The potential of in vitro neuronal networks cultured on micro electrode arrays for biomedical research

The beginning of in vitro neuronal network electrophysiology using MEAs can be reconducted to the pioneering studies of Thomas et al, Gross et al, and Pine [15–17]. In 1972, Thomas et al introduced the first MEA device consisting of approximately 30 platinized gold micro-electrodes integrated into a glass substrate (two rows of 15 micro-electrodes each, spaced 100 µm apart), and succeeded in recording the activity of cultured chick cardiomyocytes [15]. Five years later, Gross and his collaborators recorded the electrophysiological activity of an isolated snail ganglion [16]. Finally, in 1980, Pine was able to record the activity of a 3 weeks-long neuronal network derived from rat neurons using a MEA device with 32 gold micro-electrodes (two rows of 16 micro-electrodes each, spaced 250 µm apart) [17]. These three hallmark studies laid the foundations for and marked the beginnings of in vitro network electrophysiology using MEAs.

Since then, MEA technology has garnered interest and contributions from a very broad cross-disciplinary research community and has continuously improved during these years. Nowadays, there is great variety of MEA technology which depends on specifications such as active or passive devices, number and density of micro-electrodes, designs, shapes or materials, and number of independent wells (figure 1(e)). With all MEAs presented, spontaneous electrophysiological activity can be recorded simultaneously from the embedded micro-electrodes. In addition, electrical stimuli can be delivered to the cells from one or multiple micro-electrodes to investigate neuronal network evoked activity [67]. In the following paragraph, we will touch upon some of these properties.

3.1.1. Standard passive MEAs

3.1.2. High-density MEAs (HD-MEAs)

3.1.3. Three dimensional MEAs (3D MEAs)

MEAs are used to measure the electrophysiological activity of networks of electrogenic cells. Upon the occurrence of electrical activity, ions (mostly sodium and potassium) travel across the cell membrane and generate an electric field which can be recorded by means of micro-electrodes placed outside of the cell membrane. In this way, the extracellular voltage that is produced by the cell when it undergoes an AP is recorded [30]. APs constitute the elementary unit associated with the transmission of neuronal signals in a network of functionally connected neurons, and are generated by significant variations of neuronal membrane potential [92]. For this reason, they can be easily recognized for their shape, which is characterized by a rapid rise and subsequent fall of the membrane potential of neurons from a baseline level. MEA systems are equipped with software configured hardware filters allowing to record at different frequency bands (i.e. up to 50 kHz), thus increasing the temporal resolution and enabling the detection of different events, such as extracellular APs and local-field potentials (LFPs). Moreover, depending on the resolution of the devices, the activity recorded by MEAs results from the contribution of more or less neurons. In the majority of MEA studies, characterizing the electrophysiological activity of 2D neuronal networks, extracellular APs (i.e. frequencies higher than 100 Hz) are recorded. Conversely, LFPs (i.e. frequencies lower than 100 Hz) are preferred when characterizing the electrophysiological activity of 3D neuronal structures. Indeed, LFPs are generated in neuronal networks by the summed and synchronous electrical activity of individual neurons, they have multiple sources and they are shaped by the spatial and temporal characteristics of these sources. Quantitative parameters describing LFPs (e.g. frequency, duration, amplitude, power) can be extracted [93, 94].

Extracellular APs are visualized in raw data as spikes, which are sudden changes in the extracellular voltage detected by MEA micro-electrodes, recognizable as actual peaks rising above the background noise, and exceeding a certain threshold [95, 96] (figures 2(a) and (b), in green). This threshold is usually defined relative to the background noise level, which is estimated from portions of the raw signals that do not contain spikes. The choice of this threshold is critical, since it determines which events are retained for further analysis. In MEA recordings, the presence of spikes defines the firing activity of the network under investigation [97].

Zoom In Zoom Out Reset image size

Since APs are triggered only when the threshold potential of neurons is reached, and their amplitude is totally independent of the stimuli amplitude, APs are commonly defined as 'all or nothing' events. Each spike indeed can be considered indistinguishable from the others produced by the same neurons, except for the instant in time at which it occurs [98]. For this reason, variations in neuronal signals are obtained by modifying, not the amplitude, but the occurring of APs over time, and, typically, we refer to neuronal signals as the temporal sequence of APs, also called spike train [99]. The time interval between two spikes is called inter-spike interval (ISI), and the variation of this parameter indicates a change in the dynamics of neuronal signals [99].

During a spike train, periods of quiescence in which spike frequency is relatively low can be interrupted by high-frequency sequences of spikes, which take the name of bursts [23, 100, 101] (figures 2(a) and (b), in red). In MEA recordings, the presence of bursts defines the bursting activity [23, 100, 101]. To date, a common agreement about the definition of bursts has not been reached [98, 102], and a large variety of burst detection methods have been proposed. The simplest approaches involve imposing thresholds on the number of spikes and the maximum allowed ISI between them, classifying any sequence of consecutive spikes satisfying these thresholds as a burst [103, 104]. These thresholds can be chosen by the user [103, 104], or derived from raw data by adaptive burst detection algorithms [105, 106]. Other methods incorporate additional thresholds on relevant parameters (e.g. the minimum interval between two bursts, and the minimum duration of a burst) [107, 108] or take alternative approaches, including the use of statistical techniques [109, 110] (e.g. hidden Markov models [111]). From an electrophysiological point of view, the generation of bursts depends on the interaction between fast, spike-generating membrane conductances, and slower mechanisms that control when spikes occur, allowing to modulate spikes frequency more abruptly [112]. However, the electrophysiological mechanisms underlying bursting activity may vary among different types of neurons [112]. Bursting dynamics are implicated in various phenomena, including synaptic plasticity [113], selective communication between neurons [114], sensory information transmission [112], and dysfunctional states such as epileptic seizures [115].

When neurons are functionally connected in a network, synchronous sequences of spikes spatially distributed across multiple recording channels can be observed [97]. These synchronous rhythmic events, normally involving the whole network, and followed by periods where activity is relatively low, take the name of network bursts (NBs) [105, 116] (figure 2(b), in light blue). NBs are characterized by a phase of increasing activity, which reaches a widespread and intense peak, followed by a relative long-lasting phase of activity, which decreases and finally ends in a refractory silent phase [117, 118]. Due to their complex dynamics, many algorithms have been proposed to identify and detect NBs in MEA recordings, and can be divided in two approaches: (i) those setting a rate-threshold to detect NBs whenever the activity rate (i.e. the number of spikes or active micro-electrodes) exceeds a specific value [97, 117, 119, 120], (ii) those setting an ISI-threshold to detect bursts whenever the ISI between consecutive spikes is less than a specific value, thereby restricting the detection of NBs only when the high activity originates from bursts in different recording channels [98, 105, 116]. Rate-threshold detectors simply bin together the spike times from all recording channels within a specified time window in order to create a firing rate histogram. In its most basic implementation, two parameters need to be set: the time window and the activity rate threshold [97, 117, 119, 120]. Conversely, ISI-threshold detectors consider that periods of low and high ISIs correspond to spikes occurring within and outside of bursts, respectively. At its most basic implementation, the ISI threshold is the only parameter required [98, 105, 116].

Compared to the activity of the adult brain, network bursting activity recorded in in vitro neuronal cultures resembles the spindles observed in the electroencephalogram (EEG) of sleeping brains, as well as epileptic activity [121]. How NBs originate in in vitro neuronal cultures is not completely clear. Experimental evidence suggests that isolated populations of oscillatory neurons within the network spontaneously synchronize and generate periodic bursts involving the whole network [122]. This particular behavior, as compared to the wide repertoire of electrical activity patterns found in wake conditions in vivo, has been partly justified by the absence of afferent inputs [123–125]. In this sense, network bursting activity would represent an exploring dynamic of close in vitro systems that, missing the natural input-output pathways of the in vivo brain, have to find a stable state with properly formed synapses [121]. As it is known from the literature, indeed, network bursting activity has an important role in establishing appropriate connections in the developing brain [126, 127]. In this sense, while spikes and bursts are already visible in the early stages of neuronal development in vitro, NBs are normally observed only later, in mature (i.e. functionally connected) neuronal networks [18, 97, 107, 117, 118, 128, 129].

The signals detected by MEAs can be graphically represented in raster plots, whose visual observation allows to qualitatively appreciate the electrophysiological activity patterns of the neuronal networks under investigation (figure 2(b)). Moreover, quantitative parameters can be extracted to describe specific characteristics of firing activity, bursting activity, and network bursting activity. Some of these parameters are easily obtainable [130], and, for this reason, they are most commonly used (figure 2(b), table 2). These include, for instance, the number of active micro-electrodes (i.e. number of micro-electrodes exhibiting a number of spikes in time exceeding a specific value), and the mean firing rate (MFR, i.e. number of spikes in time per electrode, averaged among all active micro-electrodes in a well), both describing the firing activity of the network. Similarly, the number of actively bursting micro-electrodes (i.e. number of micro-electrodes exhibiting a number of bursts in time exceeding a specific value), and the mean burst rate (MBR, i.e. number of bursts in time per electrode, averaged among all actively bursting micro-electrodes in a well), can be extracted to describe the bursting activity. Bursts can be further characterized by calculating the mean ISI within bursts (i.e. mean time interval between two consecutives spikes within bursts), the mean burst duration (MBD, i.e. average duration of all bursts detected), the mean inter-burst interval (IBI, i.e. mean time interval between two consecutive bursts), and the percentage of spikes within bursts or of random spikes (i.e. not incorporated in bursts). Also NBs, after detection, can be described by quantitative parameters extracted from raw data, including the network burst rate (NBR, i.e. number of NBs detected in time in a well), and the network burst duration (NBD, i.e. average duration of all NBs detected).

Besides these most common parameters, also others can be extracted ad hoc by using specific algorithms in order to describe specific characteristics of network activity that may arise, for instance, by the visual observation of raster plots. As an example, network bursting activity can be further characterized by extracting the coefficient of variation of the network burst interval (CV_NIBI), which allow to better describe the patterns of synchronous rhythmic activity that can be easily appreciated in raster plots [19]. Data from MEA recordings can also be analyzed to estimate functional connectivity [131]. Correlation-based methods (including independent components analysis and various measures of synchrony [132], cross-correlation [133], correlation coefficient [134], and partial correlation [135]) are the most commonly used, and enable to evaluate not only the interactions among the elements of a neuronal network, but also strength of connections, represented in connectivity maps (figure 2(c)). Other algorithms allow to analyze the propagation of neuronal signals (i.e. to identify one or more sources of APs, and to characterize their propagation in the neuronal network) [22]. With HD-MEAs, neuronal signals propagation can be characterized even at sub-cellular level, by tracking APs along neurons axons [136, 137] (figure 2(d)).

The electrophysiological activity of in vitro neuronal cultures is deeply shaped by the underlying physiological characteristics of the neuronal networks under investigation. As it will be largely discussed in the following paragraphs, MEA recordings appear to be markedly different according to (i) the species of origin of neuronal cultures, (ii) the genetic background, (iii) the developmental stage, (iv) the culturing conditions, (v) the type of neurons, and (vi) the region of the brain to which they belong. For instance, by comparing rodent and hiPSCs-derived neuronal networks, differences in spikes rate, network synchronicity and bursts features can be observed [20]. The same applies when comparing the raster plots and the parameters extracted from the same cultures at different days during maturation or differentiation [19, 97] or in different culturing conditions [138, 139], from hiPSCs-derived neuronal networks containing different ratios of excitatory and inhibitory neurons [18, 21], or from neuronal cultures composed of cortical or hippocampal neurons [140]. For this reason, neuronal cultures on MEAs have been identified as optimal platforms for physiology studies and disease models, since by analyzing MEA recordings, precious information about the cellular and molecular mechanisms occurring in the neuronal networks under investigation can be easily obtained.

