销售热线

13818239648
主营产品:试剂,耗材,血清,细胞,抗体
  • 技术文章ARTICLE

    您当前的位置:首页 > 技术文章 > An organic artificial spiking neuron for in situ neuromorphic sensing and biointerfacing

    An organic artificial spiking neuron for in situ neuromorphic sensing and biointerfacing

    发布时间: 2023-03-07  点击次数: 964次

    nature electronics

     

     

     


     




    An organic artificial spiking neuron for in situ neuromorphic sensing and biointerfacing

     

     

     

    Received: 2 March 2022

     

    Accepted: 29 September 2022

    Published online: 7 November 2022

     Check for updates

     

    Tanmoy Sarkar1,7,8, Katharina Lieberth1,8, Aristea Pavlou1, Thomas Frank  2, Volker Mailaender 1,3, Iain McCulloch 4,5, Paul W. M. Blom  1,

    Fabrizio Torricelli  6  & Paschalis Gkoupidenis  1 

     

    The efective mimicry of neurons is key to the development of                     neuromorphic electronics. However, artifcial neurons are not typically       capable of operating in biological environments, which limits their ability   to interface with biological components and to ofer realistic neuronal        emulation. Organic artifcial neurons based on conventional circuit             oscillators have been created, but they require many elements for their       implementation. Here we report an organic artifcial neuron that is based    on a compact nonlinear electrochemical element. The artifcial neuron can  operate in a liquid and is sensitive to the concentration of biological species (such as dopamine or ions) in its surroundings. The system ofers in situ       operation and spiking behaviour in biologically relevant environments      including typical physiological and pathological concentration ranges        (5150 mM) —and with ion specifcity. Small-amplitude (1 –150 mV)               electrochemical oscillations and noise in the electrolytic medium shape the neuronal dynamics, whereas changes in ionic (≥2% over the physiological    baseline) and biomolecular (≥ 0.1 mM dopamine) concentrations modulate the neuronal excitability. We also create biohybrid interfaces in which an     artifcial neuron functions synergistically and in real time with epithelial cell biological membranes.

    Neurons are the fundamental units of the nervous system and are used to transmit and process electrochemical signals. They operate in a liquid electrolyte and communicate with each other via gaps (syn- apses) between the axon of pre-synaptic neurons and the dendrite of post-synaptic neurons (Fig. 1a). Neuromorphic computing uses hardware-based implementations to mimic the behaviour of synapses and neurons1, for efficient brain-inspired computing. The approach could also be used to interface biology with electronics that share similar biocomputational primitives27. Synaptic phenomena the gradual and activity-dependent coupling between neurons —are typi- cally mapped onto memory devices, which can be binary, multistate

    or analogue. However, to emulate neuronal spiking and oscillatory dynamics, electronic oscillatory circuitry is required8,9.

    Neuron-like dynamics can be created with conventional micro- electronics using oscillatory circuit topologies to mimic neuronal behaviours. For example, neuromorphic electronic circuits consisting of ring oscillators have been used for the implementation of mechani- cally flexible, skin-inspired electronics and neuro-inspired mechanore- ceptors10,11. Many-element artificial neurons based on solid-state silicon or organic devices have also been reported9,1113. However, although these approaches can mimic specific aspects of neuronal behaviour, the integration ofa large numbers oftransistors and passive electronic

    1 Max Planck Institute for Polymer Research, Mainz, Germany. 2 Max Planck Institute of Neurobiology, Martinsried, Germany. 3 Dermatology Clinic, University Medical Center of the Johannes Gutenberg-University Mainz, Mainz, Germany. 4 KAUST Solar Center, Physical Science and Engineering Division, King        Abdullah University of Science and Technology (KAUST), Thuwal, Saudi Arabia. 5 Department of Chemistry, University of Oxford, Oxford, UK. 6 Department of Information Engineering, University of Brescia, Brescia, Italy. 7 Present address: Infineon Technologies AG, Dresden, Germany. 8These authors                 contributed equally: Tanmoy Sarkar, Katharina Lieberth.

    Nature Electronics | Volume 5 | November 2022 | 774–783                                                                                                                              774


    b

     


    Fig. 1 | A biological neuron and the OAN. a, Simplified schematic of a biological neuron. Action potentials, the basic cell-to-cell communication events, are         generated by rapid transmembrane ion exchanges through ion channels, and      they propagate across the axon. In myelinated cells, alternate myelin/non-myelin domains (nodes of Ranvier) contribute to the fast and long-range action              potential propagation. Biological neurons are immersed in an electrochemical   environment, such as an aqueous electrolyte. This extracellular space is              a common reservoir containing various biological carriers for signalling             and processing (ions, biomolecules and so on). Noise is also present in this          environment. Ionic channels on the membrane endow neurons with ionic/          molecular specificity and recognition. b, Circuit diagram ofthe OAN. The main   part is an OEND that displays S-shaped negative differential resistance (S-NDR)


    phenomena and is responsive to ionic and biomolecular species common to        biological environments. The OEND consists oftwo OECTs, namely, T1 and T2,       that are connected via the R1 = 5 kΩ and R2 = 10 kΩ resistors in a cascade-like          configuration with feedback. The OAN is formed when the OEND is connected to  an RC element (R = 10 kΩ; C = 6 nF to 10 μF) and voltage source Vin. Here Vout and Iout are the resulting output voltage and current, respectively, ofthe OEND under the influence of Vin. G, gate; S, source; D, drain. c, Schematic ofthe OECT that forms    the OEND. The channel ofthe OECT consists of an organic mixed ionic–electronic conductor (OMIEC), such as PEDOT:PSS. d, Sensing mechanism in PEDOT:PSS.     Ionic or polyatomic ions interact with PEDOT:PSS and modulate the doping level  and hole conductivity of PEDOT:PSS, which results in a change in the OECT drain  current and threshold voltage.


