Recent progress in translational engineered in vitro models of the central nervous system

Advances in the complexity and functionality of in vitro models of the CNS mean that such models now have the potential to overcome many of the limitations of traditional model systems. Nikolakopoulou et al. provide an overview and user guide for in vitro CNS models, with a focus on ‘Organ-on-a-Chip’ technologies.

Abstract

The complexity of the human brain poses a substantial challenge for the development of models of the CNS. Current animal models lack many essential human characteristics (in addition to raising operational challenges and ethical concerns), and conventional in vitro models, in turn, are limited in their capacity to provide information regarding many functional and systemic responses. Indeed, these challenges may underlie the notoriously low success rates of CNS drug development efforts. During the past 5 years, there has been a leap in the complexity and functionality of in vitro systems of the CNS, which have the potential to overcome many of the limitations of traditional model systems. The availability of human-derived induced pluripotent stem cell technology has further increased the translational potential of these systems. Yet, the adoption of state-of-the-art in vitro platforms within the CNS research community is limited. This may be attributable to the high costs or the immaturity of the systems. Nevertheless, the costs of fabrication have decreased, and there are tremendous ongoing efforts to improve the quality of cell differentiation. Herein, we aim to raise awareness of the capabilities and accessibility of advanced in vitro CNS technologies. We provide an overview of some of the main recent developments (since 2015) in in vitro CNS models. In particular, we focus on engineered in vitro models based on cell culture systems combined with microfluidic platforms (e.g. ‘organ-on-a-chip’ systems). We delve into the fundamental principles underlying these systems and review several applications of these platforms for the study of the CNS in health and disease. Our discussion further addresses the challenges that hinder the implementation of advanced in vitro platforms in personalized medicine or in large-scale industrial settings, and outlines the existing differentiation protocols and industrial cell sources. We conclude by providing practical guidelines for laboratories that are considering adopting organ-on-a-chip technologies.

Introduction

The human brain has a complex physiology that distinguishes it from other animal brains. Features that are unique to the human brain include its large size, its highly folded cortex, and the unique pathophysiology of human neural diseases. The hierarchical complexity of the human brain (Supplementary Fig. 1A) poses a substantial challenge for the development of models for research of CNS pathologies: Animal models lack many essential human characteristics and are associated with many specific interspecies differences; in addition, they are expensive, their throughput is low, and they raise ethical concerns. Conventional in vitro models, in turn, cannot provide information regarding behavioural responses, many functional responses, or systemic responses (organ–organ interactions), and are therefore considered to be too simplistic for many practical applications. The lack of adequate CNS models is one explanation for the low success rate of CNS drug development (Kesselheim et al., 2015; Gribkoff and Kaczmarek, 2017); this low success rate has led many major pharmaceutical companies to limit their research and development investments in the neurological domain (Wegener and Rujescu, 2013).

The lack of adequate CNS models has spurred academic and industrial researchers to seek out new technologies for mimicking brain physiology and functionality in health and disease, by using tools such as induced pluripotent stem cells (iPSCs) (reviewed in Shi et al., 2017), organ-on-a-chip (OoC) systems, organoids (Pașca, 2018), 3D printed gels (Hopkins et al., 2015) and neuronal machine interfaces (Moxon and Foffani, 2015). Though none of these methods can fully capture the complex physiology, anatomy and functionality of the human brain (or of any other whole organ, for that matter), they are nonetheless showing very promising results in terms of their capacity to recapitulate certain human functions or pathological mechanisms, as well as to reveal new physiological interactions that could not have been identified with current standard tools in vitro or in vivo. To give the reader a quick overview of different models of the CNS we have compared the use of rodent models versus conventional cultures, organoids and OoCs, rating each model's usability for specific CNS studies (Table 1). Advanced in vitro platforms are rapidly becoming more accessible in terms of cost, ease of use and availability. So far, however, use of these tools has tended to be limited to the laboratories in which the technologies were developed, suggesting that their adoption by members of the wider CNS research community has lagged behind. This gap suggests a need to raise CNS researchers’ awareness of the variety of novel CNS models that are currently available and that might serve to complement traditional in vitro and in vivo models, thereby enhancing the overall accuracy of preclinical evaluations aimed at predicting clinical outcomes.

