Full-bandwidth electrophysiology of seizures and epileptiform activity enabled by flexible graphene microtransistor depth neural probes
通过柔性石墨烯微晶体管深度神经探针实现的癫痫和癫痫样活动的全带宽电生理学

Pathological electrophysiological signals resulting from epilepsy span a wide frequency range, from fast activity, at hundreds of hertz or higher, to less explored signals below 0.1 Hz, including DC-shifts (sustained over seconds to minutes) and infraslow oscillations^1,2,3. In contrast to the widely used AC-coupled recordings, DC-coupled recordings allow more comprehensive assessment of the pathophysiological processes associated with either an increase in neuronal activity during seizures or a loss of neuronal activity during a spreading depolarization (SD)⁴. Clinically, electrophysiological biomarkers have been explored to guide surgical resection of the seizure onset zone^5,6; however, improvement and refinements are still needed^7,8. Indeed, intracranial clinical studies using large metallic electrodes suggest that the efficacy of current electrophysiological biomarkers could be improved by the inclusion of DC-shifts⁹. However, broad adoption of DC-shifts as a novel biomarker is limited by the poor capabilities of microelectrode technology to reliably record this type of electrophysiological signal in vivo with high spatiotemporal fidelity. This is due to polarization-induced drift and signal attenuation causing distortion of the measured signal by metallic electrodes^10,11,12. Currently, the gold standard for high-fidelity DC-coupled recordings is the glass micropipette; however, its use is limited to few-point measurements, hampering the spatial resolution required for seizure localization, and is not suitable for clinical applications. An alternative to commonly used passive electrodes are field-effect transistors (FETs), which are active transducers offering important advantages in electrophysiology¹³. Among the few FET technologies that have been validated for in vivo electrophysiology, graphene-based FETs are particularly attractive because of the combination of material properties, including chemical and electrochemical inertness, high electrical mobility, biocompatibility, as well as facile integration into flexible and ultrathin substrates¹⁴. Recent reports demonstrate the potential of graphene solution-gated field-effect transistors (gSGFETs) for neural interfacing^15,16. A proof-of-concept of high-fidelity DC-coupled recordings enabled by gSGFETs was demonstrated by mapping chemically induced cortical SDs in anaesthetized rats¹⁷. To explore its usefulness for epilepsy research, we have developed implantable, flexible graphene depth neural probes (gDNPs) capable of recording localized full-bandwidth neuronal activity through cortical columns and subcortical structures. Here, we present a wafer-scale microtechnology process to fabricate ultrathin gDNPs consisting of a linear array of graphene microtransistors imbedded in a polymeric flexible substrate. To penetrate through the mouse cortex and reach the hippocampus without buckling, we adapted an insertion protocol that uses silk fibroin (SF)^18,19 to temporarily stiffen flexible gDNPs. The ability to detect electrophysiological biomarkers of epileptiform activity in preclinical rodent models of induced seizures and chronic epilepsy has been validated experimentally²⁰. These included seizure events such as spike-and-wave discharges (SWDs)²¹ and high-frequency oscillations (HFOs)^22,23. Additionally, we highlight the suitability of graphene transistor technology to record concurrently fast and slow brain electrophysiological signals by benchmarking against a solution-filled glass micropipette recording, demonstrating the high fidelity of the DC-coupled recordings obtained by the gDNPs. Here, we demonstrate the capability of gDNPs to record both active and passive ictal baseline shifts^24,25,26 as well as SD associated with seizure activity. Altogether, our results illustrate the potential of graphene microtransistor technology to reveal mechanistic insights into the involvement of infraslow oscillations and DC-shifts associated with seizures in vivo in awake brain.
癫痫引起的病理生理电信号跨越广泛的频率范围，从数百赫兹或更高的快速活动，到尚未充分探索的低于 0.1 赫兹的信号，包括直流偏移（持续数秒至数分钟）和极慢振荡^1,2,3。与广泛使用的交流耦合记录相比，直流耦合记录允许更全面地评估与癫痫发作期间神经元活动增加或去极化扩散（SD）期间神经元活动丧失相关的病理生理过程⁴。在临床上，已经探索了电生理生物标志物来指导癫痫发作起始区的手术切除^5,6；然而，仍需改进和细化^7,8。实际上，使用大型金属电极的颅内临床研究表明，通过包括直流偏移，可以改善当前电生理生物标志物的有效性⁹。 然而，微电极技术在体内以高时空保真度可靠记录这种电生理信号的能力有限，这限制了 DC 偏移作为一种新型生物标志物的广泛应用。这是由于金属电极引起的极化诱导漂移和信号衰减，导致测量的信号失真^10,11,12。目前，高保真度直流耦合记录的金标准是玻璃微管；然而，其使用仅限于少量点测量，阻碍了癫痫定位所需的空間分辨率，并且不适合临床应用。一种替代常用被动电极的方法是场效应晶体管（FETs），它们是主动传感器，在电生理学中提供了重要的优势¹³。 在为数不多的已被验证用于体内电生理学的 FET 技术中，基于石墨烯的 FETs 因其材料特性的组合而特别吸引人，这些特性包括化学和电化学惰性、高电迁移率、生物相容性，以及易于集成到灵活和超薄基板中¹⁴。最近的报告表明，石墨烯溶液栅场效应晶体管（gSGFETs）在神经接口方面具有潜力^15,16。通过映射麻醉大鼠中化学诱导的皮质 SDs，展示了 gSGFETs 实现高保真度直流耦合记录的原理证明¹⁷。为了探索其在癫痫研究中的实用性，我们开发了可植入的、灵活的石墨烯深度神经探针（gDNPs），能够通过皮质柱和皮层下结构记录局部全带宽神经元活动。在这里，我们提出了一种晶圆级微技术工艺，用于制造由嵌入在聚合物柔性基板中的线性阵列石墨烯微晶体管组成的超薄 gDNPs。 为了穿透小鼠皮层并到达海马体而不弯曲，我们采用了一种插入协议，该协议使用丝素蛋白（SF）^18，19来暂时使柔性的 gDNPs 变硬。在诱导性癫痫和慢性癫痫的啮齿动物模型中检测癫痫样活动的电生理生物标志物的能力已经通过实验得到验证²⁰。这些包括癫痫事件，如尖波和波放电（SWDs）²¹和高频振荡（HFOs）^22，23。此外，我们通过将 gDNPs 与溶液填充的玻璃微管记录进行基准测试，突出了石墨烯晶体管技术同时记录快速和慢速脑电生理信号的适用性，证明了 gDNPs 获得的直流耦合记录的高保真度。在这里，我们展示了 gDNPs 记录活动性和被动性癫痫发作基线偏移的能力^24，25，26以及与癫痫活动相关的 SD。 总的来说，我们的结果表明，石墨烯微晶体管技术在揭示清醒大脑中与癫痫发作相关的亚慢振荡和直流偏移的机制方面具有潜力。

