Microelectrode arrays cultured with in vitro neural networks for motion control tasks: encoding and decoding progress and advances

Abstract

Microelectrode arrays (MEAs) cultured with in vitro neural networks are gaining prominence in bio-integrated system research, owing to their inherent plasticity and emergent learning behaviors. Here, recent advances in motion control tasks utilizing MEAs-based bio-integrated systems are presented, with a focus on encoding-decoding techniques. The bio-integrated system comprises MEAs integrated with neural networks, a bidirectional communication system, and an actuator. Classical decoding algorithms, such as firing-rate mapping and central firing-rate methods, along with cutting-edge artificial intelligence (AI) approaches, have been examined. These AI methods enhance the accuracy and adaptability of real-time, closed-loop motion control. A comparative analysis indicates that simpler, lower-complexity algorithms suit basic rapid-decision tasks, whereas deeper models exhibit greater potential in more complex temporal signal processing and dynamically changing environments. The review also systematically analyzes the prospects and challenges of bio-integrated systems for motion control. Future prospects suggest that MEAs cultured with in vitro neural networks may leverage their flexibility and low energy consumption to address diverse motion control scenarios, driving cross-disciplinary research at the intersection of neuroscience and artificial intelligence.

Introduction

The brain, being among the most vital organs in the human body, serves as the central hub for advanced activities such as cognition¹, language², and learning³. At its core, neurons form intricate neural networks through synaptic connections, enabling the processing of complex information⁴. In essence, the brain functions as a highly efficient “biological computer,” characterized by its compact size and immense computational capacity.

To explore and harness the functional properties of neural networks, the use of microelectrode arrays (MEAs) has significantly advanced as a tool for studying neuronal activity and interactions. MEAs facilitate concurrent stimulation and recording from extensive neuronal populations, providing valuable insights into neural function and connectivity. Researchers have utilized MEAs technology to investigate neuronal spontaneous activity, along with their responses to external stimuli, allowing for an enhanced understanding of neural behavior^{5, 6–7}. Studies have shown that neural networks cultivated on MEAs exhibit strong learning and task execution capabilities⁸.

A substantial body of research has demonstrated the intelligence and application potential of MEAs-based in vitro neural networks. Bakkum et al.⁹ utilized the adaptive capabilities of in vitro neural networks on MEAs to control a robotic arm for drawing and subsequently investigated their neurodynamic properties in car navigation tasks⁸. Llofriu et al.¹⁰ employed reinforcement learning via the Q-Learning method and the discharge characteristics of hippocampal multi-scale spatial cells to reduce the number of steps required for robot navigation. Yada et al.¹¹ utilized the reservoir computing capabilities exhibited by in vitro neuronal networks cultured on MEAs to train and execute navigation tasks under steady-state network conditions. Recent studies have further expanded the application scope of MEAs-based neural networks. Kagan et al.⁵ trained in vitro neural networks cultivated on MEAs to play the arcade game “Pong,” while Cai et al.⁷ leveraged the unsupervised learning capabilities of these networks for speech recognition and nonlinear equation prediction.

Although artificial intelligence (AI) hardware has increasingly evolved towards lower energy consumption in recent years¹², its extensive computational requirements during training still depend on large-scale integrated circuits, consuming substantial energy. Compared to artificial neural networks supported by silicon-based transistors and other electronic components, in vitro neural networks possess the advantages of low power consumption, strong learning capabilities, and high flexibility. Computing with brain tissue holds promise for achieving genuine intelligence rather than merely “artificial intelligence”⁷. Furthermore, MEAs execute intelligent motion control tasks by collecting signals from in vitro neural networks. In this process, neural signal encoding and decoding technologies form a crucial bridge enabling communication between researchers and the in vitro neural network. These technologies allow the understanding and refinement of neural network outputs, making the entire workflow functional. Without effective encoding and decoding methods, neural signals would degrade into meaningless “noise,” limiting their practical applications.

It is worth noting that decoding algorithms for in vivo brain signals are relatively mature^{13, 14–15} and have been extensively reviewed^{16, 17, 18–19}, providing systematic guidance in fields such as disease diagnosis. Nevertheless, comprehensive and systematic reviews focusing on encoding and decoding algorithms for intelligent task control using in vitro neural networks cultured on MEAs remain scarce. Existing reviews in the field of this type of bio-integrated system focus on specific aspects. For instance, Bisio et al.²⁰ reviewed closed-loop control systems using in vitro neural networks for neuro-controlled prosthetics; Chen et al.²¹ summarized the intelligent properties and theoretical foundations of in vitro neural networks; Soriano et al.²² emphasized the cultivation and biological characteristics of in vitro neural networks; Kagan et al.⁶ discussed the ethical challenges associated with synthetic biological intelligence. In contrast to these existing reviews, this paper focuses on the critical technologies underpinning this type of MEAs cultured with neural networks for motion control tasks, including every component of the system, and encoding and decoding technologies. This work aims to fill a gap in the review literature of this field, providing guidance and direction for future research.

Fundamental components for an intelligent motion control system

The intelligent motion control system based on MEAs cultured with in vitro neural networks consists of three main components: MEAs cultured with neural network, a bidirectional communication system, and an actuator (Fig. 1).

Fig. 1 [Images not available. See PDF.]

The three components of the intelligent motion control system: MEAs cultured with a neural network, a bidirectional communication system, and an actuator (created with http://BioRender.com)

In this system, the MEAs cultured with in vitro neural networks serve as the core, responsible for training and providing the initial signals for decision-making. MEAs facilitate the interaction of signals with the bidirectional communication system, recording spontaneous neural activity from biological networks and delivering electrical stimuli back to these networks for training purposes. The spontaneous firing of the biological neural network serves as the source for decoding decisions, and its plasticity provides an environment conducive to learning and adapting to the environment. In the subsequent two chapters, biological mechanisms underlying the learning and decision-making capabilities of in vitro neural networks cultured on MEAs are critically examined, with specific attention to synaptic plasticity dynamics and emergent computational properties, as well as providing a comprehensive review of in vitro MEAs. The bidirectional communication system is responsible for processing the neural electrical signals received by the MEAs. Only after the neural electrical signals are processed and decoded by the bidirectional communication system can they be applied to the subsequent decision-making of the actuator. Meanwhile, the bidirectional communication system dispatches encoded electrical stimulation to the MEAs cultured with neural networks, in accordance with the sensor readings. A section will be dedicated to discussing existing bidirectional communication system and their development trends. The actuator, also known as the task end for motion control, is the component that executes the motion, including sensors. This can be either a simulated movement or a real-world entity like a vehicle or robot, which will be introduced in the following text.

In vitro neural network culture and learning mechanism

Compared to the neural networks of living organisms, the in vitro neural networks are artificially cultivated rather than innate. Fortunately, the intelligent behaviors of in vitro neural networks, such as learning and memory, largely align with established findings in neuroscience. Moreover, the encoding and decoding algorithms of in vitro neural networks are significantly inspired by neuroscience and the processing algorithms for in vivo brain signals. Therefore, this chapter aims to introduce the biological theoretical foundations of encoding and decoding in in vitro neural networks, as well as the procedures for obtaining in vitro neural networks, providing a theoretical basis for the subsequent discussion of encoding and decoding algorithms.

Neural plasticity as the basis of intelligence

Artificial neural networks achieve learning by adjusting network weights, whereas biological intelligence relies on neural plasticity, a process governed by the well-known Hebb’s Learning Rule: “Fire together, wire together”²³. This principle reveals that the weights of connections between neurons can be modulated through changes in electrical activity intensity. The alterations in the characteristics of synaptic connections among neurons primarily include long-term potentiation (LTP) and long-term depression (LTD) (Fig. 2a). LTP refers to the strengthening of the relationship between two neurons, or even the formation of new connections, following high-frequency (Tonic stimulation above 20 Hz²⁴) stimulation or synchronous firing of the two neurons²⁵. This occurs through increased activation of N-methyl-D-aspartate synaptic receptor proteins²⁶ or the incorporation of more α-Amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid receptors into the membrane²⁷. In contrast, LTD involves the reduction of receptors or the weakening and elimination of unused connections in response to low-frequency stimulation and negatively correlated spike timing between two neurons (1 Hz, Negative correlation refers to the injection of current into a synapse so that its action potential occurs before the presynaptic²⁸).

Fig. 2 [Images not available. See PDF.]

Theories of encoding and decoding in in vitro neural networks. a LTP induced by strong stimulation or near-coincident discharges and LTD caused by weak stimulation or asynchronous firing peaks. b In vivo encoding and decoding process. c Two encoding theories. d Tissues derived from the cerebral region in neonatal rodents. e Human pluripotent stem cell culturing process. f A typical modular culture medium schematic, divided into four compartments (the figure is created with http://BioRender.com)

Neurons cultured in vitro spontaneously form neural networks with abundant nodes, providing a structural basis for neural plasticity. Research has shown that neural networks cultured in vitro exhibit capabilities for both short-term memory²⁹ and learning³⁰. The plasticity of neural networks, whether in vivo or in vitro, depends on specific forms of stimulation. In living organisms, stimuli are derived from the sensory input of the body’s organs, with signals transmitted by sensory organs reshaping neural networks to better adapt to tasks^31,32. Since in vitro neural networks lack sensory organs, they rely on artificial stimulation. Furthermore, the core of controlling intelligent tasks using in vitro neural networks lies in regulating in vitro neural networks based on feedback from target task completion. By altering the structural connections between neurons, these networks achieve the functions of learning and adaptation.

Bio-inspired encoding and decoding for motion control

In biological systems, external information is perceived through sensory organs and converted into electrical or chemical signals, which are transmitted to the brain for processing. This process is referred to as the encoding of external information. Decoding, on the other hand, refers to the transmission of processed signals from the brain to effectors, where they are translated into specific behaviors (Fig. 2b). This encoding and decoding mechanism in biological systems reflects the high degree of adaptability developed through long-term evolution for processing complex external information. It enables efficient interaction and information exchange with the external environment. Although the specific mechanisms of biological encoding and decoding are not yet fully understood, multiple hypotheses concerning these processes can offer meaningful insights and inspiration for investigating encoding and decoding mechanisms within AI systems.

The progress in understanding the encoding and decoding mechanisms in biological systems has been largely attributed to the discovery of spike signals. Alan Hodgkin and Andrew Huxley were the first to detect spike signals in the giant squid³³. This discovery revealed the fundamental characteristics of neuronal electrical activity. Subsequently, researchers identified a strong correlation between the electrophysiological responses of certain cells and features of external stimuli, a form of encoding referred to as feature encoding^{34, 35–36}. In the study of encoding mechanisms in single neurons, there are two main perspectives on how spikes achieve encoding: rate coding and temporal coding (Fig. 2c). In 1926, E. D. Adrian^{37, 38–39} proposed that neuronal spike encoding depends solely on the firing rate within a specified temporal interval, irrespective of the specific temporal sequence. This perspective resembles an analog-to-digital conversion unit, where continuous physiological electrical signals are transformed into discrete firing rate signals. In contrast, from the viewpoint of temporal coding, the specific timing sequence of neuronal spikes constitutes another critical dimension for information encoding^{40, 41, 42–43}, which is characterized by significant additional correlation between stimulus parameters and higher-order temporal patterns of spikes (such as inter-spike intervals or phase relationships with sub-window fluctuations) within the encoding window, beyond what can be explained by spike count alone⁴⁴. However, the information gain from time coding is far outweighed by its waste of resources⁴⁵, so most motion control systems are coded using issue rate coding.