     

     


    components results in bulky biomimetic circuits that are not suitable for direct, in situ biointerfacing.

    Volatile and nonlinear devices based on memristors or spin torque oscillators can be used to increase the integration density and emulat- ing neuronal dynamics14,15. Metal-oxide memristive devices based on metal insulator transitions exhibit negative differential resistance phenomena that are suitable for the emulation ofneuronal dynamics14. Such devices with a negative differential resistance are locally active, and therefore, electrical input stimuli trigger voltage or current spikes in analogy with biological neurons. Artificial neurons based on memris- tive devices with a negative differential resistance have the potential for high integration density1,16. Nevertheless, the intrinsic sensitivity of solid-state memristive devices to moisture prevents in situ biointerfac- ing in biologically relevant host environments17. Although memristive arrays have been used for pre- and post-acquisition biosignal process- ing, they have not been used for in situ biointerfacing18,19.


    Spin torque oscillators are magnetic nanodevices compatible with silicon technology1,15. Their nonlinearity and dynamics have been recently leveraged for spoken language and audible source recogni- tion15. However, there is no viable route for biointerfacing with spin torque oscillators, as their oscillatory dynamics are too fast (around gigahertz frequencies) for interacting in real time with biological processes. Their operation also requires the presence of magnetic fields. Other approaches for spiking or oscillatory devices and cir- cuits —including Mott-transition-based memristive devices20,21, fer- roelectrics22, photonics23and two-dimensional materials24have been developed, but all of them encounter similar problems. By omitting various aspects ofactual biological wetware, artificial neurons based on electronics are insufficiently capable ofemulating/handling the biosig- nal diversity and thus of operating in situ in biological environments.  Organic electrochemical devices based on organic mixed ionic electronic conductors offer an alternative approach to neuromorphic


     


     

    a

     

    V(I) (OEND)

    I(V) (OEND)

    Vout, Iout (OAN)

    OAN: OEND + V  + RC

     

     

     

     

     

     

    Voltage

    b

     

     


    C = 1 µF

     

    C = 5 µF


     

     

    0       0.2      0.4      0.6      0.8      1.0

    Time (s)

    c

    54

    52

    50

    48

    46

    0              20             40             60             80

     Vin (mV)

    150

    100

    50

    50       17,000   35,000   50,000  105,000

    Number of spikes

     


     


    Fig. 2 | Nonlinear phenomena ofthe OEND and bifurcation. a, Nonlinear          phenomena ofthe OEND in the V(I) mode (application of current sweep and         measurement ofthe resulting voltage) and I( V) mode (application ofvoltage       sweep and measurement ofthe resulting current). Parametric oscillations           (Lissajous-like plot) ofthe OAN (OEND + Vin + RC) and their projection in the         current versus time plane are shown. Here Iout ( Vout) is the output current (voltage) between the OEND and ( Vin + RC). b, Regular (tonic) current spiking Iout for


    various capacitances C = 110 μF and Vin = 1.75 V. They axes are the same for all     the subpanels. c, Frequency response of a voltage-controlled oscillator (C = 1 μF). This response shows the firing frequency versus input voltage difference (ffir             versus ΔVin), where ΔVin = Vin Vth is the input voltage above the threshold voltage Vth for spiking. d, Continuous OAN firing under fixed Vin. The insets show the        stability ofthe spiking waveform for 50, 3.5 × 104 and 1.05 × 105 spiking cycles. All the measurements are performed in an aqueous electrolyte (100 mM NaCl).


     

     

     


    electronics25. Organic electronics can operate in close proximity to biology due to their soft nature and the ability to directly interact with ions in aqueous electrolytes3,26,27. Organic synaptic transistors have developed rapidly, showing outstanding analogue memory phenomena with low-voltage operation and linearity in weight update (in contrast, nonlinear phenomena are required for the implementation of neu- ronal dynamics)2831. In addition, organic artificial synaptic devices have demonstrated neurotransmitter-mediated plasticity when cou- pled with dopaminergic cells32,33 and have also been interfaced with non-electrogenic cells34. However, such devices are passive elements and therefore are not able to emulate neuronal dynamics. Recently, organic electrochemical neurons have been reported13. However,this approach is based on many-element conventional oscillatory circuitry and does not possess the complexity in neuronal dynamics that nonlinear elements can potentially display. The integration of neuromorphic electronics with biology requires artificial synapsesthat can interfacewith biological ones, as well as artificial spiking neurons that can operate and respond to local biological signals in situ, that is, in the wet biological environment.  In this Article, we report an organic artificial spiking neuron with electrobiochemical degrees of control that enable local and in situ neuromorphic sensing and biointerfacing. The organic artificial neuron (OAN) can operate in a liquid and shows inherent biosensing primitives. It consists ofa compact nonlinear electrochemical element that exhib- its negative differential resistance that is sensitive to the biological environment that hosts the OAN. The OAN responds to ionic species commonly found in the extracellular space, and its spiking response is sensitive to typical physiological and pathological ionic concentra- tion ranges (5–150 mM). Small-amplitude (1 –150 mV) electrochemical oscillations and noise in the electrolytic medium shape the neuronal firing properties. Therefore, the artificial neuron exhibits spiking properties (which are stable for >105 spiking cycles) that depend on the local ionic, biomolecular or neurotransmitter species ofthe aqueous environment behaviour that is analogous to a biological neuron that is surrounded by the extracellular space containing various biological carriers for signalling and processing (Fig. 1a). In particular, in situ changes in ionic (≥2% increase) and biomolecular (≥0.1 mM of dopa- mine) concentrations modulate the neuronal excitability and trigger spikes. Ion-specific oscillations for sodium (Na+) and potassium (K+) are also possible, providing a pathway for emulating ion channel dynamics. Furthermore, we create a biohybrid interface in which artificial neurons synergistically function with membranes of epithelial cells and where

    the biological membrane barrier modulates the spiking properties of the artificial neuron in real time.