A graphene-based SGFET is a three-terminal device in which single-layer graphene prepared by chemical vapour deposition (CVD) is used as the channel material in contact with the drain and source terminals. Graphene is the sensing part of the device directly exposed to the neural tissue. The current in the graphene channel can be modulated or pinned by a third terminal given by a reference electrode (gate), which is also in contact with the conducting neural tissue (Fig. 1a). Thus, variations in the electrical potential in the tissue can be transduced into variations in the channel current; this transduction mechanism has been shown to offer a very versatile sensing platform for electrophysiology^27,28. The flexible gDNP prepared in this study consists of a linear array of 14 recording transistors, each with an active area of 60 × 60 μm², and a pitch of 100 μm. The probe’s tip design consists of a polyimide shank of 200 μm width and 1.6 mm length (Fig. 1b). A gDNP design with smaller SGFETs (20 × 20 μm²) and narrower shank (150 μm width) has also been fabricated (Supplementary Fig. 1). The gDNPs were fabricated on a 10-μm-thick flexible polyimide (PI) substrate using a wafer-scale microfabrication process previously reported¹⁶ (Methods). A two-level metallization strategy, with the metal levels interconnected by via holes (Fig. 1b), reduces track resistance and improves sensor performance. To characterize gDNPs in saline solution we simultaneously measure the drain–source current (I_DS) versus the applied gate–source voltage (V_GS) for all the transistors on the shank with a fixed drain–source voltage (V_DS) using customized electronics (Methods). gSGFETs exhibit very high transconductance (g_m) values due to direct contact with the tissue, possible as a result of its biocompatibility and electrochemical inertness, and to the very high charge carrier mobility of graphene²⁹. Figure 1c shows the transfer curves as well as the normalized transconductance (|g_m| = g_m/V_DS) for all 14 gSGFETs of an exemplary gDNP device. The small dispersion of the charge neutrality point (CNP), defined as the value of V_GS where the I_DS reaches a minimum, attests to the homogeneity of the gDNP. Furthermore, g_m shows a very stable response over a wide applied gate frequency range (up to 5 kHz), both in the hole regime, V_GS < V_CNP, and in the electron regime, V_GS > V_CNP (Fig. 1d). Such a constant frequency response is important for a proper calibration of the recorded signals³⁰. The detection limit of the sensors was evaluated by means of the effective gate noise (that is, root-mean-square gate voltage noise, V_RMS) integrated between 1 Hz and 2 kHz, with averaged values of 25–30 µV for all fabricated gDNPs (Supplementary Fig. 2).
基于石墨烯的 SGFET 是一种三端器件，其中使用化学气相沉积（CVD）制备的单层石墨烯作为与漏极和源极端子接触的通道材料。石墨烯是直接暴露于神经组织的器件的传感部分。石墨烯通道中的电流可以通过由参考电极（栅极）提供的第三个端子进行调制或固定，该参考电极也接触导电的神经组织（图1a）。因此，组织中的电势变化可以被转换为通道电流的变化；这种转换机制已被证明为电生理学提供了一个非常通用的传感平台^27,28。本研究中制备的柔性 gDNP 由一个线性阵列的 14 个记录晶体管组成，每个晶体管的活性面积为 60×60μm²，间距为 100μm。探针的尖端设计由 200μm 宽和 1.6mm 长的聚酰亚胺支架组成（图1b）。还制造了具有更小 SGFET（20×20μm²）和更窄支架（150μm 宽）的 gDNP 设计（补充图。 1) gDNPs 是在 10 微米厚的柔性聚酰亚胺（PI）基板上使用之前报道的晶圆级微加工工艺制造的¹⁶ (方法)。一种两级金属化策略，金属层通过通孔（图1b）互连，降低了轨道电阻并提高了传感器性能。为了表征盐水溶液中的 gDNPs，我们同时测量了所有在支架上的晶体管的漏源电流（I_DS）相对于施加的栅源电压（V_GS），使用固定的漏源电压（V_DS）和定制的电子设备（方法）。gSGFETs 表现出非常高的跨导（g_m）值，这是由于与组织的直接接触，可能是由于其生物相容性和电化学惰性，以及石墨烯非常高的电荷载体迁移率²⁹。 图 1c 展示了所有 14 个 gSGFET 的传输曲线以及归一化跨导（|g_m| = g_m/V_DS）的示例 gDNP 器件。电荷中性点（CNP）的小分散性，定义为 V_GS 的值，其中 I_DS 达到最小值，证明了 gDNP 的一致性。此外，g_m 在宽范围的施加栅极频率（高达 5 kHz）下表现出非常稳定的响应，无论是在空穴区，V_GS < V_CNP，还是在电子区，V_GS > V_CNP（图 1d）。这种恒定的频率响应对于正确校准记录的信号³⁰非常重要。 传感器的检测限通过有效门噪声（即均方根门电压噪声，V_RMS）在 1 Hz 和 2 kHz 之间的积分来评估，所有制造的 gDNPs 的平均值为 25–30 µV（补充图 2）。

a, Schematic of a gSGFET and biasing. b, Optical microscope image of a gDNP containing 14 transistors with a pitch of 100 μm on a 200-µm-wide PI shank (left) and a magnification of one gSGFET (top right). The area highlighted in red shows the graphene sensing area (60 × 60 μm²) of the transistor. The schematic of the cross-section of one transistor shows the interconnected metal track strategy to reduce the shank width of the gDNP (bottom right). c, Electrical characterization of all 14 gSGFETs in a gDNP in a 150 mM saline solution (V_DS = 50 mV) by measuring I_DS and g_m versus V_GS. The lines represent mean values and the shadows the standard deviation. d, Transconductance spectroscopy of the gSGFET bias at the point of maximum g_m in the electron (V_GS > V_CNP, orange line) and hole (V_GS < V_CNP, blue line) regimes. The squares in c (right) are the values of maximum g_m as measured in steady-state mode. The decay observed in the grey shaded areas is due to the filtering of the interfacing electronics. e, Coloured scanning electron microscopy images of the gDNP: uncoated (left) and back-coated with ~80-µm-thick (middle) and ~150-µm-thick (right) SF. Scale bars, 100 µm. f, Mechanical assessment of the gDNP coated with the two thicknesses of SF showing the averaged compression force versus displacement (n = 10 trials; the lines represent the mean, the coloured shadows the standard deviation). The optical images correspond to the two different conditions of the experiment. The image in the green box corresponds to the point marked by the green square in the force–displacement plot, a point before buckling, and the image in the orange box to the point marked by the orange square, representing a point after buckling. g, Functionality assessment after insertion in agar. The boxplots (n = 14 gGFETs; the median, quartile box and minimum and maximum values are presented) show the distribution of the transconductance (g_m) and V_RMS noise level for all gSGFETs in a gDNP measured in phosphate-buffered saline (PBS) solution, inserted and measured inside an agarose gel brain model, and measured in PBS solution after removal from the agarose gel. h, Image sequence of a SF-coated gDNP at different timepoints after insertion in the agarose gel brain model, revealing the timeframe of the degradation of the SF coating upon insertion.
Translated Text: a，gSGFET 及其偏置的示意图。 b，一个包含 14 个晶体管的 gDNP 的光学显微镜图像，间距为 100 μm，在 200-µm 宽的 PI 支架上（左侧）和一个 gSGFET 的放大图（右上）。红色高亮区域显示晶体管的石墨烯传感区域（60 × 60μm²）。一个晶体管的横截面示意图显示了减少 gDNP 支架宽度的互连金属轨道策略（右下）。 c，在 150 mM 盐水溶液中，通过测量I_DS和g_m相对于V_GS的电气特性，对 gDNP 中的所有 14 个 gSGFET 进行电气表征（V_DS = 50 mV）。线条代表平均值，阴影代表标准偏差。 d，在电子（V_GS > V_CNP，橙色线）和空穴（V_GS < V_CNP，蓝色线）区域中，gSGFET 偏置在最大g_m点的跨导光谱。 在 c（右侧）中的方块是稳态模式下测量的最大 g_m 的值。灰色阴影区域观察到的衰减是由于接口电子设备的滤波。 e，gDNP 的彩色扫描电子显微镜图像：未涂覆（左侧）和背面涂覆约 80-µm 厚（中间）和约 150-µm 厚（右侧）的 SF。比例尺，100 µm。 f，对涂覆两种厚度 SF 的 gDNP 进行机械评估，显示平均压缩力与位移的关系（n = 10 次试验；线条代表平均值，彩色阴影代表标准偏差）。光学图像对应于实验的两个不同条件。绿色框中的图像对应于力-位移图中绿色方块标记的点，即屈曲前的点，橙色框中的图像对应于橙色方块标记的点，代表屈曲后的点。 g，在琼脂中插入后的功能评估。 箱线图（n = 14 gGFETs；展示了中位数、四分位数箱以及最小值和最大值）显示了在磷酸盐缓冲盐水（PBS）溶液中测量的所有 gSGFETs 的跨导（g_m）和V_RMS噪声水平，这些 gSGFETs 被插入到琼脂糖凝胶脑模型中测量，并在从琼脂糖凝胶中取出后测量了 PBS 溶液中的值。h，展示了在琼脂糖凝胶脑模型中插入后不同时间点的 SF 涂层 gDNP 图像序列，揭示了 SF 涂层在插入后的降解时间框架。