For the same principle, the electrophysiological activity of neuronal cultures recorded by MEAs appears to be highly sensitive to small changes to the chemical environment, in particular to the presence of any kind of compound able to influence the physiological mechanisms of neurons. A simple example is set by the two main excitatory and inhibitory neurotransmitters (i.e. glutamate and gamma-aminobutyric acid (GABA), respectively) (figure 3). It was observed that the application of rising concentrations of glutamate, or of agonists of glutamate receptors such as α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid (AMPA) and N-methyl-D-aspartate (NMDA), finely modulated the excitatory synaptic transmission in a concentration-dependent manner, resulting in the electrophysiological activation of networks (i.e. increase of firing and bursting activity) at low concentrations, and in the loss of electrophysiological activity, due to excitotoxic effects, at higher concentrations [23, 140] (figure 3). Conversely, the addition of high concentrations of GABA in hiPSCs-derived neuronal networks including both excitatory glutamatergic and inhibitory GABAergic neurons, completely abolished firing and bursting activity immediately, reflecting the upregulation of inhibitory synaptic transmission over excitatory one [18] (figure 3).

Zoom In Zoom Out Reset image size

Pioneering works from Gross et al [24], and many others following them [23, 25–27], saw the potential to utilize in vitro neuronal networks coupled to MEAs as biosensors. Their studies aimed to identify the most suitable parameters for characterizing their unique substance-specific profiles in response to neurotransmitters, blockers, signaling molecules, drugs, and other neuroactive compounds [23–27], paving the way for today's high-throughput neurotoxicity screenings on MEAs.

MEA recordings are frequently combined with other electrophysiology techniques, including patch-clamp and calcium imaging, biological methods, such as immunostainings and transcriptome analysis, pharmacological treatment and genetic manipulation, with the aim to obtain more comprehensive information about the characteristics of the neuronal cultures under investigation.

Electrophysiology techniques, such as patch-clamp and MEAs, measuring the electrophysiological activity at different levels and with different resolution, can be combined in order to use the advantages and overcome the limitations of each. MEAs offer the opportunity of recording the electrophysiological activity of neuronal networks, simultaneously and at many sites, allowing a detailed investigation of electrophysiological activity and dynamics at the network level. Conversely, patch-clamp enables to detect those subthreshold signals which could not be recorded by MEAs (e.g. excitatory and inhibitory post-synaptic potentials), and to investigate the contribution of specific ion channel currents to the phenotype observed at the network level, for instance, in disease models [60, 141, 142]. By fine-tuning the setup settings to avoid electrical interference, patch-clamp and MEAs can be combined in the same experiment: while MEA micro-electrodes are recording or stimulating (e.g. to induce synaptic plasticity), single neurons are patched, and intracellular activity is measured [143]. Alternatively, data from MEAs and patch-clamp experiments can be integrated, allowing to obtain a more comprehensive overview of the electrophysiological activity of the neuronal networks under investigation [142].

MEAs can also be combined with calcium imaging to investigate the contribution of intracellular calcium transients to neuronal activity measured at the network level. For this purpose, ad hoc MEA devices and setups have been developed to combine MEAs with optical microscopy [144], allowing to simultaneously record neuronal network activity and measure intracellular calcium transients [145]. Alternatively, data from independent MEAs and calcium imaging experiments can be integrated. This allows, for instance, to study the specific effect of drugs on neuronal network activity and calcium signaling [146], or to evaluate the involvement of calcium-dependent pathways in both physiological (e.g. synaptic plasticity [147]), and pathophysiological mechanisms [148, 149].

Electrophysiological recordings from MEA experiments are frequently integrated also with data obtained through other biological methods, such as immunostainings and transcriptome analysis. Immunostainings techniques can be used in combination with MEA recordings to evaluate, for example, neuronal morphology [20, 150, 151], cell viability [140, 148, 152, 153], and structural connectivity (i.e. anatomical synapses) [131, 154]. Transcriptome analysis, conversely, help researchers shed light on the molecular mechanisms underlying the electrophysiological phenotype observed at the network level, by revealing, for instance, the expression profile of receptors, ionic channels, and other determinants of neuronal electrophysiology [155–157]. Often, neuronal networks on MEAs are pharmacologically treated to block the activity of specific receptors, ionic channels and enzymes, and investigate their contribution to the observed electrophysiological phenotype, hence their involvement in the physiological or pathological mechanisms under investigation [18, 155, 158, 159].

Neuronal cultures on MEAs can also be genetically modified with different purposes. Commonly used techniques of genetic manipulation include genome editing (e.g. CRISPR-Cas9) which allow to insert, replace, or delete DNA sequences, and technologies aiming to induce the expression of heterologous genes (i.e. genes which are not normally expressed in neurons), or to control the expression of receptors, ionic channels, enzymes and other proteins which are normally expressed in neurons but at different levels. These techniques can be used, for instance, to investigate how mutations, genetic variants, or the level of expression of certain proteins affect the electrophysiological phenotype observed at the network level [148, 158, 160–162], but also to label or target specific neuronal populations (e.g. for optogenetic stimulation which allows to activate or inhibit specific neuronal populations [88, 131, 137, 163]).

Numerous examples of studies in which MEA recordings were combined with electrophysiology techniques, biological methods, pharmacological treatment, and genetic manipulation, will be provided in the following chapters.

Rodent neuronal cultures are generally obtained from the embryonic or newborn brain of rats or mice. Neurons and glial cells from the desired brain region (e.g. cortex, hippocampus, thalamus) are first dissociated, and then plated on MEAs on top of the micro-electrodes. After plating, neurons grow out dendrites and axons, and form new synapses including glutamatergic excitatory synapses and GABAergic inhibitory synapses. After a maturation period of 3 weeks, neuronal networks reach a steady state of activity and are considered mature. From that moment they can remain viable for up to several months [164]. In most of the studies with MEAs, glial cells are maintained in culture with neurons to sustain neuronal networks maturation and function. However, their overgrowth can be limited by controlling the composition of the culture medium [159]. In other techniques, such as patch-clamp, it is preferred to deplete them by using anti-glial cell agents (e.g. cytosine arabinoside).

Methods to isolate neurons and glial cells from the brain of adult rats and mice have also been developed. However, protocols are more challenging to perform, and the resulting neuronal networks are more susceptible to stressors and less suited for long-term studies. For this reasons, in this review we will mainly focus on rodent neuronal cultures dissociated from embryonic or newborn rodents, which are the most used in combination with MEAs [165].

4.1.1. Activity exhibited by 2D rodent neuronal networks on MEAs

Regardless of the brain region from which neurons and glia are dissociated, spontaneous activity emerges in rodent neuronal cultures on MEAs toward the end of the first week in vitro, when a few random spikes and almost no bursts can be observed on a small portion of the recording micro-electrodes [97, 118, 128, 166]. At the end of the second week, neuronal networks exhibit a rich and stable activity behavior characterized by an evident increase in firing and bursting activity, both temporally (i.e. firing and bursting rate) and spatially (i.e. number of involved micro-electrodes) [97, 118, 128, 166]. Bursts progressively synchronize on a large number of recording micro-electrodes into long and periodic NBs, and the percentage of spikes within a burst reaches its maximum, indicating that the bursting activity dominates this stage [97, 118, 128, 166]. However, network synchronization is restricted to a few sites and does not involve all the active micro-electrodes. After three weeks in vitro, the network dynamics change dramatically, and long NBs start to be substituted by shorter NBs, with a reduced IBI and increased CV_NIBI, denoting a higher variability in the burst generation process. At this age, the network displays frequent bursting patterns and high random firing rate (i.e. spikes not incorporated in bursts), which starts to strongly decrease in the weeks thereafter [97, 118, 128, 166]. During the fourth and the fifth week in culture, the network reaches a stable condition of maturation, exhibiting rich and elaborated temporal patterns of bursting activity, characterized by NBs involving most of the recording micro-electrodes, with nearly similar values for the burst features, and almost no random spikes [97, 118, 128, 166] (figure 4(a), table 3).

Zoom In Zoom Out Reset image size

Table 3. Representative MEA studies providing a characterization of the spontaneous activity of cortical and hippocampal neuronal networks during development and at maturity, with the most relevant parameters and respective values describing mature activity. DIV = days in vitro, MFR = mean firing rate, MBR = mean burst rate, MBD = mean burst duration, IBI = mean inter-burst interval, PRS = percentage of random spikes, NBR = network burst rate, NBD = network burst duration.

	Neuronal type	Reference	Days in vitro for recording				System and plate format	Micro-electrodes per well	Relevant parameters and values characterizing mature activity
			Random spikes	Bursts	Network bursts	Mature activity
2D	Cortical neurons	Chiappalone et al [103]	7 DIV	14 DIV	14 DIV	28 DIV	Multichannel System swMEA	60	MFR ∼ 1,5 spikes s−1
									MBR ∼ 3 bursts min−1
									MBD ∼ 500 ms
									IBI ∼ 35 s
	Cortical neurons	Charlesworth et al [166]	7 DIV	14 DIV	14 DIV	28 DIV	Multichannel System swMEA	60	MFR ∼ 1 spikes s−1
									MBR ∼ 6 bursts min−1
									MBD ∼ 1000 ms
									PRS ∼ 25%
	Cortical neurons	Martens et al [142]	10 DIV	13 DIV	13 DIV	17 DIV	Multichannel System swMEA	60	MFR ∼ 1,5 spikes s−1
									PRS ∼ 25%
									NBR ∼ 12 NB min−1
	Hippocampal neurons	Frega et al [22]	—	—	—	28 DIV	Multichannel System swMEA	60	MBR ∼ 20 bursts min−1
									MBD ∼ 250 ms
									PRS ∼ 15%
									NBR ∼ 20 NB min−1
									NBD ∼ 600 ms
	Hippocampal neurons	Charlesworth et al [166]	7 DIV	14 DIV	14 DIV	28 DIV	Multichannel System swMEA	60	MFR ∼ 1,2 spikes s−1
									MBR ∼ 5 bursts min−1
									MBD ∼ 1000 ms
									PRS ∼ 15%
3D	Cortical neurons	Smith et al [93]	—	—	—	21 DIV	Multichannel System swMEA	60	MFR ∼ 0,2 spikes s−1
									Mean local field potential power ∼ 1 × 10−6 dB mV−2
	Hippocampal neurons	Frega et al [22]	—	—	—	28 DIV	Multichannel System swMEA	60	MBR ∼ 5 bursts min−1
									MBD ∼ 300 ms
									PRS ∼ 35%
									NBR ∼ 5 NB min−1
									NBD ∼ 900 ms

The developmental profile of spontaneous activity is similar in rodent neuronal cultures dissociated from the cortex and the hippocampus. Nevertheless, slight differences have been observed both during development and at maturity, and have been reported in the comparative study of Charlesworth et al [166]. Qualitative differences in activity patterns and in the burst shape were already appreciated from a preliminary observation of raw data and raster plots, moreover quantitative differences were found by comparing the parameters extracted from MEA recordings of cortical and hippocampal neuronal networks (figures 4(a) and (b), table 3). In particular, statistically significant differences in firing rate and in the fraction of spikes occurring within bursts (both higher in hippocampal networks), were observed during development, although these differences were no longer significant by the fourth week in vitro [166]. At maturity, three features showed to be critical to differentiate cortical from hippocampal cultures: CV_IBI (i.e. hippocampal spike trains tended to fire in bursts that were more regularly spaced than spike trains from cortical neurons), the percentage of theta bursting (i.e. 4–10 Hz oscillations), and the mean correlation between pairs of micro-electrodes, which were both higher in hippocampal networks [166]. However, overall, the spontaneous activity of rodent neuronal networks, dissociated either from the cortex or from the hippocampus, can be considered comparable, except for a few slight differences.