    Organic artificial spiking neuron

    The OAN consists ofa compact nonlinearbuildingblock made ofonlytwo organic electrochemical transistors (OECTs), namely, T1 and T2 (Fig. 1b). Both OECTs are p-type transistors: T1 is a depletion-mode transistor, whereas T2 is an enhancement-mode transistor. The mixed ionic  electronic conductor poly(3,4-ethylenedioxythiophene) (PEDOT) doped with poly(styrene sulfonate) (PSS) and poly(2-(3,3′ -bis(2-(2- (2-methoxyethoxy)ethoxy)ethoxy)-[2,2′ -bithiophen]-5-yl) thieno [3,2-b] thiophene) (p(g2T-TT)) are used for the T1 and T2 channel, respec- tively35,36. The electrical characteristics of T1 and T2 are presented in Supplementary Fig. 1. The OECTs operate in aqueous environments and are sensitive to ionic species and polyatomic ions (Fig. 1c). For instance, the channel ofT1 consists ofthe organic mixed ionic–electronic conduc- tor PEDOT:PSS (ref. 26). Both channel and gate of an OECT are in direct contact with the electrolyte. In the case of a p-type OECT, when a posi- tive gate voltage VG is applied, cations drift into the polymeric channel and reduce the hole concentration, and consequently, drain current ID is lowered. When a negative VG is applied, cations are removed from and anions drift into the polymeric channel, the hole concentration increases, and this results in a larger ID. In OECTs, ions can penetrate the bulk ofthe polymer and the volumetric nanoscale ionic–electronic charge compensation results in a large current modulation. This high gate voltage to drain current modulation yields a high transconduct- ance, which is the hallmark of OECTs. Another key feature of OECTs is the dependence of ID on the ion concentration in the electrolyte. More precisely, the fixed charges in the ion-conducting phase of the polymeric channel are electrostatically compensated by the mabileions provided by the electrolyte. This Donnan equilibrium results in a concentration-dependent voltage drop at the polymer/electrolyte interface, which is, in turn, mirrored by the OECT threshold voltage. As shown in Fig. 1d, ionic or polyatomic cations (anions) interact with PEDOT, a hole conductor, and PSS, an ionic conductor, and decrease (increase) the doping level of PEDOT, resulting in a decrease (increase) in ID and the OECT threshold voltage (Fig. 1d). Here T1 and T2 are con- nected in the cascade-like configuration with feedback resistors R1 and R2, obtaining a two-terminal organic electrochemical nonlinear device (OEND) (Fig. 1b).

    The currentvoltage V(I) and voltagecurrent I( V) characteristics ofthe OEND are displayed in Fig. 2a. The OEND is accessed either in the V(I) or I( V) mode with current I or voltage Vas the independent variable, respectively. An S-shaped negative differential resistance is accessible only in the V(I) mode, by applying current at the single-valued nega- tive differential resistance characteristic. In the I( V) mode, the OEND


     


    shows m*lued characteristic with unstable points of operation to be directly accessed by applying a voltage. A detailed analysis of the OEND response with experimental and simulation data is shown in Supplementary Figs. 1 and 2. Consecutive V(I) or I( V) scans display highly reproducible responses of the OEND (Supplementary Fig. 3). The time response τOEND of the OEND element for square-wave input pulses is τOEND ≈ 11 ms (Supplementary Fig. 3).

    When the OEND is coupled to an RC element (Fig. 1b) forming an OAN, its response bifurcates, producing voltage or current oscillations (Fig. 2; load-line analysis is shown in Supplementary Fig. 4). These spike-based oscillations represent the‘action potentials’of the OAN (Supplementary Fig. 5). Figure 2b displays the current response Iout ofthe OAN for different values of capacitor C. The firing frequencyffir can be finely tuned between 6 and 40 Hz by changing C, a range that is consistent with physiological levels of instantaneous firing rates in biological neurons37. The firing frequency range can be further extended by modifying C or by using high-frequency-response OECTs38. In the case of polymer-based capacitors, capacitances in the range of nanofarads to microfarads can be reached with micrometre-scale PEDOT:PSS-based capacitors26. The parametric oscillatory response in the IV plane and the instantaneous power dissipation Pinst of the OAN is shown in Supplementary Fig. 6. The instantaneous power dis- sipation is given by Pinst = Vin/R( Vin Vout), where Vin, Vout and Iout refer to the input voltage, output voltage and output current of the whole OAN, respectively. The calculation includes all the OAN components (for example, transistors, resistors and capacitor). The mean power dissipation Pmean = 143 μW. As shown in Fig. 2cand Supplementary Fig. 7, the OAN behaves as a voltage-controlled oscillator andffir is modulated by increasing ΔVin Vin = Vin Vth, where Vth is the OAN oscillation thresh- old) within the oscillation window. For ΔVin = 0 –70 mV, the relative increase in firing frequency is Δffir/fmin 18%, with a voltage-controlled oscillator sensitivity of Δffir/fminVin 260% per volt.

