Monolithic three-dimensional neural probes from deterministic rolling of soft electronics

Abstract

Cognition and behaviour rely on coordinated activity from neural circuits distributed across three-dimensions. However, typical probes for recording neural activity in the brain are limited to two-dimensional interfacing due to the planar semiconductor fabrication process. Here, we report a rolling-of-soft-electronics approach to create monolithic 3D neural probes with high scalability and design flexibility. Compared to previous stacking or assembly methods, the approach directly transforms a planar device into a 3D probe by leveraging the softness of flexible electrodes. The electrode shanks are initially fabricated in a single plane and then connected to flexible spacer. By varying features of the planar design, such as shank pitch and spacer layer thickness, the device can then be deterministically rolled into several versatile 3D probe designs containing hundreds of electrodes. With the system, we demonstrate single-unit spike recording in vivo in rodent and non-human primate models. We also show that the probe can provide microscopy-like 3D spatiotemporal mapping of spike activities in the rodent visual cortex, with five-week-long recording stability and promising 3D decoding performance of visual orientation

The brain is comprised of highly interconnected networks of varying length scales, which consist of spatially distributed and functionally linked regions that continuously communicate with each other^1,2. Consequently, fully decoding cognition and underlying brain functions requires recording neural activity with high spatiotemporal resolution across three dimensions^2-6. Over the past few decades, single-unit (SU) recording has been established as the gold standard for measuring neuronal activity, with penetrating neural probes serving as the primary tools for measuring SU activities². These are typically silicon probes which, due to the maturity of the manufacturing technologies developed for the semiconductor industry, can be made reliably and at reasonable cost^7,8. However, semiconductor fabrication is a planar process and monolithic silicon probes are therefore typically constrained to two-dimensional (2D) interfaces, resulting in a dimensional mismatch with the three-dimensional brain. For example, multi-shank Michigan-type probes typically contain only a planar electrode array orthogonal to the cortical surface when implanted. Utah-type probes similarly have only a planar array parallel to the cortical surface.

Stacking multiple individual planar probes to create a 3D system has been recently explored for both silicon and polymer devices^3,8-16. Additionally, other approaches — such as 3D popup assembly^4,17,18, ultrathin threads^19,20 , Paradromics array²¹, and direct-print 3D probes²² - have provided valuable advancements. However, it remains challenging to create flexible 3D neural probes that are monolithic and have similar scalability as modular silicon probes due to the superiority of planar semiconductor fabrication technologies.

In this Article, we report a scalable, rolling-of-soft-electronics (ROSE) approach to create monolithic 3D neural probes with high scalability and design flexibility. Unlike stacking or other assembly techniques, this approach transforms a planar device into a 3D probe by leveraging existing semiconductor fabrication technologies and the bendability of flexible electrodes. In addition, the use of flexible devices reduces the mechanical mismatch between the implanted probes and surrounding brain tissue compared to rigid silicon probes^23,24, which can contribute to device failure from immune response or even neuron death.

ROSE starts with a fully-fabricated device on a flexible substrate, with all the electrode shanks on the same plane. Using a PDMS spacer and micro-mandrel, the device is rolled from a planar design to a 3D design, with the final device geometry predetermined via varying the number of shanks, the shank pitches and spacer thickness. We demonstrate several geometries of 3D probes containing up to 256 channels from rolling of different planar designs. Systematic insertion studies and mechanical modelling reveal that avoiding slipping-induced instability is critical to achieving successful insertion and preventing buckling, which was used to guide the shank design.

We validate the ROSE probe in vivo in rodent and non-human primate (NHP; rhesus monkey) models. Neuronal recordings from the ROSE probes with hundreds of electrodes show high yields of SU spike recording and clear temporal responses across distant areas of the visual cortex. Immunohistological studies in rodent models also show that ROSE probes reduce tissue stress and inflammatory reactions compared to traditional stiff silicon probes. Semi-chronic recording of the visual cortex in mice over five weeks illustrates the long-term recording capability of the ROSE probes. Additionally, using ROSE probes in awake rodent models reveal 3D-distributed orientation tuning and higher decoding performance using the 3D-recorded units compared to those from a single electrode plane.

Rolling of soft electronics (ROSE) approach

As illustrated in Fig. 1a, the ROSE approach starts with a fully-microfabricated, planar neural probe device on a flexible Kapton substrate (thickness ~ 25 μm), designed to include all the electrode shanks in a row with the desired number of electrodes and their positions on each shank, separated by a chosen pitch. To facilitate rolling, all interconnects are routed and grouped at one end of the shank row and eventually form a connector pad matrix (e.g., 16×16 for 256-ch devices) on the same device substrate. Our prototype rolling leverages a 500-μm-diameter micro-mandrel to stick to the far end of the shank row and roll towards the interconnect end. To control the radial electrode spacing among the rolled loops, we bond a Polydimethylsiloxane (PDMS) layer on the planar device to the base region connecting the shanks to serve as a soft spacer. The planar device eventually rolls into a monolithic 3D neural probe (ROSE probe) with a cylindrical form factor where all shanks evenly distribute along a spiral (Fig. 1b-d). The detailed rolling process is described in Methods and Supplementary Fig. 1. Due to the straightforward rolling process, the ROSE probe offers unique scalability advantages from its monolithic nature compared to existing 3D probe systems that deploy a stacking approach (Supplementary Table 1). While some existing work (e.g., Paradromics array²¹, direct-print 3D microelectrodes²²) has demonstrated impressive scalability with high-density arrays, the ROSE probe’s monolithic and flexible construction provides several distinct benefits. First, ROSE allows for a customizable, true 3D arrangement of electrodes, enabling tailored volumetric recordings for different applications. Secondly, beyond the specific layout of the electrode shanks and interconnects, ROSE leverages thin-film flexible electronics with rather conventional circuit design and fabrication (Supplementary Fig. 2), poised to a robust and cost-effective manufacturing that can be advantageous for volume production, as we have witnessed in Si chips. Thirdly, the use of flexible materials like Polyimide, with Young’s modulus two orders of magnitude lower than Si and tungsten, holds great promise in mitigating the mechanical mismatch between implanted probes and surrounding brain tissue. For example, our Finite-element analysis (FEA) simulation shows significantly lower strains induced by brain micromotion in shanks made of Polyimide compared to those made of Si (Supplementary Fig. 3 and Supplementary Note 1), an important factor known to reduce immune responses²⁵. Another interesting method for transforming planar precursors into 3D devices is the 3D ‘popup’ approach, which has demonstrated complex 3D structures and the ability to record neural activities from dorsal root ganglion (DRG) and spheroids^4,17,18, but is not currently suitable for intracortical neural recordings.

Fig.1 ∣. Rolling of soft electronics (ROSE) leads to monolithic 3D neural probes.

a, Schematic of a monolithic 3D neural probe (ROSE probe), with its final structure of ROSE probe (top right), and the explosive view of a shank structure schematic (bottom right). b, Optical image of a ROSE probe with 64 shanks and 256 electrodes. Inset: magnified view of the electrode shanks. c-d, SEM images of the ROSE probe in $b$. Inset: magnified view of an individual shank which has 4 microelectrodes. e, Photograph of a ROSE probe connected with a 256-ch MagMatrix connector. f, Impedance histograms of a 64-shank ROSE probe (256-ch) before and after rolling with a 200-μm-thick PDMS spacer. g, 3D impedance colormap of the ROSE probe in $f$, with the Z axis being the electrode number (from bottom to top).

