64-Channel Carbon Fiber Electrode Arrays for Chronic Electrophysiology

A chief goal in neuroscience is to understand how neuronal activity relates to behavior, perception, and cognition. However, monitoring neuronal activity over long periods of time is technically challenging, and limited, in part, by the invasive nature of recording tools. While electrodes allow for recording in freely-behaving animals, they tend to be bulky and stiff, causing damage to the tissue they are implanted in. One solution to this invasiveness problem may be probes that are small enough to fly under the immune system's radar. Carbon fiber (CF) electrodes are thinner and more flexible than typical metal or silicon electrodes, but the arrays described in previous reports had low channel counts and required time-consuming manual assembly. Here we report the design of an expanded-channel-count carbon fiber electrode array (CFEA) as well as a method for fast preparation of the recording sites using acid etching and electroplating with PEDOT-TFB, and demonstrate the ability of the 64-channel CFEA to record from rat visual cortex. We include designs for interfacing the system with micro-drives or flex-PCB cables for recording from multiple brain regions, as well as a facilitated method for coating CFs with the insulator Parylene-C. High-channel-count CFEAs may thus be an alternative to traditional microwire-based electrodes and a practical tool for exploring the neural code.

In order to understand information processing in the brain, scientists must be able to take reliable measurements from the central nervous system (CNS). Ideal methods would allow us to record from all neurons in a brain (10 8 in a rat or zebra finch 1 at cellular spatial resolution and a temporal resolution on which neurons operate (kHz range) in behaving animals 2,3 . While established methods for whole-brain monitoring (e.g. fMRI or PET) typically give poor spatial and temporal resolution, state of the art optical methods for imaging of genetically-encoded calcium indicators have yielded cellular spatial resolution while recording all neurons in a larval zebrafish brain (on the order of ~10 5 neurons) simultaneously at low temporal resolution (~0.5 Hz) 4,5 , although faster dynamics are possible with optical methods, including with genetically-encoded voltage indicators or lightsheet microscopy 6,7 .
In contrast to most imaging of calcium indicators, recording neuronal activity using electrodes provides spike-time temporal resolution but low cellular yield (state of the art simultaneously recorded neurons using Neuropixels probes is ~3000 neurons using 8 probes, each of which contains 384 recording sites, yielding roughly one unit per recording site 8,9 . A further limitation of electrode methods is their invasive nature: implanting large electrode arrays damages the tissue the electrodes are meant to record from 10,11 , thus limiting the yield of recorded neurons, the longevity of the recording quality, and the ability to track the activity of individual neurons over long timescales. These are particularly pressing problems for investigations into how the neural code changes during learning, which unfolds over days, weeks or months [12][13][14][15][16][17][18] ; or for the creation of practical brain-machine-interfaces (BMIs), which rely on the stability of recorded signals as well as the underlying neural code in order to decode a patient's thoughts or intended actions 13,19 .
Metal electrodes have been used for recording single neurons since the mid-20th century 20 . These are typically tungsten or PtIR wires, on the order of 5-50 m µ in diameter [21][22][23] . In rodent work, the typical approach involves arrays composed of individual metal wires spun into groups of four (tetrodes) and bundled into groups of up to 40 tetrodes, or 160 channels 24 . Monolithic silicon devices are a prominent alternative 8,25,26 that allows for high-density recording sites and high channel counts at the expense of a large cross-section (e.g. µ × m 70 20 in Neuropixels probes) and high stiffness, which make it a challenge to record on long-timescales (however, see Okun et al. 27 ).