In decoding mechanisms, from the perspective of mapping relationships, neuronal decoding can be categorized into individual decoding and population decoding. Single-cell decoding pertains to interpreting the activity of an individual neuron as indicative of particular behavioral events^46,47, meaning that certain behaviors are controlled exclusively by the activity of a single neuron. The advantage of individual decoding lies in its high precision, but it demonstrates limited stability when subjected to noise interference. In contrast, population decoding integrates the firing information from multiple neurons to achieve behavioral decoding on a larger scale^48,49. This approach offers higher robustness and can reduce the impact of noise. Additionally, Bayesian brain decoding has emerged as one of the more recent popular hypotheses⁵⁰, This theory posits that the brain integrates prior knowledge with real-time information to interpret external information through probabilistic inference. However, in the context of synthetic intelligence research, the prior knowledge within in vitro neural cultures is relatively limited, and the computational demands for posterior probability are significant, resulting in fewer applications of this decoding method. Nevertheless, its potential merits further exploration.

Advanced in vitro neural network cultivation

In vitro neural networks are primarily structured in two dimensions (2D). The cells available for in vitro culture mainly include primary cells and induced pluripotent stem cells (iPSCs)²². Primary cells are obtained directly from cerebral tissue (Fig. 2d), while iPSC technology enables somatic cells to acquire the ability for division and differentiation by introducing retroviruses⁵¹ (Fig. 2e).

Primary neural cells are typically harvested from rat embryos or neonatal rats^11,24,52. When transplanted into MEAs culture media enriched with nutrients, these cells can develop neural nodes and networks similar to those of the brain^53,54. This method has the advantages of being a well-established and standardized procedure⁵⁵, and has been shown to maintain cell viability even after transportation⁵⁶, enabling the formation of mature commercial cultivation systems. Most in vitro neural cell cultures used in synthetic intelligence projects originate from this approach^{24,52,57, 58–59}. However, its limitation lies in the need for direct sampling from brain tissue, which is invasive and destructive to experimental animals. In comparison, pluripotent stem cells can be derived from somatic cell samples, causing less harm to the organism. Human-induced pluripotent stem cells (hiPSCs), in particular, form in vitro neural networks that better highlight human-specific intelligence characteristics and comply with ethical standards. Research demonstrated that human neural cultures could consistently complete gaming tasks correctly for longer durations, suggesting superior performance of human neural cells in intelligent tasks⁵. Additionally, neural cells derived from human somatic cells via reverse transcription exhibit high homogeneity, containing few or no contaminant cells, such as immune cells^60,61. Nevertheless, variability in cells during the reverse transcription process remains a major challenge for hiPSCs. This variability may lead to the proliferation of tumor cells, thereby affecting the connectivity characteristics of the neural network and its performance during learning processes^51,62.

In cell cultivation, the use of modular culture methods (Fig. 2f) to constrain the unnecessary connectivity of neural networks, thereby controlling network topology and achieving up to 100-fold enhancement in directional information flow, is considered an effective approach to improving their performance in intelligent tasks⁶³. JUN et al. explored the use of Polydimethylsiloxane (PDMS) molds to fabricate culture media⁶⁴. This material is biocompatible and easy to process, laying a foundation for subsequent modular culture techniques. Subsequently, Tessadori et al. demonstrated experimentally that a PDMS-based modular culture medium with two modules could prevent “crosstalk” between neurons. This setup allowed precise closed-loop feedback from stimuli, improving the obstacle avoidance performance of the actuated system²⁴. Li et al. further increased the number of modules to four⁵², and used statistical methods (Kolmogorov–Smirnov test, wherein P values less than 0.05 were deemed statistically significant) and neurodynamic to illustrate the advantages of modular culture media. The obstacle avoidance accuracy of the actuated system reached as high as 80%.

In addition to the aforementioned 2D cell culture methods, emerging three-dimensional (3D) brain organoid technology has been introduced into the field of biological intelligence research in recent years⁷. 3D brain organoids possess more complex and human brain-like network structures, but their application is accompanied by technical challenges related to data acquisition and decoding. With the advancement of 3D MEAs technology and data processing techThe patch-clamp technniques, 3D brain organoids are expected to be further utilized in biological intelligence projects.

MEAs as essential tools for acquiring motion control signals and enabling closed-loop feedback

MEAs serve as the key part to obtain original input data for decoding, and accurately input pre-designed encoded information into the in vitro neural network via appropriate mediators. This chapter focuses on the critical device enabling information interaction in in vitro neural networks, specifically the interfaces between the biological and artificial components.

Neuronal electrophysiological signals represent the primary signal modality for regulating in vitro neural networks. The main techniques for detecting electrical signals include patch-clamp technology, microfilament electrode technology, and MEAs technology. The patch-clamp technique employs voltage or current clamp modes to record ionic currents and membrane potentials across cellular ion channel⁶⁵. While it provides high signal-to-noise ratios, this technique is technically complex, requires specialized training, and is limited to single-cell measurements^66,67. Additionally, its invasive nature can interfere with neuronal firing⁶⁵. Microfilament electrode technology, on the other hand, analyzes extracellular electrical signals⁶⁸. It is less invasive but has limited signal sensitivity. MEAs technology, with its superior biocompatibility, high detection precision, and multi-channel recording capability, has gradually become the mainstream approach for recording and stimulating neuronal electrophysiological signals⁶⁹.

For the context of in vitro neural networks, MEAs are implanted onto glass-based surfaces, where adhesive cells are cultured on the MEAs⁷⁰. This setup enables the recording of electrical signals from neural cultures or the electrical stimulation of these networks. The application of MEAs to in vitro neural networks dates back to the 1970s⁷¹. Over more than half a century of development, mature commercial systems have emerged⁶⁹. MEAs implanted with cultured neuronal systems have been extensively employed in motion control tasks (Table 1).

Table 1. Comparison of MEAs with in vitro neural networks for motion control tasks

MEAs material	Fabrication technology	Number of electrodes	Spatial resolutionb (μm)	Form of the culture	Decoding algorithmc	Execution end	Post-training performance improvementd	Refs
TiN	MEMS	60	200	Primary cultures(rat)	CFR Algorithm	Robot arm movement	95%	9
TiN	MEMS	60	200	Primary cultures(rat)	CFR Algorithm	Virtual motion	288%	8
TiN	MEMS	60	200	Primary cultures(rat)	FR Mapping	Obstacle avoidance	260%	24
TiN	MEMS	60	200	Primary cultures(rat)	AI Model	Path planning	--	11
TiN	MEMS	60	200	Primary cultures(rat)	AI Model	Robot arm movement	90%	59
ITO-pa	MEMS	64	150	Human fetal neural stem cells	AI Model	Direction identification	240%	142
Platinum	CMOS	26000	17.5	hiPSCs	FR Mapping	Virtual motion	--	5

^aITO-p Indium tin oxide–platinum

^bSpatial resolution: spatial resolution refers to the minimum distance that electrodes can distinguish, i.e., the distance between two adjacent electrodes

^cCFR central firing rate, FR firing rate

^dPost-training performance improvement: although the tasks performed vary, it is not valid to compare the results of the tasks, but we can calculate the accuracy improvement value of each task before and after learning (e.g., the probability that the car can avoid obstacles before and after the training is improved), and this performance index can illustrate the learning results of the cultures to a certain extent. However, whilst metrics for all tasks have been normalized to dimensionless growth rates, different experiments vary due to differences in the form of MEAs, cultivation methods of cultures, decoding algorithms, and execution of the tasks, and learning outcomes need to be considered in combination with all of the above factors

With advancements in technology, MEAs in motion control tasks are evolving towards 3D structuring, advanced material integration, and high-density miniaturization (Table 1).

3D MEAs for multi-point electrical signal acquisition from a brain organoid in space

Traditional MEAs primarily adopt a 2D planar structure (Fig. 3(i)), designed to detect the electrical signals of cells on the bottom surface of culture dishes. These MEAs have been extensively studied and are technically mature⁷², making them the mainstream method for detecting electrical signals in in vitro networks. Some 2D MEAs feature protruding contact points, creating a pseudo 3D electrode structure capable of penetrating dead cell layers in acute brain slices to detect neuronal signals⁷³(Fig. 3a(ii)). However, such MEAs are still unable to measure electrophysiological signals in the vertical dimension and thus cannot be considered fully 3D. In recent years, with advancements in 3D organoid cultivation techniques, fully 3D MEAs have emerged. These vertically structured MEAs can better detect the signal activities of 3D neural networks^{74, 75, 76–77}(Fig. 3a(iii), (iv)). Compared to 2D MEAs, 3D MEAs capture signals more consistent with the natural structure of neural networks, offering broader possibilities for studying the functions of complex neural networks.

Fig. 3 [Images not available. See PDF.]

MEAs development trends. a (i) traditional 2-dementional MEAs⁷². Copyright 1998 by Elsevier. Reproduced with permission. (ii) mimetic 3D MEAs, which can pierce cell membranes, protruding electrodes can pierce cells to measure high signal-to-noise signals, but do not have a 3D structure⁷³. Copyright 1998 by Elsevier. Reproduced with permission. (iii) 3D MEAs used polyimide materials, less rigid, less aggressive to cells. Reproduced with permission from ref. ⁷⁴ (2020, CC BY 3.0) (iv) product of 3D shell-shaped MEAs, can be wrapped around the surface of brain-like organs to measure a full range of signals. Reproduced with permission from ref. ⁷⁷ (2022, CC BY 4.0). b (i) typical TiN MEAs. Reproduced with permission from ref. ²⁴ (2012, CC BY 3.0) (ii) 128-channel MEAs coated with TPU and PU utilized a liquid metal-polymer conductor, with high light transmission and flexibility. Reproduced with permission from ref. ⁷⁶ (2024, CC BY 4.0) (iii) high transparency MEAs enhanced with PEDOT:PSS, calcium imaging, and fluorescent staining supported. Reproduced with permission from ref. ⁸⁴ (2021, CC BY 4.0). c (i) traditional 60 channels MEAs⁹. Copyright 2007 by Frontiers Research Foundation. Reproduced with permission. (ii) high-density MEAs with PDMS structure fabricated using complementary metal-oxide-semiconductor technology. Reproduced with permission from ref. ⁸⁸ (2022, CC BY 4.0) (iii) 64 × 64 electrode pixels CMOS MEAs platform. Reproduced with permission from ref. ⁸⁹ (2016, CC BY 4.0). d (i) early use of PDMS-isolated culture chambers to customize neural network growth patterns by microfluidic systems⁹⁸. Copyright 2006 by Elsevier. Reproduced with permission. (ii) Graphene perforated electrodes with integrated microfluidic environment enable precise delivery of chemical drugs to localized areas¹⁰². Copyright 2024 by Royal Society of Chemistry. Reproduced with permission. (iii) Perforated MEAs with microfluidic structure¹⁰³. Copyright 2025 by Elsevier. Reproduced with permission

While 3D MEAs are used for brain organoid signal acquisition, the rigid structure may cause damage to biological tissues, and flexible materials with little rigidity are valued. Encapsulation with polyimide allows the electrodes to be bent, allowing a better fit to the measurement object and significantly reducing mechanical damage to tissue and metal leakage to cultured cells^74,78. Much of the manufacturing of flexible 3D electrodes relies on 3D printing technology, and 3D printing manufacturing methods (e.g., direct laser writing) can also allow for faster manufacturing, allowing for automated production of flexible MEAs in batch quantities⁷⁹. Moreover, highly customizable 3D printed MEAs can be adapted to different morphologies of neural cultures for a more flexible target audience⁸⁰. In a novel study, 2D MEAs templates were produced and then folded into vertical structures by drawing on Kirigami techniques, and the folded vertical structures were stable and reliable for more than half a year⁸¹. However, it should be noted that the stresses faced by 3D flexible MEAs when implanted into cells may shift the measurement sites, and the relationship between rigidity and stability needs to be balanced.