Due to the lack of compact and reusable connectors for large-scale soft devices, we designed a printed-circuit-board (PCB)-based connector named MagMatrix to connect ROSE probes and rigid data acquisition electronics (Fig. 1e, Supplementary Fig. 4). Details of the MagMatrix connector are described in Supplementary Note 2. Impedance characterization has revealed negligible difference before and after rolling (Fig. 1f, Supplementary Figs. 5 and 6) for our prototype 64-shank, 256-ch devices with an average impedance of 239.1±70.2 kΩ at 1 kHz and with a yield of 92.2%, suggesting that the ROSE probe possesses the same performance as its planar device precursor. The impedance of functional channels in a ROSE probe has also shown excellent uniformity (Fig. 1g).

The design flexibility of the ROSE approach

A unique advantage of the ROSE approach is its general applicability to a wide variety of planar designs of flexible devices, which through rolling, realize 3D soft probes with precisely controlled dimensions and electrode configurations in three dimensions. This property highlights the robustness of this methodology for highly customizable designs based on experimental and anatomical demands. We successfully achieved ROSE probes with different sizes and densities by varying device parameters, including the number of shanks, shank pitches, and the PDMS spacer thickness, all at the 256-ch level (Fig. 2a-d, Supplementary Fig. 7 and Supplementary Table 2). Moreover, the design can be further extended by varying the length of individual shanks, forming customizable 3D geometries to serve specific recording needs (Supplementary Fig. 8).

Fig.2 ∣. The ROSE approach leverages conventional microelectronic design flexibility and is highly deterministic.

a, Schematic illustration of the ROSE probe showing the shank pitch and PDMS spacer. b-c, Optical images of 256-ch, 64-shank ROSE probes with versatile shank pitches and PDMS thicknesses. d, Optical images of 256-ch ROSE probes with custom shank design, which includes 8, 4, 2 and 1 electrode per shank, respectively. All probes are rolled with 200-μm thick PDMS spacers. Insets: detailed microscope images of individual shanks (top left) and top views (with respect to the perspective in $a$) of the probes (bottom right). Scale bars: 200 μm for individual shanks and 500 μm for the top views. e-g, Comparisons of the rolling loops and offsets between experimental and theoretical results from rolling using different shank pitch. Data was derived from 32 shanks in a single probe for each pitch, shown as mean ± SD. ($e$), PDMS spacer thickness ($f$), and different shank design/number ($g$) for the 256-ch ROSE probes. h, Maximum principal strain contour (${\epsilon }_{\text{max}}$) of the electrodes shown in three individual shanks (left) and in the interconnect layer within the rolled base (right). The left shows three individual shanks (Shank No. 1, 32, and 64, see Supplementary Fig. 10b) in a same 64-shank ROSE probe. i, Average electrode impedance and probe yield comparison of the same device before and after rolling with different PDMS spacer thicknesses. Data was derived from 256 electrodes in a single ROSE probe, shown as mean ± SD.

As shown in Fig. 2e-g, the rolling process is highly deterministic, yielding a predictable number of loops and shank positions. Based on the theoretical prediction by an “Archimedean spiral” model, we approximately calculate the top view structure of the ROSE device (Supplementary Note 3). By knowing the 2D design of shanks, the initial rolling radius, and the PDMS spacer thickness, the coordinates of each specific shank in the spiral structure, and hence all electrode positions, can be determined and designed before rolling. The exact method to attain the offsets between experimental and theoretical results is described in Supplementary Fig. 9.

The ROSE process is also reversible and non-destructive. To ensure that small radii of curvature do not negatively influence the recording quality, we further investigate the strain originated from rolling. FEA results revealed that rolling influences only the base region of the shanks. In contrast, the strain of electrodes in the shanks is negligible (Fig. 2h) due to the narrow and long nature of the shanks, which mechanically isolates the electrode region from being severely deformed during rolling. Even in the rolled base region and innermost region of the ROSE structure (with the largest curvature), the highest strain in the electrode interconnects is still less than 3% (compressive, Supplementary Fig. 10), far less than the fracture strain of gold thin films on a flexible substrate^26,27. Theoretical modeling also confirms the gradual decrease of strain from the inner to the outer area of the spiral base due to the increasing radius of curvature, in excellent agreement with the FEA results (−2.82%, −1.38%, and −1.03% in theory versus −2.66%, −1.19%, and −0.94% in FEA for the #1, #32, and #64 shank, respectively, see Supplementary Note 4 for details). Since the rolling operation results in negligible strain, the ROSE process preserves electrode performance, further vindicated by no impedance degradation from the planar device to ROSE probes with various PDMS spacer thicknesses (Fig. 2i). Next, we characterized the robustness of the ROSE probe. Device failures might occur under extreme mechanical stress, with the ‘neck’ of the ROSE probe, which connects the rolling base and the pad area, being most likely undertaking such stress. FEA simulation results indicate that the device ‘neck’ is safe to experience a bending radius of up to 1.12 mm and can be twisted up to 180° (Supplementary Fig. 11 and Supplementary Note 5). Besides simulation, we also validated the reliability of the device ‘neck’ with cyclic bending tests up to 10,000 cycles (Supplementary Fig. 12) and demonstrated a consistent device yield and average impedance value after 10,000 cycles of bending. In addition, after being soaked in a phosphate-buffered saline (PBS) solution at 77°C for 10 days (equivalent to 160 days at 37°C according to the accelerated aging calculation method specified in ASTM F1980), the device’s yield dropped less than 10% (N = 4, Supplementary Fig. 13), indicating the strong reliability of the ROSE probe. Maintaining stable electrode impedance is a crucial first step for achieving chronic brain recordings.

Insertion dynamics of the ROSE probe

To shed light on the implantation process, we used 64-shank ROSE probes with varying shank pitch and PDMS spacer as a model system and studied their insertion dynamics with brain phantoms. A representative quantitative force dynamic measurement in 0.6% agarose gel models reveal the whole insertion process with distinctive phases: dimpling, puncturing, inserting, resting, and retracting (Fig. 3a, Supplementary Fig. 14). Upon touching the gel surface, the contact force initially increases almost linearly as the vertical probe displacement dimples the gel surface. Once the probe displacement reaches a certain threshold²⁸, the contact force first saturates then enters a ‘force drop’ region, indicating the shank tips start puncturing into the gel. Notably, for ROSE structures with appropriate shank density, each 1.5-mm-long shank has been able to insert into the gel without any insertion aids at an insertion speed of 0.5 mm/min. After all shanks successfully puncturing the gel, the contact force increases almost linearly again as shanks insert deeper, presumably resulting from the increase of friction between the gel and the shanks²⁹. The high-density shanks in the ROSE probe all successfully penetrate the brain phantom without any visually observed buckling. After stopping the 2-min long insertion at a final displacement of 1 mm, the force exhibits stress relaxation of the hydrogel material during a 3-minute resting phase (Supplementary Fig. 15) due to the recovery of the deformed gel, dropping approximately by 2/3 at the end of this phase. Upon retracting, the force eventually decreases and crosses over zero before reaching the minimum, then slowly goes back to zero when the shanks withdraw from the gel at the same speed. Details of the insertion force dynamics are also displayed in Supplementary Movie 1. When comparing the force dynamics of a ROSE probe and its planar device precursor under the same insertion conditions (Fig. 3b), the ROSE probe exhibits a larger puncture force (defined as the peak of the contact force in the puncturing phase) with a later puncture. On the other hand, the insertion forces are significantly smaller, suggesting strong effects from the inter-shank coupling. We further studied the insertion dynamics, especially for puncture forces from ROSE probes with different PDMS spacer thicknesses (100 – 400 μm, Fig. 3c and Supplementary Fig. 16a) and shank pitches (200 – 500 μm, Fig. 3d, and Supplementary Fig. 16b). Overall, ROSE probes with lower shank density (from wider shank pitch or larger PDMS spacer thickness) enter the puncturing phase earlier than those with denser shanks, consistent with the early puncture from the planar probe device. While the shank pitch doesn’t significantly affect the slope of contact force increase during the inserting phase, larger PDMS spacer thickness increases this slope, eventually leading to larger puncture forces. Except for the cases of 100-μm-thick spacer (Fig. 3c) and 200-μm shank pitch (Fig. 3d), which failed the insertion and damaged the gel seriously, all probes succeeded in the insertion. A theoretical model is developed to estimate the puncture forces, as shown in Supplementary Note 6. The puncture forces of all inserted cases are in great agreement with the theoretical estimations (Fig. 3c-d). We note that those successful designs correspond to exceedingly high shank densities (Supplementary Table 2), and well meet high-spatial-resolution extracellular unit recording requirement¹⁷.