Full size image

gDNPs are highly flexible, compared with traditional rigid depth electrodes, and although flexibility is advantageous once inserted into tissue, this provides a challenge during insertion. To insert these probes we temporarily stiffened the gDNP using SF^18,19. The stiffening technique (Methods) consisted of a moulding process in which the gDNP is back-coated with SF, allowing the preparation of a rigid shank with a defined shape and thickness. We tuned the thickness of the SF by controlling the mould’s trench depth, achieving two typical thicknesses of 80 ± 10 and 150 ± 12 μm (Fig. 1e). The mechanical properties of the SF-coated gDNPs were assessed using a buckling test, in which perpendicularly positioned probes were driven against a flat, hard surface (Fig. 1f). An initial linear increase in force was observed for both coating thicknesses tested, the probes remaining straight before buckling. Continued application of force resulted in buckling and bending, characterized by a peak in the force–displacement curve. The obtained peak forces, 101 ± 21 mN for the 150-µm-thick SF and 29 ± 13 mN for the 80-µm-thick SF, are in good agreement with the previously reported values of peak forces for similar SF-coated neural probes^19,31. We electrically characterized the gDNPs before and after the SF stiffening process, as well as before and after insertion into and removal from an agarose gel brain model. Figure 1g shows the distribution of the normalized g_m as well as V_RMS for all 14 transistors in a gDNP, confirming that neither the stiffening process nor insertion in an agarose brain model impairs gDNP performance in terms of transconductance or noise (Supplementary Fig. 3).
gDNPs 具有高度灵活性，与传统刚性深度电极相比，尽管在插入组织后灵活性具有优势，但在插入过程中这带来了挑战。为了插入这些探针，我们使用 SF^18,19暂时使 gDNP 变硬。硬化技术（方法）包括一个成型过程，其中 gDNP 背面涂覆 SF，允许制备具有定义形状和厚度的刚性柄。我们通过控制模具的沟槽深度来调整 SF 的厚度，实现了两种典型的厚度，分别为 80±10 和 150±12μm（图1e）。使用屈曲测试评估了 SF 涂覆的 gDNPs 的机械性能，其中垂直放置的探针被驱动到平坦、硬的表面上（图1f）。观察到两种测试的涂层厚度在初始阶段力线性增加，探针在屈曲前保持直线。继续施加力导致屈曲和弯曲，在力-位移曲线上表现为峰值。 获得的峰值力，150-µm 厚的 SF 为 101±21 mN，80-µm 厚的 SF 为 29±13 mN，与之前报道的类似 SF 涂层神经探针的峰值力值^19，31非常吻合。我们在 SF 硬化过程前后，以及插入和从琼脂糖凝胶脑模型中取出前后，对 gDNPs 进行了电学表征。图1g显示了所有 14 个 gDNP 中归一化的g_m以及V_RMS的分布，证实了硬化过程或插入琼脂糖脑模型中都不会损害 gDNP 在跨导或噪声方面的性能（补充图3）。

Figure 1h depicts video frames of a SF-coated gDNP after insertion (insertion speed 400 μm s^–1) in an agarose brain model; the images help to visualize the bending-free insertion of the gDNP shank and the fast water absorption (<10 s) by SF. Moreover, it can be seen that the SF was completely delaminated from the polymeric shank after removal from the agar gel (180 s).
图 1h 展示了在琼脂脑模型中插入（插入速度 400 μm s^–1）后的 SF 涂层 gDNP 的视频帧；这些图像有助于可视化 gDNP 腿部的无弯曲插入和 SF 的快速吸水（<10 s）。此外，还可以看到，在从琼脂凝胶中取出后（180 s），SF 完全从聚合物腿部剥离。

These observations confirm the suitability of the SF back-coating stiffening procedure to enable the flexible gDNPs to reach the target position in a brain-like tissue without bending; the fast dissolution of the shuttle once inserted enables the probe to quickly regain flexibility. This insertion strategy is only suitable for single-time use; for further insertions, the flexible gDNP needs to be recoated. We have verified that the gDNP technology can withstand multiple coating and insertion procedures (Supplementary Fig. 4).
这些观察结果证实了 SF 背涂加固程序适用于使柔性 gDNPs 在类似大脑组织中到达目标位置而不弯曲；一旦插入，穿梭的快速溶解使探针能够迅速恢复灵活性。这种插入策略仅适用于单次使用；对于进一步的插入，柔性 gDNP 需要重新涂覆。我们已经验证了 gDNP 技术可以承受多次涂覆和插入程序（补充图4）。

We assessed full-bandwidth recording capability by implanting a gDNP into awake, head-fixed mice. The electrophysiological signal measured by the graphene transistors was acquired with a customized electronic system that enables simultaneous recording in two frequency bands with different gains preventing amplifier saturation (Fig. 2a and Methods). gDNPs were implanted in the right hemisphere visual cortex (V1) and lowered until the tip reached hippocampal tissue. Baseline activity was recorded for 10–20 min. To induce network discharges and synchronicity of neuronal bursting, 200 nl of 4-aminopyridine (4-AP, 50 mM), a selective blocker of Kv1 potassium channels^32,33, was focally injected into the cortex adjacent to the gDNP (Fig. 2b).
我们通过在清醒、头部固定的老鼠中植入 gDNP 来评估全带宽记录能力。由石墨烯晶体管测量的电生理信号是通过一个定制的电子系统获得的，该系统允许在两个具有不同增益的频率带中同时记录，以防止放大器饱和（图2a和方法）。gDNPs 被植入右侧半球视觉皮层（V1）并降低，直到尖端到达海马组织。记录了 10-20 分钟的基线活动。为了诱导网络放电和神经元爆发的同步性，将 200 nl 的 4-氨基吡啶（4-AP，50 mM），一种选择性阻断 Kv1 钾通道的阻断剂^32，33，聚焦注射到 gDNP 旁边的皮层中（图2b）。

