks. The better adhesion of the lower graphene to parylene-C layer than the upper graphene contributes to the sloped internal strain which initiates the self-folding. In addition, the inner graphene has good cell affinity and can be utilized as an electrode surface. Notably, the cell encapsulation approach is advantageous for long-term and multisite electrophysiological measurements from 3D aggregates thanks to the spontaneous formation of 3D aggregates inside the folded electrode. The unique measurement capability showed that the modular 3D neuronal culture that recapitulated inherent brain architectures exhibits dynamically rich activity patterns compared with a 2D homogenous culture. We believe that this integration of graphene bioelectronics with 3D assembly techniques will pave the way towards precise measurements of complex 3D brain-like tissues to gain insight into their functional dynamics associated with brain development and neurological disorders.

Experimental Section
Fabrication of Self-Folding Graphene/Parylene-C/Graphene Film: The fabrication process is illustrated in Figure S1 (Supporting Information). A 1 wt% solution of alginate-Na (Sigma-Aldrich) was spin-coated on a glass substrate at 2000 rpm for 50 s. The sample was immersed in a 100-mM solution of calcium chloride for 30 min to induce the gelation of Ca-alginate. Polycrystalline monolayer graphene that had been CVD-grown on Cu(copper)-foil was purchased from Graphene Platform. The Raman spectrum of pristine graphene contained no apparent peak in the D-band, indicating there were few defects in the graphene. The graphene was transferred onto the Ca-alginate layer by the conventional poly(methyl methacrylate) (PMMA)-assisted method. 6 wt% PMMA in anisole was spin-coated on the graphene at 2000 rpm for 40 s. The PMMA-assisted graphene was released from the Cu-foil by floating on Cu etchant (45% FeCl 3 solution, Sigma-Aldrich) and then transferred to a water bath to wash off remaining etchant. The floating PMMA-assisted graphene was transferred onto the Ca-alginate layer. The PMMA was removed by immersing the sample in acetone for 1 h. Subsequently, the graphene was coated with 50-to 200 nm thick parylene-C using CVD (SCS, LABCOATER PSD2010). In this process, dichloro-di(p-xylylene) was vaporized at 175 °C and then pyrolyzed at 690 °C to generate a chloro-p-xylylene monomer. Chloro-p-xylylene was condensed on the graphene surface with the vacuum pressure of 35 mTorr at room temperature. In the case of fluorescent labeling of self-folding films, fluoresceinisothiocyanate-conjugated PEI (FITC-PEI) was coated on the surface of parylene-C. The surface was covered with an additional parylene-C thin layer to avoid fluorescence quenching with the upper graphene. A solution of FITC-PEI was prepared by dissolving 10 µg mL −1 FITC-isomer in a solution of 0.1 wt% PEI. After that, another graphene sheet was transferred onto the parylene-C layer using the PMMA-assisted wet transfer method. Finally, the stacked layers were patterned by RIE with O 2 plasma through a photoresist mask to create the array of rectangular films. Positive photoresist (S-1813, Rohm and Haas) was spin-coated on the upper graphene at 2000 rpm for 40 s and then heated on a 100 °C hot plate for 2 min. The photoresist was photolithographically patterned using a mask aligner (MA-10, Mikasa) with a Cr mask or a maskless exposure system (PALETTE, Neoark). The UV-exposed photoresist was developed by a solution of MICROPOSIT 351 developer (Rohm and Haas). Two cycles of etching with power of 100 W and an O 2 gas flow rate of 30 sccm for 6 min were conducted to etch the stacked layers. The residual photoresist mask was removed with an acetone solution.