Table 1.

Overviewing comparison of rodent in vivo models

	Human relevance	Disease models	Systemic effects	Brain regions	Behaviour	Electrophysiology	Mechanistic studies	ADME/ TOX	HTS	Cost
Standard 2D cultures
Human primary	+++	++	–	++	–	++	++	+	+	++
Human iPSC	+++	+++	–	++	–	++	+++	+	+++	+++
Rodent primary	–	++	–	+++	–	+++	+++	+	+	++
Cell lines	+	+	–	–	–	+	+++	+	+++	++
Organoids
Human primary	+++	+++	–	++	–	–	++	–	–	++
Human iPSC	+++	+++	–	+	–	+	++	+	++	++
Rodent primary	–	+	–	+++	–	+	++	–	–	++
Cell lines	+	+	–	–	–	+	++	–	++	+
OoC
Human primary	+++	++	++	++	–	++	++	++	–	++
Human iPSC	+++	+++	++	++	–	++	+++	++	–	++
Rodent primary	–	++	++	+++	–	+++	+++	+	–	+
Cell lines	+	+	++	–	–	+	+++	++	–	+
Rodent in vivo	–	++	+++	+++	+++	++	+	++	–	+++

Table shows rodent in vivo models (the most commonly used mammal), standard 2D cell culture models, organoid cultures and OoC for their human specificity and their capacity to model human diseases, systemic effects, brain regionality, behaviour, drug absorption, distribution, metabolism and excretion, and toxicity (ADME/TOX). We also rate the possibility for electrophysiological studies, detailed mechanistic studies, high throughput studies (HTS), and the cost of the model. For the three in vitro models, we divided them into the accessible cell sources, human primary cells, rodent primary cell and hiPCS, and cell lines. Notably, we want to emphasize that human primary cells from the CNS are scarce. We further wish to highlight that this rating, the appropriateness of each model, varies for each specific study, and our rating should be used as a general guideline of what is possible to achieve with each model. – = poor/non-existent; + = OK; ++ = good; +++ = excellent.

Accordingly, in this review we will report on recent developments of in vitro CNS models, and will discuss how these models can address some of the challenges associated with current in vivo models. We focus our discussion on engineered in vitro models, i.e. models that involve a technical approach to control the organization of cells, and briefly mention self-assembled structures such as spheres and organoids. We begin by providing an overview of some of the main recent developments in in vitro CNS models. Next, given that a key goal of advanced in vitro platforms is to enhance the translatability of experimental results—which implies reliance on human tissue sources and, specifically, human-derived iPSCs (hiPSCs)—we discuss some of the challenges associated with integration of hiPSCs into these platforms. We then provide a detailed explanation of some of the fundamental principles underlying engineered in vitro models of the healthy CNS and discuss several applications of these models in practice. We subsequently discuss the most relevant models of CNS disease. We conclude by providing practical guidelines for biomedical labs that are considering adopting these technologies and give an outlook for future developments.

Overview of recent developments in in vitro CNS models

The CNS is an intricate cellular network, consisting of neurons, astrocytes, oligodendrocytes, pericytes, immune cells, and vascular endothelium embedded in a tissue-specific extracellular microenvironment (Rauti et al., 2019). Understanding the physiological cellular mechanisms of the CNS is essential for identifying potential drug targets, as well as for predicting drug side effects and the pathogenesis of neurological diseases.