In table 3, we reported some representative MEA studies providing a characterization of the spontaneous activity of cortical and hippocampal neuronal networks during development and at maturity, with the most relevant parameters and their respective values describing mature activity.

According to the research aim, MEA recordings can be performed during network development (e.g. to investigate physiological and pathophysiological mechanisms occurring during development), or later, when neuronal networks are fully mature. Moreover, rodent neuronal cultures dissociated from the cortex or from the hippocampus can be preferred, on the basis of the relevance of the brain region in the physiological mechanism, or disease, under investigation. For instance, cortical neuronal networks represent a good system for physiology studies, and for disease modeling of most neurological disorders. Conversely, neuronal cultures isolated from the hippocampus might be preferred to investigate synaptic plasticity phenomena occurring in the hippocampal circuits, and to model neurodegenerative disorders in which the hippocampus is involved.

4.1.2. Neuronal network physiology studies

4.1.3. Neurotoxicity screening

4.1.4. Disease modeling

4.1.5. Drug studies

2D neuronal cultures are relatively simple to obtain and show clear advantages related to controllability and observability. However, they also have major limitations, as they are inherently unable to exhibit certain characteristics of in vivo systems. Indeed, it has become clear that morphological and electrophysiological properties of neuronal cells are substantially influenced by their immediate extracellular surroundings, yet the features of this environment are difficult to mimic in vitro [9–11].

In 2D neuronal models the structure is constrained by the presence of a rigid planar substrate, and this implies that the morphology of neurons is unrealistically flat, the interactions with glial cells and components of the ECM are limited to 2D, and axons-dendrites outgrowth cannot occur in all directions [9–11]. Moreover, various criticisms have been introduced on the validity of 2D models in order to investigate neuronal network activity. One of the major issues is related to the poor dynamics exhibited by 2D neuronal networks, often dominated by bursting activity encompassing most of the neurons in the network [121, 128], which is in contrast to the wide repertoire of electrical activity patterns found in in vivo studies [123, 124].

From this perspective, it is clear that 2D models for studying the characteristic of in vivo systems are extremely reduced, and that there is a tremendous need to develop culture systems that more closely model the complexity of the brain. These reasons justified the combined push in the neurophysiology community to introduce in vitro models based on 3D neuronal networks, which would allow the investigation of cellular behavior and network activity in a more physiologically relevant environment. A growing body of evidence, indeed, suggests that neurons grown in 3D cultures better represent in vivo cellular behavior than cells cultured in monolayers, as regards gene expression, differentiation, morphology, cellular signaling, and viability [9–11]. 3D neuronal cultures may also be more appropriate for electrophysiological studies than 2D counterparts. For instance, in 3D culture, Na⁺/H⁺ exchangers have been shown to have polarized expression to the apical membrane, which is difficult to maintain in monolayer cultures [262]. Moreover, neurons in a 2D environment show exaggerated Ca²⁺ dynamics in comparison to 3D cultures [263]. Lastly, it is worth underlying that 3D neuronal cultures preserve the primary advantages of traditional in vitro systems, such as control of cellular environment, and accessibility for repeated imaging.

To obtain 3D neuronal cultures from dissociated rodent neurons, biocompatible scaffolds mimicking the ECM are employed. This allows neurons and glial cells to grown within it, reproducing the 3D cytoarchitecture of the neuronal tissue in the brain. In the past years, two main approaches have been developed: (i) ECM-based scaffolds, and (ii) microbeads-based scaffolds.

For the design of ECM-based scaffolds, numerous substrates have been successfully developed, including biopolymers [93, 264], agarose [265], collagen [88, 266], chitosan [267, 268], silk proteins [269] and gel-like substances [267]. The majority of them are biocompatible polymer gels or solid porous matrices, that can be further coated with specific ECM components to support cell development, and guide axons-dendrites outgrowth [270]. It has been shown that the use of these materials induces the spontaneous formation of 3D neuronal networks with arborizations in the 3D space, closely representing the in vivo conditions [93, 265, 271]. Although this approach is very versatile, only a few studies used MEAs to characterize the electrophysiological activity of 3D neuronal networks grown within ECM-based scaffolds [88, 93, 266, 268], and some of them pointed out some limitations.

As an example, Smith et al utilized a 3D cell culture scaffold made of Alvetex, a polystyrene-based material, comprised of voids of variable sizes with interconnecting pores in which rodent mixed neuronal-glial cultures were seeded [93]. Interestingly, astrocytes morphology was notably different (i.e. extensive, intricate processes), more consistent with the one observed in vivo in comparison to 2D cultures. Moreover, by using planar MEAs, the authors demonstrated that primary neuronal-glial cultures could successfully grow in Alvetex 3D scaffolds, producing 3D neuronal networks that exhibited a spontaneous electrophysiological activity with features closer to the ones of brain slices, rather than 2D cultures of the same cells [93]. In particular, the firing rate observed in 3D cultures was significantly lower than that seen in 2D cultures, although immunostainings showed that this lower rate was not a result of a loss of viability, and LFPs occurred more frequently outside burst events. In the discussion, they suggested that the lower firing rate may arise from technical limitations related to the use planar MEAs to record the activity of 3D neuronal networks (i.e. signal attenuation due to a greater distance between cells and MEA micro-electrodes) and the possibility that some viable neurons within the 3D scaffolds lay outside of the MEA micro-electrodes receptive fields [93].

Another approach to obtain 3D neuronal cultures from dissociated rodent neurons is based on the use of microbeads. This innovative technique was proposed for the first time in 2008 by Patout et al, and involves the use of microbeads to build a modular 3D scaffold for the growth of 3D neuronal networks [266]. In Patout et al work, silica microbeads coated with ECM components were designed to be large enough to provide an adhesion surface for neuronal cell bodies, and to support growth and differentiation of axons and dendrites in the 3D space [272]. Microbeads were added in succession and spontaneously assembled into ordered 3D hexagonal arrays. Layers containing distinct subsets of neurons were formed, and neuronal processes could grow creating synaptic connections between neurons on different beads and from different layers. Layers of beads coated with chemical attractants were included to promote axonal growth and orient functional neuronal connections between different layers [272]. In the resulting 3D neuronal cultures, neuronal processes grew between the beads over the course of 3 weeks in culture to form highly interconnected networks. Despite the small number of glial cells on each bead, they were sufficient to maintain neurons health and support uniform and robust neuronal processes outgrowth throughout the bead layers. Moreover, the void spaces between the beads permitted the necessary medium exchange to maintain healthy growing conditions. Both the number of cells per bead and the number of neuronal processes were similar in all layers of the array, sign of uniform neuronal cultures [272].

The main advantage of using microbeads-based scaffolds, as compared to other methods (such as ECM-based scaffolds), is that they allow to have a high control over the final structure, organization and cell density of 3D neuronal networks, depending on the density of cells during plating and on the dimension of beads. Moreover, the final density of cells achieved in 3D neuronal cultures within microbeads-based scaffolds (i.e. 75 000 cells mm⁻³ in Patout et al [272].) is close to the 91 000 cells mm⁻³ observed in vivo [273], and definitely higher than the density for optimal viability in biocompatible polymer gels or solid porous matrices (i.e. ∼4000 cells mm⁻³).

Regarding electrophysiological recording on MEAs, microbeads-based scaffolds demonstrated to be well-suited to record the electrophysiological activity of 3D neuronal networks [22, 274, 275]. In a study of Frega et al, the electrophysiological activity of 3D cortical networks grown on beads was recorded and compared with the one exhibited by 2D cultures [22]. Rodent hippocampal neurons were cultured in 5–8 interconnected layers of glass microbeads coated with adhesion molecules on planar MEAs. MEA recordings showed that the quasi-synchronous network bursting observed in 2D neuronal cultures was maintained also in 3D networks, while global frequency was decreased and global synchrony was lost [22] (figure 4(c), table 3). In the discussion, they speculated that global asynchronous activity patterns may result from the higher complexity of the network. In particular, the wider and longer interactions may contribute to desynchronize and temporally differentiate the network activity, producing bursting activities confined to sub-populations in the network [22]. More recently, chitosan microbeads were used as 3D scaffold for rodent neuronal cultures on MEAs [274]. Chitosan is a biocompatible polymer with a structure similar to that of glycosaminoglycans, one of the components of ECM. Recently, it has been found that chitosan itself (i.e. without pre-treatment with adhesion molecules) is able to sustain primary neurons growth, allowing the formation of functional neuronal cultures [275]. Moreover, chitosan microbeads show to have internal micro-porosities that support not only the exchange of nutrients and other molecules, but also the outgrowth of neuronal arborizations with characteristics more similar to in vivo conditions [274]. Similarly to what observed by Frega et al, MEA recordings of 3D neuronal networks grown on chitosan microbeads showed a rich repertoire of complex and asynchronous activity patterns, which seems to better recapitulate activity dynamics of in vivo brain regions, when compared to the traditional networks' dynamics from 2D cultures [274, 275].

It is worth underlying that in all these reported studies, researchers employed traditional planar MEAs to characterize the electrophysiological activity of 3D neuronal networks. This is probably due to the fact that even though 3D models have become more common, almost all commercially available MEAs are designed for 2D cultures. Although this approach is valid and can provide a general view of the activity of 3D neuronal networks, the full benefit of having a 3D model is not achieved, as planar MEAs can only provide data from a single 2D plane. As discussed at the end of 3.1.3, several prototypes of MEAs with non-planar micro-electrodes and of 'true' 3D MEAs (i.e. recording simultaneously from multiple 2D planes) have recently been developed. To our knowledge, to date only Shin et al used their 'true' 3D MEA device with a top-down approach to record the electrophysiological activity of 3D rodent neuronal networks grown in ECM-based scaffolds [88]. Certainly, 3D MEAs represent a promising approach to provide the high level of spatial and temporal resolution necessary to fully harness the potential of 3D models.

In addition to dissociated neurons, which can be cultured in 2D or 3D networks, thin slices can be isolated from the whole brain or from different brain regions (e.g. cerebellum and hippocampus) of rodents. While in dissociated neuronal cultures, the in vivo organization of the neuronal tissue (i.e. the relative in vivo positions and connections between cells) is no longer preserved, brain slices maintain some of the 3D structural and functional synaptic connections within local brain regions, since, after isolation, there is no opportunity for cells to alter their relative organization from the initial in vivo state. Brain slices are cut at a thickness ranging from approximately 100–400 μm, and are normally prepared from the brain of embryonic and newborn mice and rats up to postnatal day 12 [54]. At this age, indeed, the cytoarchitecture of the neuronal tissue is established, the brain is larger and easier to manipulate, and neurons are likely to survive explantation, and the mechanical trauma caused when neuronal processes are cut, thanks to the high levels of plasticity [276].

