Cellular-scale silicon probes for high-density, precisely localized neurophysiology

Abstract

Neural implants with large numbers of electrodes have become an important tool for examining brain functions. However, these devices typically displace a large intracranial volume compared with the neurons they record. This large size limits the density of implants, provokes tissue reactions that degrade chronic performance, and impedes the ability to accurately visualize recording sites within intact circuits. Here we report next-generation silicon-based neural probes at a cellular scale (5 × 10 µm cross section), with ultra-high-density packing (as little as 66 µm between shanks) and 64 or 256 closely spaced recording sites per probe. We show that these probes can be inserted into superficial or deep brain structures and record large spikes in freely behaving rats for many weeks. Finally, we demonstrate a slice-in-place approach for the precise registration of recording sites relative to nearby neurons and anatomical features, including striatal µ-opioid receptor patches. This scalable technology provides a valuable tool for examining information processing within neural circuits and potentially for human brain-machine interfaces.

NEW & NOTEWORTHY Devices with many electrodes penetrating into the brain are an important tool for investigating neural information processing, but they are typically large compared with neurons. This results in substantial damage and makes it harder to reconstruct recording locations within brain circuits. This paper presents high-channel-count silicon probes with much smaller features and a method for slicing through probe, brain, and skull all together. This allows probe tips to be directly observed relative to immunohistochemical markers.

INTRODUCTION

Much of our current understanding of neural functions was gained through electrophysiological recording from individual neurons in behaving animals, one-at-a-time. Yet recording neurons one-at-a-time provides only a very limited view of information processing, which involves rapid interactions and coordination between neurons (Fujisawa et al. 2008). Simultaneous recordings from ensembles of many neurons has been achieved using large numbers of wires (Schwarz et al. 2014), microfabricated electrode grids (e.g., the Utah array; Mitz et al. 2017), or microelectrode arrays shaped through photolithography (Berényi et al. 2014; Jun et al. 2017; Merriam et al. 2011; Scholvin et al. 2016; Shobe et al. 2015). These important approaches have yielded valuable results but nonetheless share substantial limitations, especially for investigating densely packed, locally connected neurons.

To achieve the stiffness needed for a millimeters-long element to penetrate the brain, each element has typically had a width of at least 25–100 µm (with cross section in the thousands of µm²). This large foreign body causes substantial direct mechanical damage, and is also detected and rejected by the brain’s immune system (Biran et al. 2005; Potter et al. 2012; Prasad et al. 2012; Saxena et al. 2013; Winslow and Tresco 2010). The immune reaction leads to loss of neurons in the vicinity of the electrodes, one factor that frequently curtails the duration of chronic neural recordings (Patel et al. 2016). The spacing between large elements must also be large to distribute tissue damage; for example, in a standard Utah array contacts are separated by 400 µm. For many brain regions and cell types, this means that the neurons monitored from different electrodes are unlikely to be in direct communication with each other (Fujisawa et al. 2008; Straub et al. 2016).

Compounding this problem, large devices must be removed from the brain before processing the tissue for histological analysis. Since the recording electrodes are not observed in situ, it has proven difficult to obtain accurate registration between electrophysiological recordings and key anatomical markers. Lack of anatomical registration has impeded research in several subfields, including the investigation of the function of striatal “patches” (irregularly shaped, ∼100-µm scale zones that are detectable through staining for µ-opioid receptors, a.k.a., striosomes; Desban et al. 1993; Graybiel et al. 1981). Although visualization of spatial patterns of ensemble activity can be achieved through other methods, such as multiphoton imaging of genetically encoded calcium indicators, these currently lack the exquisite temporal resolution of electrophysiology (although see Adam et al. 2019).

