To achieve therapeutic interventions for stroke beyond the state-of-the-art we need to consider stem cell therapy in the context of the extremely complex, highly interconnected structural and functional circuitry of the human brain, especially the altered, highly plastic, structure-function relationships triggered by an ischemic lesion. Endogenous neuroplasticity after stroke and its relevance for the functional integration of morphogenetically engineered transplants to elicit recovery of motor, sensory, or cognitive deficits as a result of brain ischemia, can be briefly discussed in the context of Hebbian and homeostatic mechanisms.

During development, the CNS is characterized by a high level of plasticity, where functional brain networks and circuits are shaped by Hebbian mechanisms, such as long-term potentiation (LTP) and long-term depression (LTD) (61, 62) constituting the synaptic basis of associative learning (63). These processes are regulated by evolutionarily conserved mechanisms and involve functional and structural changes which occur in a well-orchestrated spatiotemporal manner. However, once development has been completed, Hebbian mechanisms alone cannot explain how the mature brain maintains normal function. In the same manner, Hebbian laws per se cannot sufficiently explain plastic network responses, such as synaptic scaling (64, 65), which are observed following externally induced perturbations. This suggested that there may be another form of plasticity underlying such responses. Compelling experimental evidence over the last two decades supports that another form of plasticity, i.e., homeostatic plasticity, does exist (64, 66–69). Indeed, homeostatic plasticity operates in tandem with Hebbian plasticity along complex temporal and spatial scales and serves to maintain normal neural network function as well as regulate network responses to perturbation (67, 70).

It is well-documented that any major disturbance to normal neural network function tends to result in severe, permanent functional deficits, but also trigger adaptive and maladaptive network responses, which are demonstrated as changes in network structure and function (12, 18, 19, 66, 69–74). Such alterations may represent various underlying forms of neuroplasticity, which involve both homeostatic and Hebbian mechanisms (67, 75, 76), and which can be described as morphology-activity, (i.e., structure-function) relationships. Such relationships are highly complex and, as a result, poorly understood. However, elucidation of how different forms of plasticity may result in adaptive or maladaptive neural network responses and, thereby, how such processes may promote or hinder functional restoration after a perturbation such as stroke is of fundamental importance for our ability to achieve functional repair.

As mentioned earlier, plastic responses to an ischemic event include spontaneous functional recovery (77), rewiring of surviving neuronal networks and axonal ramification, and the recruitment of intact synapses post-lesioning (78), demonstrated as intra- and interhemispheric neural network remodeling (74). Furthermore, there is increasing consensus as to the likelihood that such responses are also influenced by evolving inflammatory processes (12). The underlying mechanisms regulating such responses largely involve structural as well as functional changes in brain circuits closely associated with the ones directly affected by stroke. These circuits can further be engaged and modified through experience-dependent plasticity (17, 79).

It is a reasonable assumption to make that there is a critical period of increased neuroplasticity after stroke, which can be harnessed to promote functional recovery, as exemplified in a number of preclinical as well as clinical studies (12, 17, 77, 80–84). Moreover, recent evidence suggests that recovery after stroke can be mediated through the engagement of neural networks in a functional hierarchy both in the contralesional as well as ipsilesional brain hemisphere and also upstream as well as downstream of the lesion (85). Furthermore, it is becoming increasingly evident that any cortical remapping that may occur after stroke is both activity dependent, but also based on competition (86).

Where does cell replacement therapy fit in this context? To seek answers to this question, it is important to consider the behavior of engineered neural transplants in the ischemic brain in terms of connectomes (87, 88) i.e., the changing structure-function relationships and complex interplay between affected and intact neural ensembles in the brain in the context of evolving stroke pathology.