While acute slice preparations allow to maintain neuronal viability and functionality up to a few hours, critically limiting the time-window for experiments, organotypic brain slices can be kept in culture for longer times, even for several weeks [54]. To achieve this goal, a perfusion system is necessary for the delivery of gases and nutrients, which are required to sustain cell viability and physiology [277]. After a few days, organotypic slice cultures stabilize with high cell viability, and can be maintained in culture up to several weeks, during which experiments are performed [94]. In this sense, organotypic slice cultures represent a compromise between the longevity of dissociated cell cultures, and the preservation of the cytoarchitecture of the brain. Organotypic slice cultures, indeed, preserve the 3D organization of the neuronal tissue, at least partially, and appear to be representative of the brain region from which they are derived. Both neuronal and non-neuronal cells are the same that are found in vivo [278], glial cells are present in similar proportions to those observed in vivo [279], and vascular cells are maintained [280]. A further advantage is that the development of cells and synapses in brain slices from embryonic and postnatal rodents mimics the development of the brain in vivo [278, 281]. For instance, developmental changes in spine density and shape, and increased connectivity recapitulate the in vivo phenotype observed in acute slices from age-matched time points [281].

4.3.1. Activity exhibited by organotypic slices on MEAs

4.3.2. Physiology studies and disease modeling

As soon as hiPSCs technology was introduced in 2006 by Yamanaka et al, several research groups focused their efforts on developing methods to differentiate hiPSCs into the cell types of the human body, including neurons and glial cells. In this chapter, we will firstly describe the most commonly used methods of differentiation of hiPSCs into neuronal cultures, giving particular attention to the ones optimized for experiments on MEAs. Afterwards, we will look at some relevant studies in which hiPSCs-derived neuronal networks on MEAs were used to study the brain physiology, model neurological disorders, investigate disease pathophysiology, and test new therapeutic treatments.

5.1.1. Protocols of neuronal differentiation of hiPSCs on MEAs

5.1.2. Activity exhibited by 2D hiPSCs-derived neuronal networks on MEAs

Similarly to that described for rodent neuronal cultures, the electrophysiological activity of hiPSCs-derived neuronal networks at maturity is generally characterized by rich and elaborated temporal patterns of bursting activity, which may vary depending on the neuronal populations included in the network under investigation. Also the development of electrophysiological activity retraces the steps described for rodent neuronal networks, but with different timescales according to the neuronal differentiation protocol which is adopted.

Differentiation protocols based on dual-SMAD inhibition result in the formation of mature and functional neuronal cultures whose electrophysiological activity can be recorded by means of MEAs. So far, the vast majority of studies characterized long-term development of network activity and connectivity of cortical neuronal networks [326, 330, 331], in some cases, with the aim of performing a direct comparison with rodent neuronal cultures, held as the 'gold standard' in the MEAs field [4, 20]. In Hyvärinen et al, for instance, hiPSCs-derived cortical networks were plated on MEAs at day 32 of differentiation, and regularly recorded over the span of 4 months [20]. Widespread activity was expressed from the first day on MEAs onwards. In particular, during the first two weeks on MEAs (i.e. days 32–46 of differentiation), spike rate increased, and uncorrelated spikes started to organize into NBs. Both the number of bursts and the percentage of spikes participating in bursts progressively increased. Moreover, while, initially, NBs were short and at high frequency, once networks matured, bursts became longer and less frequent [20]. After 66 d of differentiation, hiPSCs-derived cortical networks reached a mature and stable state of activity, characterized by a rich repertoire of bursting activity which was still apparent after 4 months [20] (figure 5(a), table 5). As compared to rodent counterparts, hiPSCs-derived cortical networks developed slowly (i.e. 66 d of differentiation versus 28), and at maturity were characterized by different dimension and morphology, bursts with longer duration and lower spike frequencies, and by a higher number of random spikes [20]. Other studies report even longer developmental timescales. For instance, Odawara et al investigated the development of spontaneous electrophysiological activity of hiPSCs-derived cortical networks obtained with the differentiation protocol from Axol Bioscience. Neuronal network activity was recorded for over 1 year, and the complete maturation of spontaneous firing, evoked responses, and pharmacological modulation by glutamatergic and GABAergic receptors antagonists and agonists was observed after 5–7 months [326] (table 5).

Zoom In Zoom Out Reset image size

Table 5. Representative MEA studies providing a characterization of the spontaneous activity of hiPSCs-derived neuronal networks obtained by using different protocols of neuronal differentiation during differentiation and at maturity, with the most relevant parameters and respective values describing mature activity. DIV = days in vitro, WIV = weeks in vitro, MFR = mean firing rate, MBR = mean burst rate, MBD = mean burst duration, IBI = mean inter-burst interval, PRS = percentage of random spikes, NBR = network burst rate, NBD = network burst duration.

	Neuronal differentiation protocol	Neuronal type	Reference	Days in vitro for recording				System and plate format	Micro-electrodes per well	Relevant parameters and values characterizing mature activity
				Random spikes	Bursts	Network bursts	Mature activity
2D	Dual-SMAD inhibition	Mixed (cortical glutamatergic neurons, GABAergic neurons)	Hyvärinen et al [20]	32 DIV	46 DIV	52 DIV	66 DIV	Axion Biosystems 12wMEA	64	MFR ∼ 2 spikes s−1MBR ∼ 5 bursts min−1 MBD ∼ 800 ms PRS ∼ 40%
	Axol Bioscience (proprietary)	Mixed (cortical glutamatergic neurons, GABAergic neurons)	Odawara et al [326]	2 WIV	7 WIV	10–13 WIV	20–30 WIV	Alpha Med Scientific swMEA	64	MFR ∼ 6 spikes s−1NBR ∼ 3 bursts min−1 NBD ∼ 350 ms
	Dual-SMAD inhibition	Dopaminergic neurons	Sundberg et al [332]	—	—	—	56 DIV	Axion Biosystems48wMEA	16	MFR ∼ 0,5 spikes s−1MBR ∼ 0 bursts min−1 MBD ∼ 300 ms
	Dual-SMAD inhibition	Motor neurons	Kim et al [150]	—	—	—	30 DIV	Axion Biosystems48wMEA	16	MFR ∼ 7 spikes s−1NBR ∼ 1 bursts min−1 NBD ∼ 5 s
	Overexpression of Ngn2	Cortical glutamatergic neurons	Frega et al [155]	17 DIV	17 DIV	24 DIV	28 DIV	Multichannel System 24wMEA	12	MFR ∼ 5 spikes s−1MBR ∼ 4 bursts min−1 MBD ∼ 800 ms PRS ∼ 30%
	Overexpression of Ngn2 and Ascl1	Cortical glutamatergic and GABAergic neurons	Mossink et al [18]	—	—	—	49 DIV	Multichannel System 24wMEA	12	MFR ∼ 3 spikes s−1NBR ∼ 3 NB min−1 NBD ∼ 750 ms PRS ∼ 50%
Brain organoids	Guided method	Cortical organoid	Trujillo et al [333]	—	—	2 months	10 months	Axion Biosystems12wMEA	64	MFR ∼ 15 spikes s−1MBR ∼ 0,3 bursts s−1 Synchrony index ∼ 0,8
Other 3D structures	Overexpression of Ngn2	Cortical glutamatergic neurons	Muzzi et al [55]	17 DIV	28 DIV	38 DIV	42 DIV	Multichannel System swMEA	60	MFR ∼ 0,5 spikes s−1PRS ∼ 95% NBR ∼ 0,5 bursts min−1 NBD ∼ 600 ms

Interestingly, the electrophysiological activity of hiPSCs-derived neuronal cultures obtained with dual-SMAD inhibition is critically different according to the neuronal type which is included in the network under investigation. While cortical glutamatergic and GABAergic neuronal cultures are characterized by elaborated temporal patterns of bursting activity, dopaminergic and motor ones show a completely different electrophysiological behavior. Differences can be clearly appreciated by visual observation of raster plots, and by comparing the quantitative parameters extracted from raw data: dopaminergic neuronal cultures are characterized by a low and sporadic firing activity, and almost no bursting activity [162, 332], while motor neurons are characterized by high firing activity, and few but very long bursts [150] (figures 5(b) and (c), table 5).

As mentioned in the previous paragraph, other neuronal differentiation protocols are based on the overexpression of lineage-determining transcription factors, such as Ngn2 and Ascl1 for differentiation into cortical glutamatergic and GABAergic neurons, respectively. In this case the developmental profile of network activity is similar, but with a different, generally shorter, timescale [18, 155]. In Frega et al [155], for instance, spontaneous activity emerged in homogeneous glutamatergic neuronal networks on MEAs after only a few weeks of differentiation (i.e. days in vitro, DIV 17). At this stage, only random spikes and almost no bursts were observed, indicating that neurons were electrophysiologically active, but still immature and not yet integrated in a functionally connected network [155]. Afterwards, electrical activity increased, reaching its highest level during the fourth week (i.e. DIV 24). By this timepoint, neuronal networks exhibited a rich bursting activity, and rhythmic synchronous events occurring across all the micro-electrodes (i.e. NBs), meaning that neurons have self-organized into a synaptically connected network [155]. Similarly to neuronal cultures obtained with dual-SMAD inhibition, initially, networks bursts appeared short and at a high frequency, while once network were fully mature by the end of the fourth week (i.e. DIV 28), less frequently occurring and longer-duration NBs became the more dominant form of activity [155] (figure 5(d), table 5).

Interestingly, when cortical glutamatergic and GABAergic neurons are co-cultured, GABA signaling shows to deeply shape network activity [18]. In this regard, Mossink et al performed a comprehensive analysis comparing different network compositions of either glutamatergic neurons alone, or in co-culture with GABAergic neurons on MEAs [18]. Over development, they detected a shortening of NBD, as well as reduced NBR and MFR, in contrast to increased percentage of random spikes, in neuronal networks in which glutamatergic and GABAergic neurons were co-cultured. All these parameters only became significantly different after 6 weeks of differentiation (i.e. DIV 42), when GABAergic neurons were fully mature, indicating that these differences in network activity were dependent on GABA signaling [18] (figure 5(e), table 5). Moreover, by recording spontaneous activity of mature networks with different ratios of glutamatergic and GABAergic neurons, they found that the NBD was negatively correlated to the percentage of GABAergic neurons, meaning that the effect of GABA signaling on network activity was scalable to the amount of inhibition in the network [18].

In table 5, we reported some representative MEA studies providing a characterization of the spontaneous activity of hiPSCs-derived neuronal networks obtained with different protocols during development and at maturity, with the most relevant parameters and their respective values describing mature activity.

To conclude, it is worth underlying that, besides the protocol of neuronal differentiation, and the neuronal populations included in the network (i.e. neuronal type and ratio), several other factors influence the functional development, and deeply shape the activity of neuronal networks cultured on MEAs, including the culturing conditions [139, 328] (e.g. the ECM substrates used for coating) and the presence of glial cells [334, 335]. For instance, a study from Tukker et al demonstrated that co-cultures of cortical glutamatergic and GABAergic neurons containing hiPSCs-derived astrocytes, developed spontaneous neuronal activity earlier, with increased bursting behavior later during development, as compared to neuronal cultures without glial cells [334]. All these factors should be always taken into account when developing a hiPSCs-based model on MEAs.

5.1.3. Neuronal network physiology studies

5.1.4. Disease modeling

The most commonly used methods of neuronal differentiation enable to differentiate hiPSCs into 2D neuronal cultures consisting of different neuronal populations, and glial cells, depending on the adopted protocol. Nevertheless, as discussed for rodent cultures in 4.2, it has become clear that 2D neuronal models have major limitations, since cultures are constrained into monolayers, thereby they are inherently unable to reproduce the 3D environment and cytoarchitecture of the human brain [9–11]. For this reason, researchers have combined their efforts to develop protocols to differentiate hiPSCs into 3D neuronal cultures, which would allow the investigation of cellular behavior and network activity in a more physiologically relevant environment. Interestingly, unlike dissociated neuronal cultures from rodents, hiPSCs offer the opportunity to obtain brain organoids, which are self-assembled 3D aggregates of neuronal cells, with cell types (i.e. neurons and glial cells) and cytoarchitecture that resemble the embryonic human brain [56]. For this reason, brain organoids represent a promising system for modeling neurogenesis and neurodevelopmental disorders.