Fig.3 ∣. Insertion dynamics of ROSE probes and robustness of aid-free insertion.

a, Insertion force dynamics of a 64-shank, 256-ch ROSE probe with a 300-μm thick PDMS spacer and a 300-μm shank pitch. Inset images are the real-time photos of the probe at different insertion stage. The concentration of the agarose gel was 0.6% and the insertion speed was 0.5 mm/min. Scale bar: 1 mm. b, The comparison of insertion force dynamics between a 64-shank planar probe (110-μm shank width, 300-μm shank pitch) attached on a glass slide and a 3D ROSE probe from the same planar probe design (rolled with a 200-μm thick PDMS spacer). c-d, Comparisons of the puncture forces between experimental and theoretical results from rolling using different PDMS spacer thickness ($c$) and shank pitch ($d$). The experimental results were extracted from the results in Supplementary Fig. 16 while the method to obtain the theoretical results are given in Supplementary Note 6. e, Schematic illustration for pre-puncturing stage of a ROSE probe, highlighting both successful shank puncturing into the gel and shank-slipping-induced failure if the force before puncturing is too large. f, The scaling law of the ROSE-probe insertion considering slipping with different PDMS spacer thicknesses and shank pitches compared to experimental observations. The blue dashed curve is calculated by the equation $F_c(a,b)=F_{total}(w,d,f_s)$, where $F_c$ is the critical contact force for slipping with inner diameter $a$ and outer diameter $b$ of a hollow cylinder, $F_{total}$ is the total insertion force of the probes with shank pitch $w$, PDMS spacer $d$ and the puncture force for a single shank $f_s$. Red (indicating failure) and green (indicating success) points are from experimental observation. g, Average electrode impedance and probe yield of a 256-ch ROSE probe (300-μm thick PDMS spacer and 300-μm shank pitch) as a function of insertion cycles without any insertion aids. Data is presented as mean ± SD.

The mechanism behind the failure in extreme cases is intriguing. Theoretical calculations indicate that the buckling force threshold of a 1.5-mm-long Kapton shank is approximately 3.58 mN (Supplementary Note 7), which is about one order of magnitude larger than the puncture force for a single shank, estimated to be F_p ~ 0.46 mN from the insertion experiments using 10 widely separated shanks (Supplementary Fig. 17). Therefore, conventional buckling is less likely to cause failure in the current design. It turns out that slipping between the shank tip and the gel surface can happen much more easily when the contact force reaches a threshold, leading to shank deformation and accelerated collapse (Fig. 3e). A theoretical model approximating the ROSE probe shank arrays’ contact behavior as a hollow cylinder (Supplementary Note 6) is developed to predict the critical contact force for shank slipping. The combination of shank pitch and spacer thickness that indicate the onset of slipping before insertion can be analytically obtained by equating the critical contact force of slipping with the puncture force. It is shown in Fig. 3f that this theoretical scaling law (dash line) strongly agrees with experimental observations (green dots for successful insertion, red dots for failed cases). This scaling law of the ROSE-probe insertion provides design guidelines for controlling the density of ROSE probes to avoid slipping between the shank tips and the gel with such shank design, thereby avoiding subsequent bending and collapse of the shanks. When designed appropriately under the guidance of this scaling law, ROSE probes are highly robust for insertion. Cyclic brain phantom insertion tests (Supplementary Fig. 18) of a ROSE probe with up to 1,000 insertion cycles led to minimal changes in the average impedance and the electrode yield, indicating the excellent reliability and reusability of the ROSE probe during implantation (Fig. 3g).

While it is challenging to fully mimic the complexity and heterogeneity of the brain using gel phantom, we believe gel insertion tests remain highly informative for studying probe penetration, especially for insertion dynamics up to the puncture point^30-32. However, it is worth noting that significant differences between the insertion dynamics of brain mimics, in vivo brains, and ex vivo brains have been reported³³. In our case, insertion of the same ROSE probe into the gel and brain showed a similar trend but also non-negligible differences (Supplementary Fig. 19), presumably due to the more severe dimpling on the brain surface compared to the gel.

In vivo electrophysiology

To validate the ROSE probe in vivo, we first implanted 128-ch ROSE probes (32 shanks with 4 electrodes per shank) into the rat visual cortex (Fig. 4a, Supplementary Fig. 20a). For this study, we maintained the ROSE shank thickness at 31 μm and shank width around 70 μm. This design consistently penetrated rat brains with negligible bending or buckling, while our insertion tests indicate that with the same thickness, ROSE shanks as narrow as 30 μm can be inserted into rat brains without stiffening aids, showing the potential for further miniaturization. Meanwhile, ROSE probes can also potentially be made ultra-thin, similar to the recent development of super compliable devices^19,34-36, and coated with temporary stiffeners ^37-39 for the insertion.

Fig.4 ∣. Demonstration of high-yield, microscopy-like 3D recordings and their durability from ROSE probes in rodents in vivo.

a, Illustration (left) and photos (right) of a 32-shank, 128-ch ROSE probe implanted in rat visual cortex. Scale bar: 0.5 mm. b, Representative broadband signals from one shank (electrodes A–D, tip to bottom); 100 sorted spike waveforms per single unit (SU) are overlaid (white: average waveform). c, Spontaneous SU firing rates across cortical layers (mean±SD, N = 5 rats). d, LFP spectrogram and spike raster during visual stimulation show 66% SU yield and increased high-gamma power and firing rates in 64 channels (p < 0.05, N = 180 trials). e, 3D spatiotemporal spike rate mapping in Layer II/III reveals stimulus-evoked response propagation across visual areas, data is averaged from 180 stimulation trials. f, SU yield over 5 weeks in 5 mice, normalized to number of functional electrodes in week 0. g, Spike amplitudes across 5 weeks from one mouse show no significant change (p = 0.11, One-way ANOVA). The box is determined by the 25th and 75th percentiles of data while the whiskers indicate 1.5 times the interquartile range. The asterisks represent the minimum and maximum values while the horizontal lines and small squares within the box indicate the median and mean. No significant changes in spike amplitude were found across 5 mice, 5 weeks (p > 0.05). h, Unit similarity (Mahalanobis distance) between within-week (gray) and across-week (orange) SU clusters; values below DM = 1.05 (5th percentile of the entire distribution) indicate consistent tracking. i, Example of 4 stable SUs tracked across 5 weeks; 150 waveforms per week are overlaid, white lines show average. j, Histology at 4 weeks shows comparable NeuN-stained neuron density around ROSE and Si probe holes, with reduced Iba-1 microglia activation near ROSE probe. Scale: 100 μm. k, Quantification of fluorescent intensities in concentric 10 μm bins. ROSE probes show similar NeuN density (p = 0.81) but significantly lower Iba-1 activation vs. Si (p = 8.23e-38, N = 13 for ROSE, 37 for Si probe, Two-way ANOVA). “ns” = not significant; ****p < 0.0001. Data shown as mean ± SEM.