In addition to the distinct cellular subtypes, organization and interconnectivity, it is imperative to recapitulate the extracellular milieu when generating translatable in vitro CNS models. Hence, all the physical, chemical, and mechanical cues of the extracellular matrix (ECM) should be considered (Frantz et al., 2010; Abdeen et al., 2016; Uwamori et al., 2017). The brain ECM has a distinct structure and composition that differentiates it from other organs and it is organized in three different compartments: the basement membrane, the perineuronal nets and the interstitial matrix (Novak and Kaye, 2000; Yamaguchi, 2000; Bonneh-Barkay and Wiley, 2009; Lau et al., 2013; Rauti et al., 2019). The basement membrane mainly consists of collagen IV, laminin–nidogen complexes, fibronectin, heparan sulphate proteoglycans such as perlecan and agrin (Baeten and Akassoglou, 2011; Xu et al., 2019) and a plethora of growth factors (Barcelona and Saragovi, 2015). The perineuronal nets are mainly composed of hyaluronan, proteoglycans, such as chondroitin sulphate proteoglycans (e.g. aggrecan, brevican, phosphacan and neurocan) and heparan sulphate proteoglycans (e.g. syndecan and glypican), while the interstitial matrix is composed of proteoglycans, hyaluronic acid, tenascins and fibrous proteins (e.g. collagen and elastin) as well as glycoproteins (Novak and Kaye, 2000; Rauti et al., 2019). The extracellular milieu of the brain serves as a physical support for cell migration as well relaying mechanical and biochemical stimuli that influences cell growth and differentiation (García-Parra et al., 2013; Levy et al., 2014; Potjewyd et al., 2018; Rauti et al., 2019).

Currently, the most extensively explored CNS-derived in vitro models consist of 2D and 3D cellular cultures of various species (Zhuang et al., 2018). The first 2D in vitro animal CNS models were established by Harrison and Hoadley (Hoadley, 1924; Waddington and Cohen, 1936), who were also the first to observe neurite extensions. Since then, improvements in the capacity to maintain healthy in vitro CNS cultures for long periods enabled additional primary models to be developed, based on cultures of hippocampal cells (Fig. 1A) (Dotti et al., 1988; Barrejón et al., 2019), cortical cells, midbrain cells (Brewer, 1995; Lingor et al., 1999; Pacitti et al., 2019), astrocytes and microglial cells (Giulian and Baker, 1986). In parallel, organotypic cultures, including cultures of hippocampal tissue (Fig. 1B), the substantia nigra, and many others also emerged (LaVail and Wolf, 1973; Knopfel et al., 1989; Robertson et al., 1997); these models were suggested to better resemble in vivo conditions compared with 2D monocultures, yet lack cytoarchitecture, physiological perfusion, and cannot be scaled to larger studies of human tissue (Humpel, 2015).

Figure 1.

In vitro CNS models. (A) Confocal micrographs showing 2D-hippocampal dissociated cultures, immunostained for the cytoskeletal component β-tubulin III (in red), the glial protein GFAP (in green) and DAPI to visualize neurons (in blue). Scale bar = 100 µm. Modified from Barrejón et al. (2019) with permission. (B) Light micrograph of a hippocampal slice [modified from Miller et al. (2015) with permission]. (C) Confocal section and 120-µm thick 3D stacks reconstruction showing a 3D hydrogel-encapsulated cortical neuronal network, immunostained for neurons (red, β-tubulin III), glia (green, S100) and nuclei (blue, DAPI). Scale bar = 50 µm [modified from Dana et al. (2014) with permission]. (D) Confocal image representing neurosphere processed for immunofluorescence against Arl13b (red) and DNA [modified from Shimada et al. (2017) with permission]. (E) A representative image of an organoid immunostained for neurons (TUJ1, green) and progenitors (SOX2, red) [modified from Lancaster and Knoblich (2014) with permission]. (F) Schematic image of a microfluidic device where vascular and neuronal networks were co-cultured [modified from Osaki et al. (2018b) with permission]. (G) Schematic representation of in vivo and in vitro cortical brain layer structures, in which each colour represents a different printed layer. In the bottom panel, confocal reconstructions of the neurons coloured for their z-axis distribution through the gel after 5 days of culture. Scale bar = 100 µm [modified from Lozano et al. (2015) with permission]. (H) Schematic sketch of a potential 3D-printing procedure to generate a mini-brain from cellular spheroids [modified from Han and Hsu (2017) with permission].

While both primary and organotypic CNS models can provide insights about cellular morphology and functionality, while retaining most of the cells’ in vivo properties (Balgude et al., 2001; Hopkins et al., 2013, 2015), they still come with many limitations, including the challenge in preserving their viability and sterility (Walsh et al., 2005), as well as the variability of cell maturation (Gähwiler, 1981). Importantly, these models are generally animal-derived, and thus do not resemble human physiology, including different degrees of circuit complexity and brain architecture (Herculano-Houzel, 2014; DeFelipe, 2015; Hopkins et al., 2015).