Advanced materials enhancing electrode performance and biocompatibility

TiN is the most common electrode material (Fig. 3b(i)). Electrodes of different materials can be applied to different scenarios; TiN materials have high mechanical stability and are suitable for implantation of primary neural cultures in mice, which has the advantage of being able to output larger voltages. Electrodes such as ITO or liquid metal-polymer conductor, which can adapt to higher sample rates and acquire more information, are preferred in the more costly hiPSCs culture (Table 1, Fig. 3b(ii)). The introduction of advanced materials has significantly enhanced the performance of MEAs. For example, incorporating nanomaterials into metallic electrodes can markedly improve their electrochemical properties, biocompatibility, and signal detection sensitivity⁸². Conductive polymers, such as poly(3,4-ethylenedioxythiophene):poly(styrene sulfonate)(PEDOT:PSS), have attracted increased interest recently owing to their superior biocompatibility and enhanced signal recording performance⁸³. By applying electroplating techniques, PEDOT:PSS can be utilized on MEAs electrodes to further optimize signal recording quality⁸⁴ (Fig. 3b(iii)). On the other hand, because electrophysiological signal recording alone may provide incomplete information, there is an increasing need to integrate optical techniques such as calcium imaging. This has driven the development of transparent electrode manufacturing processes. Transparent electrodes can be fabricated using cost-effective materials like indium tin oxide⁸⁵, graphene electrodes with excellent biocompatibility⁸⁶, or low impedance conductive polymer electrodes⁸⁷. These advancements enable MEAs to combine with optical methods like calcium imaging, meeting the requirements for multimodal signal recording.

High-density electrodes for increased decoding data

The number of channels in the MEAs determines the amount of space in which information can be exchanged with culture, and for tasks similar to binary classification, a high level of accuracy can be achieved with only one pair of electrodes (one stimulating and one detecting electrode)⁵⁹. However, the number of neural cultures required is much higher than 2 because the optimal “stimulus-response pair” has to be found before performing the task.60 electrodes is a standard configuration for in vitro neural networks performing action control tasks (Table 1), and fewer electrodes can be used for actual experiments after the appropriate stimulus and response sites have been found. When performing more complex tasks, a higher number of channels can lead to more information. Traditional MEAs generally possess electrode counts varying between tens to several hundreds of channels (Fig. 3c(i)). However, following the advent of complementary metal-oxide-semiconductor (CMOS) technology, electrode counts have swiftly expanded, now reaching up to tens of thousands of channels^{88, 89, 90–91} (Fig. 3c(ii), (iii)). Additionally, 3D printing technology has been utilized in the fabrication of high-density electrodes, achieving channel densities of up to 2600 per square centimeter and enabling the design of 3D electrode structures⁹². The advantages of high-density MEAs include the capacity to generate a greater volume of informative data, providing enhanced input for neural signal decoding algorithms and improving decoding precision, and providing more stimulus sites allows for a higher spatial resolution of the stimulus (Table 1). However, this also introduces increased complexity in data processing and imposes higher computational burdens^93,94.

Perforated MEAs with microfluidic structures

The microfluidic structure is a key technology for achieving modular culture, and the microfluidic tubing enables neural network synapses to be directionally guided to connect in the different chambers of the culture dish^95,96. Modular culture is important because the different brain regions of a biological brain have different topologies and connectivity properties, and in motor control tasks of neural bio-integrated systems, cultivating groups with different topologies (perceptual and motor control regions) helps to mimic the structure of the brain and achieve closed-loop control of motor control⁹⁷. In addition, compared to the traditional closed chamber petri dish, the microfluidic system can provide nutrients more conveniently and prolong the survival time of cells⁹⁸ (Fig. 3d(i)). Adding microfluidic channels to the petri dish complicates the structural design of the MEAs and culture chambers; nevertheless, studies have been conducted to combine high-density MEAs with microfluidic systems for modular regulation and high-resolution recording⁹⁹.

On the other hand, perforating the electrodes with pneumatic or hydraulic pressure allows the nerve cells to fit more tightly onto the electrodes, allowing the measurement sites to be anchored and the measured signals to have a higher signal-to-noise ratio¹⁰⁰. Perforated MEAs are mostly used in experiments on the retina in vitro due to the large curvature of the retina, which makes it more difficult to fixate¹⁰¹. At the same time, the shape of the advanced 3D brain organoid is also spherical, and the surface has a very similar structure of large curvature with the retina, so the perforated microfluidic MEAs are very suitable for the signal acquisition of the 3D brain organoid. The perforated MEAs can be well integrated with the microfluidic system, where the culture is pressed onto the electrodes by the hydraulic control of the microfluidic system. The rational structural design ensures that the culture fits tightly on the electrode while the pressure difference is not too large to cause tissue tearing¹⁰² (Fig. 3d(ii)). Combining perforated MEAs with flexible materials, the flexible MEAs are wrapped around the surface of the spherical retinal culture, while the negative pressure inside the holes makes them forcefully anchored, which provides a better idea for the fixation of the measurement site¹⁰³ (Fig. 3d(iii)).

In summary, MEAs are continually advancing towards 3D structuring, the use of advanced materials, high-density integration, and microfluidic integration. In cutting-edge research, MEAs technology can simultaneously incorporate multiple advantages and adapt to the development of brain organoids, becoming a critical tool for exploring the functions of complex neural networks⁷⁷. Moreover, the information expressed by in vitro neural networks is not limited to electrophysiological signals. Combining MEAs with multimodal detection techniques, such as optical and chemical methods, can provide a more comprehensive analysis of network information²¹.

Signal processing system for raw signal processing and encoding-decoding

The signal processing bidirectional communication system serves as the core interface devices that establish electrical connections with MEAs culture media, enabling signal generation and reception. First, because the neural signals collected by MEAs are weak and challenging to analyze directly, a bidirectional communication system provides high-gain amplifiers^24,58 and employs filtering, analog-to-digital conversion^58,59, and other modules to transform the signals into a more readable format. Second, these systems supply the stimulus sources required for encoding, allowing for the stable input of stimuli with adjustable frequencies and waveforms. Finally, a bidirectional communication system provides a data processing platform, enabling the implementation of decoding and encoding feedback according to temporal and logical sequences.

The “bidirectional” nature of a signal processing system means that it needs to process the received signal as well as have the ability to emit electrical stimuli. The application of brain signal decoding technology can be traced back to 1999¹⁰⁴. At that time, many communication systems were designed as open-loop systems, focusing solely on collecting and decoding neural signals and outputting commands to external devices. These output signals had no direct impact on the input¹⁰⁵. However, pure decoding could not facilitate learning or intelligent behavior in biological networks. The introduction of closed-loop feedback stimulation of closed-loop systems created opportunities for teaching and learning in neural networks, enabling the full utilization of their structural and functional plasticity.

Electrical stimulation, a critical component of closed-loop systems, was first demonstrated in 1792 by Luigi Galvani, who discovered its ability to enhance neural activity during experiments on nervous tissue¹⁰⁶. Later, Jackson et al. found that electrical stimulation could remodel the structure of neural networks by exploiting their plasticity, thereby making them more controllable¹⁰⁷. Incorporating electrical stimulation—essentially the “encoding” process of brain signals—into a bidirectional communication system to create closed-loop setups allows for the teaching of neural networks. This, in turn, enables the use of neural network learning outcomes for executing intelligent tasks. However, as electrophysiological instruments for in vitro neural networks must simultaneously provide stimulation and detection in a closed-loop control system, the high development costs and implementation challenges have become significant barriers to further advancement in this field²⁰.

The earliest closed-loop signal processing system for in vitro neural networks was proposed by Marom et al., aiming to explore the learning properties of neural networks³⁰. Subsequently, researchers have developed various open-source designs for signal processing closed-loop systems. Wagenaar et al. designed MEABench¹⁰⁸, a system equipped with an interface for data acquisition cards and a modular communication program, facilitating easier debugging and future upgrades. Rolston et al. developed the NeuroRighter system (Fig. 4a), which features 1000-fold signal amplification, filtering capabilities, a sampling rate of up to 25 kHz, and dual-mode stimulation for current and voltage. Remarkably, its cost is only one-fourth of commercial systems¹⁰⁹. Subsequently, Newman et al. improved the system by adding features such as clamping population discharges and incorporating an API interface¹¹⁰. More recently, the Neuroplatform system has advanced the field further by integrating sophisticated microfluidic systems (Fig. 4c). It supports a sampling rate of up to 30 kHz, an electrical stimulation precision of 0.15 μA, and includes optical stimulation and detection functionalities, significantly enhancing multimodal experimental capabilities¹¹¹. On the other hand, the microfluidic system also allows for uniform neuronal growth and facilitates patterning of the culture protocol⁹⁵.

Fig. 4 [Images not available. See PDF.]

Advanced bidirectional communication system. a Components of the closed-loop electrophysiology system: NeuroRighter. Copyright 2010 by Frontiers Research Foundation¹⁶⁷. Reproduced with permission. b General architecture of the open source Neuroplatform system, which has a microfluidic system to ensure the supply of nutrients. Reproduced with permission from ref. ¹¹¹ (2021, CC BY 4.0). c High-speed FPGA electrophysiological system, applied to CMOS-MEAs¹⁶⁸. Copyright 2010 by Frontiers Research Foundation. Reproduced with permission. d The Neurochip3 system achieving miniaturization and integration. Reproduced with permission from ref. ¹¹³ (2024, CC BY 4.0)

Systems for brain signal applications must emit and receive multi-channel signals, resulting in substantial data processing demands during decoding. Traditional systems typically rely on computers for data processing. However, characterized by their capability for parallel processing, enabling real-time signal processing and resource optimization, providing a pathway for developing standalone in vitro neural networks signal processing systems integrated with decoding algorithms. Park et al. implemented an electrophysiological closed-loop system using a Field Programmable Gate Array (FPGA), achieving efficient real-time processing and optimized hardware resource utilization¹¹². Shupe et al. developed a bidirectional brain machine interface closed-loop system (Fig. 4b, d), successfully inducing neural plasticity and enhancing neural activity¹¹³. These studies provide valuable references for the future development of an integrated, high-performance hardware-software closed-loop bidirectional communication system.

Advancing motion control using MEAs cultured with in vitro neural networks

The capacity of in vitro neural networks to perform motion control tasks demonstrates their intelligence-related capabilities. Through decoding, neuronal activity can be mapped to actuators for task control, while closed-loop feedback encoded stimuli provide information for the learning of intelligent behaviors. Tessadori et al. proposed that the tasks of in vitro neural network actuators can be implemented in either physical or virtual environments²⁴. In virtual environments, software programs simulate actual control tasks, and decoding results are input into the program to execute tasks (Fig. 5a). In physical actuators, sensors are required, which pose challenges due to errors but more closely resemble the human brain’s task execution in real-world environments (Fig. 5b). Today, tasks such as obstacle avoidance, navigation, and prediction are the stage of AI, while in vitro neural networks cultured in vitro acting directly on actuators represent a form of true intelligence, distinct from traditional AI.

Fig. 5 [Images not available. See PDF.]