a, Schematic of the recording set-up and concept of a graphene transistor as a transducer for neural recording. V_sig, voltage signal; I_signal, current signal; V_D, drain voltage; V_S, source voltage. b, Schematic of a mouse skull showing the location on the brain of the gDNP, the glass micropipette, the reference electrode (Ref.) and the needle to inject the chemoconvulsant drugs (left). Coronal view of the mouse brain showing the localization of the gDNP (right).V1, primary visual cortex; V2M; medial secondary visual cortex; CA, Cornu Ammonis area of the hippocampus. c, Long electrophysiological recording (120 min) of two channels on the gDNP array (the visual cortex (top) and hippocampus (bottom)), showing the full-bandwidth signal (frequency, f  > 0.001 Hz) and the HP-filtered signal (f  > 0.5 Hz). Baseline activity, pin-prick SD, increased neuronal activity after 4-AP injection and seizures, some of them followed by a SD, are evident. Below, different events are shown in higher resolution: baseline (yellow) revealing theta activity in the lowest channel, profile visualization with recording from all 14 transistors following pin-prick SD (beige) and seizure activity shown for the uppermost and lowest channels of the gDNP (green). d, Full-bandwidth recording obtained with the gDNP (black) and glass micropipette (green) after the injection of 4-AP (left). The recording shows the DC-shift concurring with seizure, followed by a SD, recorded by all 14 transistors. The low-frequency cross-correlation analysis of the signals (<5 Hz) from the micropipette and all transistors on the shank shows a very high correlation (right). e, Averaged PSD over the electrophysiological recordings of all transistors during baseline, epileptiform activity, epileptiform activity HP-filtered at 0.5 Hz and post-mortem (n = 14 gSGFETs; the lines represent the mean, the coloured shadows the standard deviation). The grey area highlights the low-frequency part (<0.5 Hz) usually cut off with conventional AC-coupled recordings. ISA, infraslow activity. f, Comparison of baseline activity and post-mortem in one channel of the gDNP for the full-bandwidth signal (top) and HP-filtered signal (>200 Hz; middle). The SNRs of four in vivo experiments performed with four different gDNPs were evaluated (bottom; n = 14 gGFETs for each boxplot; the median, quartile box and minimum and maximum whiskers are shown). The SNR was calculated for different frequency bands (low frequency, 1–70 Hz; high frequency, 80–200 Hz; very high frequency, 200–4,000 Hz; see Methods).

Full size image

Full-bandwidth recordings

Figure 2c displays an uninterrupted electrophysiological recording over 2 h (only the uppermost and the lowest channels of the implanted gDNP are displayed); the complete dataset is shown in Supplementary Fig. 5. The ability of the graphene transistors to provide long and stable full-bandwidth recordings without the need for electronic off-set readjustments contrasts the limitation of DC-coupled passive electrodes³⁴. The recording shows the full-bandwidth signal (>0.001 Hz) and the signal high-pass (HP)-filtered above 0.5 Hz (which is the expected signal recorded by AC-coupled electrodes)³⁵. The recording is broken down into different experimental conditions during the recording: baseline, needle-induced pin-prick SD³⁶, interictal activity⁵ induced by chemoconvulsant drugs and seizures. During baseline recording (Fig. 2c), the lower channels exhibit theta activity, correlated with animal movement, indicating that the gDNP reached the hippocampus, confirmed post hoc by histological analysis of fixed brain sections (Supplementary Fig. 6). After the injection of 4-AP, epileptiform spiking evolved and five seizures (over 60 min) were elicited in this example, one of which was followed by a SD. A second cortical injection of 4-AP induced two additional seizures, both followed by SDs that were detected first in the hippocampus. In five different mice injected with 4-AP, an average number of 7 ± 3 seizures were recorded in 60 min after drug injection. In this chemoconvulsant model, SDs could be observed initially either in superficial cortical layers or in the hippocampus (Fig. 2c and Supplementary Fig. 7).

The fidelity of the DC-coupled recordings was validated by simultaneous recordings using a solution-filled glass micropipette. Figure 2d shows the full-bandwidth recording obtained with the gDNP and micropipette after the injection of 4-AP. Both recordings reveal DC-shifts preceding seizure, and SD occurring after seizure. The DC-shifts measured by the gDNP have a similar shape, magnitude and temporal duration to the signal recorded by the glass micropipette. A cross-correlation analysis (signal filtered <5 Hz) of the signals recorded by the glass micropipette and the 14 gDNP transistors demonstrated a very high correlation (above 90%) for the channels located at the same cortical depth as the micropipette. The same high correlation was observed during hour-long recordings and for small, sudden DC-shifts (Supplementary Fig. 23).

Assessment of the detection limits of gDNP

Post-mortem recordings were acquired to characterize the electrical noise level of the gDNP in the activity-free brain state and, consequently, to quantify the detection limit of the gDNP. Figure 2e shows the averaged power spectral density (PSD) calculated using the recordings of all channels in a gDNP, obtained from different brain states (baseline, after injection of 4-AP and post-mortem). Compared with the baseline PSD, the large amplitude of the PSD at low frequencies (<1 Hz) after 4-AP injection provides an indication of the interictal and DC-shifts in the brain. We calculated the signal-to-noise ratio (SNR) in three different bands, that is, 1–70, 80–200 and 200–4,000 Hz (Fig. 2f). This analysis shows that the gDNPs are able to record electrophysiological signals in typical bandwidths with a SNR ratio higher than 1 dB.

Electrophysiological recording of characteristic epilepsy biomarkers

Epilepsy alterations of physiological brain activity include interictal spikes, fast activity at hundreds of hertz (such as HFOs) and pre-seizure active DC-shifts. We used two different pro-convulsive drugs (4-AP or picrotoxin (PTX)) to induce and evaluate epileptic activity. Figure 3a shows the interictal activity and associated HFOs (>80 Hz)^22,23 recorded by three of the transistors of a gDNP, each located at a different depth in the mouse brain. Filtering between 80 and 600 Hz revealed layer-specific bursting of HFOs and sharp wave ripples during interictal spikes with characteristic oscillations of 200–300 and 400–600 Hz in the cortical and hippocampal channels, respectively³⁷ (Supplementary Fig. 8). Entrained interictal epileptiform activity was found in all channels before each seizure. Figure 3b illustrates characteristic examples of sharp wave ripples and HFOs induced by 4-AP and PTX recorded by the lowest channel of the gDNP (hippocampal CA1 region). The HFO and ripple traces shown in Fig. 3b exhibit high-frequency tones up to 600 Hz. The filtered traces (>200 Hz) are compared with the original traces (full bandwidth) for verification of the ripples. The advantage of the gDNPs for monitoring concurrent DC-shifts and high-frequency activity is illustrated in Fig. 3c, which shows an epileptic spike-triggered SD arising from the hippocampus. The layer-dependent silencing of the neural activity by the hippocampal SD is represented in Fig. 3d in terms of the percentage of activity variation. The layer-dependent amplitude of the SD and the following hyperpolarization, also shown in Fig. 3d (more details of this analysis are provided in the Methods and Supplementary Fig. 9), reveal that the silencing of the neural activity in the hippocampus is correlated with the amplitude and subsequent hyperpolarization wave of the SD. The silencing of neural activity in the hippocampus by the SD is visualized with more clarity in Fig. 3e, where the spectrograms for the uppermost and lowest channels are compared.

a, Interictal activity in three different channels of a gDNP (Ch01, Ch07 and Ch14), showing full-band and BPF (80–600 Hz) signals. b, Sharp wave ripples and HFO recorded in the hippocampus induced by 4-AP and PTX for the full-bandwidth and HP-filtered (>200 Hz) signals. The figure also shows the corresponding spectrograms (range 10–800 Hz). c, Electrophysiological full-band recordings and HP-filtered at >0.5 Hz from the cortex (top channel) to the hippocampus (bottom channel) illustrating a SD arising from the hippocampus after a seizure event. d, Percentage of neural activity variation calculated for each channel by comparing the activity before and during the SD (corresponding to the grey areas in c; left). Absolute amplitudes of the SD and hyperpolarization waves concurrent with the seizure (right). e, Hippocampal neural silencing during the SD illustrated by the spectrograms (range 10–1,000 Hz) of the uppermost and lowest channels of the gDNP. f, DC-shifts precede seizures at the injection site after administration of 4-AP in awake mice. Representative traces recorded by the gDNP showing a seizure event and the concurrent DC-shift after injection of 4-AP (top). Data are shown for two different frequency bands: DC-coupled (<0.1 Hz) and BPF (4–12 Hz) recordings. T, time. The vertical bars indicate the DC-shift onset (blue), the seizure onset (time T_s, red) and seizure end (green). Depth profile of the peak DC-shift amplitude for each transistor (n = 4 seizures; bottom left). The line represents the mean value of the DC-shift and the horizontal bars the standard deviation. Spatiotemporal dynamics of the DC-coupled voltage for the pre-ictal to seizure (ictal) transition (−5 to 2 s; bottom right). The data correspond to the average of n = 4 seizures in one mouse (Supplementary Fig. 10). Sup. cortex, superior cortex. g, CSD analysis of the low-frequency activity (<70 Hz) corresponding to the recording shown in c. Magnifications of the pre-ictal to seizure transition (<70 Hz) are presented, showing dipoles in the different layers of the cortex and hippocampus. The two graphs correspond to the CSD analysis performed with (left) and without (right) the contribution of the DC-coupled signal.