In this section, we will review some relevant protocols to differentiate hiPSCs into brain organoids and other 3D structures on MEAs, afterwards we will look at some first studies in which brain organoids in combination with MEAs were used to model neurological disorders.

5.2.1. Brain organoids and other 3D structures

5.2.2. Disease modeling

In MEA studies, a priori considerations are required to obtain in vitro neuronal models which can provide the most valuable and accurate information on the physiological mechanisms or the diseases under investigation, in compromise with the availability of materials, time, and expertise. In this sense, the experimental aim is critical to guide the decisions to be taken, in particular, regarding the choice of (i) the neuronal source (i.e. rodent or hiPSCs), (ii) the neuronal network composition (i.e. heterogeneous or homogeneous cultures, which cell types and neuronal populations, from which brain region), and (iii) the structural complexity (i.e. 2D or 3D neuronal networks). For each of these decisions, advantages and disadvantages should be considered.

The first decision to be taken regards the neuronal source (i.e. whether to isolate neurons from rodent brains or to differentiate them from hiPSCs lines). Compared to hiPSCs, rodents are an accessible source of mammalian neurons beyond comparison for many laboratories. Since their introduction (more than a century ago), protocols to isolate and culture rodent neurons have been highly optimized and standardized, and countless lines of transgenic mice and rats, with well-defined genotypes and phenotypes, have been made commercially available. This provided neuronal cultures with reduced variability, and more reproducible and reliable results [148, 228, 234, 235, 237, 240–242]. For this reason, 2D rodent neuronal networks represent the 'gold standard' in the MEAs field, and studies using MEAs to characterize the neuronal activity of healthy rodent cultures, both during development and at maturity, are largely available in the literature [97, 118, 128, 166].

Several studies have demonstrated the validity of results provided by rodent models when it comes to investigate physiological and pathological mechanisms that occur in similar ways in both rodent and human brains. Frequently, indeed, results obtained from rodent neuronal networks on MEAs have been confirmed by following works based on hiPSCs-derived neuronal networks, or by in vivo observations. This is the case, for instance, of studies investigating synaptic plasticity phenomena [143, 169–176] and neuromodulation of neuronal activity [187–193], screening neurotoxic compounds [203–207], and characterizing the phenotype of certain neurological disorders [225, 227, 243–245]. In all these cases, rodent neuronal cultures still represent the most rapid, economical, and accessible way to gain valuable insights about the human brain, in both physiology and pathology.

Nevertheless, the use of rodent neurons for biomedical research shows several limitations. Firstly, the use of rodent cells requires frequent euthanasia of live animals and arises substantial ethical concerns. Secondly, besides the genomic, developmental, and physiological inter-species differences [2–5], problems of translatability are also related to the fact that behavioral and cognitive functions, typical of the human brain, can poorly be reproduced in rodent models. By extension, many neurological disorders, particularly those correlated to complex behavioral and cognitive impairments such as psychiatric and neurodevelopmental disorders, can hardly be modeled by using rodent models. In this sense, hiPSCs-derived neurons are believed to represent a promising attempt to reduce the use of animals in biomedical research, and to overcome problems of translatability into human patients.

A key advantage of the use of hiPSCs-derived neurons on MEAs for biomedical research is that hiPSCs technology enables to derive neuronal cultures directly from patients with a well-defined genetic background and clinical phenotype [152, 155, 162, 342, 343, 346, 357]. On one side, this gives the opportunity to develop patient-specific in vitro platforms to test which drugs and therapeutic treatments are more likely to be effective in the specific patient, paving the way toward personalized medicine. On the other side, it enables to investigate the pathophysiology of complex disorders more deeply, in particular those with a relevant genetic component or with unclear causes, by linking specific alterations of the electrophysiological activity of neuronal networks recorded on MEAs, to specific pathogenic mutations, and phenotypic traits [152, 161, 340, 342, 352, 353]. Unluckily, the complexity of neurological disorders, which are usually correlated to an elevated level of heterogeneity in patients' population, makes it difficult to establish this link. For this reason, it is preferred to have access to a cohort of patients sharing, for instance, pathogenic mutations in the same gene [155, 162], or a common phenotypic trait [152, 352], rather than a single patient, to avoid the risk to model the patient-specific phenotype, hence to gain insights about pathophysiological mechanisms and candidate therapeutic targets which cannot be generalized to other patients.

Another advantage of the use of hiPSCs-derived neurons is that hiPSCs lines can be easily genetically manipulated by means of genome editing techniques. This allows to differentiate affected neuronal networks carrying specific pathogenic mutations and genetic variants from healthy hiPSCs lines [18, 150, 156, 160, 314, 327, 353], and isogenic controls from patients' hiPSCs lines [155, 161, 346–348], which simplifies the characterization of the pathophysiological phenotype and enable to more clearly isolate and study the contribution of specific mutations and variants to the disease.

To date, the main limitations of hiPSCs-derived neurons are related to the fact that hiPSCs technology is still in its early stage. Although several protocols of neuronal differentiation are available nowadays, most of them have not been fully optimized and standardized yet. For this reason, often, a large amount of materials, time, and expertise is required, and it is not easy to compare neuronal cultures obtained through different protocols.

Once the neuronal source is decided, a choice should be made regarding the composition of neuronal networks. In the case of rodent neuronal cultures, dissociated neurons or organotypic slices can be isolated from different brain regions according to their relevance to the physiological or pathological mechanism under investigation. In most of MEA-based studies for biomedical research, rodent neurons are isolated from the cerebral cortex, which is involved in complex brain functions, such as memory and learning, and is frequently affected in neurological disorders. Therefore, cortical neuronal networks represent a good system for physiology studies, such as the investigation of synaptic plasticity phenomena [143, 169–175], signal propagation [185, 186], and neuromodulation [187–193], and for disease modeling of neurological disorders, including AD [159, 222], epilepsy [158, 228–231, 233], neurodevelopmental disorders [142, 236, 237], TBI [247, 248, 254], and encephalopathies [250–252]. In other cases, neuronal cultures isolated from the hippocampus are preferred. For instance, thanks to the partial preservation of structural and functional synaptic connections, organotypic slices from the hippocampus are particularly well suited to investigate synaptic plasticity phenomena occurring in the hippocampal circuits [294], whereas dissociated hippocampal neurons can be used to model neurological disorders in which the hippocampus might be involved, such as AD [141, 148, 223], DLB [224], epilepsy [226, 227, 232] and psychiatric disorders [239].

The large availability of protocols to differentiate hiPSCs into different cells of the nervous system represents a great advantage for both physiology studies and disease modeling. Researchers can choose the cell types and neuronal populations which are the most relevant to the physiological mechanisms or the pathophysiology of the disease under investigation, including glial cells (i.e. astrocytes, oligodendrocytes, and microglia) and specific neuronal populations (i.e. cortical glutamatergic and GABAergic neurons, dopaminergic neurons, motor neurons) [316–325].

Different cell types and neuronal populations can be combined into heterogeneous cultures, aiming to mimic the complexity of the human brain and to model the pathophysiological phenotype resulting from the interplay of different neuronal populations. This is the case of models based on highly heterogeneous cortical cultures, including neurons and glial cells obtained with dual-SMAD inhibition protocols [8, 309]. Other models, conversely, are based on the generation of homogeneous populations of cells (i.e. neurons obtained with Ngn2 or Ascl1 induction, glial cells) that can be cultured alone or together in co-cultures. The aim, in this case, is precisely to reduce the complexity of the model, in order to characterize the contribution of specific neuronal populations [18, 314], or of the interactions between specific neuronal populations and glial cells [151, 341] to the pathophysiological phenotype.

Similarly to rodent models, hiPSCs-derived cortical cultures are the most used for modeling neurological disorders, including ASD [18, 152, 160, 340–342], TSC [151, 315, 346], FXS [156, 347, 348], psychiatric disorders [154, 161, 352, 357], epilepsy [327, 330, 353, 354], and AD [153]. Other disorders, such as PD [314] and amyotrophic lateral sclerosis (ALS) [150, 314], conversely, have been modeled by using homogenous cultures of dopaminergic and motor neurons, respectively, since they represent the neuronal populations which are specifically affected in these neurodegenerative diseases.

Another choice that should be based on the characteristic of the disease under investigation and on the experimental aim is the differentiation protocol. Several protocols of neuronal differentiation are nowadays available, and are based on two approaches: (i) the dual-SMAD inhibition principle, and (ii) the overexpression of lineage-determining transcription factors. Differentiation protocols based on dual-SMAD inhibition allow to obtain heterogenous cultures of neurons and glial cells in about 2 months [8, 309–311]. Since they mimic the timescale of human cortical development in vivo [8, 309], they are particularly well-suited to investigate pathophysiological mechanisms occurring in the cerebral cortex during development [152]. However, they are long protocols, requiring a large amount of materials and time, and providing a limited control over the final density and composition of the networks. Conversely, protocols based on the overexpression of transcription factors enable to obtain homogeneous cultures of glutamatergic neurons [7, 50], or co-cultures of glutamatergic and GABAergic neurons [18, 312, 313], in a shorter time (i.e. few weeks), with a reduced variability over the final density and composition of the network (e.g. neuronal networks with controlled percentage of glutamatergic and GABAergic neurons can be obtained [18]). Since, in many cases, it is not necessary to use long protocols based on dual-SMAD inhibition to obtain the desired pathophysiological phenotype, protocols based on the overexpression of transcription factors often represent the most rapid and convenient way to obtain human in vitro models of neurological disorders [18, 155, 161, 356].

Finally, besides the composition of neuronal networks in terms of cell types, a decision should be taken regarding the structural complexity of the model (i.e. whether culturing neurons and glial cells into 2D or 3D neuronal networks). 2D neuronal networks are the most widely used, since they are simpler to culture, they make changes in morphology easy to monitor with various imaging techniques, and they provide valuable results in both physiological studies and disease modeling. Nevertheless, the lack of 3D cytoarchitecture makes them inherently unable to model certain aspects of the brain, such as the interactions cell-cell and cell-ECM in the 3D space. As a consequence, also the electrophysiological behavior of 2D neuronal cultures is less complex, and farther than what observed in vivo [22]. On the contrary, 3D models allow to study neuronal networks in a more physiologically relevant environment. These include (i) 3D neuronal cultures, from both rodents and hiPSCs, (ii) organotypic brain slices isolated from rodent brains, and (iii) hiPSCs-derived brain organoids. 3D neuronal cultures are obtained by culturing neurons, dissociated from rodents or differentiated from hiPSCs, within a 3D scaffold, consisting of biocompatible polymer gels, solid porous matrices, or microbeads, to mimic the ECM [55, 88, 93, 266, 268, 274, 388]. This allows neurons and glial cells to grow arborizations in the 3D space, partially reproducing the cytoarchitecture of the neuronal tissue. The electrophysiological activity of 3D neuronal cultures recorded by MEAs showed to be deeply shaped by the 3D cytoarchitecture, less stereotypical and similar to what observed in vivo [22, 55, 274]. Undoubtedly, neuronal cultures grown within a 3D scaffold do not fully mimic in vivo conditions, but they allow to have a good reproducibility, and full control over the final size, structure, cellular composition (i.e. cell type and ratio) and density. Another 3D model which can be combined with MEAs is represented by organotypic slice cultures isolated from rodents. Organotypic slices partially preserve the 3D organization of the neuronal tissue, and the local structural and functional connections of the brain region from which they are derived [54, 278–281], however they cannot be maintained in culture for long periods of time (i.e. few weeks). In this sense, they represent a compromise between the longevity of dissociated cell cultures, and the preservation of the 3D cytoarchitecture of the brain. Undoubtedly, hiPSCs-derived brain organoids represent the most intriguing option to model the complexity of the human brain in both physiology and pathophysiology [56, 358–360]. hiPSCs, indeed, are able to self-assemble without the use of a 3D scaffold, into spheroidal aggregates with an organized cytoarchitecture, composed of NPCs, neurons and glial cells, resembling the embryonic human brain at different developmental stages [56, 333, 358–360, 379]. However, they are a relatively new technique with critical limitations related to protocols lack of standardization, high variability, and inaccessibility in terms of time and materials (i.e. extremely long protocols, up to several months) [329]. Despite the undeniable advantage of having a 3D model which so well mimic the cytoarchitecture, development, and electrophysiology of the in vivo brain, the above-mentioned limitations fully explain why hiPSCs-derived brain organoids in combination with MEA technology are not widely used for biomedical research.