Motion control applications of MEAs cultured with neural networks. a Applications in virtual environments (i) path planning tasks conducted within computer software. Reproduced with permission from ref. ²⁴ (2012, CC BY 3.0) (ii) training in vitro neural networks to play simple arcade games such as “Pong,” collecting motor area potential information to map to game play decisions, and giving sensory areas electrical stimulation based on game play effects. Reproduced with permission from ref. ⁵ (2022, CC BY 4.0). b Application in physical environments (i) controlling a robotic arm for artistic creation, converting photographs into electrical stimuli via a CCD camera to feed back to the neural network so that it can draw as it learns⁹. Copyright 2007 by Frontiers Research Foundation. Reproduced with permission. (ii) navigating a robot in a physical environment¹¹. Copyright 2015 by Elsevier. Reproduced with permission. (iii) controlling a robotic car for obstacle avoidance. Reproduced with permission from ref. ¹⁰ (2021, CC BY 4.0) (iv) controlling a robotic arm to mimic human hand movements. where the mechanical values are used as a feedback reference for the stimulus, training allows the system to discriminate between grasping patterns. Reproduced with permission from ref. ⁵⁹ (2024, CC BY 4.0)

One of the most widely applied areas is obstacle avoidance and navigation tasks. As early as 2001, DeMarse et al.¹¹⁴ controlled robots to avoid obstacles in a virtual world. Subsequent studies have developed obstacle avoidance and navigation tasks into the physical world (Martinoia et al.⁵⁷, Novellino et al.¹¹⁵, Warwick et al.⁵⁸, Tessadori et al.²⁴, Li et al.⁵², Llofriu et al.¹⁰, Yada et al.¹¹, Bakkum et al.⁸). Although these studies differ in sensor selection, encoding methods, decoding algorithms, and actuator design, their core principle is consistent: distance information is obtained through sensors to control motors and adjust the movement direction of vehicles or robots. Based on the excellent performance in navigation and obstacle avoidance tasks, in vitro neural networks may offer new solutions for problems such as autonomous driving and path planning. Furthermore, in vitro neural networks have also been used to control the motions of mechanical arms. Bakkum et al. proposed an innovative work⁹ where an in vitro neural network acted as an artist, controlling a mechanical arm to draw. Ades et al. designed a bio-integrated system combining an in vitro neural network and a mechanical arm, using tactile sensors to perceive the external environment⁵⁹. While these actuators focus on “motion-based” tasks, in vitro neural networks can also execute other types of intelligent tasks. Kagan et al. connected an in vitro neural network to the classic game “Pong”⁵, letting the network play the game like a human. In recent years, MEAs have also been used in areas other than motion control, such as MEAs equipped with neural networks: Cai et al. utilized in vitro neural networks for speech recognition and chaotic equation prediction tasks, during which the in vitro neural network demonstrated strong unsupervised learning capabilities⁷. Compared to the actuator tasks in AI applications, the application areas of MEAs with neural networks are far less extensive and remain largely unexplored, offering significant room for future investigation.

Encoding and decoding algorithms for motion control

With the continuous development of neural networks-based motion control research, understanding the encoding and decoding mechanisms of neural electrical signals has become increasingly important. The core of this mechanism lies in developing suitable algorithms that can both generate stimulation information recognizable by neural networks, known as encoding, and accurately interpret the electrophysiological signals expressed by the neural networks, known as decoding. (Fig. 6 illustrates the closed-loop process of in vitro neural networks performing intelligent tasks.)

Fig. 6 [Images not available. See PDF.]

Closed-loop experimental workflow for task execution in a neural networks-based motion control system. After MEAs and a bidirectional communication system collect raw neural signals, the signals undergo preprocessing, where complex waveforms are converted into datasets related to firing rates through analog-to-digital conversion, and neuronal firing is classified. The dataset is then input into a decoding model, which maps the neural data to decision-making for the actuator. After executing a decision-driven action, the discrepancy between the performed action and the intended target is computed and subsequently delivered back to the in vitro neural network as electrical stimulation signals. This feedback enables the network to continuously learn and evolve, aligning more closely with the requirements of task execution. (Fig. 6 is created with http://BioRender.com)

The specific forms of neural information encoding and decoding algorithms differ between in vivo neural networks and in vitro neural networks used in intelligent tasks. Decoding algorithms for in vivo neural networks are primarily applied in disease classification and can be processed offline^{13, 14–15}. Unlike in vivo neural networks, in vitro neural networks are typically tasked with goal-oriented activities, necessitating higher levels of real-time performance and dynamic responsiveness. The signal processing of in vitro neural networks often compresses the data collected by MEAs into the key feature of spike rate, allowing the system to fulfill real-time task requirements.

Decoding algorithm for mapping neural network signals from MEAs into motion decisions

Decoding algorithms determine how information is extracted from neural network activity and mapped to decision-making for intelligent behavior. Based on different decoding algorithms, these methods can be categorized into three types: mapping firing rates directly to decisions (hereafter referred to as FR Mapping), determining decisions based on the central firing rate (CFR) of population encoding, and decoding using AI models (hereafter referred to as AI Model). The existing encoding and decoding algorithms are presented in Table 1. It is worth noting that the evaluation of decoding performance varies for different tasks. We believe that the metric of growth rate of accuracy before and after training allows for side-by-side comparisons. We have found that control tasks of varying complexity and decoding algorithms dominate the degree of performance improvement, the task of controlling a robotic arm for the tasks of drawing and grasping mode recognition is more complex, the cultures are more difficult to learn, and the accuracy improvement is lower (Table 1); and in the case of identical decoding algorithms, differences in performance also arise from different sources of cells (human-derived neuronal cells perform better than mice)⁵, sites stimulated by MEAs, and electrode density. The combination of high-density electrodes, advanced organoid cultures helps to increase the effectiveness of the data sources and the amount of data for the decoding algorithms. We will begin this section with an overview of preprocessing algorithms, followed by several of the most classical as well as decoding algorithms and forms of encoding that may be put to use.

Preprocessing of the data collected by the MEAs into valid signals

To extract valid spike signals, bandpass filtering is typically employed to eliminate low-frequency and high-frequency noise. Neural spike responses are concentrated around the peaks of depolarization, making spike extraction crucial for analyzing neural activity (Fig. 7b). Neural signals collected by MEAs undergo amplification and analog-to-digital conversion, generating raw spike data that contains substantial background noise. Bandpass filtering is commonly used to remove noise¹¹⁶, and the selection of appropriate detection thresholds is crucial for effectively identifying spike signals from neural recordings. Thresholds that are too high risk losing valid signals, while thresholds that are too low may introduce noise, necessitating careful selection¹¹⁷. Typically, a lower threshold is used for initial detection, and false positives are removed using signal shape characteristics, such as repolarization detection, where signals must exceed six times the noise value within a defined time after reaching the threshold^118,119.

Fig. 7 [Images not available. See PDF.]

Neural signal processing and mapping algorithms for robot control. a Typical process of the FR Mapping algorithm. Reproduced with permission from ref. ¹¹⁵ (2022, CC BY 4.0). b Schematic diagram of spike detection algorithm based on threshold detection¹²³. Copyright 2021 by Elsevier. Reproduced with permission. c Typical process of the AI Mode. Reproduced with permission from ref. ¹¹ (2021, CC BY 4.0). d An optimized wavelet parameter-based spike sorting method flowchart. Reproduced with permission from ref. ¹¹⁸ (2015, CC BY 4.0). e Typical process of the CFR Algorithm⁸. Copyright 2021 by IOP Publishing Ltd. Reproduced with permission

Despite filtering and initial processing, single electrodes often struggle to directly capture the firing signals of individual neurons (Fig. 7d). This is due to the size of electrodes (10–30 μm diameter) and electrode spacing (30–200 μm)^57,59, which, although smaller than typical neuron dimensions, are affected by background noise and complex network structures¹²⁰. The process of classifying and reconstructing single-neuron firing data from multi-neuronal signals is known as spike sorting. Traditional spike sorting involves two main steps: feature extraction and clustering¹²¹. Feature extraction reduces data dimensionality while preserving key information, enabling accurate classification with minimal data loss. Principal component analysis (PCA) is a widely applied technique that selects the principal components linked to the top k covariance matrix eigenvalues to construct a transformation matrix W_k, Spike times T are projected onto W_k, as shown in Eq. 1 (ref. ¹²²):

1

$$T_{Afte{r}_{PCA}}=T·W_k$$

In addition to PCA, other methods like wavelet transforms¹²³, Independent Component Analysis¹²⁴, or hybrid approaches combining multiple techniques can be used¹²⁵.

Once features are extracted, spikes are clustered. Common algorithms include Gaussian mixture clustering¹²⁶, which classifies spikes into single neuron firing signals based on the principle of minimum residuals, k-means clustering^127,128, t-distribution-based clustering¹²⁹, and Bayesian methods¹³⁰. However, with increasing MEAs channel counts and decreasing electrode spacing, the complexity of neuron-to-electrode relationships challenges traditional methods¹³¹ To address these challenges, template matching and deep learning methods have emerged. Template matching algorithms, akin to wavelet transforms^{132, 133, 134–135}, utilize a predefined waveform template which often derived from clustering. This template slides through MEAs data using iterative greedy algorithms¹²¹, enabling differentiation of neuron firing activities. Deep learning approaches¹³⁶, with their end-to-end classification capabilities, efficiently handle large datasets but require substantial training data, presenting challenges for practical application.

Spike sorting algorithms can be implemented using software such as Python, C++, or Matlab on CPUs or GPUs^135,137,138. Recently, hardware-based spike sorting has gained attention for improved real-time performance and integration. Chen et al. deployed PCA on an FPGA, pioneering FPGA-based spike sorting¹³⁹. Xu et al. developed CMOS hardware for spike sorting, with a core area of only 0.0175 mm² (ref. ¹⁴⁰). Valencia et al. used an FPGA for template matching with a latency below 1 μs. These hardware solutions enhance the real-time efficiency and compactness of the bidirectional communication system, promoting integration. Implementing spike sorting algorithms on hardware further optimizes system performance, making it more compact and efficient.

Efficient decoding via firing rate projection

FR Mapping is a simple and efficient decoding algorithm (Fig. 7a) widely used in neural networks-based bio-integrated research. A method proposed by Kagan et al. is a classic example of this type of algorithm⁵. This method involves culturing neurons in different regions of a dish and optimizing the layout of stimulation and detection areas using the EXP3¹⁴¹ algorithm from the multi-armed bandit problem. The dish is ultimately divided into two regions: stimulation and detection. The in vitro neural network receives two different encoded electrical stimuli (75 mV, 100 Hz, 100 ms/ 150 mV, 5 Hz, 4 s). Through closed-loop feedback, the network optimizes its connection structure to improve decision-making capabilities. By calculating the average firing rate in the active region ${\partial }_{\mathrm{AFR}}$ , the prediction error ℇ is derived from the product of the difference between the previous state and the target and the accuracy error. The next action (a) is then predicted using Formula 2 (ref. ⁵):

2

$$a={\partial }_{AFR}\epsilon$$

Decisions are made based on the “a” value of the two regions, and the next electrical stimulation is determined, completing the closed-loop control. By comparing the firing rates of the two regions, decisions are mapped to two outputs, a strategy referred to as “winner-takes-all”¹¹⁵. Additionally, linear mapping was used to map CFRs to decisions due to the presence of more than two possible outputs for the actuator^24,58. The greatest advantage of FR Mapping lies in its simplicity and low computational resource consumption, allowing for high-speed online training.

Central firing rate mapping decoding algorithm inspired by biological neural decoding theory

The population coding theory posits that neuronal cells convey specific sensory and motor information through the combined firing of individual neurons (Fig. 7e). The CFR method serves as a central measure of population activity, representing the overall firing activity of the network to some extent. Its expression is given as Eq. 3 (ref. ⁸):

3

$$CFR=\frac{{\sum }_{k=1}^N{f}_k\left(X_k-R_x,Y_k-R_y\right)}{{\sum }_{k=1}^N{f}_k}$$ Here, CFR represents the central firing rate, N denotes the total count of detection channels, F_k denotes firing rate captured by the kth electrode, X_k and Y_k represent spatial positions of the kth electrode, and R represents the position of the reference point.