    The amplitude and window of the current or voltage oscillations can be precisely designed (Supplementary Fig. 8), for instance, by engi- neering the threshold voltage oftransistors T1 and T2. In this case, the oscillatory window is shifted to a lower voltage level when decreasing the threshold voltage ofT1 by doping PEDOT:PSS with the amine-based molecular dopant N-methyl-2,2′ -diaminodiethylamine (ref. 39), and Pmean is decreased from 143 to 24 μW. For a capacitor with C = 6 nF and by neglecting the static power dissipation, the energy consumption per spike is Espike = 57 nJ. The amplitude profile of the current oscil- lations can also be engineered by varying R1 and R2 of the OAN (the effect ofR1 and R2 on the nonlinear properties ofthe OEND is shown in Supplementary Fig. 2). It should be noted that doping causes perma- nent threshold-voltage shifts and dynamic reconfigurability can be induced by introducing synaptic transistors instead of volatile ones. In the case of PEDOT:PSS, the threshold voltage can be tuned on the fly by changing the ion concentration40, a phenomenon that is exploited here to obtain an ion-concentration-dependent firing frequency. The OAN displays fully consistent stability when continuously operated at various amplitude and frequency conditions, as the firing response is practically unaffected for >105 spiking cycles (Fig. 2d). The OAN stability as a function of the number of spikes is also evaluated. The amplitude of Iout reduces by ~2.8 × 10−5% per spike. As a result, after 106 spikes, the OAN current amplitude is equal to approximately 71% ofthe initial amplitude.

    The OAN shows the key characteristics observed in the spiking response of biological neurons. The OAN operates in a liquid, a prop- erty that is reminiscent of the extracellular environment of biologi- cal neurons in the cerebrospinal fluid. The excitability of the OAN, that is, the tendency of a neuron to fire spikes, can be modulated by the presence of electrochemical oscillations transmitted by means of ionic fluxes in the electrolytic medium. Figure 3a shows that the in-liquid electrochemical oscillations shape the firing properties of the OAN, mimicking the characteristic features ofbiological neurons41.


    An increase of only a few millivolts at the potential of the electrolytic medium elicits spikes with high temporal precision (Supplementary Fig. 9), and a forced bursting activity is phase locked with the ionic signal in the electrolyte. During the time window of the input signal that is above the OAN threshold voltage Vth, the OAN fires and there- fore the input phase coordinates firing. Variation in the electrolyte potential as small as 1 –2 mV at the very edge of the OAN threshold Vth results in stochastic firing, thus reproducing the behaviour observed in biological neurons42. Such small variations in the electrolytic medium potential are in the same range ofthe biopotentials ofthe extracellular electrolytic space (microvolts to millivolts)43. The OAN threshold allows for additional bioplausible behaviours, such as in-liquid all-or-nothing spiking and subthreshold oscillations (Supplementary Figs. 9 and 10, respectively). The spiking properties of the OAN for input voltage pulses is shown in Supplementary Fig. 9.

    Due to its finite response time, the OAN displays a stimulus  response delay (Fig. 3b), as well as behaves as a temporal integrator (Fig. 3c). The stimulusresponse delay in biological neurons, known as spike latency, can provide a rapid and efficient neural coding scheme beyond simple rate coding, as latency can be a faster differentiator than the mean firing rates44. We further investigate this delay and reproduce the spike latency characteristic ofbiological neurons. Figure 3bshows that the stimulus-firing phase difference Δφ is modulated by the input voltage difference ΔVin (time-domain response is shown in Supple- mentary Fig. 11). Stronger stimulation induces shorter latencies, as observed in biological sensory systems45. Due to the finite and relatively slow time constant of biomembranes, biological neurons temporally integrate inputs and display firing under certain interstimulus timing conditions46. This integration lowers the timing precision and offers temporal buffering windows for ongoing neuronal inputs, ensuring the consolidation and stability of neuronal sequences46. Analogous to biological neurons, the OAN integrates time, buffers input stimuli and fires for short non-overlapping interstimulus intervals (Fig. 3c).

    Biological environments are characterized by seemingly random fluctuations at a range of spatiotemporal scales, and therefore, neu- rons are constantly operating under noisy conditions47. This noise couples with neuronal dynamics, influencing neuronal excitability and firing properties. Although counterintuitive, noise can be benefi- cial for neuronal communication and processing. For instance, noise can enable the transmission of weak subthreshold signals, smooth- ens subthreshold-to-threshold nonlinearities or even enhances the communication efficiency by increasing the signal-to-noise ratio or by altering the neuronal coding schemes47. The noise-induced activ- ity of the OAN is presented in Fig. 3d. The OAN is biased with a d.c. input voltage at the subthreshold regime, and white noise of variable amplitude ( Vpp = 5−150 mV) is injected in the electrolytic medium to emulate extracellular noise/fluctuations. As the amplitude ofthe noise increases, a gradual transition from tonic ( Vpp = 0−25 mV) to irregular firing ( Vpp = 50 mV) is observed. For low noise levels, the frequency of tonic firing remains practically constant, that is,ffir 6.5 –7.0 Hz. As the noise level increases, packets of spikes are observed. Recurrence plots of the interspike intervals and spike-to-spike amplitudes (Sup- plementary Fig. 12) indicate a change in the coding scheme from tonic to noise-induced bursting activity, as well as the resilience of spiking against noise in the electrolytic medium.