Several approaches have been used to insert smaller electrodes into the brain. For example, flexible, cellular-scale electrodes can be infused through a micropipette (Liu 2018) and linear polymer probes can be dragged into the brain using a stiff needle or silicon support as a shuttle (Chung et al. 2019; Hanson et al. 2019). However, these insertion devices are typically at least as large as traditional electrodes (25 µm+ in width; although see Luan et al. 2017). The resulting acute “stab-wound” has been shown to cause substantial, permanent neuron loss even if no foreign body remains in the brain (Potter et al. 2012). As an alternative, very thin (<10 µm) electrodes made from carbon fibers can be directly inserted, either one at a time (Kozai et al. 2012), in bundles (Guitchounts et al. 2013), or mounted on silicon supports to increase overall stiffness and penetration (Patel et al. 2015). The latter approach shows promising chronic performance (Patel et al. 2016) but has not yet demonstrated the feasibility of large channel counts.

To overcome these limitations, we designed neural probes with high channel counts but considerably smaller physical shank dimensions than standard silicon devices. We reasoned that the force required to insert an element without buckling depends on its length; by coupling a relatively short (500 µm) electrode-bearing lower portion with cellular-scale cross section (5 × 10 µm) to a more robust upper portion (far from the recording sites), we obtained reliable insertion even deep into the brain. The smaller size probes allow much higher density neural recording, with excellent chronic performance for months (at least) after implant. We further developed a method whereby probes are left in place during brain sectioning and histological processing by slicing through the entire head after decalcifying the skull. This was made possible by the very thin probe shanks that can be easily sliced through without distorting the surrounding tissue. By cutting thick (∼300 µm+) slabs of tissue and using immunohistochemistry to detect key anatomical features, we found that we can directly visualize identified recording sites within intact micro- and mesoscale neural circuits.

METHODS AND MATERIALS

Nanofabrication process flow and probe assembly.

The probes were fabricated at the Lurie Nanofabrication Facility at the University of Michigan. The process begins with thermally growing and patterning a 1.2-µm-thick oxide on a silicon wafer as a masking layer. A 20-µm-deep and 1.2-µm-wide trench is etched using deep reactive ion etching (DRIE) where stiffeners are placed. The etch recipe was optimized to achieve vertical sidewalls and high-aspect ratio trenches. Boron is thermally diffused at 1,175°C, 12-µm deep into the silicon surface in places that were not masked. The mask is removed except for a small section around the probe tip ends. A second, 5-µm shallow boron diffusion is performed to define the thickness of the tips. A stack of electrically insulating silicon oxide and silicon nitride is deposited using low pressure chemical vapor deposition. A 1.4-µm-thick layer of electrically conductive, phosphorous-doped polysilicon is deposited and patterned to form interconnects. For this step, patterns formed by stepper and e-beam lithography (for 256-channel probes) are stitched together. The e-beam was only used to keep the narrow traces separate. The polysilicon traces routing along the tip of the probes were designed to have a 470-nm minimum feature size. Another stack of insulators (identical to those previously mentioned) is deposited. After etching vias, Cr/Pt/Au is sputtered and patterned via lift-off to form bonding pads and electrodes. Lastly, the outline of the probe is defined using DRIE, and any undoped silicon is wet etched in ethylenediamine pyrocatechol. For handling and protection of the fine tips, the devices remain attached to the wafer frame by small tabs. The probes were mounted onto printed circuit boards (PCBs) using Crystalbond 509 (SPI supplies) and wire bonded.

Animal surgery.