ROSE probes demonstrated high-fidelity electrophysiological recordings, capturing both local field potentials (LFP) and extracellular SU spikes (Fig. 4b, Supplementary Fig. 21). Generally, the recordings exhibited a noise floor of ~ 7 μV_rms and captured ~1.2 well-isolated SUs per channel. When recording from different animals, we varied the insertion depths from 0.8 to 1.5 mm, which spanned a depth range from Layer II/III to Layer V. By averaging the spontaneous firing rate at various depths, we observed a lower spontaneous firing rate in Layer II/III, with relatively higher rates in Layer IV and V, except at the interface between these 2 layers (Fig. 4c). These observations are consistent with the previously reported laminar recordings in rat visual cortex⁴⁰, thereby verifying the precise placement of ROSE probes in the targeted locations.

The ROSE probes also captured robust visually evoked responses. A typical response to visual stimuli includes increased gamma-band oscillations in LFP and a higher spike firing rate⁴¹. From the visually evoked recordings, we isolated 79 SUs across 71 channels, while statistical analysis revealed that data in 64 channels exhibited significantly higher high-gamma power and increased spike firing rates following the onset of visual stimulation (Fig. 4d). Such a high-yield recording performance has been consistently achieved using ROSE probes, with a 64% SU yield on average (Supplementary Fig. 22), a competitive performance with the leading results from state-of-the-art neural probes (Supplementary Table 3). The remarkable performance of ROSE probes in acute settings is particularly noteworthy given its large channel count and high shank number relative to the size of the rodent brain^15,42-46. For instance, when compared to recordings in rodents using 3D probes assembled from stacking, the ROSE probe demonstrated greater SU yield at a much higher shank density (Supplementary Table 1).

To our knowledge, ROSE probe is the first monolithic device that achieves simultaneous layer-specific and multi-regional SU recordings. In previous studies, shank-based neural probes are typically restricted to mapping spike activities within a linear or cross-sectional brain region^2,47-50. To capture higher dimensional information, multiple insertion sites with several independent probes are needed^5,20,51,52, but often with low probe density and requiring careful alignment and accurate electrode localization. Another way of realizing multi-regional recording is to assemble multiple planar probes together^3,14,16,19. However, it requires the stacking of multiple individual probes and their connecting PCBs. Meanwhile, monolithic devices such as the Utah array (UEA) have demonstrated compelling simplicity and have advanced in regulatory approval for human studies, but UEAs lack depth profiling and are limited by electrode density and throughput⁶. For example, UEAs used for rodents are usually with only 4×4 channels^42-44. In contrast, the ROSE probe’s monolithic, flexible design supports 128 high-density electrodes implanted in rodents simultaneously, covering a 3D volume of the brain while maintaining high-yield in vivo performance.

Based on the position of the implanted probe, we mapped the spike activities recorded in Layer II/III onto the visual cortex of rat⁵³. Initially, we observed a latency of about 100 ms between the stimulation onset and the initiation of evoked responses in the monocular area of V1 (V1m). From 100-ms onward, we observed a gradual increase in the number of spiking channels and spike firing rates in the monocular area of the secondary visual cortex (V2m) and binocular area of V1 (V1b) (Fig. 4e). Eventually, at about 300 ms, the evoked spiking activity progressed to the lateral area of V2 (V2l). This phenomenon is consistent with the observation made previously by voltage-sensitive dye imaging⁵¹ that visually evoked events are usually initiated from V1m and then propagate in both directions to the V1 binocular area (V1b) and V2 (Supplementary Fig. 23). To some extent, ROSE probe has demonstrated microscopy-like 3D spatiotemporal mapping of spike activities (Supplementary Movie 2). The results indicate the potential use of the ROSE probe to enable high spatiotemporal resolution in three dimensions and multi-region hierarchical processing of neural activity.

After achieving consistently high acute performance, we went further to investigate the long-term recording capability of ROSE probes. We implanted 16-shank, 128-channel ROSE probes into the visual cortex of mice for a semi-chronic recording study spanning 5 weeks (N = 5). Similarly, we collected visually evoked activities over 5 consecutive weeks following the probe implantation. Overall, considering the initial number of functional electrodes in each probe, the SU yield across all 30 recording sessions was 43% ± 10% (mean ± SD). The mean number of SUs identified per session was 45 ± 13 (mean ± SD). Upon analyzing the results, we observed no decline in SU yield between week 0 (33 ± 10%, mean ± SD) and week 5 (39 ± 7%), with recording performance peaking at week 3 (51 ± 4%) (Fig. 4f). Spike amplitude and the waveform shape are established metrics for assessing the stability of the electrode-tissue interface⁵⁴. Here, statistical analysis revealed no significant differences in spike amplitudes across 5 weeks from individual mice (Fig. 4g). The baseline noises also remained consistently low throughout all weeks and animals (Supplementary Fig. 24). We envision such a semi-chronic stability will enable numerous behavior studies in animals and even temporary use in humans. Previous studies indicate that chronic inflammation around microelectrodes persists, leading to neuronal cell death and process loss by 4 weeks post-implantation^55,56. Therefore, the stable SU yield, consistent spike amplitudes, and stably low noises observed over 5 weeks suggest significant promise for extended chronic recordings.

To further evaluate the validity and stability of these over-month-long recordings, we conducted a unit similarity analysis (see Methods). Unit similarity was quantified using Mahalanobis distances (D_M) between the centroids of SU clusters⁵⁷. The distribution of DM between all SU pairs represents the overall similarity of sorted units in each session (Fig. 4h). Across the 5 weeks of recordings from 5 animals, the 5th percentile of these distributions consistently exceeded 1, suggesting a good separation of the sorted and curated SUs. Such a similarity measure is critical, especially in chronic recordings, to ensure the reliability of recorded SUs, as device shunting in the brain can lead to overcounting of units. Using the 5th percentile of the resulted distribution as a threshold, we identified trackable SUs from week 1 to week 5 with consistent spike amplitudes and waveforms (Fig. 4i). The ability to maintain stable, month-long SU recordings facilitates many long-term neuroscience studies, as experiments investigating sensory responses or behavioral tasks often rely on stable SU activities over weeks to uncover novel insights^58-60. We also note that these stable SU recordings were made possible through numerous iterations of ROSE probes and in vivo neural engineering efforts, which are documented in this work to share with the research community (device engineering: Supplementary Table 4, in vivo neural engineering: Supplementary Table 5).

We hypothesize that the over-month-long recording stability is attributed to the flexibility of the ROSE shanks. To investigate it quantitatively, we conducted a histological study for up to 4 weeks comparing the cortical tissue reactions to both ROSE probes and a traditional Michigan-style Si probe (NeuroNexus) (Supplementary Fig. 25). Representative histology images of the tissue around the probe show that ROSE array resulted in similar neuron density in both the 1-week and 4-week brain samples compared to the traditional Si probes, with no statistical significance found when comparing individual bins between two groups (Fig. 4j, Supplementary Fig. 26a). For Iba-1 marked microglia intensity, although both groups have elevated microglia presence within tissue approximately 100 μm from the probe hole, ROSE group exhibits overall significantly lower microglia intensity than the Si group (Fig. 4k, Supplementary Fig. 26b). As first responders in the brain, microglia are recruited to the implant site as a response to the implant injury and the presence of the foreign body. Higher Iba-1expression usually indicates more microglia accumulation and activation and is commonly associated with more tissue damage and inflammation. Based on this comprehensive study, we conclude that the flexibility of the ROSE array allows for less tissue stress, thereby reducing inflammatory reaction compared to traditional stiff Si probes.

While 16 shanks appear to be the current maximum for rodent arrays⁴², higher shank numbers are required for NHP and human applications⁶¹. To shed light on the potential brain damage caused by a scaled-up version of the ROSE probe, we conducted tissue volumetric displacement analysis in an ex vivo pig brain (Supplementary Note 8). Overall, we observed comparable insertion footprints to the original probe design, suggesting a minimal tissue disruption caused by the implantation of a 256-shank ROSE probe.