As an example, Bakkum et al. mapped the CFR to changes in the steering angle $\Delta \theta$ of a robotic vehicle, thereby controlling the actuator’s behavior. The relationship is expressed as Formula 4 (ref. ⁸):

4

$$\mathrm{\Delta }\mathrm{\theta }=M·\mathrm{CFR}$$ where M is the mapping matrix. After calculating the angular error between the decision and the target, two electrical stimuli with different amplitudes and frequencies are used to encode feedback, forcing the neural network structure to adapt in a direction conducive to task execution. Similarly, Martinoia et al. employed the CFR Algorithm, but with a different encoding method that used stimulation frequency as the encoding parameter⁵⁷. Population decoding methods align well with neuroscientific findings on brain cognitive mechanisms and may represent a plausible approach to electrical signal decoding in the brain. Additionally, this method is relatively straightforward to implement, requiring less computational complexity.

End-to-end black-box decoding model based on artificial intelligence algorithms

Pizzi et al.¹⁴² were among the first to use AI models to decode neural network electrical signals. They employed a Self-Organizing Map¹⁴³, characterized by its lack of convergence requirements, where neural competition influences the output. Ades et al. proposed an intriguing approach⁵⁹ that applied transfer learning. They utilized a convolutional neural network designed for time-frequency image processing, transforming the neural signal decoding problem into a computer vision detection problem.

In a neural networks-based motion control system, one of the most extensively utilized decoding models based on AI is the reservoir computing approach (Fig. 7c)^{144, 145–146}. The reservoir is a specialized form of a recurrent neural network renowned for its exceptional performance in handling temporal signals, making it particularly suited for complex spatiotemporal electrical signals. Its internal connections are highly random and disordered, resembling the physical connectivity of the brain to some extent. This characteristic offers a unique advantage for simulating the dynamic behavior of neural systems. Cai et al.⁷ employed an in vitro neural network as a physical reservoir. By applying logistic regression to the collected electrical signals, they successfully performed speech classification tasks. Yada et al. designed a single-layer readout neuron in the reservoir’s readout layer, extracting signals to control a robotic vehicle¹¹.

AI models provide an effective interface for understanding the information encoded in neural network electrical signals. Using AI models as decoding tools integrates artificial “intelligence” with biological “intelligence,” enabling both networks to learn and evolve through training, thereby enhancing their mutual adaptability. As evident in Table 1, the application of AI models for neural signal decoding has become a prominent trend.

Additional algorithms with potential applications in decoding

Spiking neural network (SNN) constitutes the third generation of neural network models. Compared to the second-generation artificial neural networks, SNNs have a lower level of abstraction from in vitro neural networks and mechanisms that more closely mimic the behavior of neural networks on MEAs, resulting in higher compatibility with biological networks. This characteristic makes SNN a promising approach for decoding, potentially leading to superior outcomes.

Taherkhani’s review provides an overview of various SNN architectures¹⁴⁷. Among these, the commonly adopted Leaky Integrate and Fire model computes the forward propagation of a single neuron as follows¹⁴⁷:

5

$${\tau }_m\frac{d{v}_m\left(t\right)}{\mathit{dt}}=-\left(\left(v_{m\left(t\right)}-E_r\right)\right)+R_m\sum _i\sum _j{w}_i\delta (t-t_j)$$ Where, v_m(t) represents the membrane potential, τ_m denotes the membrane time constant, E_r is the constant membrane resting potential, R_m corresponds to the membrane resistance, $w_i$ represents the input weight of ith neuron, $\delta$ denotes the Dirac function, t_j denotes the jth spike time. As the complexity of SNN structures increases, their training methods can be categorized into three types: One category involves unsupervised learning^{148, 149–150}, another adopts supervised learning similar to artificial neural networks, such as using the backpropagation algorithm^{151, 152–153}, and a third employs ANN weight mapping methods¹⁵⁴.

SNNs have been applied to decoding tasks for in vivo neural signals, such as electroencephalogram decoding^155,156 and brain signal enhancement¹⁵⁷, demonstrating their superior performance in capturing and processing bioelectrical signals. However, the application of SNN in in vitro neural networks decoding algorithm remains limited, presenting vast potential for future development.

Neural dynamics for analyzing neuronal connectivity at the network level

Neural dynamics research can be divided into two levels: the single cell level¹⁵⁸ and the network level¹⁵⁹. In the context of in vitro neural networks, analyzing the spatiotemporal firing activity patterns of neurons at the network level facilitates statistical assessment of firing characteristics and overall performance. This analysis provides valuable insights for developing encoding and decoding methods in in vitro neural networks. Li et al. used neural dynamics to compare the effects of isolated and non-isolated structures on firing rates. The research employed the Kolmogorov–Smirnov method¹⁶⁰ to demonstrate significant differences in the performance of two types of culture media⁵². Additionally, cross-rank correlation is an effective method for measuring connectivity in neural networks. Assuming the firing rate at one point in the network space is $F_X$ , and at another point is $F_Y$ , heir correlation coefficient ${\rho }_{X,Y}$ is defined as¹⁶¹:

6

$${\rho }_{X,Y}=\frac{Cov\left. (F_X,F_Y\right)}{\sqrt{Var\left. (F_X\right)·Var\left. (F_Y\right)}}$$ Here, Cov denotes the covariance matrix, and Var represents variance. A higher correlation coefficient indicates a tighter connection between the two points. Mohseni Ahooyi et al.¹⁶¹ used this method to measure the functional connectivity of network cultures and to infer the direction of signal propagation within the network. Regression analysis of firing rates is another approach for measuring connectivity. Kagan et al.⁶ utilized this method to perform regression analysis between the firing rates of the stimulation and detection regions. The results showed low p values, indicating a significant correlation between the two regions.

Reservoir computing, which is widely applied in decoding algorithms, can also be explained through neural dynamics. Cai et al.⁷ investigated the fading dynamics of neurons following electrical stimulation, observing that the firing rate decayed in a sigmoid-like manner and could be modulated by different stimulation methods. This discovery provides a crucial reference for selecting appropriate encoding strategies.

Encoding methods for sensor information feedback in a closed-loop system

Encoding functions as a closed-loop feedback mechanism in the control of motor systems. After the effector completes its movement, sensors collect real-world data and compare them with the decoded decisions to identify any discrepancies. These discrepancies, in turn, prompt electrical stimulation of the neural network, inducing structural changes—reliant on neural plasticity—that enable the network to progressively adapt to its environment.

Currently, there are no definitive quantitative metrics that specify how particular parameters of electrical stimulation alter the structure of neural networks. However, Warwick et al. have proposed an experimental approach⁵⁸: under identical experimental conditions, they partitioned a dish containing neural networks cultured on MEAs into distinct regions and applied stimulation to each region individually. If, when stimulating a specific region, another particular region displayed a response rate exceeding 60% while the response rates of the remaining regions fell below 20%, this result indicates that the stimulated region and the region exhibiting the highest response rate form an “input–output” pair. In such a configuration, stimulation of the input region can substantially modify the connection weights of the output region. Consequently, when decision errors are significant, sustained stimulation to the input region can be administered to enhance the network’s adaptability to its environment.

Electrical stimulation for encoding typically employs biphasic rectangular pulse stimulation, which can be readily generated by a stimulation chip. Bidirectional voltage stimulation is effective in balancing charges. The stimulation amplitude is generally on the order of 100–500 mV, ensuring safety with effective stimulation; however, in certain experiments aimed at testing neural networks or under specific conditions, amplitudes exceeding 1 V may be applied. High-voltage stimulation can break through the interfacial impedance between the electrode and the cell contact, allowing rapid depolarization of the nerve cells, but also potentially causing damage to the cells. The stimulation amplitude can be fixed, such that a reward stimulus is delivered when errors are within an acceptable range. The reward stimulus has a relatively fixed pattern, such as stimulating the same region in the same form each time, thereby reducing the entropy of the network and reinforcing existing connections. The punishment stimulus, employed when errors are substantial, is random with stimulation sites or parameters randomly selected each time, prompting structural changes in the neural network that guide it toward better environmental adaptation⁵. In some studies^58,63, the encoding parameters vary continuously with sensor output; that is, the sensor values are linearly mapped to the encoding parameters. While this approach may provide the neural network with additional information, it undoubtedly increases the complexity of stimulation.

Trends and challenges

With the continuous advancement of in vitro cell culture techniques, the structure of in vitro neural networks is becoming increasingly complex, progressing towards 3D configurations^162,163. This structural evolution provides decoding algorithms with larger and more biologically relevant raw neural data while simultaneously posing significant challenges for data interaction. The core issues can be divided into two main aspects: For hardware, the primary challenge lies in improving the performance of hardware systems, such as MEAs and bidirectional communication systems, to accommodate the 3D advancements in in vitro cell culture techniques. These systems must be capable of providing high-density, high-throughput neural signal data to support decoding algorithms. The development of 3D MEAs allows for full-dimensional acquisition of electrophysiological signals from 3D brain organoids. However, most current MEAs cultured with neural networks for motion control tasks rarely utilize 3D MEAs. Even in the study by Cai et al.⁷, which employed 3D brain organoids, MEAs were still limited to planar signal acquisition. Miniaturized and high-density MEAs can significantly increase the volume of data obtained from brain organoids. While this increased data volume offers the potential for higher decoding accuracy, it also imposes significant computational demands on decoding algorithms. As discussed earlier, FPGA acceleration and terminal CPU/GPU computations are currently sufficient for processing 2D neural networks data in real time. However, when electrode counts rise to tens of thousands and the structure transitions to three dimensions, real-time decoding faces performance bottlenecks. For software, with the rapid growth in data volume, more efficient preprocessing of signals for decoding—such as feature extraction and data compression—becomes critical. Condensing valid information to reduce data dimensionality not only alleviates computational pressure but also improves algorithmic efficiency. Additionally, the information representation mechanisms of large-scale neural networks remain challenging to fully understand, complicating the selection of optimal decoding algorithms. It is worth noting that recent advancements in AI, including deep learning and reinforcement learning, have seen increasing application in the decoding of in vitro neural networks. These AI techniques may offer a more seamless integration with biological intelligence, potentially bridging gaps in decoding performance and enabling novel applications.

On the other hand, ethical and moral issues in this field need to be taken into account. In contrast to in vivo experiments, in vitro-cultured neural networks have fewer ethical challenges. The core ethical challenges are primarily manifested in two aspects: First, brain organoids may mature to a developmental stage approaching consciousness during cultivation, leading to controversies over whether they can generate self-awareness and how to define their corresponding moral status¹⁶⁴. Both the academic community and the public generally hold that if self-awareness is present, they should be regarded as moral individuals and granted special protection, which poses severe challenges to the boundaries of experiments¹⁶⁵. In this regard, there is a need for researchers to intensify the discussion on the determination of the production of self-awareness in organoids, to introduce a clear indicator, and to stop experiments as soon as they approach that indicator. Second, the ambiguous rights of human cell donors and the discrepancy in public acceptance constitute another ethical dilemma. Studies on hiPSCs have shown that public acceptance of cells used for therapeutic purposes is significantly higher than that for commercial applications such as biocomputing¹⁶⁶. However, issues such as whether donors have the right to use biohybrid systems derived from their cells remain unclear, necessitating the establishment of specialized institutions for regulation and supervision. To address ethical challenges, establish a comprehensive framework integrating consciousness assessment and donor rights tracking, while promoting transparent research logs and interdisciplinary ethics reviews. Additionally, implement international regulatory mechanisms like collaborative ethical reviews and dynamic guideline updates to prevent technology abuse and ensure alignment with societal values.