The isolation of human-derived stem cells and hiPSCs created exciting new opportunities for the development of scalable human models in neurobiology (Dubois-Dauphin et al., 2010; Hopkins et al., 2015; Pacitti et al., 2019; Silva and Haggarty, 2020). Moreover, the development of 3D in vitro culture systems (Fig. 1C) can recapitulate more complex cell–cell interactions, opening up the door to more closely resemble in vivo cell environments (Dana et al., 2014; Langhans, 2018; Pașca, 2018). Examples range from cell-self-assembly-based approaches (e.g. spheroids) to scaffold-based platforms (e.g. hydrogels).

Novel in vitro CNS 3D-models, such as neurospheres (Hogberg et al., 2013) and organoids (Lancaster et al., 2013), began to be developed based on stem cells. Neurospheres (Fig. 1D) are self-assembled, dense structures, mainly composed of neural stem cells, neuronal and glial-restricted progenitor cells, postmitotic neuronal cells and dead or dying cells. Neurospheres constitute valuable systems for studying neurogenesis and neural development, and also serve as an almost unlimited source of neural and progenitor cells. A major limitation of such systems is that, because of low access to oxygen and nutrients, cells growing in the centre of the neurospheres die (Bez et al., 2003; Jensen and Parmar, 2006). Other drawbacks that have limited their use include continuous loss in neurogenic potential after some rounds of subculturing, in addition to difficulties in merging findings between different laboratories. Lancaster and Knoblich (2014) introduced, for the first time, the concept of human brain organoids (Fig. 1E), which are self-assembled cells, derived from pluripotent stem cells, featuring lineages and structures of the early embryonic CNS. There has been an exponential surge in the use of this technology, especially as a tool for in vitro disease modelling (Lancaster et al., 2013; Mariani et al., 2015; Jo et al., 2016; Raja et al., 2016; Lee et al., 2017). Organoids are particularly useful for the identification and testing of novel therapeutic approaches, in light of their capacity to adapt to genome editing techniques or gene therapy (Yin et al., 2016; Gonzalez-Cordero et al., 2018). Pellegrini et al. (2020) recently published an in vitro organoid model, which shows key functions of the human choroid plexus: barrier formation and CSF secretion. Drawbacks of organoids include the fact that their spontaneous formation makes it difficult to create reproducible systems in terms of cell types and organization. Furthermore, they lack many features, such as vascular perfusion, mechanical cues and circulating immune cells, which are critical for the physiological functionality of organs (Ingber, 2016).

To combine the advantages of current in vitro and in vivo models (Table 1), a novel platform began to be developed: OoCs, microfluidic devices in which cells and tissues can be cultured in micro-scale volumes and in a controlled microenvironment, designed to mimic specific in vivo cues (Meyvantsson and Beebe, 2008; Meer and Berg, 2012; Halldorsson et al., 2015; MacKerron et al., 2017, Osaki et al., 2018c; Sosa-Hernández et al., 2018; Oddo et al., 2019). These microfluidic cell chambers are also collectively referred to as microphysiological systems. The simplest platform is a single, perfused microfluidic chamber in which one kind of cell type or cell mixture can be grown. In a more complex design, two or more cell-containing chambers in the same chip are separated by membranes, channels or gel regions to allow direct contact or secretome-mediated interactions (Figs 1F and 2), as further described the ‘Engineered in vitro models to mimic CNS physiology’ section. Yet another level of complexity is to connect two or more microdevices (Fig. 1F), containing different cell types, allowing the interaction of different tissues or tissue regions (e.g. blood–brain barrier) (Bhatia and Ingber, 2014; Phan et al., 2017; Oddo et al., 2019). Such systems might ultimately provide the possibility of studying multi-organ physiological systems (Esch et al., 2014; Maschmeyer et al., 2015; Ingber, 2016).

Figure 2.