Results
Assembly and tip preparation. In this report, we tested two types of CF arrays: a 16-channel version (assembly of which was reported in 30 , and an expanded 64-channel version. For the latter, we designed a 3D-printed plastic block to hold the 64 CFs (Fig. 1a). As in the 16-channel arrays, carbon fibers were threaded through wells at the top of the block, and down through the exit holes on the bottom. The pictured design features four exit holes, arranged in a 2 × 2 grid, but this arrangement can be modified to an 8 × 2 grid of exit holes (designs available at github.com/guitchounts/electrodes/) or a single exit hole to bundle the 64 CFs into one. In the implants described here, fibers were bundled into one set of 64. After threading, fibers protrude from the top wells and from the bottom exit holes. To connectorize the array, the top-well-protruding wires are first de-insulated (by a flame passed swiftly over the wires, to burn off the Parylene-C), then pushed farther into the wells until they no longer protrude through the top. Then, a dab of silver paint is applied to the wells (Fig. S1d); once it spreads through the wells, excess paint is brushed off with a cotton-tipped applicator (Fig. S1e,f), and a connector is placed into the wells (Fig. S1g,h). In the 64-channel version, this is a 70-pin Hirose connector whose pins were folded down manually before insertion (Figs. 1a; S1a-c). The connector is then glued to the plastic block with super glue. The design is configurable to single-block 64-channel arrays (Fig. 1b) or four 16-channel arrays on flex-cable PCBs for multi-brain-region recordings (Figs. 1c; S1i; S2), and is compatible with micro-drive devices (Fig. 1d) 35 .
The impedance of untreated tips is too high for single-unit recordings in a fixed-implant preparation 36 , requiring treatment before recordings can be made. In a previous report, tips were prepared by individual fire-sharpening, which de-insulated ~80 µm along the shaft of each wire, simultaneously sharpening it 30 . While successful, this technique is difficult to scale to high-channel-count arrays. To get around this problem, we turned to chemical treatment of the tips, first etching the tips using sulfuric acid in order to roughen the tip surface and increase the surface area 31,32 , then electroplating the tips with PEDOT-TFB, which was previously shown to reduce impedance and increase conductivity and stability compared to the more conventionally used PEDOT-PSS 33,34 (Fig. 1e). The resulting tips had significantly reduced impedances (before treatment: 4 84 0 68 . ± . MΩ median ± s.e.m. (n 224 = tips from two 64-channel and six 16-channel arrays), and . ± . 0 17 0 86 MΩ in the same tips after treatment with the acid and PEDOT-TFB (two-sided t-test on impedance before/after treatment of tips: t 3 88 = . , = . p 0 00014)) (Fig. 1f). The bulk of the impedance change following the acid and PEDOT-TFB treatments came from the electroplating rather than etching (Fig. 1g). Following acid etching and electroplating with PEDOT-TFB, ~15% of tips retained an impedance of at least 4 84 . MΩ (the pre-treatment median), which likely include dead channels. We visualized the effect of this treatment on the tips using scanning electron microscopy on 16-channel arrays (Fig. 2).
The CFEA requires individual carbon fibers to be insulated with parylene-C before assembly, which in our previous design involved a painstaking manual procedure in which individual fibers were mounted on sticky notes and loaded into the parylene vapor deposition machine (PDS Labcoter 2010). To facilitate this process, we designed coating jigs ( Fig. 3a-c) that hold multiple cartridges (Fig. 3d,e), each of which in turn hold several dozen carbon fibers splayed harp-like across two pieces of sticky notes. This design (Fig. 3f) can be laser-cut out of plastic and modified to various sizes of cartridges (to hold longer fibers, for example) or jigs (to hold more or fewer cartridges).
Long-term recordings. We tested the longevity of acid/PEDOT-TFB-treated arrays by recording chronically from a 16-channel version of the array implanted in rat primary visual cortex (V1). Figure 4 shows examples from one such implanted rat. Fig. 4a shows two seconds of bandpass-filtered activity four days after the implant surgery ('Day 4'), while Fig. 4b shows activity from the same array 55 days later. Spiking activity was prominent across most channels in both recordings. We performed spike sorting on both recording sessions using automated clustering based on the ISO-SPLIT algorithm (MountainSort 37 ), and found a minor decrease in signal to noise ratios of well-isolated sorted units ( . ± . . ± . 55 days later ( = − . = . t p 0 76, 0 46, t-test)). We repeated the spike sorting on two more recording sessions from another rat, for a total of = n 37 isolated units across the four sessions in two rats (Fig. 4d).

64-Channel recording and responses to visual stimulation.