Fabrication of Self-Folding Graphene-Based Microelectrode Array: The fabrication process is illustrated in Figure S12 (Supporting Information). Preparation of the sf-gMEA began with the formation of a Ca-alginate layer on a 0.7 mm thick 50 × 50 mm 2 glass substrate. It was followed by the transfer of 8 × 8 mm 2 graphene onto the Ca-alginate twice to form the lower graphene layer. The sample was coated with 100 nm thick parylene-C using the CVD process. Sputtering with a power of 100 W and a pure Ar flow rate of 20 sccm (ULVAC, VTR-150 M/SRF) formed thin films of Cr (10 nm) and Au (100 nm) on the parylene-C layer. Photolithographical resist patterning and metal yielded patterns of interconnect lines and connection pads. Positive photoresist (S-1813) was spin-coated and then photolithographically patterned using a UV exposure system (UV Kub, BEAMS). The thin films of Cr and Au were etched by a ceric ammonium nitrate-based Cr etchant (Sigma-Aldrich) and an iodide-based Au etchant (Kanto Chemical), respectively. The remaining photoresist was removed with an acetone solution. The connection pads were aligned with intervals of 1275 µm to match the position of 64-ch probes built into the connector of the MEA control system. A 12 × 12 mm 2 graphene sheet was transferred onto the surface of the sample to form the upper graphene layer. The graphene surface was coated with a phosphate buffered saline (PBS) solution of 0.1 wt% PEI at 4 °C overnight to enhance cell adhesion to electrodes. The stacked layer structure composed of Ca-alginate, lower graphene, parylene-C, and upper graphene was etched by RIE with oxygen plasma through a S1813 photoresist mask to form the array of self-folding electrodes. The electrode position was aligned with the end of interconnect lines using a mask aligner (PEM-800, Union Optical). The shape of electrode was a 200 × 400 µm 2 rectangle with 50 × 50 µm 2 cross-shaped hinges which overlapped the interconnect lines. The array had 64 recording electrodes (1 mm in inter-electrode distance) and four reference electrodes. It was coated with 1 µm thick parylene-C for passivation of interconnect lines. The top parylene-C layer was etched by RIE with O 2 plasma through a S1813 photoresist mask to expose the regions around the electrodes and connection pads. The photoresist mask was thickly formed by spin-coating at 1000 rpm for 40 s because the thick passivating parylene-C layer had to be etched. The mask pattern was aligned with the electrode pattern on the substrate using the mask aligner. Three cycles of RIE with power of 100 W and an O 2 gas flow rate of 30 sccm for 5 min were performed. Immersion in acetone dissolved the photoresist mask. A glass ring (20 mm in diameter) was attached to the substrate to create a reservoir for the culture medium. For bonding of the glass ring, polydimethylsiloxane (PDMS, Silpot 184, Dow Corning Toray) prepolymer was coated on its edge. After mounting the class ring at the center of the substrate, the PDMS was cured at 70 °C for 2 h.
Raman Spectroscopy: The state of strain in the graphene was characterized using Raman spectroscopy (Renishaw, in Via Qontor) under 532 nm excitation with a standard gating (1800 lines mm −1 ). The microrolls were suspended in a PBS solution in a glass bottom dish (Matsunami glass). The microrolls were viewed through a 50× lens. A laser power of 100 mW was used; the spectra were recorded with an acquisition time of 0.25 s and accumulated five times. Spectra in the 1275 to 2825 cm −1 range were obtained at 900 points (1 × 1 µm spacing) from six independent microrolls for each condition.
To evaluate the strain in graphene, shifts in G (Δω G ) and 2D (Δω 2D ) modes of the spectra were analyzed using a customized MATLAB (Math Works) code. The spectra were smoothed by a five-point moving average and then fitted with a Lorentzian. The positions of positive peaks in the ranges from 1560 to 1660 and 2625 to 2725 were detected as ω G and ω 2D , respectively. Vector decomposition in a ω g -ω 2D space was used to evaluate the uniaxial strain ε g in graphene. [40] The origin of the ω g -ω 2D space was set at a point O (ω g 0 = 1581, ω 2D 0 = 2677). To extract the contribution of the strain for a given point, P (ω g , ω 2D ), OP (Δω G , Δω 2D ) was decomposed: OP = a e T + b e H , where a and b are constants, and e T and e H are unit vectors for tensile strain ((Δω 2D /Δω G ) ε = 2.2) and hole doping effects ((Δω 2D /Δω G ) n = 0.70), respectively. The uniaxial strain ε g was estimated using the uniaxial strain-sensitivity of the G mode, Δω G / Δε g = −23.5 cm −1 /%, obtained in previous work. [61] The estimated strain was averaged from the 900 observed data to evaluate the magnitude of strain for each film.
Cell Culture: Cell culture was performed as described in previous work. [36,62] Primary hippocampal neurons and cardiomyocytes were dissected from the hippocampi and hearts of 18 d old Wistar rat embryos (Charles River Laboratories). The tissues were dissociated into single cells using trypsin. The dissociated cells were suspended at a cell density of 5 × 10 6 cells mL −1 in a culture medium. The culture medium for neurons was Neurobasal medium (Thermo Fisher Scientific) supplemented with 0.5 × 10 −3 m glutamine (Sigma-Aldrich), 25 × 10 −6 m glutamine (Sigma-Aldrich), 50 µg mL −1 gentamicin (Thermo Fisher Scientific), and 2% B-27 supplement (Thermo Fisher Scientific). The culture medium for cardiomyocytes was Dulbecco's modified Eagle's medium (DMEM, Thermo Fisher Scientific) containing 10% fetal bovine serum (FBS, Thermo Fisher Scientific) and 1% penicillin-streptomycin (pe-st, Corning). All animal experiments were approved by the Biological Safety and Ethics Committee of NTT Basic Research Laboratories (approval ID 2019-05), which are in compliance with the Guidelines for the Proper Conduct of Animal Experiments of the Science Council of Japan (Kohyo-20-k16-2, 2006). C2C12 cells (mouse myoblast cell line) were purchased from ECACC. The cells were cultured in DMEM containing 10% FBS and 1% pe-st while maintaining the undifferentiated state.