In vitro culturing of multiple brain regions. (A) Representative image of a microfluidic device (top) and immunofluorescence micrographs (bottom) of cortical (in green) and striatal (in red) neurons growing inside the chips [modified from Peyrin et al. (2011) with permission from The Royal Society of Chemistry]. (B) Novel multielectrode array device used for co-culturing primary rodent hippocampal and cortical neurons [modified from Soscia et al. (2017) with permission]. (C) Schematic representation of a novel brain-on-a-chip model comprising the three different brain regions, prefrontal cortex, hippocampus and amygdala, shown via confocal images (bottom), stained for β-tubulin III (in green), and GFAP (in red), [modified from Dauth et al. (2017) with permission]. (D) Schematic representation of a microfluidic device allowing to metabolically couple neuronal and endothelial cells [modified from Maoz et al. (2018) with permission].

Microfluidic models offer several key advantages, including flexibility in design and low-cost fabrication compared to custom-made large-scale cell culture, fluidics, and robotic systems. Additional benefits over traditional cell culture formats can include lower risk of contamination, lower consumption of reagents, and efficient experimental throughput (Halldorsson et al., 2015). Importantly, OoC technology presents the possibility to apply mechanical forces to recreate physiological movements (Huh et al., 2010; Kim et al., 2016), as well as fluid flow and shear stresses (Bhatia and Ingber, 2014; Bischel et al., 2015; Benam et al., 2016; Ingber, 2016). Nevertheless, there are some challenges limiting the use of OoC systems, such as a fairly long prototyping time, the lack of standardized protocols, the requirement of specialized equipment and complex and time-consuming fabrication processes (Coluccio et al., 2019).

To overcome these issues, researchers have been developing methodologies for 3D printing (Fig. 1G and H) in vitro brain models (Lozano et al., 2015; Han and Hsu, 2017; Hampson et al., 2018; Sivandzade and Cucullo, 2018). This approach enables CNS models to be fabricated along the z-axis using different materials (even living cells), thereby creating 3D structures that are biologically active (Xu et al., 2006; Gu et al., 2016, 2018; Bishop et al., 2017; Han and Hsu, 2017; Thomas and Willerth, 2017; Knowlton et al., 2018; Potjewyd et al., 2018; Oliveira et al., 2019). These in vitro models offer the opportunity to provide a more reliable representation of in vivo nervous tissue with customized design and precise fabrication, facilitating the generation of platforms with great consistency that can be used for drug testing and in clinical applications. Currently, in 3D printing, the most important limitations that need to be addressed are the development of 3D printing methods with minimal impact on cellular stress, reproducible cell technology, a better understanding of molecular gradients in the native nervous tissue (Zhuang et al., 2018), design of the ECM and validation of cellular function (Rauti et al., 2019).

It should be noted that, despite the benefits of 3D OoC setups, the 3D character of these systems makes imaging and data evaluation challenging (Booij et al., 2019). Nevertheless, today, most microfluidic devices are relatively thin (100–1000-µm height and thinner cell layers within these structures). As such, unlike conventional 3D cultures, these devices are often compatible with live imaging experiments, such as cell migration assays, and conventional immunohistochemistry assessment. For example, Deosarkar et al. (2015) used confocal imaging to image the independent vascular channels of their device with dimensions of 200 μm × 100 μm × 2762 μm (width × height × length) (Deosarkar et al., 2015). Recent technological developments in two-photon microscopy (Rakotoson et al., 2019) and advanced 3D imaging, which uses artificial intelligence and machine learning (Joshi et al., 2018; Masullo et al., 2018; Puls et al., 2018; Scheeder et al., 2018; Booij et al., 2019), integration of biosensors (Misun et al., 2016; Maoz et al., 2017), and hiPSC-derived cellular components are expected to advance 3D in vitro modelling. A recent Sciencemag Technical feature comment suggested that the incorporation of biosensors and microfluidics with tissue culture could soon reduce (if not replace in many cases), animal-based research (Dove et al., 2018).