We then proceeded to implant the 64-channel arrays into rat V1. Figure 5 shows example traces from one such recording while the animal was behaving freely in the recording chamber. Having determined that we could record spontaneous neuronal activity using chronically-implanted 64-channel CF arrays, we sought to determine whether these arrays could capture sensory evoked activity as well. For this, we presented a V1-implanted rat with a visual stimulus: flashed cage lights. The rat was free to behave in the recording chamber in the dark, while the cage lights were flashed (On for 500 milliseconds, Off for a uniformly random time between 400 and 600 milliseconds, for 225 trials.). This strong visual stimulus produced evoked multiunit (MU) responses (Fig. 6a,b) and LFPs (Fig. 6c). Figure 6a shows one trial (grey patch indicates time when lights were On); Figure 6b,c show trial-averaged MU and LFP responses, respectively. The mean LFP response showed a strong modulation following stimulus onset, as well as a smaller response following stimulus offset, as has previously been reported [38][39][40] . www.nature.com/scientificreports www.nature.com/scientificreports/ Finally, to test for potential cross-talk among channels in the CFEA, we measured correlations among pairs of bandpass-filtered 30-second traces and compared these pairwise correlations to data obtained from recordings using nichrome wire tetrodes implanted in rat V1 (Fig. S3). The nichrome tetrode arrays were arranged in an × 8 2 grid, with ~300-micron spacing among tetrodes. This spacing allowed us to compare correlations among pairs of carbon fibers to those of closely spaced nichrome wires (i.e. within a given tetrode) and distant nichrome wires (i.e. across tetrodes). We reasoned that making these comparisons would provide a better understanding of the degree of cross-talk within the closely-spaced carbon fibers. Correlations among channels in the CFEA were comparable to, albeit slightly lower than, those within nichrome tetrodes, consistent with the idea that the devices presented here are capable of recording independent neural signals despite the close bundling of the fibers.

Discussion
CFEAs have shown promise in recording from neuronal populations over long timescales but have so far been prepared in small channel-count configurations 28,30,[41][42][43] . It is becoming increasingly clear that in order to understand the brain, neuroscientists will need to measure larger populations of neurons spread out across multiple brain regions 8 (however, see Gao et al. 44 ). Our results show the feasibility of producing high channel-count, high-density CFEAs and using them to record evoked activity in rat visual cortex. Further, we have shared designs to implement simultaneous recordings from multiple brain regions using four 16-channel flex-PCB CFEAs (Fig. 1c) and micro-drive-enabled 64-channel arrays (Fig. 1d). We have also designed a jig for setting up Parylene-C deposition onto the bare CFs (Fig. 3), which we hope will facilitate the implementation of CFEAs by other labs.
The 64-channel CFEAs are faster to assemble (per channel) than their 16-channel predecessors (~2 hours for the former and ~1 hour for the latter: a 2X improvement from 3.75 to 1.875 minutes/channel). Although etching and electroplating with PEDOT-TFB are not by themselves new techniques, neither has previously been applied to CFEAs. The method of preparing the recording sites using acid etching and electroplating with PEDOT-TFB further sped up the assembly process over the previous fire-sharpening technique, which was done on individual fibers and required extensive practice 30 .
CFEA designs have for the most part required manual assembly. This will obviously make recording extremely large populations of neurons (10 5−6 ) a challenge. As such, manually-assembled CFEAs may be a more appropriate alternative to other manually-made electrodes like tetrodes, rather than arrays manufactured using nanotechnology 8,25,26 , which may soon be able to accommodate thousands of recording sites. CFEAs are smaller and more flexible than many silicon probes, although still stiffer than brain tissue; efforts to further reduce the stiffness of electrodes are proving to be a successful strategy for improving recording technology 45 . Previous reports have shown CFEAs to produce a smaller immune response than larger implants 11,42,46 , and further experiments will be www.nature.com/scientificreports www.nature.com/scientificreports/ required to assess the immune responses of large channel-count arrays such as those described here. While manual assembly of CFEAs limits their scalability, they can, in principle, be manufactured using automated systems, and some efforts are underway to implement those 47 .