Full size image

DC-shifts are typically associated with seizures²⁴ and can be differentiated into active shifts that precede the seizure, and passive shifts that follow the intense neuronal firing at seizure onset³⁸. For these chemically induced seizures events, we observed that the DC-shift preceded the seizure onset (T_s) by 2–3 s (n = 4 seizures; Fig. 3f and Supplementary Fig. 10). Moreover, we observed a layer-dependent amplitude of the DC-shift, with negative amplitudes in the cortical layers (close to the 4-AP injection site) and positive amplitudes recorded in the channels close to the hippocampus (Fig. 3f, bottom left). The spatiotemporal dynamics of the DC-coupled extracellular potential for the pre-ictal to ictal transition is depicted in Fig. 3f (bottom right).

The inversion of DC-shifts can be used to identify current sources and sinks through the cortical laminae. Current source density (CSD) analysis is a technique to identify source activation in a variety of focal neurological disorders, including epilepsy^39,40. CSD analysis (Methods) of the data in Fig. 3c revealed a large net ionic sink in the hippocampal extracellular space after the seizure, followed by a large source at the beginning of the hyperpolarization wave (Fig. 3g). Enlarging the seizure onset region, four sink and source regions can be identified from the laminae profile. CSD analysis computed with a HP-filtered signal (0.5 Hz < f < 70 Hz) failed to report the ionic sinks preceding and during the seizure in the bottom layers (Fig. 3g, AC-coupled), illustrating the importance of using full-bandwidth recordings for CSD analysis to avoid misinterpretation of the CSD sinks and sources (Supplementary Fig. 11 illustrates additional examples of CSD analyses).

We implanted gDNPs in the right-hemisphere somatosensory cortex of WAG/Rij rats (n = 4), a rodent model of absence epilepsy²⁰, and obtained chronic full-bandwidth recordings over a 10-week period (Fig. 4a). The WAG/Rij rats exhibited frequent spontaneous SWDs, a characteristic thalamocortical oscillation of 8–10 Hz (ref. ²⁰; see Methods). Implanted animals were connected 1–2 times per week for tethered recordings (using a commutator to enable free movement of the rats; Fig. 4a and Supplementary Fig. 12).

a, Cartoon of the chronic experimental set-up indicating the approximate location of the gDNP. b, Averaged transfer curves (n = 14 gSGFETs) for one gDNP over the implantation time. c, Averaged maximum transconductance (green) and CNP (blue) of the transfer curves shown in b (n = 14 gSGFETs; the lines represent the mean, the coloured shadows the standard deviation). d, Normalized SWD activity (red), calculated as the ratio of the SWD and baseline activity, and average V_RMS noise (black) of all channels over the implantation period (n = 14 gSGFETs; the lines represent the mean, the coloured shadows the standard deviation). e, SWDs in the uppermost and lowest channels. The DC potential (<0.5 Hz) overlaps the 0.005–100 Hz signal; the signal filtered above 0.5 Hz is also shown. f, SWD events in the same channel at days 1, 28, 57 and 72. g, Amplitude–phase correlation between ISO and SWD (left). The recording (Ch02, black, 0.01–100 Hz) shows two SWD events (green shadow) with ISO (yellow, 0.01–0.1 Hz; top), RMS (red, 5–9 Hz) with HP filtering (grey, >1 Hz; middle) and the ISO phase (blue, 0.01–0.1 Hz; bottom). The red dotted lines mark the SWD ISO phase. Density distribution of the ISO phase and SWD RMS in the same channel is also shown (right), distinguishing low density (white) and high density (red) regions. h, Density distribution for all channels. i,j, Chronic biocompatibility of gDNPs in Sprague Dawley rats (n = 20) at time points of 2, 6 and 12 weeks (n = 6 for each time point), analysing for astrocyte cells (i) and microglia activation (j). The fluorescent image of GFAP in i is a measure of positive astrocyte cells, in the area of insertion, at 12 weeks (the star indicates the insertion point). The evolution of fluorescence intensity (150 μm from the probe site) obtained for probes with and without graphene is shown. Example fluorescent image of activated microglia (j). The evolution of activated microglia (0.7 mm² around the insertion site) is shown in j for both types of probes. The data are presented as mean values with standard deviation.

Full size image

Transistor curves were measured in each recording session to assess device stability, changes at the device–tissue interface and importantly, to permit selection of an optimal V_GS to maximize the SNR, a feature possible with active sensor devices²⁸. Figure 4b shows the averaged transfer curves of a gDNP measured over 10 weeks (Supplementary Fig. 13). The stability of the transistors’ performance is illustrated in Fig. 4c, which depicts the position of the CNP and the maximum value of g_m over the implantation period. The averaged g_m value remains approximately constant over the whole study. Because g_m is directly related to the tissue–graphene interfacial capacitance and to the carrier mobility in graphene, the stability of g_m strongly suggests little or no variation of these two parameters. The CNP (Fig. 4c) shows a notable shift (200 mV) during the few first days after implantation, and then remains stable for the rest of the experiment. We tentatively attribute the initial shift to the adsorption of negatively charged species, which reduce the intrinsic p-type doping of the graphene transistors²⁹.

The long-term functionality of the gDNP was assessed by evaluating the quality of the recorded signals over the implantation period using two parameters, normalized SWD power amplitude and the transistor noise (V_RMS). For further details, see Methods and Supplementary Figs. 13 and 14. The variation of these two parameters over time is shown in Fig. 4d, demonstrating the ability of the implanted devices to monitor seizure activity with high fidelity during the whole implantation period. The slight increase in the normalized SWD activity could result from a strengthened coupling between neural tissue and the gDNP or from an increase in the seizure power and duration as the animal ages⁴¹.

Correlation between SWDs and infraslow oscillations

The WAG/Rij rat chronic model of absence epilepsy offers the possibility to investigate correlations between infraslow oscillations (ISOs) and SWD events⁴². Owing to the full-bandwidth capability of the gDNP, we were able to observe these correlations as well as a dipole across the neocortex upper and lower layers. Figure 4e shows the uppermost and lowest channels having opposite phases during each SWD (Supplementary Fig. 15). The information contained in the DC-shift is lost after application of a HP filter (>0.5 Hz), as typically used with conventional AC-coupled microelectrodes (Fig. 4e, right). This observation persists over the implantation period, as shown in Fig. 4f. To further evaluate the relation of ISOs and SWDs, we calculated the phase of the signals (limited to the 0.01–0.1 Hz band) and the power of the neural activity associated with the SWDs (5–9 Hz). Figure 4g displays the full band and the ISO component (0.001–0.1 Hz) of the signal recorded from one of the upper channels, together with the RMS of the SWD signal (5–9 Hz) and the ISO phase. Prior to the SWD events, whose onset is marked by the abrupt increase of the SWD RMS, the ISO phase is consistently observed to change (Supplementary Figs. 15 and 16). The correlation between the ISO phase and the SWD can be better illustrated by the joint probability distribution plot (Fig. 4g, right) obtained from a long recording (1,600 s, >35 SWDs; see Methods). Figure 4h depicts the correlation between the ISO phase and the RMS of the SWDs for each channel, revealing an inversion in the lower layer of the neocortex (Supplementary Fig. 16). To further confirm the capability to observe these ISOs preceding the seizure events, we calculated the depth profile of the DC-coupled signal (averaged for n = 9 SWD events) in a timeframe around the start of the SWDs. Preceding the SWD (with its characteristic positive DC-shift) by a couple of seconds, a small negative DC-shift was observed at superficial to medial layers (Supplementary Fig. 17).