To demonstrate the impact of ROSE probes’ 3D recording capability, we investigated layer-specific neural biology on orientation tuning in the mouse V1 by employing a 16-shank ROSE design with 8 electrodes per shank, spanning from Layer II/III to Layer VI (Supplementary Fig. 27). Similar to previous analysis, we identified 67 SUs (61% SU yield), with 57 units having significantly higher firing rates post-stimulation (Supplementary Fig. 28). We then examined the orientation preferences by analyzing evoked spike firing at each stimulation angle. The recorded units displayed diverse orientation preferences and tuning bandwidths (Fig. 5a, b). We adopted the previously reported metrics to characterize the orientation tuning effect^62,63, including the orientation selectivity index (OSI) and the direction selectivity index (DSI). In summary, we identified 52 units showing clear orientation tuning curves (OSI > 0.25), with a major preference for 90° and 270°, consistent with previous reports highlighting the dominance of cardinal orientations⁶³ (Fig. 5c). The OSI of recorded units ranged from 0.25 to 0.9, following a log-normal distribution (Fig. 5d). Meanwhile, most of the recorded units exhibited a low DSI, with only 4 out of the 52 units showed a DSI over 0.5. The low DSIs are expected as only ~20% of cells in the mouse visual cortex are direction selective⁶⁴.

Fig.5 ∣. 3D ROSE probes facilitate neural decoding of visual orientations in the awake mouse.

a, Peri-stimulus time histograms (PSTHs) from 4 representative SUs reveal heterogeneous temporal responses to visual stimulation (Dashed line: stimulation period, data are averaged from 1,440 trials). b, Left: orientation-tuning curves for the same units; OSI = (R_pref – R_orth) / R_pref, DSI = (R_pref – R_oppo) / R_pref, where R_pref stands for the response to the preferred angle; R_orth stands for the response to the angle orthogonal to the preferred angle; Roppo represents the response to the angle 180° from the preferred angle. Right: mean (white) and individual (grey) spike waveforms (100 spikes per unit). c, Histogram of orientation preference shows a cardinal-angle bias. d, Orientation-selectivity histogram follows a log-normal profile, confirming unbiased SU capture by ROSE probe. e, 3D mapping of preferred orientations (left) and selectivity (sphere size, right) illustrates a “salt-and-pepper” organization. Electrodes are placed from the shank bottom (layer H) to the tip (layer A). f, Latent variables learned from 1,200 trials (52 orientation-selective units, two 500-ms bins per trial) clusters stimulus orientation distinctly. g, Predicted angles versus true angles for held-out trials (n = 240) align closely except for 180° reflections (opposite moving directions), consistent with limited direction selectivity of recorded SUs. h, Mean decoding error using all 52 units versus units confined to a single electrode plane (10-fold cross-validation). Full 3D sampling yields a 6.4° mean error (excluding 180° errors), while any one-plane subset exceeds 20° errors. i, Fraction of predictions within 0° and 15° error for the same training paradigms: the full 3D dataset achieved 71 % exact and 96 % ≤ 15° accuracy on average (10-fold cross-validation), outperforming any of the planar subsets.

Leveraging the 3D nature of ROSE recordings, we examined how orientation tuning is distributed across 3D coordinates (Fig. 5e). After analyzing the data from each unit, our results supported the “salt and pepper” organization^65,66 over the existence of local orientation biases⁶⁷, with neurons tuned to different orientations interspersed across all spatial axes. We also observed some units in proximity were tuned for similar orientations, but concluding a local orientation bias would require higher-density recordings and more extensive statistical analysis. Meanwhile, we observed a similarity of OSIs (0.46 - 0.53) recorded from different depths, suggesting that orientation discrimination tasks were performed across cortical layers from Layer II/III to Layer VI (Supplementary Fig. 29). As the first monolithic 3D probe, the ROSE probe has proven its exceptional utility in revealing the orientation selectivity of 67 single neurons in a 3D volumetric distribution. Within one single insertion, we captured 52 orientation-selective units across 5 visual cortical layers and mouse V1, LM, and PM areas, forming a diverse, unbiased dataset.

The rich information content from ROSE 3D recording also facilitates neural decoding, which translates neural activities into real-world outputs by leveraging deep learning algorithms. The decodability of a recording serves as advanced benchmarks for neural recording technology - the more useful information a recording contains, the more precisely it can be decoded. Hence, we investigated how the 3D recording capabilities of the ROSE probe perform in decoding the orientations of visual stimuli^68,69. In this study, we adopted a state-of-art neural encoding method⁷⁰ to perform dimensionality reduction on neural data and generate learnable neural latent embeddings. After substantial optimization of the training parameters (Supplementary Fig. 30), the learned latent variables from 1200 trials of data formed distinctive clusters for each orientation (Fig. 5f), with angles 180° apart clustered in proximity, reflecting the similarity of spike firing patterns under the same orientation but opposite directions of stimuli. Interestingly, the clusters formed a circular shape in the angle sequence despite the model not being informed of the physical meaning of the training labels.

The inferred latents were then used for decoding purposes. Our results show that the orientation of visual stimuli can be decoded with a low average error of 6.4°. While the orientation decoding is accurate, we noticed that for certain angles, the decoder differentiates the moving directions of stimuli poorly, leading to a prominent 180° error (Fig. 5g). This error is presumably due to the small proportion of direction-selective neurons recorded in this experiment. Overall, 71% of the predicted orientations had 0° error, and over 96% of the predictions fell within a 15° error margin (55% and 72% if considering the directions). Considering the number of units (52) used for the decoding, we have achieved a promisingly low error compared to the existing literature^19,71. When compared to any single shank, plane, or cross-sections of units, we demonstrated a significantly higher decoding performance using the 3D recorded units, indicating the non-redundant information these units contain and highlighting the advantage of 3D recording for high-precision neural decoding (Fig. 5h, i, Supplementary Fig. 31, 32). This result is significant because, with the same number of shanks and similar implant volumes, the UEA can only record from a single electrode plane. We envision this capability to be highly impactful for advancing brain-computer interface (BCI) applications and understanding the intricate dynamics of the brain.

In addition to rodents, we tested a large-scale ROSE probe (256-channel with 32 shanks and eight electrodes per shank) in the visually responsive cortex (V4/Inferior Parietal Lobule) of a rhesus macaque monkey (Fig. 6a, Supplementary Fig. 20b). Similarly, we analyzed the recorded signals across multiple frequency bands (Fig. 6b, Supplementary Fig. 33). The 256-channel ROSE probe provided a high-resolution 3D view of visually evoked LFP responses (Fig. 6c, Supplementary Fig. 34). Each shank’s 8 electrodes captured laminar profile of the evoked LFPs, revealing a detailed picture of current flow through multiple cortical depths (Fig. 6c). By examining these laminar profiles, we localized the peak evoked response and identified a region of strong current sinks at electrode layer E. Thus, we constructed 2D current source density (CSD) plots focused on the specific electrode layer (Fig. 6d), which highlights how stimulus-driven currents propagate across the cortical surface. This spatial visualization underscores the complexity of cortical circuitry and the advantage of the ROSE probe’s dense 3D sampling capabilities.