In the future, such bio-integrated systems may become a low-power alternative to silicon-based chips, not only by using plasticity to become the computational core of intelligent algorithms, but also by storing information through the structure of neural networks, and furthermore, by becoming an ultra-low-power “bio-computer” that integrates storage and computation. Continued development, under strict ethical and moral constraints, may also create truly intelligent bio-mechanical hybrid robots, so that the robots are more like human beings, and better serve human beings, which is a step forward in the exploration of the field of “intelligence.” Furthermore, such systems hold the potential to facilitate in-depth investigations into the dynamics and firing patterns of neural networks, thereby contributing to a more profound understanding of the principles underlying brain electrical activity.

Conclusion

This article centers on the encoding and decoding techniques used in MEAs bearing cultured neural networks for motor control, providing a systematic overview of the latest research advances in closed-loop learning and decision-making with in vitro neural networks. Owing to the biological property of neural plasticity, these networks can spontaneously adjust synaptic connections under electrical stimulation, exhibiting a certain degree of learning and adaptive capacity to accomplish motor control tasks. On the hardware front, the rapid development of 3D MEAs and high-density electrodes enables the acquisition of richer, more precise neural signals and supports closed-loop control on a larger scale. However, this also heightens the challenges of data processing and real-time decoding, necessitating a careful balance between hardware integration and algorithmic efficiency.

In terms of decoding algorithms, traditional approaches, such as FR Mapping and CFR, are favored for fundamental real-time motor control scenarios because of their relatively low implementation and computational complexity. With the emergence of AI, models such as reservoir computing, deep neural networks, and SNN-inspired architectures have significantly advanced the ability to process complex temporal data and enhanced adaptive control mechanisms. Nevertheless, the trade-off among accuracy, interpretability, and real-time hardware performance remains a key focal point in current research.

From an application perspective, MEAs cultured with neural networks have demonstrated preliminary feasibility in areas like robotic navigation, path planning, and robotic arm control. Compared with conventional AI methods, in vitro neural networks hold potential advantages in energy efficiency and adaptability; still, more rigorous experimentation and engineering-based practice are needed to refine the associated technologies. Looking ahead, as 3D brain organoids and novel MEAs designs continue to mature, combined with efficient data preprocessing and hardware acceleration, in vitro neural networks may attain higher levels of plasticity and real-time performance for larger-scale and finer-grained motor control. This development is poised to establish them as a pivotal branch of bio-artificial hybrid intelligence research, offering new opportunities and challenges for the next generation of interdisciplinary studies in AI and neuroscience.

Acknowledgements

This work was sponsored by the National Key R&D Program of China (2022YFC2402500, 2022YFB3205602), the National Natural Science Foundation of China (No. 62121003, T2293730, T2293731, 62333020, 62171434 and 62471291), the Major Program of Scientific and Technical Innovation 2030 (2021ZD02016030), the Joint Foundation Program of the Chinese Academy of Sciences (No. 8091A170201), the Scientific Instrument Developing Project of the Chinese Academy of Sciences (No. PTYQ2024BJ0009) and the National Natural Science Foundation of Beijing (F252069).

Author contributions

S.H., X.C., Y.L., J.L., and S.L. investigated the literature, conceived and wrote the manuscript. L.J, P.W., S.S., L.S, C.L., and K.Z. prepared the figures. X.C., H.S., J.L., M.W., Y.L., and J.L. checked the manuscript and revised it. All authors have given approval to the final version of the manuscript.

Conflict of interest

The authors declare no competing interests.

References

1. Tognoli, E; Kelso, JAS. The Metastable Brain. Neuron; 2014; 81, pp. 35-48.

2. Li, P; Legault, J; Litcofsky, KA. Neuroplasticity as a function of second language learning: anatomical changes in the human brain. Cortex; 2014; 58, pp. 301-324.

3. Chanaday, NL; Cousin, MA; Milosevic, I; Watanabe, S; Morgan, JR. The synaptic vesicle cycle revisited: new insights into the modes and mechanisms. J. Neurosci.; 2019; 39, pp. 8209-8216.

4. Farhy-Tselnicker, I; Allen, NJ. Astrocytes, neurons, synapses: a tripartite view on cortical circuit development. Neural Dev.; 2018; 13, 7.

5. Kagan, BJ et al. In vitro neurons learn and exhibit sentience when embodied in a simulated game-world. Neuron; 2022; 110, pp. 3952-3969.

6. Kagan, BJ et al. The technology, opportunities, and challenges of Synthetic Biological Intelligence. Biotechnol. Adv.; 2023; 68, 108233.

7. Cai, H et al. Brain organoid reservoir computing for artificial intelligence. Nat. Electron.; 2023; 6, pp. 1032-1039.

8. Bakkum, DJ; Chao, ZC; Potter, SM. Spatio-temporal electrical stimuli shape behavior of an embodied cortical network in a goal-directed learning task. J. Neural Eng.; 2008; 5, pp. 310-323.

9. Bakkum, DJ; Gamblen, PM; Ben-Ary, G; Chao, ZC; Potter, SM. MEART: the semi-living artist. Front. Neurorobotics; 2007; 1, 12.

10. Llofriu, M et al. Goal-oriented robot navigation learning using a multi-scale space representation. Neural Netw.; 2015; 72, pp. 62-74.

11. Yada, Y; Yasuda, S; Takahashi, H. Physical reservoir computing with FORCE learning in a living neuronal culture. Appl Phys. Lett.; 2021; 119, 173701.

12. Theis, TN; Wong, H-SP. The End of Moore’s Law: a new beginning for information technology. Comput. Sci. Eng.; 2017; 19, pp. 41-50.

13. Hosny, M; Zhu, M; Gao, W; Fu, Y. Deep convolutional neural network for the automated detection of Subthalamic nucleus using MER signals. J. Neurosci. Methods; 2021; 356, 109145.

14. Lu, H; Liu, S; Wei, H; Chen, C; Geng, X. Deep multi-kernel auto-encoder network for clustering brain functional connectivity data. Neural Netw.; 2021; 135, pp. 148-157.

15. Cui, X et al. Deep feature fusion based childhood epilepsy syndrome classification from electroencephalogram. Neural Netw.; 2022; 150, pp. 313-325.

16. Li, Z. Decoding methods for neural prostheses: Where have we reached?. Front. Syst. Neurosci.; 2014; 8, 129.

17. Zhao, Z-P et al. Modulating brain activity with invasive brain–computer interface: a narrative review. Brain Sci.; 2023; 13, 134.

18. Zhao, Y; Chen, Y; Cheng, K; Huang, W. Artificial intelligence based multimodal language decoding from brain activity: a review. Brain Res. Bull.; 2023; 201, 110713.

19. Mathis, MW; Perez Rotondo, A; Chang, EF; Tolias, AS; Mathis, A. Decoding the brain: from neural representations to mechanistic models. Cell; 2024; 187, pp. 5814-5832.

20. Bisio, M et al. Closed-Loop systems and in vitro neuronal cultures: overview and applications. Adv. Neurobiol.; 2019; 22, pp. 351-387.

21. Chen, Z et al. An overview of in vitro biological neural networks for robot intelligence. Cyborg Bionic Syst.; 2023; 4, 0001.

22. Soriano, J. Neuronal cultures: exploring biophysics, complex systems, and medicine in a dish. Biophysica; 2023; 3, pp. 181-202.

23. Cooper, SJ; Donald, O. Hebb’s synapse and learning rule: a history and commentary. Neurosci. Biobehav Rev.; 2005; 28, pp. 851-874.

24. Tessadori, J; Bisio, M; Martinoia, S; Chiappalone, M. Modular neuronal assemblies embodied in a closed-loop environment: toward future integration of brains and machines. Front. Neural Circuits; 2012; 6, 99.

25. Bliss, TV; Lomo, T. Long-lasting potentiation of synaptic transmission in the dentate area of the anaesthetized rabbit following stimulation of the perforant path. J. Physiol.; 1973; 232, pp. 331-356.

26. Tsien, JZ; Huerta, PT; Tonegawa, S. The essential role of hippocampal CA1 NMDA receptor-dependent synaptic plasticity in spatial memory. Cell; 1996; 87, pp. 1327-1338.

27. Choquet, D; Opazo, P. The role of AMPAR lateral diffusion in memory. Semin. Cell Dev. Biol.; 2022; 125, pp. 76-83.

28. Bi, G; Poo, M. Synaptic modifications in cultured hippocampal neurons: dependence on spike timing, synaptic strength, and postsynaptic cell type. J. Neurosci.; 1998; 18, pp. 10464-10472.

29. Dranias, MR; Ju, H; Rajaram, E; VanDongen, AMJ. Short-term memory in networks of dissociated cortical neurons. J. Neurosci.; 2013; 33, pp. 1940-1953.

30. Shahaf, G; Marom, S. Learning in networks of cortical neurons. J. Neurosci.; 2001; 21, pp. 8782-8788.

31. El-Boustani, S et al. Locally coordinated synaptic plasticity of visual cortex neurons in vivo. Science; 2018; 360, pp. 1349-1354.

32. Fischler-Ruiz, W et al. Olfactory landmarks and path integration converge to form a cognitive spatial map. Neuron; 2021; 109, pp. 4036-4049.e5.

33. Hodgkin, AL; Huxley, AF. Action potentials recorded from inside a nerve fibre. Nature; 1939; 144, pp. 710-711.

34. Kuffler, SW. Discharge patterns and functional organization of mammalian retina. J. Neurophysiol.; 1953; 16, pp. 37-68.

35. Hubel, DH; Wiesel, TN. Receptive fields and functional architecture of monkey striate cortex. J. Physiol.; 1968; 195, pp. 215-243.

36. Desimone, R; Albright, TD; Gross, CG; Bruce, C. Stimulus-selective properties of inferior temporal neurons in the macaque. J. Neurosci.; 1984; 4, pp. 2051-2062.

37. Adrian, ED. The impulses produced by sensory nerve endings: Part I. J. Physiol.; 1926; 61, pp. 49-72.

38. Adrian, ED; Zotterman, Y. The impulses produced by sensory nerve-endings: Part II. The response of a Single End-Organ. J. Physiol.; 1926; 61, pp. 151-171.

39. Adrian, ED; Zotterman, Y. The impulses produced by sensory nerve endings: Part 3. Impulses set up by Touch and Pressure. J. Physiol.; 1926; 61, pp. 465-483.

40. Ahissar, E et al. Dependence of cortical plasticity on correlated activity of single neurons and on behavioral context. Science; 1992; 257, pp. 1412-1415.

41. Kara, P; Reinagel, P; Reid, RC. Low response variability in simultaneously recorded retinal, thalamic, and cortical neurons. Neuron; 2000; 27, pp. 635-646.

42. Tiesinga, P; Fellous, J-M; Sejnowski, TJ. Regulation of spike timing in visual cortical circuits. Nat. Rev. Neurosci.; 2008; 9, pp. 97-107.

43. Ince, RAA et al. Information-theoretic methods for studying population codes. Neural Netw.; 2010; 23, pp. 713-727.

44. Theunissen, F; Miller, J. Temporal encoding in nervous systems—a rigorous definition. J. Comput. Neurosci.; 1995; 2, pp. 149-162.