To conclude, when it comes to developing a neuronal model on MEAs, advantages and disadvantages of the different possibilities should be critically evaluated according to the experimental aim, with the purpose of obtaining a simplest model providing the most valuable and accurate information in compromise with the availability of materials, time, and expertise.

Nowadays, several MEAs are available, including (i) traditional low-density swMEAs and mwMEAs, (ii) HD-MEAs, (iii) MEAs with non-planar micro-electrodes, and 'true' 3D MEAs. Right after the choice of a neuronal model, according to the characteristics of the chosen model and to the information required for the experimental aim, a MEA device should be selected from the available ones, which is not always trivial. In this paragraph, we provide some guidelines for the choice of MEA devices. In this sense, strengths and weaknesses of each device should be evaluated, with the aim to choose the MEAs providing the most valuable and accurate information, in compromise with the availability of materials, time, and expertise.

Traditional swMEAs and mwMEAs are classified as low-density devices. swMEAs are available in different formats, with a number of micro-electrodes ranging from 60 to 256, organized in different layouts, whereas mwMEAs are available in plate formats with up to 96 independent wells, with a limited number of micro-electrodes. mwMEAs represent an outstanding high-throughput platform, enabling to measure the electrophysiological activity of several neuronal networks at the same time. This is particularly advantageous in neurotoxicity and drug screenings, when it comes to screen a large number of neurotoxic chemicals and drugs [200], to characterize their dose-response curve [201], or to test the same compound in neuronal networks with different characteristics, for instance, representative of different brain regions [195]. Similarly, mwMEAs can be used in disease modeling to characterize the electrophysiological phenotype of neuronal networks derived from the hiPSCs of different patients, which are often characterized by high variability. In all these cases, mwMEAs represent the most convenient and rapid way to obtain a general, but valuable, overview of the network activity of neuronal cultures under different conditions. However, the downside of mwMEAs is that the number of micro-electrodes per well decreases with the increase in the number of wells (i.e. Multi-Channel Systems: 12 micro-electrodes for 24-well plate and 3 micro-electrodes for 96-well plate; Axiol Biosystems: 16 micro-electrodes for 24-well plate and 8 micro-electrodes for 96-well plate). Although the spatial resolution is adequate to observe the overall neuronal network activity, it is not good enough to obtain other types of information (i.e. functional connectivity, signal propagation). In this case, devices with higher number of micro-electrodes should be chosen (swMEAs or HD-MEAs).

In particular, the recently developed HD-MEAs present a much higher number and density of micro-electrodes (up to thousands of micro-electrodes per mm²), allowing to record the activity of neuronal networks with spatial-temporal resolution and signal quality once unimaginable, and to integrate the extracellular recordings at subcellular, cellular, and network levels from the same culture [74–80]. For these reasons, HD-MEAs are particularly well-suited to investigate particular physiological and pathological mechanisms with high spatial-temporal resolution, including the contributions of individual presynaptic synapses to postsynaptic potentials in short- and long-term synaptic plasticity [143], or how AP propagation along axons is affected in neurodegenerative diseases such as PD and ALS [314]. Recently, Ito et al used a 512-channel HD-MEAs to measure the activity of organotypic cultures of cortical and hippocampal brain slices [288]. In this case, the use of HD-MEAs enabled the analysis and characterization of functional connectivity between neurons at different frequency ranges, pointing out differences of network architecture in different brain regions, made evident thanks to the high spatial-temporal resolution provided by HD-MEAs [288]. HD-MEAs should be also preferred to record the electrophysiological activity of hiPSCs-derived brain organoids, given their small size [372, 373]. However, the increased spatial and temporal resolution reached with HD-MEAs results in a huge amount of raw data (e.g. a 10 min recording with a 4096 electrode MEA device and sampling rate of 20 kHz is approximately 90 GB in size [30]), which on one hand are easily obtainable, but on the other hand are difficult to handle as regards to data storage and analysis. In addition, the use of HD-MEAs requires the careful evaluation of possible cross-talk artifacts, which can affect the electrophysiological recordings. Channel interference due to cross-talk is can be partially addressed, via continuous interleaved sampling or post-data acquisition spike sorting [81]. For these reasons, it is preferred to use HD-MEAs only when high spatial and temporal resolution is needed to investigate particular physiological and pathological mechanisms [143, 314].

Lastly, MEAs with non-planar micro-electrodes and 'true' 3D MEAs (i.e. recording simultaneously from multiple 2D planes) are nowadays available. These devices are particularly well-suited for recording the electrophysiological activity of 3D neuronal cultures, including 3D neuronal networks from both rodents and hiPSCs grown within a biocompatible scaffold, organotypic brain slices isolated from rodent brains, and hiPSCs-derived brain organoids. In these cases, electrophysiological recordings acquired by non-planar micro-electrodes, or by micro-electrodes distributed on multiple 2D planes, allow to better characterize the complex dynamics of 3D neuronal networks, similar to those observed in vivo [22, 55, 274]. However, to date, only a few studies used 3D MEA devices in combination with 3D neuronal cultures [87, 88, 149], while the vast majority of studies were limited to a general, even if valuable, overview of 3D neuronal networks activity provided by traditional planar MEAs. This is probably due to the fact that even though 3D models have become more and more common, 3D MEA technology is still in its infancy, and almost all commercially available MEAs are designed for 2D cultures. However, without any doubt, 3D MEA devices should be preferred to characterize the network activity of 3D neuronal cultures, to fully benefit of having a 3D model.

Despite the increasing popularity of MEAs, there is little insight into how MEA-based neuronal models should be used (i.e. which recommendations should be followed in regards of control neuronal networks, experimental design, data analysis and reporting) to fully harness the potential of MEA technology in the field of biomedical research.

In this sense, a recent work from Mossink et al provided a set of guidelines for the evaluation of control networks and the design of experiments for functional phenotyping on MEAs [19]. In particular, they used hiPSCs derived from 10 healthy subjects which were differentiated into excitatory neuronal networks through Ngn2 overexpression, and cultured on MEAs by different researchers over a period of several years. They showed that different control networks were highly comparable when used in the time window between DIV 27 and 35, as networks generated by Ngn2 overexpression presented stable activity at this stage [19]. However, many factors can influence the timing of this stable network activity, including the differentiation protocol and coating. Therefore, one must define the stable developmental period depending on each protocol, before pooling and comparing data [19]. They found that astrocytes and MEAs batch introduced variability, and advised to use at least 12 wells per condition, divided over two MEA batches with the same astrocytic batch [19]. In addition, they suggested to exclude from analysis wells with low or uneven cell density, or with low activity, indicating a set of values of MFR, NBR and active channels, indicative of the level of activity of healthy neuronal networks [19]. In regard of disease modeling, they advised to use hiPSCs lines derived from multiple healthy subjects and patients, or isogenic sets, to reliably characterize the disease phenotype, since differences in genetic background between hiPSCs donors dominate the variability at the transcriptional level [19].

Although these guidelines are valid only for homogeneous neuronal networks of excitatory neurons obtained through a very specific protocol of differentiation (i.e. Ngn2 overexpression), we believe that the work of Mossink et al is extremely inspiring. To our knowledge, clear indications about how the electrophysiological activity of healthy neuronal networks on MEAs should look like are largely missing. Even if the electrophysiological activity of rodent neuronal networks on MEAs, both during development and at maturity, has been characterized in the past 50 years, we noticed a great variability in the values of MFR, MBD and NBR reported for control neuronal networks (table 3) [22, 103, 142, 166]. This issue is even more critical with regard to 3D neuronal networks [22, 55, 93], hiPSCs-derived networks of specific neuronal populations (such as dopaminergic and motor neurons) [150, 332], and brain organoids [333] (tables 3 and 5), since studies characterizing the electrophysiological phenotype of these neuronal networks on MEAs are scarce or none. Although this can be partially explainable with the relative newness of these techniques, which can be in themselves related to protocols lack of standardization, high variability, and inaccessibility in terms of time and materials, this scenario can be extremely confusing to those who want to approach MEA technology for the first time. Undoubtedly, the present review represents a step in the right direction. However, we believe that more studies like the one of Mossink et al, providing a set of guidelines for the evaluation of healthy neuronal networks, and for the experimental design of functional phenotyping assays, should be carried out and made available to the scientific community in order to fully harness the potential of MEA technology.

The huge amount of data that can easily be obtained from MEA recordings shifts the challenges from the experimental to the analysis side. Improvements in data analysis have greatly simplified the extraction of quantitative parameters from raw data. However, data analysis itself can remain a hurdle, especially for researchers with no background in programming. For example, it is possible that the observed phenotype cannot be described with any of the most commonly used parameters, and new parameters need be introduced to capture these signatures and reveal previously unseen phenotypes [19]. In this case, new ad hoc algorithms should be written, in order to describe specific characteristics of network activity which may arise, for instance, by the visual observation of raster plots. Then, special attention should be given for the choice of different analysis settings and algorithms for the detection of spikes and bursts, or for the identification of NBs, since they largely influence the results. These settings need to be fine-tuned, depending on the visual observation of raster plots to accurately detect the phenotypic signatures of the network under investigation. For instance, NBs exhibited by affected neuronal networks might be incorrectly detected with commonly used settings, since these were conventionally chosen based on NBs in control networks [19]. Finally, the complex spatiotemporal organization of the electrophysiological activity generated by neuronal networks on MEAs represents a relevant challenge when it comes to selecting meaningful phenotyping metrics. In literature, MEA-based studies often report one single parameter (i.e. the MFR) as it is easily extracted from raw data. However, in the study from Mossink et al investigating the variability of MEA parameters in hiPSCs-derived neuronal networks, the MFR was found to be one of the most variable parameters, thereby not trustworthy to describe, for instance, the phenotype of disease models [19]. In addition, it only describes the general level of activity and is largely dependent on cell density. Describing the electrophysiological phenotype by any single parameter in isolation is likely to omit crucial information about how neuronal network activity is affected in pathological conditions, which could provide important clues about the underlying disease mechanisms. Thus, to adequately capture the disease phenotype of neuronal networks different quantitative parameters are needed (i.e. multiparametric approach). In particular, to use MEA parameters with high variability, such as the MFR, to describe the disease phenotype, one should include multiple supporting MEA parameters to describe the network activity characteristics [19].