    In situ spike-based neuromorphic sensing

    It is estimated that the extracellular electrolytic space occupies a volume fraction of ~15 –30% of the brain tissue. This extracellular space is an aqueous electrolyte comprising various ionic species (mostly Na+, K+, Cland Ca2+) and represents a reservoir by maintain- ing homoeostatic balance of the ion concentrations under physi- ological conditions. In mammalian cells, the range of physiological concentrations for Na+ is cext = 130−150 mM and cint = 10−15 mM and for K+, cext = 3−12 mM and cint = 150−160 mM, where cext and cint are the


     

     

     

    th

     



     

    Vth + 2 mV

     






    Vth + 10 mV




    4           5           6

    Time (s)

     

     

     

    100

     

    50

    80            60

     

    10

    5

    0

    40                60

    Time (s)

     


    Fig. 3 | An organic artificial spiking neuron. a, Electrochemical

    oscillation-mediated neuronal firing. The OAN is biased with a time-varying input voltage Vin via the electrolytic medium ofT1 and T2, with a d.c. bias of 1.68 V and    an a.c. signal of50 mVpp (≤ Vth ofthe OAN). Phase-dependent firing is observed      for electrolyte voltage differences of 1−40 mV above the firing threshold

    Vth. b, In-liquid spike latency (firing phase delay Δφ versus ΔVin). c, In-liquid  temporal integration. Firing occurs for short interstimulus time intervals or high duty cycles. d, Noise-induced neuronal firing ofthe OAN with different


    levels of amplitude ofwhite noise in the electrolytic medium. Iout versus time   (top) and the corresponding timefrequency analysis (bottom) or short-time Fourier-transform spectrogram for representative amplitudes ( Vpp = 25, 50     and 150 mV). White noise at the electrolytic medium ( Vpp = 5−150 mV) induces activity transition from tonic firing to bursting, with constant firing frequency ffir ≈ 6.5−7.0 Hz. All the measurements are performed in an aqueous electrolyte (100 mM NaCl).


     

     

     


    extracellular and intracellular concentrations, respectively48. Although homoeostatically balanced, these ion concentrations can change in different spatiotemporal scales. Minute concentration variations in the proximal extracellular space of a neuron are induced during an action potential (fluctuations of approximately 5% for Na+ and 20% for K+ from physiological baselines)49. Many pathological conditions (for example, spreading depression, epilepsy and migraines) are mani- fested as homoeostatic imbalance; for instance, extracellular Na+ and K+ concentrations can decrease to 60 mM and increase to roughly 55 mM, respectively50. Moreover, the extracellular medium contrib- utes to intercellular, non-synaptic communication via variations in the extracellular electric fields (for example, ephaptic coupling)51, neurotransmitter spillover (such as diffusion to adjacent synapses)52 or via extracellular-mediated diffusion of neuromodulators53.

    The OAN, due to its in-liquid operation, exhibits firing proper- ties that depend on the ionic concentration of the electrolyte host. As the electrolyte concentration increases, the ionic conductivity of the electrolyte also increases; thus, T1 (or T2) has a lower response time andffir increases54. Figure 4a shows the dependence of the firing waveform on the NaCl concentration (in an aqueous solution), cNaCl. Hereffir increases from 25 to >50 Hz for cNaCl = 80 –150 mM, a range that is on par with common physiological and pathological extracel- lular Na+ concentrations (Fig. 4b). This results in a relative increase in the firing frequency of Δffir/fmin 110% and agrees with the behaviour


    of biological neurons in whichffir increases with the extracellular Na+ concentration55. As further confirmation, Supplementary Fig. 13 shows that the OAN also operates under the common physiological/patho- logical range ofextracellular K+ concentrations in an aqueous solution, namely, cKCl = 5 –50 mM; furthermore, in this case,ffir increases with the K+ concentration. The responsiveness ofthe OAN at biophysically relevant ionic concentration ranges is essential for the in situ operation ofthe OAN with biological membranes and neurons.

    Changes in ionic concentration gradients between the intracellular and extracellular medium ofbiological neurons alter their excitability/ threshold, and firing can be initiated by varying these concentrations56. As an example, Fig. 4c shows that small variations (~2 –10%) in NaCl concentration over a biologically relevant baseline (100 mM NaCl) can increase the OAN excitability and induce firing. Bothffir and time delay between the increase in concentration and neuronal excitation Δtfir correlate with the ion concentration (Supplementary Fig. 14). Moreover, the behaviour shown in Fig. 4c is in agreement with the excitation profiles for similar extracellular Na+ concentrations of the Hodgkin–Huxley neuron model (Supplementary Fig. 14).

    Dopamine is a modulatory neurotransmitter that regulates essen- tial brain functions including cognition, learning, motivation, motor control, mood regulation and addiction57. At the cellular level, dopa- mine can impact neuronal excitability in a multitude ofways (resulting in excitatory or inhibitory and time- and concentration-dependent


    a                                                              b                                                                                 c


     

    150

    100

    50

    150

    100

    50

    150

    100

    50

     

     

     

    250

    Time (ms)

    50

     

    40

     

    30

     

    20

        c

    Physiological range

    80              100              120              140              160

    cNaCl (mM)

    150

    100

    50

    150

    100

    50

    150

    100

    50

     

    100 mM

    NaCl

     

    +5% NaCl

     

    +10% NaCl

    0                    2.5                   5.0                   7.5

    Time (s)

     


     


     

    d

     

    150

    100

    50

    150

    100

    50

    150

    100

    50

     

     

     

     

     

    Excitation

     

    0.25 mM dopamine

     

    0                     20                    40                    60

    Time (s)

    e

    ion-selective transistor (s) (K+, Na+ …)

     

     

    120

     

    100

     

    80

     

    60

     

     

    Selectivity, K+

     

    KCl (selective)           NaCl (non-selective)

     

    6

    Time (s)

     


     