This correlation between ISO and SWD was also observed in the other implanted WAG/Rij rats (n = 4; Supplementary Fig. 18).

Chronic biocompatibility of implanted gDNP

We also conducted an extensive chronic biocompatibility study to assess any potential neuroinflammation caused by the invasive nature of the penetrating neural probes, the presence of CVD graphene or by the release of SF following implantation. Neural probes with or without graphene at the recording sites were implanted in adult, male Sprague Dawley rats (n = 20). Histological and immunohistochemical studies were conducted at 2, 6 and 12 weeks post-implantation and compared with the contralateral hemisphere without device implantation (Supplementary Figs. 19 and 20). An additional small cohort of animals (n = 3) were implanted with commercially available rigid depth neural probes with similar dimensions (NeuroNexus, Supplementary Fig. 19). Figure 4i shows a fluorescence image of glial fibrillary acidic protein (GFAP) immunostaining (positive marker for astrocyte cells) in the area of insertion 12 weeks post-implantation (brain sections at 800 μm from the pia). There was no significant difference in the number and morphology phenotype of astrocyte cells (typically associated with inflammation) in the area directly surrounding the implanted probe site, or on comparison between probes with or without graphene (Fig. 4i, right). No significant difference was observed at 2, 6 and 12 weeks post-implantation, and values are in the same range as in the contralateral hemisphere where no probe was inserted (Supplementary Fig. 19). The activation state of microglia, assessed by morphological analysis of cells stained for ionized calcium binding adaptor molecule 1, showed no significant increase in the abundance of activated microglia present in the area surrounding the implantation site (Fig. 4j). Additional immunohistochemical analysis showed no sign of an inflammatory response over the 12-week period for either device used, comparable to commercially available rigid probes (Supplementary Fig. 19). Recently published biocompatibility data from some of our laboratories using flexible gFET devices designed with large graphene surface areas implanted epicortically also showed no significant adverse tissue response chronically⁴³. Altogether, the chronic biocompatibility study reported here indicates that gDNPs are suitable for deep brain structure chronic implantations, inducing no significant tissue damage nor neuroinflammatory responses.

We have demonstrated in this study the capability of gDNPs based on linear arrays of gSGFETs to record electrophysiological brain activity in a wide temporal range, with fidelity similar to solution-filled glass micropipettes but with superior spatial sampling. The gDNP is a novel technology that allows concurrent recordings of DC-shifts, ISOs and HFOs in preclinical epileptic models through the cortex laminae to upper hippocampal layers.

Together with the validation of the chronic functionality of implanted gDNPs and their biocompatibility, our work underlines the very distinct advantages of this technology for in vivo epilepsy research. In particular, the gDNPs have allowed us to reveal in rodent models of drug-induced seizures and chronic epilepsy²⁰ associations between DC-shifts (including ictal baseline shifts and SDs^{24,25,26,44,45}) and higher-frequency activity (including SWDs²¹ and HFOs^22,23).

The reliability, stability and high fidelity of the recordings observed in this work suggest that this technology could be eventually used to record clinically relevant electrophysiological biomarkers. In this respect, we envision that clinical development of this technology could be directed towards neural probes capable of simultaneous high-quality wide-band (DC to HFO) recordings from multiple brain regions during pre-surgical monitoring. Moreover, this technology could be applied combining depth and subdural gSGFET arrays capable of stable DC-coupled recordings with high spatial resolution, which will offer advantages over current technologies for epilepsy and neurocritical care^11,46. In particular, it is expected that this technology can help improve the identification of seizure onset zones and ultimately surgical outcome⁵. Currently, long-term seizure freedom after surgery is only around ~50% (ref. ⁴⁷); thus, improvements in a technology that more precisely defines surgical margins is warranted. The few studies that have investigated DC-shifts report that these signals co-localize with seizure onset zones and are more spatially restricted than conventional electroencephalogram (EEG) recordings⁴⁸. Therefore, routine inclusion of accurate DC-coupled recordings in surgical monitoring, which can be enabled by the technology described in this work, could result in less extensive, yet more effective surgical resections⁹. Beyond its use for epilepsy research, we envision that this technology can be applied to advance our understanding of other neurological diseases associated with SD, including traumatic brain injury, stroke and migraine⁴⁹.

Microfabrication of flexible gDNPs

Flexible neural probes were fabricated using standard microelectronic fabrication technology on a rigid 10-cm sacrificial Si/SiO₂ wafer. A 10-µm-thick PI (PI-2611, HD MicroSystems) layer was spin-coated and cured at 350 °C in a N₂ atmosphere. To reduce the shank width of the depth neural probes, we used a two-metal-level strategy in which the metal tracks, separated by PI, are interconnected by via holes. After evaporation and definition (by lift off) of a first metal layer of Ti/Au (10/100 nm), a 3-µm-thick PI layer was spin-coated and cured. A protective mask of Al (200 nm) was used to etch the second PI layer using oxygen plasma and form the via holes. On top of the via holes a second metal layer of Ti/Au (10/100 nm) was applied to interconnect the two metal layers. CVD graphene was transferred onto the patterned wafer as described in the Supplementary Information Methods. The graphene active areas were defined by oxygen-based reactive ion etching. A sandwich-like contact strategy was used to improve the contact at the drain and source terminals; the top metal structure consisted of Ni/Au (20/200 nm). For passivation, a 2-μm-thick chemically resistant polymer was deposited (SU8-2005 MicroChem) with open windows in the channel region. Finally, the gDNP structure was defined in a deep reactive ion etching process using a thick AZ9260 positive photoresist (Clariant) as etching mask. The PI probes were then released from the SiO₂ wafer and placed in a zero insertion force connector to interface our custom electronics. The transconductance of the flexible gSGEFTs could be improved by ultraviolet ozone cleaning of the contact region at the graphene–metal interface, as recently demonstrated for flexible graphene neural sensors⁵⁰. For CVD graphene growth, transfer and its characterization we refer to the Supplementary Information.

Characterization of gDNPs in saline

The gSGFETs on the neural probes were characterized in PBS solution (150 mM). Drain–source currents (I_DS) were measured varying the gate–source voltage (V_GS) versus a Ag/AgCl reference electrode, which was set to ground. Steady state was ensured by acquiring the current only after its time derivative was below a threshold (10⁻⁷ A s⁻¹). The detection limit of the gSGFET was assessed by measuring the PSD of the DC current at each polarization point V_GS. Integrating the PSD over the frequencies of interest (1 Hz–2 kHz) and using the transconductance allowed us to calculate V_RMS. The noise measurement was performed in a Faraday cage, with DC-batteries powering the amplifiers, to reduce any 50 Hz coupling or pick-up noise. Additionally, the frequency response of the device’s transconductance was measured by applying a sum of sinusoidal signals to the electrolyte solution through the reference electrode and by measuring the modulation of the drain current. The acquired signals were split into two bands: low frequencies (~0–10 Hz), at which the drain–source current was simultaneously acquired for all transistors in an array, and higher frequencies (10 Hz–30 kHz), at which each transistor was recorded individually.