All animal procedures were approved by the Institutional Animal Care and Use Committees at the University of Michigan and the University of California, San Francisco. The probes were implanted into adult male Long-Evans rats, weighing 300–350 g. Anesthesia was initialized with 5% isoflurane (vol/vol). The rats were maintained under isoflurane anesthesia, which was continuously monitored using toe pinch and breathing rate, and the flow of isoflurane was adjusted accordingly. The head was shaved at and around the area of the incision site. The shaved area was swabbed using alternating applications of betadine and 70% ethanol. Ointment was applied to the eyes to keep them from drying during surgery. Ear bars were mounted in both ears and fixed in a stereotaxic frame (Kopf Instruments, Model 900). After making an incision, the skin flaps were pulled apart using hemostats and the skull surface was cleaned using cotton swabs and 2% hydrogen peroxide (vol/vol). A burr bit (19008-07, Fine Science Tools, Foster City, CA) was used to drill holes around the periphery of the skull for bone screws (19010-00, Fine Science Tools, Foster City, CA). Reference and ground wires originating from the implant were attached to bone screws using MillMax pins and placed 1 mm caudal to lambda and over the contralateral cerebellum, respectively. Next, a 2 × 4 mm craniotomy was made over the target brain region. The dura was gently resected using dural forceps and hook, and insertion was performed within minutes of resection to avoid excessive swelling. Special care was taken to avoid larger blood vessels on the surface of the brain and if necessary, specific shanks were broken off. Just before the probe tips contacted the surface of the brain, excess liquid was removed to prevent shanks from wicking together. Shanks were lowered at ∼100 µm/s into the brain. Striatal shanks were lowered to ∼4.2 mm below the surface of the brain, leaving a ∼800-µm shank length above the brain surface. Motor cortical shanks were lowered ∼1.5 mm below the surface (at +0.5 anterior-posterior, 3.5 mediolateral relative to bregma). For the cohort of rats used for chronic stability testing, and the pallidal implant, the exposed surface of the brain was sealed with Kwik-Sil (World Precision Instruments, Sarasota, FL; Patel et al. 2015). The motor cortex implant was sealed with cyanoacrylate glue (Loctite SuperGlue Gel control). The rat implanted in striatum for slice-in-place was sealed with a thin layer of DOWSIL (3-4680, Dow Corning, Midland, MI) and petroleum jelly (Vaseline) coated along the shanks (Jun et al. 2017). Finally, the skull was covered with dental acrylic (Hygenic Acrylics, Switzerland).

Recordings and impedance measurements.

PCBs with 64 channel probes were connected to RHD2216 64-channel headstages, and PCBs with 256-channel probes were connected to two RHD2000 128-channel headstages (Intan Technologies, Los Angeles, CA). In some cases, these connections were made using custom flexible polyimide cables (MicroConnex/Carlisle Interconnect), consisting of 30-µm copper trace/spaces, sandwiched between 25-µm-thick polyimide layers, with bonding pads plated with immersion gold. Recordings were made using the Intan interface software (version 1.5.2). The sampling rate was 30 kS/s for the striatal implants, 25 kS/s for the globus pallidus (GP) implant, and 20 kS/s for the motor cortex implant. The sample depth was always 16-bit. Semiautomated clustering was performed using KiloSort2 (Pachitariu et al. 2016) followed by manual curation using the “phy” gui (https://github.com/cortex-lab/phy). Impedances (at 1 kHz) were measured using the Intan headstages. For assessment of spiking across repeated sessions, we examined brief (10 min) epochs as rats were resting quietly. Signals were first subjected to common average referencing and then high-pass filtered at ∼234 Hz using a wavelet transform (Wiltschko et al. 2008). In line with prior studies (Jun et al. 2017), spike events were detected using a threshold of six times the median absolute deviation of the signal, with a 0.5-ms dead time between possible spikes. To prevent contributions from occasional large movement artifacts, portions of the filtered signal that crossed a 1-mV threshold were discarded, as were spike events that fell within a millisecond window in which the mean amplitude across all channels was more than three times the average for the whole session. Analyses did not include channels that were not connected to brain (impedance outside the range 500–4,000 kΩ) due to mechanical failure (e.g., detachment of connector pins after multiple plug-in, plug-out cycles). We estimated the physical location of recorded single-units in each of two dimensions, by finding the peaks of Gaussians fit to average peak spike amplitudes across the electrodes (averaged along, or across, shanks, respectively).

Slice-in-place, histology, and imaging.

Rats were transcardially perfused with 4% paraformaldehyde (PFA) in phosphate-buffered saline (PBS). After perfusion, the rats were decapitated, the skull was stripped of gross tissue, and the jaw was removed. The remaining bone, brain, and probes were immersed in a solution of 0.25 M tetra-sodium ethylenediaminetetraacetic acid (EDTA) in PBS, pH balanced to 7.4, at 4°C. The solution was exchanged daily, with the soak continuing for 2–3 wk until the bone turned rubbery. The skulls were soaked in 30% sucrose in PBS for 72 h for cryoprotection, then the solution was progressively exchanged with OCT (optimal cutting temperature compound). Once the sample was immersed in 100% OCT, it was brought into a vacuum to facilitate penetration into ventricles and then frozen using dry ice. Slicing was performed on a cryostat at −16°C. The skull was mounted such that the most ventral portion was accessed first. The tissue was sliced together with the bone and the probes at 300-µm thickness. The resulting slabs were washed, blocked, and incubated for 7–10 days at room temperature with both primary antibodies Rb ∝ mOR (ImmunoStar 24216) and Ms ∝ NeuN (Millipore MAB377). The slabs were then incubated with secondary antibodies for 3–5 days at 4°C. Before imaging on a confocal microscope (Nikon TI), the slabs were soaked in a refractive index matching agent (TDE; Staudt et al. 2007). The obtained images were processed using ImageJ.