    Fig. 4 | In situ spike-based neuromorphic sensing. a, Representative waveforms ofthe OAN firing response for different ionic concentrations (here aqueous         NaCl). b, Electrolyte-controlled oscillators ofthe OAN. Firing frequency as a         function of ionic concentration,ffir versus cNaCl. The range of extracellular Na+             concentrations cext under common physiological and pathological conditions in  the brain is indicated by the red dashed line. c, Electrolyte-induced excitability     ofthe OAN. Minute perturbations in ionic concentrations (2−10% NaCl) over        the physiologically relevant baseline (100 Mm NaCl) enhance the neuronal           excitability and initiate tonic spiking. d, Iout as a function oftime. The presence


    of dopamine (0.1−2.0 mM) in the cell culture solution (phosphate-buffered         saline) excites the OAN and induces tonic spiking, demonstrating                       neurotransmitter-induced excitability. e, Ion-selective OANs induce ion-specific oscillations, emulating the dynamics of biological ion channels in liquid and on   chip (channel conductancegK,gNa). As an example, the OAN is selective to K+ ions and insensitive to Na+ interfering ions. In all the experiments (ion concentration  changes, exposure to dopamine and incorporation of ion-selective membranes), the sensing device is T1 and the same electrolyte conditions (100 mM NaCl) are    maintained for T2.


     

     


    effects), both via synaptic and non-synaptic activation of dopamine receptors58. For example, dopamine has a net excitatory effect on primate pyramidal neurons59. As shown in Fig. 4d, the presence of dopamine in the electrolyte increases the excitability of the OAN and initiates tonic firing. Excitation is observed in shorter time delays Δtfir for higher dopamine concentrations, whereas a slight decrease inffir is observed (Supplementary Fig. 15). The behaviour shown in Fig. 4d highlights that the initiation ofspikes can be biochemically triggered, a property that resembles the biological neuronal signalling phenomena. It should be noted that the OAN also exhibits biorealistic diversity in signalling, as dopamine-mediated inhibition can be induced by altering the OAN biasing scheme (Supplementary Fig. 15).

    The ion channels of biological membranes pass inward and out- ward ionic currents, a hallmark of neuronal signalling60. The dysregu- lation of these processes is the consequence of a large number of channelopathies that can lead to serious pathological conditions such as cystic fibrosis and myotonia congenita61. Another layer ofbiophysi- cal realism is added to the OAN response by implementing on-chip selectivity and specificity characteristics, akin to biological ion chan- nels. Figure 4e demonstrates that the OAN function directly incorpo- rates aspects ofthe molecular machinery that are responsible for the selective processing of the biological carriers of information. OANs displaying ion-specific (Na+ or K+) oscillatory activities are realized by incorporating ionophore-based selective membranes62at the channel/ electrolyte interface of T1 (Fig. 1b). As an example, a K+-selective OAN


    shows oscillations in the case ofthe KCl electrolyte, withffir increasing with the ion-selected concentration, but does not show oscillations in a control experiment with NaCl electrolyte (Supplementary Fig. 16). The selectivity library can be further extended with membranes that are selective to other biologically relevant ions such as Na+, Ca2+ and ammonium (NH4+) (refs. 63,64).

    Biohybrid neuron

    The OAN is capable of direct biointerfacing in a biologically relevant environment, and a biohybrid neuron is formed by incorporating a biomembrane between the gate and channel of T1 (Fig. 5). The sys- tem consists of a biological and artificial compartment. As a relevant biomembrane model system, the prototypical epithelial cell line Caco- 2, which is a model of the intestinal epithelial barrier and widely used for in vitro toxicology and drug delivery studies6567, is incorporated with the OAN. The biomembrane comprises epithelial cellsjoined with tightjunctions, thus forming a natural barrier for ion passage (Fig. 5a). The biohybrid neuron functions in situ and in real time (Fig. 5b). Initially, the OAN is operated in a plain cell culture medium; in the absence of the biomembrane barrier, electrochemical oscillations are sustained atffir 12 Hz. A barrier functionality is induced with the incorporation ofthe biomembrane. This biomembrane barrier blocks the gate-to-channel ion passage and suppresses the oscillations ofthe biohybrid neuron, withffir 0 Hz. The addition ofa chemical agent that attacks the biomembranes tightjunctions —here the toxin hydrogen


     


    a


    b

    Biohybrid neuron

     

    out

    Toxin-induced disruption

    Vin

    C

    100

    T

    1

     

     

     


    100

    Time (s)


     


    Fig. 5 | A biohybrid neuron for in situ spike-based neuromorphic                        biointerfacing. a, Optical microphotograph (representative of 10–15 similar     experiments) ofthe cell culture ofthe Caco-2 cell line as a biomembrane model   for the implementation of biohybrid neurons. Immunostaining of occludin        (coloured in green) as a relevant tightjunction protein confirms the barrier         function ofthe biomembrane. Day 0 of cell culturing indicates the absence of     tightjunctions. Day 14 of cell culturing indicates the presence oftightjunctions  that form a barrier for ion passage through the biomembrane. b, Schematic of    the biohybrid neuron consisting of an OAN with the incorporated biomembrane.


    Iout versus time and the corresponding timefrequency analysis ofthe biohybrid  neuron is displayed. The presence of an ion-blocking epithelial biomembrane     between the gate and channel ofT1 suppresses the neuronal excitability and        spiking that is initially observed in the plain cell culture medium. Toxin-mediated opening ofthe biomembrane by attacking its tightjunctions results in the           recovery ofthe spiking activity over time. In the biohybrid neuron experiment,    the sensing device is T1 (that is, incorporation ofthe biomembrane) and keeping  the same electrolyte conditions in T2 (100 mM NaCl).