Back-coating of gDNP with silk fibroin

Compared with other natural biopolymers, SF offers excellent mechanical properties, extremely good biocompatibility, biodegradability and the versatility of structural readjustments^51,52,53. Further, the byproducts of SF degradation by enzymes (for example, proteases) have low antigenicity and non-inflammatory characteristics^54,55. Temporary stiffening of the gDNPs with SF was achieved using a microstructured polydimethylsiloxane (PDMS) mould with the shape of the neural probes. To fabricate the moulds, PDMS was cast on a standard 4-inch silicon wafer with prepatterned 100- and 200-μm-thick SU8 (SU8-2050) epoxy resin. The back-coating procedure was as follows: first, the probe was placed in the mould trench filled previously with water, with the transistor side facing down. Through surface tension effects, the probe self-aligned in the mould. After evaporation of the water, SF (Silk, Fibroin Solution 50 mg ml^–1, Sigma Aldrich) was applied through a syringe to the mould’s trench. We double-coated the shank in drying intervals of 20 min and then slowly increased the temperature on a hotplate to 80 °C, leaving the SF to cure for 90 min. To have a delayed dissolution time compared with the SF curing time at room temperature, we increased the duration of the water annealing step. After curing, the coated probe can be easily removed from the PDMS mould (Supplementary Fig. 3). In all the in vivo experiments presented here, we implanted the flexible gDNPs with a 150-µm-thick SF back-coating.

Assessment of mechanical properties of the stiffened gDNPs

Standard compression tests against a hard Si substrate were performed to assess the mechanical properties of our SF-coated probes. Buckling experiments were carried out in a UMIS nanoindenter (Fischer-Cripps Laboratories). A custom clamp was fabricated to fix the probes at the end of the indenter shaft, which, in turn, was connected to an actuator and load cell. Buckling tests were carried out at a loading rate of 8.8 mN s^–1. Once the indenter detected noticeable buckling the test was automatically stopped. The maximum applied load that the indenter can apply is 500 mN. Applied force versus displacement was measured until the probe started to buckle and eventually broke down. We additionally measured the Young’s modulus of the SF cured at 80 °C by nanoindentation tests. SF was dropcast on a 2 × 2 cm² Si chip and cured. Indentation was measured using an NHT2 Nanoindentation Tester (Anton-Paar) equipped with a Berkovich pyramidal-shaped diamond tip. A maximum applied load of 5 mN was applied with a loading segment of 30 s followed by a load holding segment of 10 s and an unloading segment of 30 s. The hardness and reduced Young’s modulus are reported as average values of at least 20 indentations, performed on top of each sample (in the central region). Young’s modulus values in the range of 10 GPa were measured for SF cured at 80 °C (Supplementary Fig. 22).

Electronics for in vivo recordings with gDNPs

The experimental set-up used to perform the in vivo recordings provided V_S and V_D bias control and current-to-voltage conversion for up to 16 channels (g.RAPHENE, g.tec medical engineering). The instrumentation split the recorded signals into two bands with different gains: low-pass filtered (LPF; <0.16 Hz, 10⁴ gain) and band-pass filtered (BPF; 0.16 Hz < f < 160 kHz, 10⁶ gain). Two custom Simulink models were used (1) to generate the transfer curve of the microtransistors once inserted and at the end of the experiment, and (2) to set the V_S and V_D bias and acquire the recorded signals. Signals were sampled at 9.6 and 19 kHz, depending on the type of experiment (Supplementary Table 1).

Ethical approval and animal handling for acute and chronic experiments

Animal experiments were conducted in accordance with the United Kingdom Animal (Scientific Procedures) Act 1986, and approved by the Animals in Scientific Research Unit at the Home Office (licence PPL70-13691). C57BL/ mice were bred (2–4-month-old males), whereas WAG rats were imported (Charles River, 6–9 months of age). Animals were housed in a 12 h/12 h dark/light cycle, and food and water were given ad libitum. The animals were group-housed prior to headbar surgery, but after this, animals were individually housed.

Acute preparation surgeries for headbar attachment and craniotomy

For both surgeries, aseptic techniques were used with mice anaesthetized using isoflurane (988–3245, Henry Schein) and placed in a stereotaxic frame (David Kopf Instruments). Viscotears were applied (Bausch + Lomb) and pain relief, which consisted of subcutaneous bupenorphine (0.5 mg per kg mouse (hereafter denoted mg kg^–1 mouse); Ceva) and metacam (15 mg kg^–1 mouse; Boehringer Ingelheim), was injected. Saline was administered just before recovery or every 45 min depending on the length of surgery. To apply the headbars for the Neurotar system, the skin on the top of the head was cut to expose the skull. The skull was cleaned and dried, which enabled drilling (RA1 008, Henry Schein) of a small hole in the left-hand visual cortex for a metal support screw (00-96 × 3-32, Plastics One). The headplate (Model 9, Neurotar) was firmly attached using vetbond (1469SB, 3M) and strengthened using dental cement before the exposed skull was covered with Kwik-Cast (WPI). Mice were checked daily to ensure recovery. After at least 5 days of recovery, habituation was performed by placing the mouse in the Neurotar frame for increasing periods of time (15–60 min) over several days. On the day of recording, a craniotomy was performed. Under isoflurane anaesthesia, with administration of pain medication and intramuscular dexamethasone (1 mg per kg of mouse, intramuscular; 7247104, MSD Animal Health), two areas were exposed. A large (2 × 2 mm²) craniotomy was performed over the somatosensory and visual cortex on the right-hand side and a small drill hole over the motor cortex on the left-hand side. Cold Cortex-buffered saline was continually applied to the craniotomies. After completion, the exposed dura was covered with Cortex-buffered saline, sterilized Sylgard (~200-μm thick) and a Kwik-Cast layer. After ~2 h, the animal was moved to the Neurotar frame and the craniotomies were exposed by removal of the Kwik-Cast and Sylguard layers. The gDNP was carefully connected to a printed circuit board (PCB) and lowered using a micromanipulator to just above the dura over the visual cortex. The dura was gently pierced using either microdissection scissors or a 26-gauge needle and the gDNP lowered ~2 mm into the brain. A reference wire (Ag/AgCl₂) was placed in the ipsilateral motor cortex and g.tec hardware (see the section Electronics for in vivo recordings with gDNPs) was used to generate the transfer curve to determine the optimal V_GS and initiate recordings (Supplementary Fig. 4). Chemoconvulsant was injected into the brain using a Nanofil injection system (WPI). At the end of the experiment, sodium pentobarbital was administered intraperitoneally.

Recording with solution-filled glass micropipette

Borosilicate capillary tubes (outer diameter 1.50 mm, inner diameter 0.86 mm; Warner Instruments) were pulled using a horizontal puller (Sutter Instruments P-97, resistance of 3–5 MΩ), filled with artificial cerebral spinal fluid (119 mM NaCl, 2.5 mM KCl, 2.5 mM CaCl₂, 1.3 mM MgSO₄, 1.25 mM NaH₂PO₄, 25 mM NaHCO₃ and 10 mM glucose) and attached to an CV-7B headstage (Axon Instruments). A micromanipulator (MM3301, WPI) was used to position the pipette above the cortical surface before insertion approximately 400 µm into the cortex. The headstage was provided with the same reference as the gDNP, a chlorinated silver wire touching the ipsilateral motor cortex. The headstage was connected to a Multiclamp 700B amplifier (Axon Instruments) operating in current clamp mode. Analogue–digital conversion and transistor–transistor logic pulse delivery for temporal synchronization was achieved using a Micro1401 MkII unit (CED). Data were acquired by WINEDR sampling at 20 kHz with a 4 kHz Bessel filter.