RESULTS

Neural probe design.

Each device consists of 32 shanks in a comb-like configuration (Fig. 1). Each shank has an electrode-bearing tip section (500 µm long) with cellular-scale cross section (10-µm wide × 5-µm thin). We designed multiple probe variations, in which each shank has lengths of 3–9 mm and holds two or eight electrode sites (64 or 256 total channels/probe). Recording sites were 15 µm × 10 µm and separated by 15-µm gaps. Adjacent shanks differ in length by 50 µm to reduce peak insertion force. To enable insertion, probes have a wider upper portion, with a 25-µm width and 11.5-µm thickness. For 6-mm- and 9-mm-long shanks, this upper portion widens gradually to 40 µm at the top.

Fig. 1.

Silicon probes with cellular-scale tips. A: 256-channel probe (6-mmshank version) on a penny for scale. Scale: 2 mm. B: close-up of tip portion showing 32 shanks, each with 8 electrodes. Scale: 200 μm. C: electron micrograph of probe tips, showing electrodes. D: comparison of shanks with 8 electrodes (left; in 256 channel probes) and 2 electrodes (right; in 64 channel probes). Scale: 50 μm. E: smooth transition to stiffener-reinforced upper section. Scale: 10 μm. F: sharpened tip. Scale: 5 μm.

Nanofabrication process.

The shanks were formed from the bulk of silicon wafers, and the interconnects, electrodes, and insulators were patterned thin films deposited onto the wafers (Fig. 2). Materials and processing steps were derived from the Michigan probes process (Tanghe et al. 1990), which uses diffusion of boron into silicon to form etch stops and pattern the probe shape. Each probe design used one or more of the following technical refinements:

Fig. 2.

A: main steps and materials of the nanofabrication process flow: 1) grow and pattern silicon oxide to form a masking layer. Etch trench using deep reactive ion etching (DRIE) for stiffeners. 2) Diffuse boron into silicon in places that were not masked. 3) Remove the mask except for a small section around the tip ends. Perform a 2nd boron diffusion to define tip thickness (a, top view; b, side view). 4) Deposit stack of electrically insulating silicon oxide, silicon nitride, and silicon oxide (O/N/O). Deposit layer of electrically conductive polysilicon and pattern to form interconnects. 5) Deposit another insulator stack (as before). Etch vias, sputter Cr/Pt/Au, and pattern via lift-off to form bonding pads and electrodes. 6) Etch probe outline using DRIE. 7) Etch undoped silicon with ethylene diamine pyrocatechol (EDP). B: close-up of polysilicon interconnects shaped with e-beam lithography (left) and transition to stepper patterning (right). Scale: 5 μm. C: completed wafer. For handling and protection of the fine tips, the devices remain attached to the wafer frame by small tabs. Scale: 2 cm.

First, we patterned very high-density traces within the tip sections via electron-beam lithography, which allows for orders-of-magnitude smaller features (Du et al. 2011) compared with optical patterning. As e-beam lithography is a serial and thus relatively expensive process, we used it only on the lower, thin tip sections of 256-channel devices to define the narrow valleys separating traces. The remaining traces were patterned using conventional optical stepper lithography.

Second, we integrated a stiffener into upper shank sections by selectively increasing thickness there (6-mm, 9-mm shank versions only). Traditionally in this process, neural probes have a uniform thickness defined by the depth of diffused boron in silicon. Here, a trench was etched along parts of the shank before boron diffusion. The sidewalls and the bottom of the trenches were exposed during boron doping, and hence, the thickness of this part of the shank is extended by the depth of the trench (increasing maximum thickness to 30 µm). The trenches were kept sufficiently narrow such that they were completely refilled in subsequent steps.