     

     

     


    peroxide lowers the barrier of ion passage and results in a gradual recovery of oscillations. This demonstration shows that the biomem- brane dynamics change the excitability ofthe OAN in real time, and this is directly reflected on the firing response of the biohybrid neuron. A more detailed schematic ofthe biohybrid neuron and a description of the biomembrane disruption mechanism are shown in Supplementary Fig. 17. Such biohybrid systems can be used as controllable in vitro models for basic research, such as to understand the underlying mecha- nisms of neuronal signalling, as well as a platform for studying the barrier integrity ofbiological tissues under various physiopathological conditions or the influence of external physicochemical cues (toxins, neuromodulators and so on). It should also be mentioned that interfac- ing the OAN with biological neurons requires that both domains have similar dimensions. Therefore, careful design ofthe OAN is necessary, as the device dimensions play a critical role in the spiking response ofthe OAN. The impact ofthe device dimension on the OAN spiking response with simulations is shown in Supplementary Fig. 18.

    Conclusions

    We have reported an OAN based on a nonlinear electrochemical ele- ment. Inspired by the properties of biological neurons functioning in wet surroundings, the OAN can mimic the biological sensitivity to ionic and biomolecular species in a surrounding aqueous environment. The artificial neuron exhibits nonlinear phenomena that depend on the composition ofbiophysically relevant host environments. We experi- mentally validated its operation with various electrolytes, including common aqueous electrolytes, buffered solutions and cell culture media. We also created biohybrid interfaces in which the OAN was modulated by the biological membrane ofepithelial cells in situ and in real time. A comparison with the state-of-the-art technology is provided in Supplementary Tables 1 and 2.

    Neuronal excitability, dynamics and spiking properties depend on the electrolytic potential and noise, as well as on the local concen- trations of specific ionic and biomolecular species. Therefore,just as in biological neurons —where sensing and actuation is merged and happens locally in the same surroundings —sensing (for example, neu- rotransmitters) and actuation/communication (via spiking, oscillating or other behaviours) are inherently embedded in the device operation, and this can enable tighter closed-loop control ofbiological substrates.


    The operation in close physical, functional and temporal proximity with biology can enable the real-time interaction between artificial and biological rhythmicity, for example, the development ofnew strategies for understanding, restoring and augmenting biorhythmic processes.  In contrast to conventional organic ring oscillators that consist of multiple transistors, only two transistors are required for the OAN.

    This compactness means that the OAN can potentially be merged into a single device a challenging venture for many-element implementa- tions. Negative-differential-resistance-based ionoelectronics can lead to much richer dynamics compared with conventional electronics. For practical applications, the integration density and variability of soft matter devices should be further developed and improved. In addition, although the OAN is externally powered in this work, biofuel-powered and self-sustainable oscillators could be developed that emulate cer- tain metabolic pathways ofbiological neurons68. Non-synaptic modes ofneuronal communication that are found in biological networks could be introduced with global electrolytes69,70. Furthermore, synaptic capabilities can be introduced at the function of the OAN circuit, by incorporating organic synaptic transistors28,31,71,72. Finally, in the case of dopamine detection, latentmemory time windowscan form the basis for on-chip learning phenomena, such as biomolecular reward prediction error coding73.

    Methods

    Fabrication of OENDs and OANs

    Standard microscope glass slides (75 mm × 25 mm) were cleaned in a sonicated bath, first in soap solution (Micro-90 (Sigma-Aldrich)) and then in a 1:1 (vol/vol) solvent mixture ofacetone and isopropanol. Gold electrodes for source and drain electrodes were photolithographi- cally patterned (with positive Microposit S1813 photoresist (DOW)) on the cleaned glass slides. A chromium layer was used to improve the adhesion of gold. Each glass slide contains a series of circuit blocks consisting of T1 and T2 OECTs. The channel dimensions of T1 and T2 are W1 × L1 = 50 μm × 20 μm and W2 × L2 = 50 μm × 10 μm, respectively. The OECTs are separately gated with Ag/AgCl electrodes via aqueous electrolytes. Two layers of parylene C (SCS Coatings) were deposited. Soap (Micro-90 soap solution, 1% vol/vol in deionized water) was used for separation between the parylene C layers to enable the peel-off of the upper parylene C layer. The lower parylene C layer insulates the


     


    gold electrodes. A promoter (Silane A-174 (γ -methacryloxypropyl trimethoxysilane), Sigma-Aldrich) was added to the lower parylene C layer to enhance adhesion. In the second photolithography step using the positive photoresist AZ 9260 MicroChemicals (Cipec Spécialités), the channel dimensions of T1 and T2 are defined. Reactive ion etching (O2/CF4 plasma, 160 W for 16 min with O2 flow rate of 50 s.c.c.m. and CHF3 flow rate of 5 s.c.c.m.) was used to define the channels of T1 and T2 throughout the photoresist mask. The channel of T1 is made with the organic mixed ionicelectronic conductor polymer PEDOT:PSS (Clevios PH 1000) mixed with 5.0 wt% ethylene glycol, 0.1 wt% dodecyl benzene sulfonic acid and 1.0 wt% (3-glycidyloxypropyl)trimethoxysi- lane. The film was spin coated in two steps at 1,500 rpm and 650 rpm for 1 min and annealed at 120 °C for 1 min in between. The devices were subsequently baked at 140 °C for 1 hour. For the implementation ofT2, the semiconducting polymer p(g2T-TT) was synthesized according to another work36. Here p(g2T-TT) was dissolved in chloroform (3 mg ml–1) inside a N2-filled glovebox and spin coated in ambient conditions at 1,000 rpm for 1 min resulting in a thickness of40 nm. The devices were baked at 60 °C for 1 min. The sacrificial upper parylene C layer was peeled offto confine the polymer to the inside ofthe channel regions. Excess soap was rinsed off with deionized water. A schematic of the OAN is shown in Supplementary Fig. 19.