In addition to advancements in cell-based in vitro platforms, it is important to acknowledge the development of non-cell-based in vitro models, which can serve as alternatives to cell-based strategies. Examples of such models include the immobilized artificial membrane assay (IAM), parallel artificial membrane permeability assay (PAMPA), and the solid supported lipid membrane assay (TRANSIL) (Vastag and Keserű, 2009; Sharma et al., 2019). Likewise, computer-based models and simulations, known as in silico models, are becoming increasingly sophisticated (e.g. integrating machine learning and deep learning methods; see Yuan et al., 2018), and can be used to supplement or even replace some experimental procedures. Computer-based models offer the possibility to synthetize, prescreen and virtually test novel drugs, limiting the need for intensive laboratory experiments and expensive clinical trials, and accelerating the drug development process (Naik and Cucullo, 2012; Alsarrani and Kaplita, 2019; Chlebek et al., 2019). Nevertheless, non-cell-based models are not yet sufficient on their own, and the results obtained with such studies must be validated by (cell-based) in vitro and in vivo studies (e.g. to determine the biological activity and the brain distribution of a specific compound) (Naik and Cucullo, 2012).

Indeed, validation, i.e. ensuring that a model faithfully recapitulates in vivo physiological and pathological processes, is essential for the translatability of any model. Such validation is highly challenging in CNS models, owing to the biological complexity of the system being reproduced. Accordingly, extensive efforts are continuously underway for determining the extent to which in vitro CNS responses are representative of their in vivo counterparts. For example, Belle et al. (2018) recently used electrophysiological measurements to detect quantifiable differences but also similarities between cortical neurons in vivo and in vitro. For engineered models, in vitro to in vivo comparisons are highly challenging-yet crucial nevertheless (Frazier, 1990; Belle et al., 2018; Jones et al., 2018). Currently, microfluidic devices are validated testing different annotated and well-known drug compounds.

Another challenge that hinders the translatability of in vitro CNS models relates to the cell populations they use. For example, though the primary purpose of OoCs is to mimic human microphysiological systems in vivo, many of the OoC studies cited in this review relied on (non-human) animal cells or on human primary cells, rather than on cells derived from hiPSCs, which are likely to offer greater translational value (Table 1). It should be noted that, though we believe that hiPSC-based OoCs hold great promise for the future of precision medicine, even hiPSC-based systems may not be perfectly translatable (see Doss and Sachinidis, 2019; Ortuño-Costela et al., 2019 for extensive reviews of the translational concerns raised by iPSCs). Reliance on non-hiPSC-based cell populations is largely driven by the substantial difficulties that scientists still face in their efforts to incorporate hiPSCs (and iPSCs in general) into in vitro systems. As these difficulties are a key hindrance to the establishment of hiPSC-based engineered in vitro models as a standard tool for brain research in both academic and industrial settings, we discuss them in detail before proceeding to describe the modelling platforms themselves.

In vitro modelling of the CNS using hiPSCs: benefits and challenges

Overview of hiPSC use in in vitro CNS models

Since Takahashi and Yamanaka reported iPSCs in 2006, iPSCs have revolutionized biomedical research and boosted hopes for personalized therapeutics (Takahashi and Yamanaka, 2006; Takahashi et al., 2007). These pluripotent stem cells resemble embryonic stem cells yet can be generated from terminally differentiated adult somatic cells such as skin fibroblasts or peripheral blood; hence, the use of human-derived iPSCs does not trigger the ethical concerns associated with the use of foetal cells in clinical trials (Yu et al., 2007; Loh et al., 2009; Polo et al., 2010; Halevy and Urbach, 2014; King and Perrin, 2014; Choi et al., 2015; Shi et al., 2017; Volarevic et al., 2018). Note that in this review we use the term iPSCs to refer to the technology itself and how it can be used in in vitro CNS models; when we refer to human-derived cells, human-oriented strategies or challenges, we use the term hiPSCs.

Success in producing hiPSCs has led to an enormous boost for cell replacement strategies; adult, somatic and patient-specific cells may now be reprogrammed and converted into immature cells, which can undergo directed differentiation into specific cell types, and then be grafted into the patient (Sánchez Alvarado and Yamanaka, 2014). Nevertheless, substantial hurdles must still be overcome before hiPSCs can be integrated into CNS therapy, even though hiPSC-based clinical trials involving non-brain tissues are underway (Bragança et al., 2019; Ortuño-Costela et al., 2019). Such challenges stem from the fact that the immense cytoarchitectural complexity of the brain tissue, coupled with immune responses (even in the case of autologous sources), often results in unsuccessful implementation of CNS transplants (Zhao et al., 2011; Nikolakopoulou et al., 2016; Garreta et al., 2018). These concerns regarding in vivo trials highlight the need for effective personalized CNS in vitro models, which capture the tissue microenvironment and cellular interactions.