Third, we produced sharper tips by masking boron diffusion toward the end of the shank (3-mm shank version only). Increasing tip sharpness can reduce compression of the brain during insertion and make buckling of the shanks less likely (Bjornsson et al. 2006). By masking boron diffusion at the very tip, the thickness of the released probe was reduced to that of the remaining insulating layers, around only 1.5 µm.

Successful implantation into shallow and deep brain structures.

After wire bonding to custom PCBs, each finalized probe design was found to successfully insert into brain without requiring specialized surgical practices. During the design stage of this project, test probes (5-mm length, not all shanks present) were inserted to a depth of several millimeters into the motor cortex of an anesthetized rat and then withdrawn. Successful insertion was defined as shanks remaining intact without obvious deflection from a straight path. Test probes without stiffeners were found to insert with some difficulty (n = 2 test probes and 26/28 and 16/28 shanks successfully inserted). Adding stiffeners and the tapering increase in top section width made insertion noticeably easier (n = 2 test probes, 30/31 and 30/31 shanks successfully inserted).

Figure 3A shows images of inserted devices with 3-mm-long shanks, with an extremely narrow pitch of 66 µm, during implantation 1.5-mm deep into rat cortex. We were also able to insert multiple devices, mounted on thin flexible polyimide cables (see materials and methods), into the same hemisphere, targeting both motor cortex with a 6-mm probe and striatum with a 9-mm probe (Fig. 3B). Once inserted, the probes were secured, the site was sealed (see materials and methods), and rats consistently recovered from surgery (Fig. 3C) without behavioral impairments.

Fig. 3.

Probe insertion. A: 2 devices (6-mm, 9-mm lengths) being implanted into the same hemisphere for simultaneous recording from cortex and striatum. Scale: 2 mm. B: device with 66-μm pitch between shanks (and 3-mm length) being inserted 1.50-mm deep into motor cortex. Scale: 1 mm. C: rat after recovery, with a 256-channel implanted device connected to 2 × 128-channel Intan headstages within a 3-dimensional-printed enclosure.

Recording from large neuronal populations.

We successfully implanted 256-channel devices into brain structures both superficial (motor cortex) and deep [striatum, globus pallidus (GP)] and obtained high-quality chronic spike recordings in freely-moving rats from each structure. Figure 4A shows an example from a 9-mm, 256-channel probe targeting GP (3 days postsurgery), yielding 259 distinct simultaneously-recorded single-unit clusters.

Fig. 4.

High-density chronic spike recording. A: left: example of signals from a 256-channel probe (100-μm-shank pitch) targeting rat globus pallidus (depth 6.5 mm), 3 days postsurgery. Traces are a 250-ms duration, unfilteered and grouped by shank (red/black). A, right: spike sorting for this recording session produced 259 clusters. Each cluster is marked by a blue circle at its estimated physical location (see materials and methods). Scale bars: 1 mV/50 ms. B: illustration of the waveform distribution across electrodes for 3 example neurons (purple, blue, green) from the same session as in A. Top: waveforms highlighted in wide-band signals (250-ms duration); bottom: corresponding average waveforms, each visible only on a small subset of electrodes. The top traces are from the middle 3 shanks at bottom. Scale bars: 1 mV/50 ms. C: as in B, but recording from a 256-channel probe (66-μm-shank pitch) placed in rat M1 cortex (18 days postsurgery). M1 cells typically generated wider electric fields, visible across many electrodes over 2–3 adjacent shanks. Scale bars: 1 mV/50 ms.

The dense electrode arrangement both within and between shanks allows examination of the spatial extent of extracellular voltage changes associated with action potentials from each neuron. In our GP example (using 100-µm shank pitch), each averaged action potential was visible only across a small number of recording sites (Fig. 4B). This is consistent with relatively small extracellular fields generated by these neurons, which typically have radially symmetric dendritic fields (Kita and Kitai 1994). By contrast, when we recorded from motor cortex (using a probe with just 66-µm shank pitch), each averaged action potential was readily visible across many sites. This includes both sites along a shank (consistent with recording along the large apical dendrite; Blanche et al. 2005) and across two or even three shanks (Fig. 4C). This suggests that, at least for recording cortical projection neurons, our probes have sufficient density to detect spikes from a large proportion of cells within the two-dimensional field of recording sites.