    Electrical characterization of OENDs and OANs

    The current versus voltage characteristics ofthe individual OECTs and OENDs were obtained using a Keithley 2400 semiconductor parameter analyser. Ag/AgCl were used as the gate electrodes with 100 mM NaCl electrolyte solution, unless otherwise stated. The nonlinear charac- teristics of the OENDs were obtained by enforcing current when the corresponding voltage was measured or by enforcing voltage when the corresponding current was measured. The OEND was coupled with RC and Vin elements (external components) to complete the OAN, which exhibits neuronal dynamics when connected to voltage source Vin (Fig. 1b). The voltage oscillation was directly recorded at the Vout ter- minal using an Agilent infiniiVision digital oscilloscope. The output current Iout ofthe OEND is measured by measuring the voltage V0 across resistor RM with a differential amplifier (based on an INA122 integrated circuit) and a digital oscilloscope (Supplementary Fig. 20). To charac- terize the OAN response in the presence ofdifferent electrobiochemical signals, a Tektronix AFG1022 arbitrary function generator was used for the input voltage. The arbitrary function generator and semiconductor parameter analyser were used together to generate an arbitrary noise signal for the noise-induced neuronal characterization, phase-locking measurements and characterization of the various neuromorphic behaviours. The spectrogram ofthe short-time Fourier transform was performed using OriginPro 2016 with a Hanning-type window.

    Data availability

    The data that support the findings of this study are available from the corresponding authors on reasonable request. Source data are provided with this paper.

    References

    1.     Zhang, W. et al. Neuro-inspired computing chips. Nat. Electron. 3, 371–382 (2020).

    2.    Vassanelli, S. & Mahmud, M. Trends and challenges in                 neuroengineering: towardintelligentneuroprostheses through brain-‘brain inspired systemscommunication. Front. Neurosci.   10, 438 (2016).

    3.    van de Burgt, Y. & Gkoupidenis, P. Organic materials and devices for brain-inspired computing: from artificial implementation to   biophysical realism. MRS Bull. 45, 631–640 (2020).

    4.    Ham, D., Park, H., Hwang, S. & Kim, K. Neuromorphic electronics   based on copying and pasting the brain. Nat. Electron. 4, 635–644

    (2021).


    5.     Mikhaylov, A. et al. Neurohybrid memristive CMOS-integrated      systems for biosensors and neuroprosthetics. Front. Neurosci. 14, 358 (2020).

    6.    Shchanikov, S. et al. Designing a bidirectional, adaptive neural    interface incorporating machine learning capabilities and           memristor-enhanced hardware. Chaos Solit. Fractals 142, 110504

    (2021).

    7.     Serb, A. et al. Memristive synapses connect brain and silicon spiking neurons. Sci. Rep. 10, 2590 (2020).

    8.    Marković, D., Mizrahi, A., Querlioz, D. & Grollier, J. Physics for neuromorphic computing. Nat. Rev. Phys. 2, 499–510 (2020).

    9.    Indiveri, G. et al. Neuromorphic silicon neuron circuits. Front. Neurosci. 5, 73 (2011).

    10.   Tee, B. C.-K. et al. A skin-inspired organic digital mechanoreceptor. Science 350, 313–316 (2015).

    11.    Mirshojaeian Hosseini, M. J. et al. Organic electronics     Axon-Hillock neuromorphic circuit: towards biologically

    compatible, and physically flexible, integrate-and-fire spiking neural networks. J. Phys. D: Appl. Phys. 54, 104004 (2020).

    12.   Abu-Hassan, K. et al. Optimal solid state neurons. Nat. Commun. 10, 5309 (2019).

    13.   Harikesh, P. C. et al. Organic electrochemical neurons and             synapses with ion mediated spiking. Nat. Commun. 13, 901 (2022).

    14.   Pickett, M. D., Medeiros-Ribeiro, G. & Williams, R. S. A scalable       neuristor built with Mott memristors. Nat. Mater. 12, 114–117 (2013).

    15.   Torrejon, J. et al. Neuromorphic computing with nanoscale spintronic oscillators. Nature 547, 428–431 (2017).

    16.   Kumar, S., Williams, R. S. & Wang, Z. Third-order nanocircuit    elements for neuromorphic engineering. Nature 585, 518–523

    (2020).

    17.   Valov, I. & Tsuruoka, T. Efects of moisture and redox reactions in    VCM and ECM resistive switching memories. J. Phys. D: Appl. Phys. 51, 413001 (2018).

    18.   Gupta, I. et al. Real-time encoding and compression of neuronal  spikes by metal-oxide memristors. Nat. Commun. 7, 12805 (2016).

    19.   Liu, Z. et al. Multichannel parallel processing of neural signals in memristor arrays. Sci. Adv. 6, eabc4797 (2020).

    20.  Zhang, X. et al. An artificial spiking aferent nerve based on Mott

    memristors for neurorobotics. Nat. Commun. 11, 51 (2020).

    21.   Park, S.-O., Jeong, H., Park, J., Bae, J. & Choi, S. Experimental       demonstration of highly reliable dynamic memristor for artificial neuron and neuromorphic computing. Nat. Commun. 13, 2888

    (2022).

    22.   Mulaosmanovic, H., Chicca, E., Bertele, M., Mikolajick, T. &    Slesazeck, S. Mimicking biological neurons with a nanoscale ferroelectric transistor. Nanoscale 10, 21755–21763 (2018).

    23.  Inagaki, T. et al. Collective and synchronous dynamics of photonic spiking neurons. Nat. Commun. 12, 2325 (2021).

    24.   Hao, S. et al. A monolayer leaky integrate-and-fire