Another important consideration for researchers involves standardization in MEA data reporting. In the present review, we found considerable variation, not only in the choice of the algorithms and analysis settings do the detection of spikes, bursts, and NBs, but also in the name, definition, and unit of measurement of the quantitative parameters extracted from MEA recordings. We believe that researchers should be sure to report a full explanation of these parameters and of the procedure to extract them. Moreover, raster plots of representative network activity should be included as a supplement to any quantitative metrics used to characterize the electrophysiological phenotype of neuronal networks under investigation. This would provide some more intuitive understanding of neuronal network activity, but also would allow other researchers to quickly assess the quality of raw data, ensuring there are no technical artifacts, inactive micro-electrodes or excessively noisy channels. The inclusion of representative raster plots would be particularly important in studies including phenotypic rescue assays. Indeed, the effectiveness of certain pharmacological interventions can easily be misrepresented by selectively reporting only those parameters which see substantial improvement with drug treatment, even if that improvement does little to change the qualitative electrophysiological phenotype observed in affected neuronal networks [389]. We believe that these recommendations are a step in the right direction towards ensuring high standards of scientific integrity and transparency in data reporting, but also accessibility of data for those who want to approach the field.

The electrophysiological activity as recorded by MEAs (i.e. network activity patterns, the features of spikes and bursts, the quantitative parameters extracted from raw data) is deeply shaped by the underlying physiology of neuronal networks. Thereby, MEA-based models can provide precious information about the cellular and molecular mechanisms occurring in neuronal networks in both physiological and pathophysiological conditions.

Neuronal cultures on MEAs, both isolated from rodents or differentiated from hiPSCs lines, represents a valid experimental model for investigating events occurring during the development of neuronal networks [97, 107, 117, 118, 128, 336]. During the development of neuronal tissue, the evolution of the electrophysiological activity can be related to neurodevelopmental processes, including synaptogenesis of both excitatory and inhibitory synapses, shaping of structural and functional connectivity, and changes in the expression profiles of receptors and ion channels during development. For instance, the emergence of network burst activity reflects the formation of synapses to form a functional connected network [97, 167]. Or else, the shortening of NBs duration in late differentiation stages of hiPSCs-derived neurons indicates the full maturation of GABAergic neurons [18].

In the same way, the electrophysiological activity of mature neuronal cultures on MEAs reflects the underlying characteristics and physiological mechanisms, such as the excitatory-inhibitory balance [18], structural and functional connectivity [97], synaptic plasticity phenomena (such as LTP and LTD) [170, 172–174, 338], neuromodulation of neuronal network activity [175, 187–190], and the expression of receptors and ion channels [168]. MEAs represent a valid experimental for investigating these physiological characteristics and mechanisms occurring in healthy neuronal networks, but also in affected ones, since all these aspects are frequently impaired in pathophysiological conditions.

In the context of disease modeling, MEAs are a powerful tool to perform phenotyping assays, aiming to characterize how the electrophysiological activity of neuronal networks is altered in specific pathophysiological conditions. However, their potential does not stop there. Once established, a well-defined disease model can give insights about the underpinning molecular and cellular mechanisms contributing to the pathophysiological phenotype, and, in some cases, revealing new candidates for treatment strategies. Indeed, by observing how the activity patterns, and the parameters describing the features of spikes and bursts, are altered in affected neuronal networks, information about the underlying disease mechanisms can be deducted. In figure 6, we provided some relevant examples taken from the literature, of neuronal network activity phenotypes observed on MEAs suggesting hypothesis about the molecular and cellular mechanisms underlying pathophysiology. For instance, high firing and bursting rates, or abnormal synchronous activity, characterize a hyperactive phenotype which can be related to different causes, such as the imbalance of excitatory and inhibitory synapses [238], the overactivation of signaling pathways involved in the regulation of synaptic function [346, 354] and neuronal network synchrony [158], or intrinsic characteristics of neurons [228, 230, 327, 353]. Conversely, a decrease in firing and bursting activity can be due to the degeneration of neuronal network functionality and/or connectivity [131, 148, 153, 162, 222, 224, 243]. Also, a change of the burst shape, in particular of the NBD can give precious hints about the underlying disease mechanisms. For instance, Mossink et al observed that the knockdown of CDH13, which have been associated with ASD, caused a significant reduction of the NBD, suggesting the involvement of CDH13 in maintaining the excitatory/inhibitory balance in neuronal networks [18]. Similarly, Frega et al, based on the observation of longer NBs in hiPSCs-derived neuronal networks differentiated from KS patients, postulated the involvement of NMDARs-mediated currents, which are known to directly influence the duration of bursts [155].

Zoom In Zoom Out Reset image size

These 'hints' can be easily obtained by characterizing the electrophysiological activity of neuronal networks in specific pathophysiological conditions, and can guide researchers in the further investigation of the underlying disease mechanisms by means of MEAs themselves and (i) other electrophysiology techniques, such as patch-clamp and calcium imaging, (ii) biological methods, including immunostainings and transcriptome analysis, (iii) pharmacological treatment (i.e. the use of inhibitors of specific receptors, ionic channels, and enzymes) and (iv) genetic manipulation. For instance, in the above-mentioned study from Frega et al, researchers combined recordings from MEAs with patch-clamp experiments, transcriptome analysis and pharmacological inhibition of NMDARs-mediated currents, in order to test their hypothesis [155]. Patch-clamp and transcriptome analysis showed increased NMDARs-mediated currents, whereas the pharmacological inhibition of NMDARs resulted in the rescue of the pathophysiological phenotype of KS patients-derived neuronal networks. This work not only demonstrated the validity of their hypothesis, but also suggested a new candidate target for treatment strategies (i.e. NMDARs-mediated currents) [155]. Besides this clear example of MEAs potential in the investigation of disease mechanisms, several other studies in which MEA devices were used to investigate the cellular and molecular mechanisms underpinning the pathophysiological phenotype of affected neuronal networks, can be found in 4.1.4 and 5.1.4.2. As previously mentioned, the complexity of neurological disorders, which are usually correlated to a high level of heterogeneity in the patient population, makes it difficult to clarify the pathophysiological mechanisms underlying the electrophysiological phenotype observed in affected neuronal networks on MEAs. However, we believe that MEA technology may represent a valuable, convenient and relatively easy tool to start with in the investigation of pathophysiology of complex neurological disorders.

In vitro neuronal models are simplified systems with the aim to provide insights about complex neurological mechanisms and diseases. Therefore, a key aspect is the possibility to translate findings obtained in vitro into in vivo animal models, and, finally, humans. In this regard several studies suggested the comparability of the results found in in vitro MEA models to observations in in vivo models and in humans.

Several in vitro models of neurological disorders on MEAs recapitulate the dysfunctional electrophysiological activity observed in vivo. For instance, models of epilepsy on MEAs, obtained through the treatment with pro-convulsant drugs or electrical stimulation, are characterized by hyperexcitability, increased activity and synchronicity, which are common features of epileptic phenotypes in vivo [158, 227, 230–233]. In one of these models, Hales et al also observed evoked high frequency oscillations, which have been described in both animal models and humans in association with epilepsy [225]. Similarly, MEA-based models of neurodegenerative diseases, such as AD and DLB, showed a progressive decrease in electrophysiological activity which has been related to the degeneration of neuronal network functionality and connectivity, and to the impairment of cognitive processing observed in vivo [141, 159, 222–224]. In an in vitro model of ischemic stroke, Pires Monteiro et al observed electrophysiological responses to transient hypoxia (i.e. suppression and subsequent restoration of network activity) on timescales similar to patients with cerebral ischemia [131]. The same was found in regard to the therapeutical time window: similarly to what observed in patients, the functionality of neuronal networks was restored only if normoxia was re-established within 24 h [131].

Neurotoxicity screenings on MEAs showed that neurotoxic compounds not only affected the electrophysiological activity of neuronal networks in a way that is qualitatively similar to what observed in vivo, but often showed a time- and concentration-dependent effect which was in line to reported studies in rodent models and humans [206, 207]. As an example, Xia and Gross found that ethanol induced a decrease of firing activity in neuronal cultures isolated from the frontal cortex of rodents at concentrations comparable to those estimated in blood to cause the typical behavioral effects in vivo [207]. Similarly, Croom et al characterized the neurotoxicity kinetics of lindane (i.e. a GABA_A receptor antagonist which causes seizures in vivo) in rodent neuronal cultures on MEAs, and demonstrated the predictivity of results in regard to doses and timing reported in the literature for human and rodents [206].

In addition, several drug studies supported the relationship between the sensitivity of neuronal networks cultured on MEAs to drugs and the expected effects in vivo. For instance, Colombi et al and Que et al investigated the pharmacological effect of commonly used antiepileptic drugs in two different in vitro models of epilepsy on MEAs [227, 353]. Colombi et al studied the effect of carbamazepine and valproate in rodent neuronal networks treated with the pro-convulsant BIC, and found that both were effective in ameliorating the epileptic phenotype at concentrations close to those clinically relevant [227]. Que et al used hiPSCs-derived neuronal networks carrying a genetic variant that in vivo causes seizures. Besides increased excitability, they observed a degree of resistance to the anti-convulsant drug phenytoin, which recapitulated aspects of clinical observation of patients carrying the same genetic variant [353]. Similarly, van Hugte et al tested four different anti-seizure drugs in hiPSCs-derived neuronal networks from patient with two different types of epilepsy, and found that in vitro responses to different drugs exactly corresponded to those observed in patients [355]. All these studies suggest that hiPSCs-derived neuronal networks from patients cultured on MEAs might be used to develop patient-specific in vitro platform to test which drugs are more likely to be effective in the specific patient, paving the way toward personalized medicine.

Several studies demonstrated that 3D neuronal cultures on MEAs could better represent in vivo cellular behavior than cells cultured in monolayers. For instance, 3D rodent neuronal networks showed a repertoire of complex and asynchronous global activity patterns, resulting from the higher complexity of the network, which seemed to better recapitulate activity dynamics of in vivo brain regions [22, 274, 388]. In this regard, a particularly relevant example is set by Trujillo et al which developed a human cortical organoid on MEAs in order to investigate the development of neuronal network activity over the span of several months [333]. Interestingly, they observed similarities in the developmental trajectory of some electrophysiological features between in vitro organoids and human preterm infants, supporting the validity of hiPSCs-derived brain organoids on MEAs as a model of the human brain [333].

With the advent of hiPSCs and the possibility to derive human neuronal networks from patients, a highly heterogeneous variety of novel signatures have been observed in MEA recordings. This leads to new challenges in the MEAs field, in particular related to the identification of the most relevant parameters to describe the observed neuronal network phenotype.

As previously mentioned, multiparametric approaches should be used to describe the phenotype of affected neuronal networks, in order to avoid omission of crucial information and misinterpretation of results. However, they can present their own issues when it comes to draw insights about the underlying pathophysiological mechanisms, or to evaluate and compare the effect of different drugs and treatments strategies in phenotypic rescue experiments. In these cases, it is crucial to understand which quantitative parameters are the most physiologically relevant to the neurological disorders under investigation, which is not always trivial. Nowadays, in most studies the choice of the meaningful parameters to be extracted from MEA raw data is dependent on the visual observation of raster plots by researchers performing the analysis. Also, depending on the disease of investigation, the set of meaningful parameters can change. Undoubtedly, this represents a time-consuming procedure, inevitably affected by human error, subjectivity, intra and inter-observer variability. For this reason, we believe that advanced statistical and computational methods might provide a promising approach for an in-depth and unbiased analysis of MEA recordings.

A clear example is set by Mossink and colleagues [19]. In their study, they extracted 17 parameters in total to describe the neuronal network activity and connectivity of hiPSCs-derived neuronal networks from healthy subjects, KS patients and MELAS patients [19]. All the parameters were combined in a principal-component analysis, a statistical technique for the analysis of large datasets containing a high number of dimensions per observation, enabling the visualization of multi-dimensional data in intuitive plots. By performing this analysis, they observed that neuronal networks from KS and MELAs patients clustered separately from controls, and were able to identify the parameters that explained the differences in the electrophysiological behavior of control and affected cultures, and to calculate the percentage of variance explained by each of these parameters [19].