Electrode arrays are currently the best tool for investigating neuronal circuits continuously (i.e. 24/7) over long periods of time in freely behaving animals. However, whether electrode arrays of any kind will be the method to record all the neuronal activity in a brain is still unknown. Arrays designed to capture the entirety of neuronal  www.nature.com/scientificreports www.nature.com/scientificreports/ activity in a vertebrate brain would need to be quite noninvasive indeed; they would also need to capture the activity of millions of neurons, which would require multiplexing and digitization near the recording sites 48 .
Alternatives to electrodes might be optical methods, which are currently limited in their ability to image deep brain structures 49 . Imaging of calcium indicators has a limited temporal resolution because of molecular binding kinetics 50 . This problem may be avoided by voltage sensors, which can be imaged at kilohertz rates 6,51,52 , but in principle suffer from the same depth imaging and bleaching problems as calcium indicators 53 . One radically different approach is molecular barcoding, in which neuronal activity is recorded by a 'molecular ticker-tape' that tracks changes in intracellular calcium concentration by making insertions into a cell's DNA 54,55 . For now, perhaps the best way to get around the limitations of each technique is to use multiple complementary tools, investigating local circuits with electrode arrays after a brain region of interest is identified using whole-brain low-temporal-resolution methods like calcium imaging or fMRI.

Methods
Animals. All procedures were performed in accordance with institutional guidelines and regulations. The care and experimental manipulation of all animals were reviewed and approved by the Harvard Institutional Animal Care and Use Committee (Protocol #27-22). Experimental subjects were female Long Evans rats 3 months or older, weighing 300-500 grams ( n 6 = , Charles River). Two rats were implanted with 16-channel arrays; two were implanted with 64-channel arrays; one was implanted with four 16-channel arrays in different brain regions; and one was implanted with a 16-tetrode nichrome wire array in visual cortex.
Carbon fibers, array assembly, and tip preparation. In this report, we used µ . 4 5 m diameter carbon fibers (UMS2526, Goodfellow USA), as reported previously 30 . Epoxy sizing was removed either by heating the fibers to 400 C in a kiln, or by soaking the fibers in acetone >24 hours 56 . To insulate the fibers with Parylene-C (di-chloro-di-p-xylylene) (Paratronix Inc, Westborough MA), we laid the fibers out onto sticky notes attached to plastic cartridges (Fig. 3) and coated the fibers with ~µ 1 m layer of Parylene-C using chemical vapor deposition in a PDS-2010 Labcoter machine (Specialty Coating Systems, http://scscoatings.com/), as described previously 30 .
The fibers were then threaded through 3D-printed plastic blocks, which were custom-designed in SolidWorks (Dassault Systémes SolidWorks Corporation, Waltham, MA) and manufactured using stereolithography (Realize Inc, Noblesville, IN) (designs available on https://github.com/guitchounts/electrodes). The fibers were left protruding ~1 mm from the top of the plastic block (i.e. at the entry holes) end and were then deinsulated at the connectorization end with a lighter flame, after which they were threaded further into the block until no longer protruding. Connectorization was achieved by flooding the block's wells with silver paint (MG Chemicals, #842-20g) and attaching a Hirose DF40 connector (Digi-key, H11630CT-ND).
The recording sites were prepared by cutting the array bundles to the desired length for implantation (~1 mm for the visual cortex implants described here) using serrated scissors (Fine Science Tools, 14058-11), then dipping the tips into 99% sulfuric acid (Sigma, 339741-500 ML) for 5 minutes, then (after washing any remaining acid off with water), and electroplating with PEDOT-TFB 33,34 by applying a 2 nA current via a 10 V power supply in series with a 5000 MΩ resistor (Digikey, SM102035007FE-ND), for 1 minute (for a total charge delivery of 120nC) per electrode 28  Surgery. Animals were anesthetized with 2% isoflurane and placed into a stereotaxic apparatus (Knopf Instruments). Care was taken to clean the scalp with Povidone-iodine swabsticks (Professional Disposables International, #S41125) and isopropyl alcohol (Dynarex #1204) before removing the scalp and cleaning the skull surface with hydrogen peroxide (Swan) and a mixture of citric acid (10%) and ferric chloride (3%) (Parkell #S393). www.nature.com/scientificreports www.nature.com/scientificreports/ Three skull screws (Fine Science Tools, #19010-00) were screwed into the skull to anchor the implant. A 0.003′′ stainless steel (A-M Systems, #794700) ground wire was inserted ~2 mm tangential to the brain over the cerebellum.