Fig.6 ∣. Intracortical recording using large-scale ROSE probes results in high-yield SU spike detection in the rhesus monkey cortex.

a, Illustration of a 32-shank, 256-ch ROSE probe implanted in the macaque monkey’s visual cortex (V4/IPL area). Inset on the top right indicates the probe’s placement within brain, while the bottom right inset details the single-shank layout, with electrodes labeled from A to H. b, Example raw data recorded from monkey brain, displaying 50 seconds data from 24 channels (1 row per channel) extracted from 3 shanks (8 channel per shank, shown from shank tip to bottom). c, Upper: laminar evoked LFP from one representative shank. Black dashed line denotes the onset of visual stimulation. Lower: current source density (CSD) laminar profile derived from the same data. Red (negative values) indicates extracellular current sinks and blue (positive values) indicates extracellular current sources. Evoked LFP responses from electrode ‘B’ and ‘E’ (locations corresponding to observed current sinks) are superimposed on the CSD plots for reference. d, 2D CSD plots from a single electrode plane (layer ‘E’), where the evoked LFP responses peak. Averaged LFP data 1s before (left) and after (right) visual stimulation was extracted from 32 electrodes, then interpolated for data visualization. The CSD plots clearly shows the local current flow patterns from the outer shanks towards the inner shanks at the specific plane. Gray dots denote the physical location of each shank in the electrode plane. e, Auto-correlogram (colored) and cross-correlograms (pink) for 8 example SUs, plotted over a −75 to + 75 ms window (bin size = 1 ms), showing diverse SU spiking patterns in a large-scale 3D recording. 150 spike waveforms per unit are overlaid, with the white line representing the average spike waveform. A 3D spike firing rate map (bottom left) visualizes captured SUs and their firing rates over a 2-minute recording window, with asterisks marking the locations of the 8 example SUs. f, spiking raster from 2 trials of visual stimulation, with spiking activity from each individual shank coded with same color. Dashed lines denote the ‘on’ and ‘off’ times of visual stimuli.

Beyond LFP, the ROSE probe also demonstrates high-yield SU and MU detections. The large-scale probe captured SUs across a 3D configuration with distinct spiking patterns (Fig. 6e). To summarize, the ROSE probe accomplished a 69% multi-unit (MU) yield and 44% SU yield out of the 223 functional channels in a total 20-minute recording (Supplementary Fig. 35). The MU band signals also confirmed the robust visually-evoked responses with a 61% evoked MU yield (Fig. 6f), consistent with the known visual responsiveness of V4/IPL. Together, these results, along with our rodent experiments, highlight the ROSE probe’s ability to reliably capture high-quality, 3D electrophysiological signals across multiple frequency bands in both small and large animal models.

Conclusions

We have reported a rolling-of-soft-electronics approach that transforms a planar monolithic array of probes into 3D neural probes by deterministic rolling. Insertion studies confirm that millimetre-length-scale ROSE electrode shanks can penetrate the brain without any temporary stiffening, and revealed the slipping-induced insertion failure mechanism, providing guidelines for ROSE design and fabrication. Layer-specific, multi-regional recordings in the brains of rodents and NHP models further validated ROSE probes with high in vivo performance and stability.

Our monolithic 3D soft probes could enable several advances through microscopy-like spatiotemporal spike mappings. For instance, in brain regions such as the prefrontal cortex, neurons extend over many millimetres in both parallel and perpendicular directions to the cortical surface^47,72, , making a 3D recording approach crucial for studying interlayer and interpatch relationships²¹. Moreover, the 3D nature of recordings will allow more precise brain–computer interfaces by sampling neural activity from superficial layers, deep layers, and sulcal regions^30,32. Due to its design flexibility, the ROSE approach allows for customizable 3D geometries to serve specific recording needs, similar to 3D printing. It also provides a highly deployable pathway towards microelectrode array scaling. For example, the ROSE approach could be integrated with complementary metal-oxide-semiconductor (CMOS) technology to achieve very large-scale integrated 3D probes, with thousands or tens of thousands of electrodes in miniaturized footprints. On the other hand, it is also applicable to optoelectronic shank-based probes, enabling multimodal, closed-loop brain interrogation from combining electrophysiology and optogenetics.

Another intrinsic advantage of the ROSE probe is the softness of its electrode shanks. This property can be further exploited to enable clinical applications such as neuroprosthetics. Flexible probes fabricated from polymer substrates, such as polyimide, Parylene C and PDMS, have been shown to provide better mechanical compliance for more reliable neural^37,44,73-77 Notably, even though the ROSE probe is flexible, we observed in several experiments low-yield brain recordings, especially when the shank width is wider than 130 μm. This is presumably due to the large volumetric displacement during the probe implantation. Future research will focus on optimizing the implantation process of the ROSE probe, upgrading probe functionality, achieving cohesive interfaces among different probe materials, miniaturizing its connector for improved chronic stability, and developing bio-integrating coatings, with the aim to enhance the chronic in vivo stability of ROSE probes and to establish it as a biomedical device for translational applications.

Methods

Design and fabrication of planar electrode arrays.

The planar array comprised five layers: a Kapton/PI bilayer substrate, Au/PEDOT:PSS electrodes and interconnects, and SU-8 encapsulation. Each tapered shank (20° tip angle) is 1.6 mm long, 65–70 μm wide, and 31 μm thick, with four 10 × 20 μm² electrode sites spaced 50 or 100 μm apart, located 90 μm from the tip. Kapton film (25 μm, DuPont) was first laminated on a PDMS-coated (30 μm, 10:1 Sylgard 184) glass slide. A PI layer (5 μm, PI-2545) was spin-coated and cured at 250 °C in N₂ for 70 min. Cr (5 nm)/Au (100 nm) was deposited via e-beam evaporation, with Cr as the adhesion layer. Interconnects were patterned using S1805 PR (3.5 μm resolution); for 2.5 μm features, LOR 1A (150 nm) and S1813 PR (1.3 μm) were used. SU-8 2005 (4.5 μm) was applied for encapsulation and hard baked at 180°C for 30 min. PEDOT:PSS was electrochemically deposited with a potentiostat (Gamry Reference 620)⁷⁸. Devices were rinsed with DI water, coated with Dextran (4:1 with DI water) for laser protection, then laser-cut (U4, LPKF). After rinsing off Dextran, samples were soaked in acetone for release from PDMS and lifted off with tweezers.

Rolling of soft electronics (ROSE) approach.

Before rolling, a micro-mandrel was glued to the end of the planar array opposite the interconnect using 3M Vetbond. A 400 μm-thick, 2 mm-wide PDMS slab (10:1 ratio) was bonded to the base to act as a spacer between loops, typically aligned under the first shank. Rolling was performed counterclockwise toward the interconnect end using a custom machine and took ~5 minutes. Kapton tape on the back (see Supplementary Fig. 1a) secured the final structure. The mandrel could be removed by 20 s of sonication in acetone (Fig. 1b).

ROSE probe insertion study with brain phantoms.

A 0.6% agarose gel (BP160-100, Fisher Scientific) was prepared by mixing the powder with DI water and heating to 140 °C until clear. The solution was poured into a 5.5 cm × 1 cm cylindrical mold and cooled overnight at room temperature to solidify. Shanks were not stiffened with any coatings. Insertion tests were performed at 0.5 mm/min using a Mark-10 test stand, with force measured by an M5-012 gauge (0.1 mN resolution). For cyclic tests, the gel was fixed on a Thor Labs LTS150/M stage with the device on the moving platform (0.5 mm/s, 1 mm insertion, 8 s/cycle). Impedance was measured every 100 cycles after rinsing with DI water. For planar arrays, the base was secured to a glass slide to avoid bending.

In vivo rat survival surgery.