Highly relevant insights as to how the brain connectomes underpin the behavior of a neural system and determine functional outcome after a perturbation can be found in an excellent review by Fornito and co-authors (89). In this paper, the authors discuss how the highly interconnected, complex topology of the neural architecture of the brain, which is characterized by a fine, coordinated balance between regional segregation and specialization of function, but also highly precise functional segregation and integration of distal neuronal ensembles across multiple temporal scales, can shape the brain's response to perturbations (89–93). Specifically, brain network topology can determine not only disease progression, but also adaptive or maladaptive responses to pathological perturbations. By applying fundamental principles from graph theory, it is possible to extract key functional and structural elements of the brain's connectome that may explain or predict such responses (94).

According to graph theory, a neural assembly consists of key elements, namely nodes and edges, and as such, metrics that define properties of the graph can be applied to make inferences about the neural connectome, for example, as it is revealed by electrophysiological activity (92, 94–96). Nodes incorporate neurons, neuronal populations, and macroscopic brain regions, while edges determine structural, functional, or effective connectivity and, as a result of the latter, the directionality of internodal neural communication. In this manner, structural connectivity can be indicative of morphology-activity relationships between neurons, whereas functional connectivity, which is independent of the physical proximity of the neurons, provides a temporal correlation of their activity. On the other hand, effective connectivity can determine whether there is a causal relationship between the activity of different neurons (94). Topological properties of particular interest include economical wiring, hierarchical modularity, small-world organization enabling functional segregation, integration and specialization through high network clustering and characteristic short path lengths (93, 95–100), as well as highly-connected network hubs. Connections between hubs can be describes as rich club, i.e., central core connections, if they extend over large anatomical distances and link distinct functional systems (101). As such, rich club connections mediate high-volume network traffic and integrated brain function.

As a result of high interconnectivity of neural networks and circuits, the damage induced by lesions such as focal cerebral ischemia dissipates in the brain and progressively spreads to affect distal loci (81). This diaschisis phenomenon constitutes a maladaptive response and may explain the deafferentiation and aberrant synchronization observed in brain regions remote from the one directly affected by the stroke (74, 89, 92). Excitotoxicity thus plays a central part in the damage sustained to these remote nodes after a stroke. Associated maladaptive responses include anterograde and retrograde neuronal degeneration, for example, as a result of neuronal loss, excitotoxicity and impaired axonal transport, but also dedifferentiation. Apart from loss of tissue at the lesion site, neuronal degeneration involves structural deterioration over time of nodes connected to the primary lesion site, while dedifferentiation refers to the gradual loss of specialized function in the affected brain region, followed by increased, non-specialized activity in associated nodes (89, 92) (Figure 1A).

On the other hand, adaptive network responses include lateralization, i.e., compensation for lost function by increased structure-function interactions and the engagement of intact neuronal ensembles within the affected system, or within other systems. Other adaptive responses may include neural reserve, a process by which remaining intact tissue in the immediately affected region can sustain the previous level of behavior, and also degeneracy, where a second system, without demonstrating apparent changes, can adequately support the activity of the affected one (89, 102). Thus, the effects of a perturbation critically depend on the topological centrality and degeneracy of the affected region, as pathology of central regions can exacerbate maladaptive responses, whereas degeneracy can facilitate adaptive ones (89, 95, 96). By the same token, spontaneous or induced gain of function after stroke may imply partial restoration of neural networks previously involved in impaired function or, alternatively, a compensatory or substitution mechanism, which is contingent on the recruitment of other networks (74) (Figure 1B).

However, in any perturbed system, not least, the ischemic brain, which is characterized by highly complex pathophysiological mechanisms and associated cellular and molecular cascades, such responses are neither unequivocal, nor mutually exclusive. Furthermore, as mentioned earlier, the resilience of the perturbed system and the manner in which it may reconfigure its structure-function relationships in response to the lesion largely depends on the topological substrates being affected. It follows that a neural transplant, effectively, a new topological element (likely, a node) within the local circuitry, becomes another integral component in the system's connectome. As such, the input and output functions of the transplant node are shaped by other key elements of the connectome, as for example, its interactions with other nodes, in an interdependent, reciprocal manner.