High-density, chronic spike recordings.

The cellular-scale silicon probes were capable of recording large spikes over extended time periods: Fig. 5A shows an example recording at 139 days postimplant. To quantify changes in chronic performance over time, we recorded from a cohort of rats (n = 4) each with a 64-channel probe implanted in dorsal striatum at 1 wk and 7 wk postimplant (Fig. 5B).

Fig. 5.

Cellular-scale silicon probes record spikes for many weeks. A: example recording from a 64-channel probe, 139 days after implantation into dorsal striatum (250 ms shown). Signals are high-pass filtered and color coded by shank (alternating light and dark gray; 7 shanks shown out of 32 total). Examples of spikes from large units are highlighted in color, with the corresponding average waveforms shown at left (peak-valley ranges: brown: 300 μV; blue: 186 μV; magenta: 190 μV). Green marks below signals indicate all detected spike events (threshold of 6× median absolute deviation, M.A.D.). Scales: left: 1 ms; (right) 300 μV/50 ms. B: performance metrics comparing 1 wk postimplant (6-7 days) with 7 wk (49–50 days). Same implant type, same 6× M.A.D. event threshold as in A. Colored lines indicate individual animals (n = 4 probes, 256 sites total), and black and gray bars indicate probe means (in left plot) or means for working electrodes (right 3 plots).

We saw a slight decline (from 55.8 to 52.0, out of 64) in the mean number of working electrodes (defined as those with impedances between 0.5 and 4.0 MΩ); the remaining sites showed reduced impedance [mean per device fell from 2.2 MΩ to 1.6 MΩ; mean per site from 2.2 MΩ (n = 225) to 1.5 MΩ (n = 211); two-tailed Wilcoxon signed-rank test, P = 1.3e-19]. The mean rate of spike events increased [per device: from 47 Hz to 60 Hz threshold crossings/s/device; per site: from 0.8 Hz (n = 225) to 1.2 Hz/s (n = 211); two-tailed Wilcoxon signed-rank test for sites recorded at both time points, P = 0.0481]. The size of detected spikes was stable [mean per device, 87.0 µV to 85.4 µV; mean per site 84.0 µV (n = 225) to 83.0 µV (n = 211); no change in size for those recorded at both time points; two-tailed Wilcoxon signed-rank test, P = 0.2425]. We conclude that these devices are able to consistently record spikes over many weeks.

Anatomical registration of implanted probes in situ.

To facilitate registration between recording sites and microcircuit features, we wished to avoid pulling electrodes out of the brain at the end of experiments and instead leave electrodes in situ during processing for histology. Many standard electrodes (e.g., tetrodes or tungsten wires) cannot be cut through by cryostat blades without extensive tissue damage (J. R. Pettibone and J. D. Berke, unpublished observations). Some recent electrode designs can be cut through (Xie et al. 2015) but do not preserve the orderly arrangement of electrodes needed for registration between specific recording channels and histological locations.

We found that, following formaldehyde perfusion and decalcification of the skull, we can cleanly cut through skull, brain, and probe shanks of rats with implanted cellular-scale silicon probes (Fig. 6, A and B). By cutting 300-µm slabs we could directly observe the electrode-bearing portion of the shank tips within brain tissue (Fig. 6, B and C), while still allowing antibodies against µ-opioid receptors to fully penetrate the tissue and reveal striatal patches (Fig. 6D). Costaining for the neuronal marker NeuN revealed many neuronal cell bodies immediately adjacent to probe tips, even after prolonged chronic implantation (Fig. 6E), consistent with minimal damage to local circuits. In three-dimensional confocal stacks we could identify the specific depth of shank tips (Fig. 6F) and thereby reconstruct the positions of electrodes that recorded single units, relative to the complex patch shapes (Fig. 6, F and G). This slice-in-place approach thus allows effective direct visualization of electrodes within intact brain circuits.