45. Rolls, ET; Treves, A. The neuronal encoding of information in the brain. Prog. Neurobiol.; 2011; 95, pp. 448-490.

46. Hartline, HK; Milne, LJ; Wagman, IH. Fluctuation of response of single visual sense cells. Fed. Proc.; 1947; 6, 124.

47. Parker, AJ; Newsome, WT. SENSE AND THE SINGLE NEURON: Probing the Physiology of Perception. Annu. Rev. Neurosci.; 1998; 21, pp. 227-277.

48. Georgopoulos, AP; Schwartz, AB; Kettner, RE. Neuronal population coding of movement direction. Science; 1986; 233, pp. 1416-1419.

49. Theunissen, FE; Miller, JP. Representation of sensory information in the cricket cercal sensory system. II. Information theoretic calculation of system accuracy and optimal tuning-curve widths of four primary interneurons. J. Neurophysiol.; 1991; 66, pp. 1690-1703.

50. Ma, WJ; Beck, JM; Latham, PE; Pouget, A. Bayesian inference with probabilistic population codes. Nat. Neurosci.; 2006; 9, pp. 1432-1438.

51. Takahashi, K et al. Induction of pluripotent stem cells from adult human fibroblasts by defined factors. Cell; 2007; 131, pp. 861-872.

52. Li, Y; Sun, R; Zhang, B; Wang, Y; Li, H. Application of hierarchical dissociated neural network in closed-loop hybrid system integrating biological and mechanical intelligence. PLoS ONE; 2015; 10, e0127452.

53. Wagenaar, DA; Pine, J; Potter, SM. An extremely rich repertoire of bursting patterns during the development of cortical cultures. BMC Neurosci.; 2006; 7, pp. 1-18.

54. Scarnati, MS; Boreland, AJ; Joel, M; Hart, RP; Pang, ZP. Differential sensitivity of human neurons carrying μ opioid receptor (MOR) N40D variants in response to ethanol. Alcohol; 2020; 87, pp. 97-109.

55. Salazar, I et al. Preparation of primary cultures of embryonic rat hippocampal and cerebrocortical neurons. Bio Protoc.; 2017; 7, e2551.

56. Sammoura, FM; Popova, D; Morris, A; Hart, RP; Richardson, JR. Methods for shipping live primary cortical and hippocampal neuron cultures from postnatal mice. Curr. Res. Neurobiol.; 2023; 4, 100069.

57. Martinoia, S et al. Towards an embodied in vitro electrophysiology: the NeuroBIT project. Neurocomputing; 2004; 58, pp. 1065-1072.

58. Warwick, K et al. Controlling a mobile robot with a biological brain. Def. Sci. J.; 2010; 60, pp. 5-14.

59. Ades, C et al. Biohybrid robotic hand to investigate tactile encoding and sensorimotor integration. Biomimetics; 2024; 9, 78.

60. Van Den Berg, A; Mummery, CL; Passier, R; Van Der Meer, AD. Personalised organs-on-chips: functional testing for precision medicine. Lab Chip; 2019; 19, pp. 198-205.

61. Donegan, JJ; Lodge, DJ. Stem cells for improving the treatment of neurodevelopmental disorders. Stem Cells Dev.; 2020; 29, pp. 1118-1130.

62. Shi, Y; Inoue, H; Wu, JC; Yamanaka, S. Induced pluripotent stem cell technology: a decade of progress. Nat. Rev. Drug Discov.; 2017; 16, pp. 115-130.

63. Forró, C et al. Modular microstructure design to build neuronal networks of defined functional connectivity. Biosens. Bioelectron.; 2018; 122, pp. 75-87.

64. Jun, SB et al. Low-density neuronal networks cultured using patterned poly-l-lysine on microelectrode arrays. J. Neurosci. Methods; 2007; 160, pp. 317-326.

65. Rubaiy, HN. A short guide to electrophysiology and ion channels. J. Pharm. Pharm. Sci.; 2017; 20, 48.

66. Neher, E. Single-channel currents recorded from membrane of denervated frog muscle fibres. Nature; 1976; 260, pp. 799-802.

67. Neher, E; Sakmann, B; Steinbach, JH. The extracellular patch clamp: a method for resolving currents through individual open channels in biological membranes. Pflug. Arch.; 1978; 375, pp. 219-228.

68. Obaid, A et al. Massively parallel microwire arrays integrated with CMOS chips for neural recording. Sci. Adv.; 2020; 6, eaay2789.

69. Liu, M-G; Chen, X-F; He, T; Li, Z; Chen, J. Use of multi-electrode array recordings in studies of network synaptic plasticity in both time and space. Neurosci. Bull.; 2012; 28, pp. 409-422.

70. Kim, GH et al. Recent progress on microelectrodes in neural interfaces. Materials; 2018; 11, 1995.

71. Thomasjr, C; Springer, P; Loeb, G; Berwaldnetter, Y; Okun, L. A miniature microelectrode array to monitor the bioelectric activity of cultured cells. Exp. Cell Res.; 1972; 74, pp. 61-66.

72. Egert, U et al. A novel organotypic long-term culture of the rat hippocampus on substrate-integrated multielectrode arrays. Brain Res. Protoc.; 1998; 2, pp. 229-242.

73. Heuschkel, MO; Fejtl, M; Raggenbass, M; Bertrand, D; Renaud, P. A three-dimensional multi-electrode array for multi-site stimulation and recording in acute brain slices. J. Neurosci. Methods; 2002; 114, pp. 135-148.

74. Soscia, DA et al. A flexible 3-dimensional microelectrode array for in vitro brain models. Lab Chip; 2020; 20, pp. 901-911.

75. Lam, D et al. Spatiotemporal analysis of 3D human iPSC-derived neural networks using a 3D multi-electrode array. Front. Cell. Neurosci.; 2023; 17, 1287089.

76. Wu, Y et al. Three-dimensional liquid metal-based neuro-interfaces for human hippocampal organoids. Nat. Commun.; 2024; 15, 4047.

77. Huang, Q et al. Shell microelectrode arrays (MEAs) for brain organoids. Sci. Adv.; 2022; 8, eabq5031.

78. Byeon, SH; Cui, G; Lee, J; Park, J-U. Liquid metal based three-dimensional microelectrode arrays integrated with implantable ultrathin retinal prosthesis for vision restoration. Nat. Nanotechnol.; 2024; 19, pp. 688-697.

79. Brown, MA et al. Direct laser writing of 3D electrodes on flexible substrates. Nat. Commun.; 2023; 14, 3610.

80. Abu Shihada, J. et al. Highly customizable 3D microelectrode arrays for in vitro and in vivo neuronal tissue recordings. Adv. Sci.11, e2305944 (2024).

81. Jung, M. et al. Flexible 3D kirigami probes for in vitro and in vivo neural applications. Adv. Mater.37, e2418524 (2025).

82. Liu, Y et al. Nanomaterial-based microelectrode arrays for in vitro bidirectional brain–computer interfaces: a review. Microsyst. Nanoeng.; 2023; 9, 13.

83. Jing, L et al. Deep brain implantable microelectrode arrays for detection and functional localization of the subthalamic nucleus in rats with Parkinson’s disease. Bio Des. Manuf.; 2024; 7, pp. 439-452.

84. Middya, S et al. Microelectrode arrays for simultaneous electrophysiology and advanced optical microscopy. Adv. Sci.; 2021; 8, 2004434.

85. Weaver, IA; Li, AW; Shields, BC; Tadross, MR. An open-source transparent microelectrode array. J. Neural Eng.; 2022; 19, 024001.

86. Du, X et al. Graphene microelectrode arrays for neural activity detection. J. Biol. Phys.; 2015; 41, pp. 339-347.

87. Susloparova, A et al. Low impedance and highly transparent microelectrode arrays (MEA) for in vitro neuron electrical activity probing. Sens Actuators B Chem.; 2021; 327, 128895.

88. Duru, J et al. Engineered biological neural networks on high density CMOS microelectrode arrays. Front. Neurosci.; 2022; 16, 829884.

89. Müller, J et al. High-resolution CMOS MEA platform to study neurons at subcellular, cellular, and network levels. Lab Chip; 2015; 15, pp. 2767-2780.

90. Xue, X; Buccino, AP; Kumar, SS; Hierlemann, A; Bartram, J. Inferring monosynaptic connections from paired dendritic spine Ca²⁺ imaging and large-scale recording of extracellular spiking. J. Neural Eng.; 2022; 19, 046044.

91. Kasuba, KC et al. Mechanical stimulation and electrophysiological monitoring at subcellular resolution reveals differential mechanosensation of neurons within networks. Nat. Nanotechnol.; 2024; 19, pp. 825-833.

92. Saleh, MS et al. CMU Array: A 3D nanoprinted, fully customizable high-density microelectrode array platform. Sci. Adv.; 2022; 8, eabj4853.

93. McCready, FP; Gordillo-Sampedro, S; Pradeepan, K; Martinez-Trujillo, J; Ellis, J. Multielectrode arrays for functional phenotyping of neurons from induced pluripotent stem cell models of neurodevelopmental disorders. Biology; 2022; 11, 316.

94. Cerina, M; Piastra, MC; Frega, M. The potential of in vitro neuronal networks cultured on micro electrode arrays for biomedical research. Prog. Biomed. Eng.; 2023; 5, 032002.

95. Yang, Y et al. Simulation and fabrication of in vitro microfluidic microelectrode array chip for patterned culture and electrophysiological detection of neurons. Nanotechnol. Precis Eng.; 2024; 7, 023001.

96. Sifringer, L et al. An implantable biohybrid neural interface toward synaptic deep brain stimulation. Adv. Funct. Mater.; 2025; 35, 12.

97. Xu, S et al. Recent progress and perspectives on neural chip platforms integrating PDMS-based microfluidic devices and microelectrode arrays. Micromachines; 2023; 14, 709.

98. Morin, F et al. Constraining the connectivity of neuronal networks cultured on microelectrode arrays with microfluidic techniques: a step towards neuron-based functional chips. Biosens. Bioelectron.; 2006; 21, pp. 1093-1100.

99. Sato, Y et al. Microfluidic cell engineering on high-density microelectrode arrays for assessing structure-function relationships in living neuronal networks. Front. Neurosci.; 2023; 16, 943310.

100. Greve, F et al. A perforated CMOS microchip for immobilization and activity monitoring of electrogenic cells. J. Micromech. Microeng.; 2007; 17, pp. 462-471.

101. Reinhard, K et al. Step-by-step instructions for retina recordings with perforated multi electrode arrays. PLoS ONE; 2014; 9, e106148.

102. Esteban-Linares, A et al. Graphene-based microfluidic perforated microelectrode arrays for retinal electrophysiological studies. Lab Chip; 2023; 23, pp. 2193-2205.

103. Ye, Y et al. A hybrid bioelectronic retina-probe interface for object recognition. Biosens. Bioelectron.; 2025; 279, 117408.

104. Chapin, JK; Moxon, KA; Markowitz, RS; Nicolelis, MaL. Real-time control of a robot arm using simultaneously recorded neurons in the motor cortex. Nat. Neurosci.; 1999; 2, pp. 664-670.

105. Lee, KJ; Jang, KS; Shon, YM. Chronic deep brain stimulation of subthalamic and anterior thalamic nuclei for controlling refractory partial epilepsy. Acta Neurochir. Suppl.; 2006; 99, pp. 87-91.

106. Grahn, PJ et al. Restoration of motor function following spinal cord injury via optimal control of intraspinal microstimulation: toward a next generation closed-loop neural prosthesis. Front. Neurosci.; 2014; 8, 296.

107. Jackson, A; Mavoori, J; Fetz, EE. Long-term motor cortex plasticity induced by an electronic neural implant. Nature; 2006; 444, pp. 56-60.