What are the exact processes that determine whether and/or to what extent neurons engineered in vitro (and by the same token, in vivo), have the capacity to become part of the local connectome in a manner that may promote gain of function after stroke?

Detailed study and elucidation of relevant mechanisms in stroke patients is not feasible and is thus de facto limited to methods and analyses based on indirect and often incomplete observations or measures. Relevant research must therefore rely on robust preclinical models. Notwithstanding various excellent studies employing cell replacement strategies, coupled with cutting-edge methods for monitoring and assessment of functional integration of the engineered neurons in experimental stroke animals (16, 60), there is currently no unequivocal evidence supporting that these neurons become fully integrated into the ischemic brain. As a result, pertinent questions remain unanswered, especially with regard to the intrinsic capacity of engineered neurons to engage with local circuitry, whether we can predict their responses to evolving pathology, as well as determine whether their morphological and functional integration into the host brain will elicit adaptive or maladaptive responses and, ultimately, promote repair or inadvertently exacerbate the lesion-induced deficits. We may start addressing such questions by combining in vitro morphogenetic neuroengineering with electrophysiology studies and computational modeling.

State-of-the-art platforms for the study of in vitro neural networks include microelectrode arrays (MEAs). MEAs represent a breakthrough in in vitro electrophysiology platforms, as they enable direct observation of evolving network dynamics through extracellular recordings, as well as disease modeling and modulation of network activity by electrical, chemical, opto- or chemogenetic manipulation to selectively perturb network function or engage its plasticity (94, 103–111).

This approach largely exploits intrinsic attributes of in vitro neural networks, including emergence/morphogenesis, as a result of cell behaviors driven by local cell-cell interactions, and self-organized criticality (SoC) (112). A universal characteristic of in vitro cultured neurons, irrespective of source (i.e., rodent or human) or specific subtype, is that they self-organize into functional complex networks. This spontaneous process, which occurs in the absence of any external instructive influence (113), involves the network's gradual transition through distinct morphology-activity states underscored by increasing structural and functional complexity as a result of emerging neuronal activity, axonal elongation and pathfinding, synapse formation and pruning (113, 114). Over time, the network gains morphological and functional maturity, ultimately reaching SoC, a critical state characterized by a fine balance between excitation and inhibition (113, 115–118). Once at the SoC state, the neural network exhibits neuronal avalanches, i.e., characteristic cascades of activity defined by rich, stable spatiotemporal patterns and power-law distributions of variables (114, 119). Criticality has been closely, yet not exclusively, associated with self-organization (112, 120–124).

How does SoC displayed by in vitro neural networks relate to normal or impaired brain function? The idea that the brain operates in a near-critical state is not new (119, 125, 126), yet the presence of SoC states in the brain has, to a certain extent, been the subject of debate until relatively recently (119, 127–129). However, emerging evidence, including findings from magnetoencephalography (MEG) studies in human subjects, strongly supports the presence of SoC in the brain, not only as a mechanism underlying spatially local or internodal measures of brain activity, for example, as relevant for cortical function, but also as a global phenomenon (123, 130).

This suggests that within the brain connectome, SoC states play a putative role in information transmission and processing, including segregation and integration of function along different spatiotemporal scales (114, 123, 124, 131). As such, critical states in the brain emerge from the complex underlying morphology-activity relationships shaping the system's behavior and are stipulated to underscore adaptation processes in the context of neuroplasticity. It is currently unknown how changes in network organization over different spatiotemporal scales, for example, as a result of ischemic insult, may affect such states (114).