All procedures were approved by Dartmouth’s Animal Care and Use Committee. Male Long Evans rats (10–12 weeks, 300–350 g; Charles River) were used and housed under a 12-hour light/dark cycle. No other experiments were conducted on the animals and surgical tools were sterilized beforehand. Rats were anesthetized with 2–3% isoflurane (ISO) and 1.5% O₂ in an induction chamber, then transferred to a stereotaxic frame with non-traumatizing ear bars. Throughout the surgery, rats were maintained on a 37 °C heating pad and received 1–2% ISO with 2% O₂ via a nose cone connected to a charcoal filter. Eye lubricant was applied. After shaving the scalp between eyes and ears, the area was cleaned with betadine and 70% IPA for three times. A straight incision was made with a No. 11 scalpel, and the scalp was held open with hemostats. Fascia and blood were cleared with sterile saline and cotton swabs. A 5 mm craniotomy centered 7.0 mm posterior to bregma and 3.5 mm lateral was drilled over visual cortex. The drill bit kept moving with less than 2 s per spot to avoid thermal damage. The skull was cooled with saline, and bone debris was cleaned with sterile swabs.The bone flap was carefully removed without damaging the cortex. The remaining debris was cleared with saline and swabs. Durotomy was performed using a 30G needle to incise the dura, which was then peeled with forceps to avoid dimpling during probe insertion. No major bleeding occurred; any minor bleeding was fully controlled before insertion.

In vivo monkey survival surgery.

ROSE probe functionality was tested in the visual cortex of an anesthetized female rhesus monkey (20-22 years) using aseptic surgical procedures. Monkeys were part of an aging study at Boston University and housed individually at the AAALAC-accredited Laboratory Animal Science Center. All procedures were approved by the BU IACUC and followed NIH guidelines. Anesthesia was induced with ketamine (10 mg/kg, IM) and maintained with propofol (0.3–0.4 mg/kg/min, IV). The head was stabilized in a stereotaxic frame, and a ~5 mm × 5 mm burr hole was drilled over the right occipital lobe. After incising the dura, the ROSE probe was inserted into V4/parietal cortex using a micromanipulator. Following recordings, animals were euthanized under deep anesthesia, and the brain was perfused and collected.

In vivo mouse survival surgery for awake, head-fixed recording.

All procedures were approved by the IACUC of Weill Cornell Medical College. Female C57BL/6 mice (9-12 months, 25-35 g; Jackson Laboratory) were used for craniotomy, implantation, and recordings. Mice were housed on a 12-hour light/dark cycle, with no additional experiments performed. Anesthesia was induced with 5% isoflurane and maintained at 1-2%. Vital signs were monitored (Starr Life Sciences oximeter), and body temperature was kept at 37 °C using a heating blanket (RWD Life Science). A 4 mm cranial window was opened over the visual cortex (2.5 mm lateral, 3 mm posterior to bregma), leaving the dura intact. The window was sealed with sterilized PDMS and dental cement, and a head plate was attached. Two weeks post-surgery, mice underwent 3 days of head-fixation habituation training.

In vivo mice survival surgery for semi-chronic recording.

The procedures used for semi-chronic mice surgeries were approved by the Animal Care and Use Committee of Dartmouth College. Both male and female C57BL/6 (10~13 weeks old, Charles River) were used for the study. All animals were housed in 12-hour light/dark cycle with ad libitum access to the food and water. 4 hours prior to surgery, dexamethasone (2mg/kg; i.p.) was injected to animals to prevent edema and reduce inflammatory response. All tools for surgery were sterilized and surgery area was cleaned with 70% Ethanol. Mice were placed in an induction chamber with 2% oxygen with 2% isoflurane. Hair was shaved from between the eyes to ears and eye ointment was applied to prevent eyes from drying. Mice were positioned on a stereotaxic frame with continuous flow of 1-1.5% of isoflurane through the nose cone. Once mice are under stable surgical plane, clean midline cut was made following the mid sagittal line. Exposed skull was cleaned with sterile cotton swab and/or back side of scalpel blade. About 4mm-by-4mm cranial window (center at AP: −3.5 mm, ML: 2.5 mm) was opened. Continuous irrigation with chilled saline was applied to reduce accumulated heat from drilling and drill bit was not stayed in one position more than 2 seconds. Cranial window was covered with soft silicone gel (dura-gel) followed by Kwik-sil (WPI). Once the skull was completely dried, chamber base was anchored on the skull with dental cement (3M). After full chamber assembly, we waited for at least one week of recovery prior to device implantation.

Device implantation.

For anesthetized rodents, the ROSE probe was connected to a MagMatrix connector and aligned perpendicularly over the visual cortex. It was inserted to 700–1500 μm depth at 0.05 mm/s using a NeuralGlider, with no bleeding observed. In acute experiments, shanks were covered with Kwik-Sil or Kwik-Cast. In semi-chronic settings, a thin silicone gel (Dow Corning) was applied between shanks for stabilization, followed by Kwik-Sil and dental cement to seal and secure the probe.

In macaques, the probe was fixed to a stereotaxic arm, aligned perpendicularly over the visual cortex via a burr hole, and inserted at 0.5 mm/min for the first 50 μm. After a 1-2 min pause, insertion continued to a final depth of 1.5 mm.

In awake, head-fixed mice, brief isoflurane anesthesia was used to remove PDMS sealing before insertion. The 16-shank ROSE probe was inserted at 0.5 mm/min using a NeuralGlider, avoiding major blood vessels. After insertion, a 5-10 min wait preceded sealing with Kwik-Cast and dental cement. Final depth was calibrated to about 1.3 mm to target cortical layers II–VI.

Immunohistology and Confocal fluorescent imaging.

Non-functional single-shank ROSE probes and A-style Si probes (3 mm length, NeuroNexus) were implanted for 1- and 4-week histology. The Si probe (10 μm thick, 150 μm wide) has a comparable footprint to ROSE shanks (31 μm thick, 70 μm wide). Immunohistology followed established protocols^42,79. At designated timepoints, animals were deeply anesthetized (Ketamine/Xylazine, 90/9 mg/kg) and perfused with 300 ml PBS, followed by 4% PFA. Skulls were removed, post-fixed overnight, and brains were dehydrated in 15% and 30% sucrose, then frozen in OCT. Frozen tissues were sectioned at 25 μm (Leica cryostat), rehydrated in citrate buffer, blocked in 10% goat serum, and permeabilized with 0.1% Triton-X for 45 min.

Slices were stained with NeuN (1:250), Iba-1 (1:500), GFAP (1:500), and Hoechst. Confocal images (20×, Olympus Fluoview 3000) were analyzed with MATLAB. Tissue response was quantified using 25 concentric 10 μm bins centered on the probe site. Neuronal density dropped near the implant and returned to baseline at ~50 μm. Fluorescence intensity or cell counts were normalized to control regions (5% corners of each image). Pixels >1 SD above the mean were excluded to define background; remaining pixels were used to calculate a background threshold (1 SD below mean). Only pixels above this threshold were analyzed and plotted by distance. NeuN-positive neurons were manually counted in ImageJ per bin.

In vivo acute electrophysiology recording.

For acute rat recordings, animals were placed on a water-based heating pad to reduce electromagnetic noise. Recordings were taken during 10 minutes of spontaneous activity (in the dark) and 10 minutes of visually evoked activity. Visual stimuli were generated using MATLAB’s Psychtoolbox and displayed on a 24” LCD monitor (30 cm from the contralateral eye), consisting of randomized sinewave black-and- white gratings (2 Hz, 3 cm bars, 100% contrast, 120 Hz) across 8 directions (0°–315°). Each stimulus lasted 0.5 s, followed by a 0.5–1.5 s gray screen. Synchronization was controlled with an Arduino Mega 2560.

For awake mouse recordings, animals were head-fixed on a treadmill. Visual setup was the same, with 1 s stimulus duration. For decoding, 24 grating angles (15° steps) were presented in random order across 60 sessions (30 s rest between), totaling 1440 trials.