Another approach which might be adapted to the analysis of MEA recordings is artificial intelligence (AI), including traditional machine learning and deep learning. These methods are widely used in image recognition. Briefly, a typical machine learning workflow for image recognition consists in the preparation of a training dataset of images associated to categories. The training dataset is presented to the model, which identify and extract the key features that can be used to differentiate the images of different categories. In this way, the model 'learns' to analyze and classify new images without being explicitly programmed to do so. Machine learning approaches have been used, for instance, for the recognition of interictal epileptiform discharges (i.e. epileptiform transients indicating an increased likelihood of seizures) in EEG recordings, providing highly accurate and robust results [390, 391]. Similarly, machine learning might be adopted for the analysis of MEA recordings, for instance, for the classification of raster plots from healthy or affected neuronal networks, or for the classification of drugs based on their effectiveness in rescuing the phenotype of affected neuronal networks. In these cases, machine learning approaches could be used to analyze and classify raster plots obtained from MEA raw data, as they were images. Raster plots, indeed, contain all the information regarding firing, bursting and network bursting activity of neuronal networks. Conventional analysis of MEA recordings is based on pre-established methods to detect spikes, burst and NBs, whose settings need to be fine-tuned to properly describe the phenotype of the neuronal network under investigation. Similarly, the quantitative parameters to be extracted from raw data are decided by researchers conducting the study on the basis of visual observation of raster plots. In addition to variability resulting from human error and subjectivity, this approach inevitably introduces a limitation in terms of number of parameters which can be extracted and further analyzed, for instance, to identify differences between healthy and affected neuronal networks. For all these reasons, analysis of MEA recordings and conventional analysis methods and parameters might fail in the characterization of the electrophysiological phenotype observed in raster plots, especially for newly established disease models. Therefore, AI might assist researchers in the extraction of meaningful features from raster plots, giving the possibility to capture changes in neuronal network activity without choosing settings to detect spikes, bursts and NBs, or pre-established parameters.

Recently, Zhao et al developed a deep learning model able to split long MEA recordings into smaller slices and classify them according to the genotype of neuronal cultures (i.e. rodent wild-type versus delta-catenin knock-out neuronal networks, and hiPSCs-derived control versus Williams syndrome neuronal networks) [392]. Zhao et al reported that their model succeeded in classifying MEA recordings with a good accuracy, and could be easily adapted to the analysis of raw data from MEAs with a larger number of probes, such as HD-MEAs [392]. Similarly, Matsuda et al proposed an analysis method using deep learning to extract 4096-dimensional features from raster plot images [393]. By using MEA recordings from hiPSCs-derived neuronal networks grown on MEAs, they showed that seizure-causing compounds and seizure-free compounds were distinguished by using machine learning algorithms trained on raster plots. With this model, features that were not detected using single-parameters analysis were identified, and detailed differences between compounds were detected even when their main effect was the same [393].

We believe that approaches as the ones presented here hold promise for the future of in vitro MEA-based model for biomedical research, helping researchers to identify the phenotypic signatures of neuronal networks under investigation, regardless of visual observation of raster plots and a priori choice of analysis settings and parameters.

The possibility to obtain 'hints' from MEA recordings that can guide researchers in the investigation of the pathophysiology of affected neuronal networks is impressive. However, the identification of cellular and molecular mechanisms underlying the neuronal network phenotype observed on MEAs remains challenging. Conventional approaches are defined as hypothesis-driven, since they aim to test the contribution of candidate mechanisms to the neuronal network phenotype. They include, for instance, the external perturbation of the network with electrical stimulation, drug administration or genetic manipulation, or the integration of MEA recordings with results from other techniques, such as patch clamp, calcium imaging and biological methods. However, these approaches can be very time-consuming and expensive, and need a priori hypothesis to test.

In silico models of neuronal networks can complement and assist in vitro experiments and observations of MEA recordings, representing a powerful tool to facilitate the identification of the cellular and molecular causes of the observed phenotype. With in silico models, the experimental setup can be recreated and simulations can be run to validate or reject a large number of hypotheses. In addition, in silico models allow to extend experimental setups and to include a virtually unlimited array of parameters, which might not be easy to setup experimentally [394].

In silico models of neuronal networks consist of a system of equations able to explain biological, chemical and physical phenomena occurring at the network level. The study of computational models started in 1952 with the pioneering work of Hodgkin and Huxley, who published five papers describing a series of experiments and an empirical model of AP in a squid giant axon [92, 395]. The so-called Hodgkin-Huxley model consists of a system of four coupled nonlinear ordinary differential equations describing the generation of APs, and it represents the starting point of many models developed in neuroscience [396]. An important result of the Hodgkin–Huxley studies is that neurons are dynamical systems, and, as such, they consist of a set of variables that describe their state and a law that describes the evolution of the state variables with time [396]. Furthermore, in silico models that are based on artificial neural networks have been produced via simulation of real neural networks according to Hebbian theory [397].

Researchers developing and testing computational models often face the challenge of finding the sensitive trade-off between computational time and model accuracy. One example is given by the model introduced by Izhikevich in 2003 [398]. Izhikevich explained that the model for a single neuron must be both computationally simple, yet capable of producing rich firing patterns exhibited by real biological neurons [398]. In this sense, using biophysically accurate Hodgkin–Huxley type models is computationally prohibitive, since only a handful of neurons can be simulated in real time. Conversely, using an integrate-and-fire model is computationally effective, but it is unrealistically simple and incapable of producing rich spiking and bursting dynamics exhibited by cortical neurons. The model presented by Izhikevich is the fair compromise within the two bounds, it depends on four parameters, and reproduces spiking and bursting behavior of known types of cortical neurons [398].

The explosive increase in computational power and resources of recent times has pushed the implementation of in silico models of neuronal network activity of cultured neurons in several applications [101, 399–407], but not much literature has been produced yet on in silico models of neuronal networks grown on MEAs. As an example, Dazza et al used in silico models to describe how periodic bursts are generated and propagated in a spatial network in various neuronal systems with a novel phase-based analysis [408]. Even if presented with simulations, they showed that this methodology can be applied to other forms of neuronal spatiotemporal data (e.g. MEA recordings with sufficiently high spatiotemporal resolution). Simulations were carried out with the adaptive Exponential Integrate and Fire model [409], via the NNGT python library [410] and NEST simulator [411]. In silico models were also used in combination with in vitro MEA recordings to study the neuronal network phenotype of hiPSCs-derived excitatory neurons carrying Rett syndrome-associated MECP2 mutations [345]. Mok et al recreated the electrophysiological phenotype in a computational neuronal network model and showed that the intrinsic adaptation currents in individual neurons explained the aberrant activity [345]. Within the same scenario, in silico neuronal networks predicting the changes in neuronal spiking activity expected to result from isolated rescue of synaptic structure have been generated [385].

Most of the studies using in silico models are based on hiPSCs-derived cardiomyocytes cultured on MEAs [412–416]. Even if not directly related with our field, they can teach us new methods and approaches adaptable to neuroscience research. In Shi et al, for instance, the applicability of an in vitro/in silico approach was investigated to predict human cardiotoxicity of candidate anti-addiction drugs [412]. In particular, they combined in vitro parameters extracted from MEA recordings from hiPSCs-derived cardiomyocytes with physiologically based kinetic in silico models [412]. In another study related to safety pharmacology in heart research, a strategy to analyze the signals acquired from hiPSCs-derived cardiomyocytes on MEAs was proposed with the aim to automatically deduce the channels affected by tested drugs [413].

To conclude, we believe that one way to optimize the identification the mechanisms underlying the neuronal network phenotype observed in MEA recordings is to integrate in vitro MEA recordings and experiments with in silico models. In silico models, indeed, allow to test a large number of hypotheses suggested by the analysis of MEA recordings, and to predict candidate mechanisms that can be further investigated with in vitro experiments. This approach holds the potential to make research less time-consuming and expensive.

To date, there is a big gap between research and clinical practice. It is indeed very challenging to relate the signals recorded in vitro to the ones recorded from the brain, and this contribute to make challenging translation of results from in vitro experiments into humans.

Electro- and magnetoencephalography (EMEG) recordings are non-invasive methods to study electrical neuronal activity from the brain with fine temporal resolution [417]. While EMEG are intensively used in the clinic to identify biomarkers of normal and abnormal brain dynamics, these so-called 'macro-scale' techniques suffer from difficulty in interpretability in terms of the underlying cellular and molecular level events [417]. Vice versa, in vitro neuronal models, such as neuronal networks on MEAs, are used in research to investigate the pathophysiology of neurological disorders at cellular and molecular levels, and to test candidate drugs and treatment strategies. Little is known about the correlation between the neuronal signals recorded by MEAs in vitro and by EMEG in vivo, since micro-scale phenotypes observed in MEA recordings are hard to relate to macro-scale events detected with EMEG. Thus, there is a huge need of bridging macro- and micro-scales in research of brain physiology and disease.

To address this, several attempts have been proposed in the literature. Hagen et al, for example, developed an open-source software (i.e. LFPy2.0, a more recent upgrade of LFPy [418]) able to build multimodal modeling of neuronal network activity [419]. LFPy2.0 allows for modeling networks of multicompartment neurons with concurrent calculations of extracellular potentials and current dipole moments. The current dipole moments are then, in combination with suitable volume-conductor head models, used to compute EMEG-like signals [419]. In LFPy and LFPy2.0, micro- and macro-scale models are combined together, making use of different toolboxes: NEURON [420], to compute transmembrane currents of multicompartment neurons, and electrostatic forward models. For the latter, volume conduction models with four compartments with different electrical conductivities (i.e. scalp, skull, CSF and brain) are considered. The level of complexity of these anatomical models can be potentially increased and the accuracy of such simulations improved both for EEG and MEG [421–424].

Similarly, the virtual brain is a neuroinformatic platform for full brain network simulations using biologically realistic connectivity [425]. The micro- and meso-scale model consists of neuronal mass models or assembles of neuronal mass models, generating the so-called mass field. This simulation environment enables the model-based inference of neurophysiological mechanisms across different brain scales that underlie the generation of macroscopic neuroimaging signals including EEG and MEG [426–429]. The Virtual Brain has been used for many applications to investigate different diseases, such as neurodegenerative disorders [430], PD [431], dementia [432], TBI [433], AD [434–436], epilepsy [437–440], brain tumor [441], chronic stroke [442], or physiological brain dynamics [431, 443, 444]. Nevertheless, to our knowledge, there is no literature about the explicit integration of MEAs models in the computation.

Another example of multi-scale model platform was presented in the work of Neymotin et al [417]. In their study, they developed the human neocortical neurosolver (HNN), a modeling tool designed to provide researchers and clinicians an easy-to-use software platform to develop and test hypothesis regarding the neuronal origin of their data [417]. At the microscopic level, a neocortical model explaining the underlying cellular- and network-level activity was considered as the source that generates the macroscopic signal recorded on the scalp (i.e. EEG), or far away from it (i.e. MEG) [417]. Also in this scenario, realistic models of MEAs were not considered.

From our perspective, relating in vivo EMEG measurements of the brain to in vitro recordings from neuronal cultures on MEAs would boost not only pharmacological investigations and personalized treatment, but also basic research in neuroscience as a whole. To the best of our knowledge, such models have not yet been developed, but represent a valid future perspective for improve translation of in vitro findings into clinic.