In vitro biological substrates that combine morphogenetically engineered neurons with MEAs can be used to obtain insights into such mechanisms. For example, neural avalanches can be observed and recorded as synchronous local field potential (LFP) spikes across multiple electrode locations. Dependent on the number of embedded recording electrodes available on the MEA, which may range from a few dozen (typically 60–120 electrodes) to several thousand, the observed LFPs may arise from neuronal clusters, effectively individual nodes within the network, but also represent single-cell activity, as in the case of very high-density (>11,000 electrodes) complementary metal-oxide semiconductor (CMOS) integrated circuit MEAs (118). This effectively means that the derived neural substrates can faithfully recapitulate fundamental aspects of neural network behavior, thus enabling the study of complex network dynamics in reductionist in vitro paradigms. Maintenance of the relevant MEA-based neural networks in optimal culture conditions enables long-term monitoring of network behavior up to several weeks or months (132) (Figure 2).

Notwithstanding the tremendous potential afforded by these advanced in vitro platforms for studying engineered neural networks in the context of stroke modeling, it must be emphasized that MEA-based electrophysiology tends to generate very high volumes of data (typically, a standard 10 min recording from a 60-electrode MEA can generate 0.5 GB of data; indicatively, the corresponding data volume increases to 1,5 GB when using a 4,096-electrode MEA). As a result, the study of connectomes, which requires extraction of functional, structural and effective connectivity and, by the same token, inference about the underlying neural network state, or prediction about its dynamic responses to stroke-related pathology, including anticipated behavior post-transplantation in vivo, cannot be achieved with standard software analysis methods. Advanced computational modeling applying deep learning principles is thus necessary.

As mentioned earlier, connectomics analyses may apply graph theory and relevant theoretical models to determine underlying structure-function relationships within a neural network and infer its critical states (94, 109, 116, 117, 172). For example, advanced computational modeling may apply principles from cellular automata (CAs) or random Boolean networks (RBNs), which examine the relationships between constituent key elements of a system based, respectively, on proximal or random connections (177). Furthermore, machine learning or deep learning approaches, such as artificial neural networks (ANNs) and artificial recurrent neural networks (RNNs), are highly relevant computational models for MEA data analysis. Such models take inspiration from biological neural networks (BNNs). Briefly, in these models the neurons, i.e., the computing elements of the system, are arranged in layers through which information flows in a feed-forward manner. Interestingly, temporal dynamics of the systems can be captured by RNNs by assigning recurrencies (cycles) within the connection topology of the neurons in the system. As a result, the system possesses a memory of previous inputs; in other words, the activation state of the network at a given time represents a function of its previous activation states. As such, an RNN can be seen as an untrained reservoir of dynamics that operates in a near-chaotic state. This enables control of the system by selecting only a single linear redout layer for training (178).

RNNs are thus well-suited to capturing dynamic network behavior of in vitro neural networks, including their critical states. These attributes of RNNs make them an excellent computational paradigm for the investigation of mechanisms by which neuronal populations, including engineered neural networks, solve various computational problems in healthy and perturbed conditions (178–182). In this way, we can start addressing fundamental questions that can help determine the functional capacity of engineered neurons. Moreover, such models can be applied to decipher as well as predict the behavior of engineered neurons in response to stroke-related pathology in vitro and also post-transplantation in vivo. Importantly, the relevant computational models can be applied to make inferences about underlying critical states, including spontaneous and induced plasticity of these networks.

Finally, closed-loop hybrid systems are highly relevant in this context as they can be used for advanced simulations of spontaneous and induced plasticity, including bi-directional learning. For example, by interfacing an engineered neural network on an MEA with a simulated agent (i.e., computer hardware) we could elicit real-time agent behavior powered by the neural network using reservoir computing (183). Furthermore, by selecting a pacemaker node within the engineered neural network as the target for electrical stimulation, we were able to alter the pacemaker activity in a manner suggestive of LTP, and at the same time demonstrate learning behavior in the simulated agent (183). Other reported studies have also demonstrated the capacity of embodied neural network cultures for goal-directed behavior and learning (184–186). Such approaches are highly promising for the future development and application of relevant experimental paradigms for in vitro and in silico assessment of the endogenous as well as induced plasticity of engineered neural network, as for example, in terms of guided closed-loop neuromodulation approaches post-transplantation.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.