For cell encapsulation within a microroll or folded electrode of the sf-gMEA, the surface of the film was treated with a solution of laminin for neurons and collagen for cardiomyocytes and C2C12 cells immediately before seeding the cells. The suspended cells were seeded 2 min prior to the dissolution of the sacrificial layer. Addition of 10 × 10 −3 m EDTA initiated the folding process, which was completed in 2 min. For imaging experiments, the cell-laden microrolls were transferred onto a glass bottom dish that had been coated with a solution of PEI/ laminin for neurons and collagen for cardiomyocytes and C2C12 cells. After the cell encapsulation, C2C12 cells were induced to differentiate into myoblast and form myotubes by replacing the culture medium with DMEM containing 2% horse serum (Thermo Fisher Scientific) and 1% pe-st. For the cell culture on the sf-gMEA, neurons were cultured in the culture medium supplemented with 10 ng mL −1 brain-derived neurotrophic factor (FUJIFILM Wako Chemicals), 10 ng mL −1 glial cell line-derived neurotrophic factor (peprotech), 25 ng mL −1 neurotrophin-3 (Sigma-Aldrich), and 50 ng mL −1 nerve growth factor (Sigma-Aldrich) during the first week of culture. From 7 DIV, the half volume of the medium was replaced with a fresh one every two or 3 d. All the cultures were maintained in a CO 2 incubator (5%; 37 °C; humidified air) except for during the recording experiments.
Electrophysiological Measurement: Electrical recording and stimulation were performed using an MEA control system (MED64-Basic System, Alpha Med Scientific), which consists of a main amplifier, a head amplifier, and a connector. The sf-gMEA was connected to the amplifiers via the connector. The recorded signals were amplified with a gain of 2000 and passed through a band pass filter at 100-2000 Hz. The analog signals were digitized at 20 kHz. The recording environment was maintained at 37 °C and 5% CO 2 using a stage-top mini-incubator (TOKAI HIT). The sample with the sf-gMEA was immediately moved from the culture incubator to the recording mini-incubator. Additionally, it was incubated for 5 min before recording experiments to stabilize the recording environment and reduce variance of neuronal activities. Spontaneous activities were recorded for 5 min every 2 or 3 d from 7 DIV. For simulation experiments, biphasic pulse (amplitude 20 mA; duration 0.2 ms) was injected to an electrode.
Spontaneous activities of neurons were analyzed using a custom MATLAB code. A 5 min trace containing spontaneous activity was used to analyze characteristics, including the number of active electrodes, firing frequency, and NB. The characteristics were calculated from the traces of 64 electrodes and then were averaged for each MEA. The representative value at each DIV was obtained by averaging the value of multiple MEAs. The process of analyzing spontaneous activities was as follows. Negative peaks beyond a threshold (5× standard deviation) were detected as spike activities. For spike sorting, the spikes were sorted into spike trains of individual neurons using a spike sorting method as previously reported. [19,63] Active electrodes were considered as those presenting a firing frequency (spikes per second) higher than 0.1 Hz. Network bursts (NBs) were detected from spike trains as previously described with minor modifications. [19,64] The spike trains were partitioned into 25 ms bins. In each bin, the product of the number of spikes and the number of electrodes exhibiting spikes was calculated. Consecutive bins in which the product exceeded a threshold were defined as NBs. The threshold was set to the 95th percentile of all products or 10. When the product was 0 in two consecutive bins, the time point was defined as the NB end time. The ratio of electrodes exhibiting spikes during NBs to the total number of electrodes was calculated as a fraction of electrodes participating in NBs. To compare it between different culture conditions, the values obtained from all detected NBs in an MEA were averaged to take a representative value for each MEA.
Statistical Analysis: All data were expressed as mean ± standard deviation (SD). In Raman spectroscopy, the sample sizes (n) were six films for each bar plot. In extracellular recording experiments, the sample size for analyzing firing frequency was more than 100 electrodes. The sample sizes for analyzing NB properties were eight (modular 3D/ folded), six (modular 3D/flat) and three (homogenous 2D/flat) MEAs. Statistically significant differences (***p < 0.001; **p < 0.01; *p < 0.05) were determined using Welch's t-test. All statistical analyses were performed using MATLAB Statistics and Machine Learning Toolbox.

Supporting Information
Supporting Information is available from the Wiley Online Library or from the author.