108. Wagenaar, D., DeMarse, T. B. & Potter, S. M. MeaBench: a toolset for multi-electrode data acquisition and on-line analysis. In Proc. 2nd International IEEE EMBS Conference on Neural Engineering 518–521 (IEEE, 2005).

109. Rolston, J. D., Gross, R. E. & Potter, S. M. NeuroRighter: closed-loop multielectrode stimulation and recording for freely moving animals and cell cultures. In Proc. Annual International Conference of the IEEE Engineering in Medicine and Biology Society 2009 6489–6492 (IEEE, 2009).

110. Newman, JP et al. Closed-loop, multichannel experimentation using the open-source neurorighter electrophysiology platform. Front. Neural Circuits; 2013; 6, 98.

111. Jordan, FD; Kutter, M; Comby, J-M; Brozzi, F; Kurtys, E. Open and remotely accessible Neuroplatform for research in wetware computing. Front. Artif. Intell.; 2024; 7, 1376042.

112. Park, J; Kim, G; Jung, S-D. A 128-channel FPGA-based real-time spike-sorting bidirectional closed-loop neural interface system. IEEE Trans. Neural Syst. Rehabil. Eng.; 2017; 25, pp. 2227-2238.

113. Shupe, LE et al. Neurochip3: An Autonomous Multichannel Bidirectional Brain-Computer Interface for Closed-Loop Activity-Dependent Stimulation. Front. Neurosci.; 2021; 15, 718465.

114. DeMarse, TB; Wagenaar, DA; Blau, AW; Potter, SM. The neurally controlled animat: biological brains acting with simulated bodies. Auton. Robots; 2001; 11, pp. 305-310.

115. Novellino, A et al. Connecting neurons to a mobile robot: an in vitro bidirectional neural interface. Comput. Intell. Neurosci.; 2007; 2007, pp. 1-13.

116. Issar, D; Williamson, RC; Khanna, SB; Smith, MA. A neural network for online spike classification that improves decoding accuracy. J. Neurophysiol.; 2020; 123, pp. 1472-1485.

117. Hennig, MH; Hurwitz, C; Sorbaro, M. Scaling spike detection and sorting for next-generation electrophysiology. Adv. Neurobiol.; 2019; 22, pp. 171-184.

118. Muthmann, J-O et al. Spike detection for large neural populations using high density multielectrode arrays. Front. Neuroinformatics; 2015; 9, 28.

119. Takekawa, T; Isomura, Y; Fukai, T. Accurate spike sorting for multi-unit recordings. Eur. J. Neurosci.; 2010; 31, pp. 263-272.

120. Lewicki, MS. A review of methods for spike sorting: the detection and classification of neural action potentials. Network; 1998; 9, pp. R53-R78.

121. Lefebvre, B; Yger, P; Marre, O. Recent progress in multi-electrode spike sorting methods. J. Physiol. Paris; 2016; 110, pp. 327-335.

122. Wold, S; Esbensen, K; Geladi, P. Principal component analysis. Chemom. Intell. Lab Syst.; 1987; 2, pp. 37-52.

123. Soleymankhani, A; Shalchyan, V. A new spike sorting algorithm based on continuous wavelet transform and investigating its effect on improving neural decoding accuracy. Neuroscience; 2021; 468, pp. 139-148.

124. Jäckel, D; Frey, U; Fiscella, M; Franke, F; Hierlemann, A. Applicability of independent component analysis on high-density microelectrode array recordings. J. Neurophysiol.; 2012; 108, pp. 334-348.

125. Bestel, R; Daus, AW; Thielemann, C. A novel automated spike sorting algorithm with adaptable feature extraction. J. Neurosci. Methods; 2012; 211, pp. 168-178.

126. Pouzat, C; Mazor, O; Laurent, G. Using noise signature to optimize spike-sorting and to assess neuronal classification quality. J. Neurosci. Methods; 2002; 122, pp. 43-57.

127. Chah, E et al. Automated spike sorting algorithmbased on Laplacian eigenmaps and k -means clustering. J. Neural Eng.; 2011; 8, 016006.

128. Fournier, J; Mueller, CM; Shein-Idelson, M; Hemberger, M; Laurent, G. Consensus-Based Sorting of Neuronal Spike Waveforms. PLoS ONE; 2016; 11, e0160494.

129. Shoham, S; Fellows, MR; Normann, RA. Robust, automatic spike sorting using mixtures of multivariate t-distributions. J. Neurosci. Methods; 2003; 127, pp. 111-122.

130. Wood, F; Black, MJ. A nonparametric Bayesian alternative to spike sorting. J. Neurosci. Methods; 2008; 173, pp. 1-12.

131. Bar-Gad, I; Ritov, Y; Vaadia, E; Bergman, H. Failure in identification of overlapping spikes from multiple neuron activity causes artificial correlations. J. Neurosci. Methods; 2001; 107, pp. 1-13.

132. Pillow, JW; Shlens, J; Chichilnisky, EJ; Simoncelli, EP. A model-based spike sorting algorithm for removing correlation artifacts in multi-neuron recordings. PLoS ONE; 2013; 8, e62123.

133. Prentice, JS et al. Fast, scalable, Bayesian spike identification for multi-electrode arrays. PLoS ONE; 2011; 6, e19884.

134. Laboy-Juárez, KJ; Ahn, S; Feldman, DE. A normalized template matching method for improving spike detection in extracellular voltage recordings. Sci. Rep.; 2019; 9, 12087.

135. Yger, P et al. A spike sorting toolbox for up to thousands of electrodes validated with ground truth recordings in vitro and in vivo. eLife; 2018; 7, e34518.

136. Meyer, L; Zamani, M; Rokai, J; Demosthenous, A. Deep learning-based spike sorting: a survey. J. Neural Eng.; 2024; 21, 061003.

137. Zhang, B; Dai, J; Zhang, T. NeoAnalysis: a Python-based toolbox for quick electrophysiological data processing and analysis. BioMed. Eng. OnLine; 2017; 16, 129.

138. Rutishauser, U., Schuman, E. M. & Mamelak, A. N. Online detection and sorting of extracellularly recorded action potentials in human medial temporal lobe recordings, in vivo. J. Neurosci. Methods154, 204–224 (2006).

139. Chen, T.-C., Chen, K., Yang, Z., Cockerham, K. & Liu, W. A biomedical multiprocessor SoC for closed-loop neuroprosthetic applications. In Proc. IEEE International Solid-State Circuits Conference-Digest of Technical Papers 434–435 (IEEE, 2009).

140. Xu, H et al. Unsupervised and real-time spike sorting chip for neural signal processing in hippocampal prosthesis. J. Neurosci. Methods; 2019; 311, pp. 111-121.

141. Nakamura, T; Hayashi, N; Inuiguchi, M. Cooperative learning for adversarial multi-armed bandit on open multi-agent systems. IEEE Control Syst. Lett.; 2023; 7, pp. 1712-1717.4600691

142. Pizzi, RMR et al. A cultured human neural network operates a robotic actuator. Biosystems; 2009; 95, pp. 137-144.

143. Kohonen, T. The self-organizing map. Proc. IEEE; 1990; 78, pp. 1464-1480.

144. Bala, A; Ismail, I; Ibrahim, R; Sait, SM. Applications of metaheuristics in reservoir computing techniques: a review. IEEE Access; 2018; 6, pp. 58012-58029.

145. Farkas, I; Bosak, R; Gergel, P. Computational analysis of memory capacity in echo state networks. Neural Netw.; 2016; 83, pp. 109-120.

146. Schrauwen, B., Verstraeten, D. & Van Campenhout J. An overview of reservoir computing: theory, applications and implementations. In Proc. 15th European Symposium on Artificial Neural Networks 471–482 (2007).

147. Taherkhani, A et al. A review of learning in biologically plausible spiking neural networks. Neural Netw.; 2020; 122, pp. 253-272.

148. Diehl, PU; Cook, M. Unsupervised learning of digit recognition using spike-timing-dependent plasticity. Front. Comput. Neurosci.; 2015; 9, pp. 1662-5188.

149. She, X; Dash, S; Kim, D; Mukhopadhyay, S. A heterogeneous spiking neural network for unsupervised learning of spatiotemporal patterns. Front. Neurosci.; 2021; 14, 615756.

150. Srinivasan, G; Panda, P; Roy, K. STDP-based unsupervised feature learning using convolution-over-time in spiking neural networks for energy-efficient neuromorphic computing. J. Emerg. Technol. Comput. Syst.; 2018; 14, pp. 1-12.

151. Jin, Y; Zhang, W; Li, P. Hybrid macro/micro level backpropagation for training deep spiking neural networks. Adv. Neural Inf. Process Syst.; 2018; 31, pp. 7005-7015.

152. Wu, Y; Deng, L; Li, G; Zhu, J; Shi, L. Spatio-temporal backpropagation for training high-performance spiking neural networks. Front. Neurosci.; 2018; 12, 137.

153. Wu, Y., Deng, L., Li, G., Zhu, J. & Shi, L. Direct training for spiking neural networks: faster, larger, better. In Proc. AAAI Conference on Artificial Intelligence Vol. 33, 1311–1318 (2018).

154. Hu, Y; Tang, H; Pan, G. Spiking deep residual network. IEEE Trans. Neural Netw. Learn Syst.; 2020; 34, pp. 5200-5205.

155. Virgilio, GCD; Sossa, AJH; Antelis, JM; Falcón, LE. Spiking neural networks applied to the classification of motor tasks in EEG signals. Neural Netw.; 2020; 122, pp. 130-143.

156. Zhan, G. et al. Applications of spiking neural network in brain computer interface. In Proc. of 2021 9th International Winter Conference on Brain-Computer Interface (BCI) 1–6 (2021).

157. Singanamalla, SKR; Lin, C-T. Spiking neural network for augmenting electroencephalographic data for brain computer interfaces. Front. Neurosci.; 2021; 15, 651762.

158. Brunel, N; Hakim, V; Richardson, MJ. Single neuron dynamics and computation. Curr. Opin. Neurobiol.; 2014; 25, pp. 149-155.

159. Ma, J; Tang, J. A review for dynamics of collective behaviors of network of neurons. Sci. China Technol. Sci.; 2015; 58, pp. 2038-2045.

160. Justel, A; Peña, D; Zamar, R. A multivariate Kolmogorov-Smirnov test of goodness of fit. Stat. Probab. Lett.; 1997; 35, pp. 251-259.1484961

161. Mohseni Ahooyi, T et al. Network analysis of hippocampal neurons by microelectrode array in the presence of HIV-1 Tat and cocaine. J. Cell. Physiol.; 2018; 233, pp. 9299-9311.

162. Knight, E; Przyborski, S. Advances in 3D cell culture technologies enabling tissue-like structures to be created in vitro. J. Anat.; 2015; 227, pp. 746-756.

163. Humpel, C. Organotypic brain slice cultures: a review. Neuroscience; 2015; 305, pp. 86-98.

164. de Jongh, D; Massey, EK; Bunnik, EM. Organoids: a systematic review of ethical issues. Stem Cell Res. Ther.; 2022; 13, 337.

165. Koplin, J; Massie, J. Lessons from Frankenstein 200 years on: brain organoids, chimaeras and other ‘monsters’. J. Med. Ethics; 2021; 47, pp. 567-571.

166. Elia, N et al. Public attitudes toward the use of human induced pluripotent stem cells: insights from an Italian adult population. Front. Public Health; 2024; 12, 1491257.

167. Rolston, JD. Closed-loop, open-source electrophysiology. Front. Neurosci.; 2010; 4, 31.

168. Hafizovic, S et al. A CMOS-based microelectrode array for interaction with neuronal cultures. J. Neurosci. Methods; 2007; 164, pp. 93-106.