The arrays were targeted to right V1, ranging ~6-8 mm posterior to bregma, 4.5 mm ML, reaching layer 2/3 at 0.6 mm DV. After electrodes were inserted into the brain, the craniotomy was covered with Puralube vet ointment (Dechra) and the electrodes were glued down with Metabond (Parkell). Post-operative care included twice-daily injections of buprenex (0.05 mg/kg Intraperitoneal (IP)) and dexamethasone (0.5 mg/kg IP) for three days.
Electrophysiology. Electrode signals were acquired at 30 kHz using either a 16-channel Intan RHD2132 headstage, or a custom Intan-based 64-channel headstage 57 and Opal-Kelly FPGAs (XEM6010 with Xilinx Spartan-6 ICs). LFPs and spikes were extracted following procedures described in Dhawale et al. 57 . The LFP signals were downsampled to 300 Hz by two applications of a fourth order 5-fold decimating Chebychev filter followed by a single application of a fourth order 4-fold decimating Chebychev filter. The 16-channel data were acquired using the OpenEphys GUI 58 . Units presented in Figure 4 were sorted using the MountainSort package 37 .
In addition to the 64-channel CFEAs, we designed a 64-channel system for recording with four 16-channel arrays in multiple brain regions. The 16-channel flex-PCB (Fig. 1c) mates with a miniature 16-channel 3D-printed plastic block that holds the CFs. The connector side holds a 20-pin Hirose DF40 connector (Digi-key, H11618CT-ND) that interfaces with a breakout PCB that holds four mating 20-pin connectors (Digi-key, H11619CT-ND) on one side and a 70-pin DF40 connector (Digi-key, H11630CT-ND) on the other. The breakout board thus connects four flex-PCB arrays to one 64-channel headstage. The four flex-PCB arrays could be mounted on microdrives as in Vandecasteele et al. 59 . The 64-channel CFEAs were also adapted for microdrive-based recordings based on the optetrode design in Anikeeva et al. 35 . In our design, the 64-channel array and microdrive are assembled separately and joined using micromanipulators. Specifically, the carbon fibers in the assembled array are bundled together, then threaded through a polyimide tube attached to the microdrive. The plastic base of the array is then glued to the vented screw in which the polyimide tube and carbon fibers reside (Fig. 1d). In this design, the entire 64-channel bundle is driven as one unit, in contrast to tetrode-based microdrives that allow for advancing of individual tetrodes 60 .
To compare the cross-talk and correlation structure of signals recorded from CFEA to those from traditional nichrome wire tetrode arrays (Fig. S3), we implanted one rat with a 16-tetrode nichrome array in V1. Tetrodes were fabricated using 12.5-micron nichrome wire (Sandvik-Kanthal) following standard procedures 58,61,62 . Tetrodes were threaded through 42 AWG polyimide guide tubes into 8x2 grids of 34 AWG tubes (Small Parts). The tetrodes were plated with a mixture of gold (Neuralynx) and polyethylene glycol (PEG) as per Ferguson et al. 63 , to an impedance of ~100-250 KΩ. The ground and reference wires were bridged and implanted through a craniotomy above the cerebellum.
Behavior and visual stimulation. The recordings were carried out in a custom recording chamber while the animals were free to behave. The headstage was connected to the FPGA board via a custom shielded cable Visual stimulation was delivered via white LED lights (Triangle Bulbs Cool White LED Waterproof Flexible Strip Light, T93007-1, Amazon) mounted in the recording chamber, and controlled with a micro-controller (Arduino). During stimulus presentation, the chamber and experimental room were otherwise completely dark. The lights were On for 500 ms and Off for a uniform-random time between 400 and 600 ms.
Imaging. Scanning Electron Microscopy images of the CF tips were taken using a Zeiss Ultra55 Field Emission Scanning Electron Microscope (FESEM) at the Center for Nanoscale Systems at Harvard University.