For acute monkey recordings, the heating pad was turned off after a 15-minute wait to reduce noise. A 10-minute spontaneous and ~10-minute visual stimulation session was performed, using the same rat stimulus setup but with 2 s stimuli and 5 s rest. As monkeys' eyes are closed under anesthesia, eyelids were manually opened every 30 s. Notably, monkey recordings showed a discrepancy between multi-unit (MU) and single-unit (SU) yields. Tip electrodes (A–C) had low SU isolation, likely due to placement beyond the cortical layers. Excluding A–C increased SU yield to 58%, while MU yield remained ~70%.

Semi-chronic electrophysiology recording.

The implanted ROSE probes had 16 shanks (71 μm wide, 31 μm thick) within a 2.2 mm diameter. Prior to implantation, probes were sterilized in 70% ethanol for >15 min. A custom 3D-printed chamber was attached to the skull and fixed with dental cement. Using a reusable MagMatrix PCB connector, only the probe and a lightweight acrylic top (< 2 g) remained on the head between sessions. For weekly recordings, animals were lightly anesthetized, then maintained at 1-1.5% isoflurane to enhance spontaneous activity. The MagMatrix connector was attached, and recordings were performed in a dark room - 5 min spontaneous activity followed by 7-8 min visually evoked responses using the same protocol as in acute mouse recordings. Afterward, the connector was removed, the probe was secured to the chamber wall with copper wires, and the chamber was sealed with medical tape before returning the animal to its cage.

Device Characterizations.

The electrochemical impedance of ROSE probes was measured through the built-in impedance measurement function in RHX recording software (Intan technology). In bench testing, we used a threshold of 500 kΩ at 1kHz to define functional electrodes (channels). In brain, the situation is more complicated. We adhered to 3-fold criteria to exclude bad channels: 1) The channel impedance exceeds 10 MΩ in brain. 2) The root-mean-square (RMS) noise in the channel is higher than 20 μV. 3) The recorded LFP coherence with a known-good, neighboring channel is significantly lower or higher than expected. If all 3 criteria are met, the channel will be excluded from the functional channel group. In this work, SU yield is defined as number of channels that recorded at least 1 well-isolated SU divided by the number of functional channels. In long-term recordings, we always adopted the number of functional channels at week 0 for the calculation.

LFP analysis.

Raw data were filtered to 1-300 Hz band for LFP analysis. LFP analysis in this work are done with custom python scripts, including the LFP spectrogram, LFP color plot, current source density (CSD) color plot, etc. 1D CSD data are derived by the second derivative approximation of the LFP data recorded in a single shank, while 2D CSD data are derived from averaged LFP data in 1-second duration on the same electrode plane. Bicubic interpolation was applied to the CSD data for visualization purposes.

Spike Sorting.

The electrophysiology data was pre-processed with custom Python scripts. Data were filtered to 300-6000 Hz band and after common median referencing prior to spike sorting. The spike sorting was performed using MountainSort 5⁸⁰, with actual ROSE channel coordinates mapped to each channel of data. We adopted sorting scheme 2, training duration = 60 seconds, phase1 detection radius = 150 μm, phase2 detection radius = 50 μm, detection threshold = 5.5 for all the sorting process. After receiving the sorting result, we manually curated the sorted unit clusters by the following criteria: Firing rate > 0.1 Hz; Inter-spike-interval (ISI) violation (< 2 ms) < 1.5%; Full-width half maximum (FWHM) of the mean waveform between 0.15-0.75 ms; Peak-to-valley time (PVT) of the mean waveform between 0.15-0.85 ms; Median spike SNR > 4. After applying these criteria simultaneously, we visually inspected the auto-correlogram of each curated unit to exclude any clusters that does not adhere to the expected spike refractory period. Lastly, we conducted a unit similarity measure to prevent over-counting of units from shunted electrodes (See the SU similarity measure and SU tracking session). The clusters that did not pass the entire curation process will be counted as MUs. All SU analyses used the curated spike clusters as described above.

SU similarity measure and SU tracking.

SU similarity was calculated using Mahalanobis distances (DM) between the centroids of SU clusters, a measure based on the vector components of spike waveforms and considers the covariance along each dimension⁵⁷. For each recording, within-session, across-unit similarity was first calculated between all curated SU pairs. If 2 units from the same recording session have unit similarity significantly smaller than 1 (having very similar spike waveforms), we then calculate the Pearson’s correlation coefficient between their auto-correlograms. If the correlation is larger than 0.75, only 1 unit with larger spike amplitude will be kept for further analysis. We then combined the data from all recording sessions to form a distribution. Then, a 5th percentile of the distribution was used as a threshold to identify stable SUs recorded on the same electrode across sessions.

Orientation tuning analysis.

SU spike count in a bin from 50ms to 550ms after stimulation onset were used for the orientation studies. Data from the same angle of stimulation (60 trials per angle) were averaged. Then, orientation selectivity index (OSI = (R_pref – R_orth) / R_pref) and direction selectivity index (DSI = (R_pref – R_oppo) / R_pref) were calculated for each unit, where R_prefstands for the response (mean SU count) to the preferred angle; R_orth stands for the response to the angle orthogonal to the preferred angle; Roppo represents the response to the angle 180° from the preferred angle. After the calculation, the data was summarized in terms of preferred angles and orientation selectivity of each neuron.

Neural decoding.

We used a neural encoding model⁷⁰ (CEBRA) to train spike data from 2 consecutive 500-ms bins after stimulation onset. 83% of data (1200 stimulation trials) were used for training and 17% (240 trials) were saved for testing. The model we adopted is: model_architecture = ‘offset10-model’, time_offsets =10, batch_size =512. Training labels were set as 0 to 23, representing the 24 different angles. Spike data were standardized before training. For decoding, we adopted the K-nearest-neighbors (KNN) algorithm, with optimal k searched between 1 to 20. When assessing the decoding performance, we first calculate the mean angle error of all predictions. Then, we derived the percentage of predictions that falls into 0 and 15° margin. All the analysis was conducted in Python.

Statistical analysis.

All statistical analysis was performed in Python. For characterizing the evoked recording yield, we conducted a Student’s paired t-test, comparing the spikes count (SU/MU) and gamma band power (30 – 100 Hz) in a 550 ms bin size, 50 ms before the onset of stimulation and another 550 ms bin size, 50 ms after the onset of stimulation. If the p-value resulting from the paired t-test is smaller than .05, we conclude the statistical significance of evoked responses. For assessing the recording stability across weeks, we performed one-way ANOVA on the spike amplitude data recorded from each week, with the hypothesis that the mean spike amplitude stays the same across weeks. We then consider a p value > .05 as no significant difference across weeks. For testing whether the 5th percentile of unit similarity distribution exceeds 1 for all recording sessions, we conducted a one-sided binomial test with Bonferroni correction, and considered p < .05 for the statistical significance. For evaluating the neuron density and Iba-1 intensity surrounding ROSE and Si probes, we conducted two-way ANOVA with Sidak’s multiple comparison, and considered p > .05 as no statistical significance between 2 groups.

Inclusion and Ethics.

All animal procedures were conducted in accordance with institutional and national ethical guidelines. Rodent surgeries and electrophysiological recordings were approved by the Animal Care and Use Committee of Dartmouth College and the Institutional Animal Care and Use Committee (IACUC) of Weill Cornell Medical College. Procedures involving non-human primates (NHPs) were approved by the Boston University IACUC and carried out in compliance with the NIH guidelines.

Data availability.

All datasets generated during and/or analyzed during the current study are included with this article. Source data for device characterizations are provided within the paper. Animal electrophysiology data that supports the findings of this study are available on Zenodo (DOI: 10.5281/zenodo.15498935).

Code availability.

The custom Python scripts used to analyze neural signals, and visualize spike activities are deposited in the following GitHub repository (https://github.com/qiangy0819/ROSE_process.git).