Chronic preparation surgery and recording

First, the gDNP was coated with fibroin, as described above, to aid insertion. The rat was anaesthetized to a surgical depth using Isoflurane. After placement in a stereotaxic frame, Viscotears were applied and pain medication, which consisted of bupenorphine (0.15 mg kg^–1 mouse, subcutaneous; Ceva) and metacam (4.5 mg kg^–1 mouse, subcutaneous; Boehringer Ingelheim), was injected. The skull was cleaned and dried. Small burr holes (~1 mm) were drilled at four positions: (1) the somatosensory cortex for gDNP (as the perioral somatosensory cortex is the focal area for SWDs⁵⁶), (2) the contralateral cerebellum for a reference Ag/AgCl wire held in place by a nylon screw, (3) the motor cortex, ipsilateral, for a support screw and (4) the visual cortex, ipsilateral, for a support screw. The metal screws were inserted and provided structural support for the dental cement. Next, the gDNP and reference wire were inserted, and a DC characterization curve confirmed that the transistors were performing optimally. Dental cement, mixed with vetbond, was applied around the PCB for support. Animals were weighed daily and their physiology was monitored to ensure a full recovery. For recording, animals were anaesthesia-free and moving, with the PCB interface on the head connected to an Omnetics cable (A79635, Omnetics) that interfaced with the g.tec recording hardware as described above. After allowing ~5 min for settling, a DC characterization curve was recorded to allow accurate calibration of the gSGFETs. A script was used to calculate the optimal V_GS based on the transfer curves. Recordings were performed for ~10–60 min twice a week for 10 weeks. After recording, the Omnetics wire was disconnected and a protective cap was applied.

WAG/Rij strain of rats

WAG/Rij rats were used as a model of absence epilepsy. Spike-wave activity in the cortical EEG of adult WAG/Rij rats has a frequency of 7–10 Hz and can last from just a few, to up to 40 s in duration. From an age of 6 months, WAG/Rij rats display several hundred SWDs a day²⁰. In addition to electrophysiological signs, mild behavioural phenotypes concurrent with SWD similar to human absence epilepsy are displayed. These include facial myoclonic jerks, twitching, accelerated breathing, head tilting and eye blinking. Drugs prescribed for human absence epilepsy suppress SWDs in WAG/Rij rats, and these rats have been used to predict the action of antiepileptic medication and other experimental treatments⁵⁷.

Data analysis

All electrophysiological data were analysed using Python v.3.7 packages (Matplotlib, Numpy, Neo and Elephant) and the custom library PhyREC (https://github.com/aguimera/PhyREC). The conversion of the recorded current signals (LPF and BPF) to a voltage signal was performed by summation of the two signals and inverse interpolation in the in vivo/chronic measured transfer curve of the corresponding gSGFET¹⁷. Even for large signals (tens of millivolts), this calibration method yields distortion-free signals, in contrast to using a fixed transconductance value³⁰. Transfer curves were always measured at the beginning and end of every recording to ensure that no significant variations were present and to detect any malfunctioning transistor. Moreover, all recordings presented in the manuscript have been calibrated with the nearest-recorded transfer curve to ensure high fidelity in the voltage-converted signals.

To assess the detection limit in the conventional frequency bands (>1 Hz), baseline recordings were directly compared with post-mortem recordings (Fig. 2f). For instance, applying a digital filter (>200 Hz) and comparing post-mortem with baseline validates the ability of gDNP to record spontaneous high-frequency activity (>200 Hz) arising from groups of neurons in a non-pathological brain state (upper plot in Fig. 2f). Beyond this qualitative comparison, the SNR shown in Fig. 2f was evaluated from the ratio of the RMS mean value over 25 s of recording of baseline (spontaneous activity) and post-mortem (no activity) signals. The signals were BPF in three different bands, corresponding to local field potential activity (1–70 Hz), high-frequency activity (80–200 Hz) and very high-frequency activity (200–4,000 Hz). RMS values were calculated with a sliding window of 500 ms for the 1–70 Hz band and with a sliding window of 10 ms for the other two bands. The whisker plots were created with a Python library (seaborn.boxplot) to show the data distribution for working transistors in the gDNP (gDNP1 = 10, gDNP2 = 12, gDNP3 = 12 and gDNP4 = 14). In the whisker plots, the line in the box corresponds to the median value, the box shows the quartiles of the dataset and the whiskers extend to show the maximum and minimum values. The SNRs for the different bands were evaluated from a total of four in vivo experiments with four different gDNPs (Fig. 2f). The SNR is expressed in decibels (20 × ln(RMS(S)/RMS(N)), where S is the ‘signal’ during the baseline recording and N is the ‘noise’ measured during the post-mortem recording. The silencing of neuronal activity shown in Fig. 3d was extracted from the AC-coupled recording (HP-filtered, >0.5 Hz). Then, the RMS values of the pre-ictal phase (calculated with a sliding window of 1 s) were averaged over 50 s. A similar analysis was performed during the SD (15 s, shaded areas in Fig. 3d). The ratio of the two averaged RMS values corresponds to the neuronal activity variation (before and during the SD). The amplitudes of the hippocampal SD and the hyperpolarization wave in Fig. 3d were evaluated using the recording LPF in the infraslow regime (<0.5 Hz) and resampled at 3 Hz (instead of at 9.6 kHz, used in the original recording). The zero of the voltage was set using the mean value of the signal 50 s before the pre-ictal phase (Supplementary Fig. 9), and the minimum and maximum values for each channel were extracted (corresponding to the SD and hyperpolarization amplitude, respectively).

For the analysis of the DC-shifts pre-seizure (Fig. 3f and Supplementary Fig. 17), 20 min recordings starting from 1 min before the first seizure following 4-AP injection were selected. Seizures with concurrent high-amplitude depolarizations such as SDs were excluded from analysis. The 20 min recordings were filtered in two bands: DC-coupled (LPF, <0.1 Hz) and seizure band (bandpass second-order filter, 4–12 Hz). Signals were downsampled to 96 Hz. The RMS value of the seizure band was calculated with a 0.5 s sliding window and used to determine seizure onset and end. Epochs around −5 to 2 s relative to seizure start of the DC-coupled signal were extracted for analysing the DC-shifts. Zero voltage was set to the mean of the first 2 s of the epoch, and the maximum amplitude of the period from −5 to 0 s was then extracted and used as the amplitude of the pre-seizure DC-shift.

CSD analysis applied to the low-frequency part of the potential (local field potential) was calculated with the Python open source Elephant library (Elephant Electrophysiology Analysis Toolkit) using the class ‘Current Source Density analysis (CSD)’, and the method 1D – StandardCSD was chosen for the linear gDNP array. A homogeneous conductivity of the neural tissue of σ = 0.3 S m^–1 across the different layers was used for the calculations.

The concurrency of ISO with the SWD shown in Fig. 4h was evaluated by performing the Hilbert transform to extract the phase of the DC-coupled signal (0.01–0.1 Hz) and associating it with the RMS in the range 5–9 Hz evaluated with a sliding window of 0.5 s (typical bandwidth for the SWD). Supplementary Fig. 16 shows in more detail the dependency of the ISO phase and SWD amplitude. The data shown in Fig. 4h illustrate the joint density probability, indicating the higher probability of a given RMS amplitude to ISO phase pair.

Chronic biocompatibility study

Details of the surgery, tissue collection, analysis of histology and data statistics used for the chronic biocompatibility study are provided in Supplementary Figs. 19 and 20.

Reporting Summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.