Fig. 6.

Slice-in-place approach for precise localization of electrodes within brain circuits. A: decalcification of the skull (shown upside-down) after removal of superficial tissue. The dashed line indicates the plane and thickness of slices. B: skull and brain are sliced together, ventral to dorsal, with the probes in situ. Faint gray line (arrow) is the row of shanks. C: example of probe tips within striatum (brightfield image) at about a 4-mm depth. Scale: 400 μm. D: shank tips (white; inverted, thresholded brightfield image) superimposed on immunofluorescence imaging of neurons (NeuN, green) and striatal patches (μ-opioid receptors, red) in the same plane. Scale: 200 μm. E: close-up histology from a chronic striatum implant 35 days survival postsurgery, staining as in D. Patch neurons are visible immediately adjacent to the shanks. Scale: 50 μm. F, i–v: stack of confocal images for the lower sections of the 2 shanks shown in E. Depths correspond to electrode locations and probe tip for one shank (circled; corresponding schematic on top right). Scale: 50 μm. G, i–iv: depth profiles for the average waveforms of 4 example single units (columns) recorded on day 35 from the same 4 electrodes as in F. Scale: 100 μV, 4 ms.

DISCUSSION

We have demonstrated that silicon probes with cellular-scale tip sections can yield long-lasting chronic spike recordings in freely moving animals, large numbers of densely recorded neurons, and direct visualization of probe tips relative to histological markers. This combination of features is likely to be very useful for many systems neuroscience studies, especially investigations of neural circuit functions over a range of spatial scales.

Probes with very thin shanks are likely to produce less damage and improved localization compared with alternative advanced silicon probe designs, which have typically emphasized large numbers of sites on wider shanks (e.g., Blanche et al. 2005; Jun et al. 2017; Rios et al. 2016; Wang et al. 2019). Those designs may nonetheless have advantages for some applications, such as when multiple targets are arranged vertically within the brain (Jun et al. 2017). In other situations, such as recording from a horizontal cell layer, the current design may be more appropriate. The current design is suitable for a range of vertebrate species of mouse size and greater (e.g., all 256 recording sites will fit within the mouse striatum), but future design variations could readily expand the range of potential targets.

Neural probes made using flexible polymer substrates (Chung et al. 2019; Felix et al. 2013; Luan et al. 2017; Xie et al. 2015) can also produce recordings with high longevity. Furthermore, their mechanical flexibility prevents breaking, and they may also be sliced-in-place (Dai et al. 2018). However, their insertion often requires transient mechanical support within the brain by removable shuttles (Chung et al. 2019; Musk and Neuralink 2019), that cause acute (stab-wound) damage and hinder narrow spacing between shanks. Once implanted, flexible devices cannot be further moved (Michon et al. 2016), preventing adjustments based on observed electrophysiological features after recovery, or recording of new sets of cells. Further advantages of silicon include the sophisticated assembly tools developed for silicon technology, and the potential to seamlessly include active electronics, for example, for multiplexed readout (Jun et al. 2017), or LEDs for optical stimulation (Wu et al. 2015).

An alternative technology for recording many neurons simultaneously is optical imaging. This can have several important advantages including direct observation of neuron locations, visualization of distinct neuronal compartments (such as dendrites; Adam et al. 2019), or wide fields-of-view (Garg et al. 2019). However, imaging of large numbers of neurons has to date typically examined superficial structures in head-fixed animals, with calcium fluctuations detected by GCaMP as the activity measure. Calcium dynamics can have a nonstraightforward relationship to spiking, and action potentials remain the gold standard measure of neural “activity.” Imaging action potentials are possible using voltage sensors instead (Adam et al. 2019), but the light intensity requirements for imaging at near-millisecond resolution make large fields of view impractical with current sensors. Recording from deep brain structures in freely moving animals has been made possible using miniaturized microscopes and gradient refractive index lenses (Cai et al. 2016). However, these lenses are typically 0.5–2 mm in diameter, and implantation involves aspirating overlying brain structures, potentially altering neural dynamics.

We deliberately placed recording electrodes far from the relatively thicker, stiffer shank portions, which presumably provoke greater tissue reactions. How far away is far enough? For solid silicon probes with 300-µm width and 15-µm thickness, Skousen et al. (2011) showed that after 2 mo neural density is low in the immediately adjacent vicinity but is normal at a 100-µm distance. In the present design the closest recording site is 250 µm away from the thicker section (for 256-channel devices; 430 µm away for 64-channel devices). We saw no indication that the electrodes further away from the tip (and thus closer to the thicker sections) record fewer neurons (Fig. 5A), suggesting that we have achieved sufficient distance from any damage caused by thicker upper sections. Although the aggregate cross section of our recording tip sections is substantial (1,600 µm², comparable to the single shank of Neuropixels probes), distributing this displacement over many small elements has been shown (Skousen et al. 2011) to avoid deleterious tissue reactions.

An important objective was to improve coregistration of recording sites with mesoscale histological markers. We and others have previously attempted coregistration with µ-opioid-receptor-defined striatal patches by observing tissue damage or gliosis produced by electrodes (White and Hiroi 1998), passing current to produce marker lesions (Berke et al. 2004; Friedman et al. 2015) or coating probes in lipophilic tracer dyes such as DiI (Fujisawa et al. 2008). We have not found these approaches to be highly reliable and effective for identifying patch neurons; for example, dye coating is typically too uneven, while lesions destroy the essential immunochemical markers just where they are most needed. Some studies of patch neurons identified using imaging have begun to appear (Bloem et al. 2017; Yoshizawa et al. 2018). However, our slice-in-place approach offers higher temporal resolution with less damage to nearby brain tissue.

We chose 300-µm-thick slabs for slice-in-place analysis. Thinner sections would include less of the recording zone and require more registration between sections and increase the chance that short shank portions fall out of tissue during slicing. Thicker slabs would avoid these concerns but are harder for antibodies to penetrate and make microscopy more challenging due to increased light scattering. Scattering can be greatly reduced by tissue clearing (Chung and Deisseroth 2013; Ertürk et al. 2012; Hama et al. 2015), but we have found it is hard to avoid measurable shrinkage/expansion as the tissue is cleared, which impedes coregistration. Future improvements to our approach could use thicker slabs (or whole brains) together with enhanced clearing methods that avoid changes in tissue volume and smaller antibodies (e.g., nanobodies; Perruchini et al. 2009) for more effective tissue penetration. The small size of the tips of the shanks might lend itself well to forming three-dimensional electrode arrangements, for example, by stacking (Rios et al. 2016). If brain structures at different depths are targeted for simultaneous recording, the length of each shank and arrangement of can be chosen individually and clusters of shanks can be precisely tailored to their targets.

Overall the cellular-scale silicon probes presented here are part of a new generation of devices for chronic monitoring of large neural populations (Chung et al. 2019; Hanson et al. 2019; Jun et al. 2017; Obaid et al. 2020). Their complementary geometry, small feature size, and the ability to achieve coregistration with histology make these probes suitable for addressing many outstanding questions in systems neuroscience.

GRANTS

This work was supported by National Institute of Neurological Disorders and Stroke BRAIN Initiative awards U01-NS094375 and UF1-NS107659, the National Institute of Mental Health award R01MH11187, the Department of Veterans Affairs Rehabilitation Research and Development Service award 1I01RX001640, the University of Michigan, and the University of California San Francisco. Probe fabrication was performed at the Lurie Nanofabrication Facility, a member of the National Nanotechnology Infrastructure Network, which is supported in part by the National Science Foundation.

DISCLOSURES

No conflicts of interest, financial or otherwise, are declared by the authors.

AUTHOR CONTRIBUTIONS

D.E., J.R.P. and J.B. conceived and designed research; D.E., J.R.P., S.M.L., and C.M.C. performed experiments; D.E. and J.B. analyzed data; P.R.P., D.C., K.G., C.A.C. and J.B. interpreted results of experiments; J.B. prepared figures; D.E. drafted manuscript; J.B. edited and revised manuscript; D.E. and J.B. approved final version of manuscript.