Nowadays, hiPSCs from both healthy donors and patients with CNS diseases are used worldwide to model the complexity of human brain tissue and shed light on the mechanisms that govern its function in health and disease. The use of hiPSCs in conjunction with advanced in vitro technologies, such as OoCs, has enabled researchers to recapitulate patient-specific complex aspects of the human CNS, such as the blood–brain barrier (Vatine et al., 2019). Indeed, hiPSC-based brain-on-chip devices are routinely used to study, among others, neurodevelopment and neurodegeneration, contributing substantially to the advancement of regenerative medicine, toxicology, and high-throughput investigations (Berg et al., 2019). Nevertheless, researchers still face multiple obstacles when attempting to integrate iPSCs into in vitro platforms. These obstacles, elaborated in detail in what follows, stem both from the innate characteristics of the brain tissue and from the biological characteristics of the cell source.

Current challenges for hiPSC-based advanced in vitro models

Donor variability and cell heterogeneity

Donor variability, a feature shared with primary cells, is a major hurdle in hiPSC-based in vitro modelling. Residual epigenetic memory, genetic background and specific characteristics acquired during the reprogramming and differentiation processes result in great diversity among hiPSC-derived cell lines (Kim et al., 2010, 2011; Polo et al., 2010; Bar-Nur et al., 2011; Boland et al., 2014). Moreover, several hiPSC lines show defective proliferation and differentiation potential due to incomplete reprogramming (Ohnuki et al., 2014). Advanced gene transduction and editing technologies [e.g. zinc-finger nucleases, transcription activator-like effector nuclease (TALEN), clustered regularly interspaced short palindromic repeats (CRISPR)/Cas-9 based gene or base editing] and next-generation sequencing are used widely to handle the variability issue in hiPSC lines (Komor et al., 2016; Brookhouser et al., 2017; Doss and Sachinidis, 2019).

Differentiation protocols, reproducibility, and maturity

Reliable translational models require both high-quality hiPSCs and effective differentiation protocols towards the desired cell fates. Thus, researchers in both academia and industry have been working continuously towards the development of CNS lineage-specific differentiation protocols. Nevertheless, the need for pure and mature cell types remains largely unmet. The cells frequently show immature functional characteristics with embryonic and foetal tissue attributes. Consequently, they can serve as good models for neurodevelopmental and early onset disease studies but cannot adequately mimic late onset disease and mature tissue (Miller et al., 2013). Interestingly, a recent study by Lu et al. (2019) highlights problems with iPSC-derived cells. The authors question the identity of iPSC-derived brain microvascular endothelial cells (BMECs) (Lu et al., 2019). According to the authors, the differentiated cells recapitulated properties from neuroectodermal epithelium rather than the blood–brain barrier. Moreover, several studies have reported biased differentiation potential of hiPSCs towards specific lineages due to residual epigenetic memory. It may be possible to reset these cells to pluripotential patterns by increasing the number of passages before differentiation (Polo et al., 2010; Bar-Nur et al., 2011; Kim et al., 2011; Boland et al., 2014; Kedziora and Purvis, 2017; Doss and Sachinidis, 2019). Still, overall, to increase the translatability of iPSC-based in vitro models, it remains a critical issue to produce high-quality hiPSCs and to develop current protocols further.

Immunogenicity of hiPSC-derived cells

Earlier studies performed in animals (de Almeida et al., 2014; Zhao et al., 2015) and in human co-culture systems (Huang et al., 2014) have shown that differentiated cells are less immunogenic than the corresponding iPSC populations. This observation suggests that hiPSC-derived tissues may replace autologous tissue transplants, since they can surpass the physiological immune responses and avoid rejection by the patients. This same feature suggests, however, that hiPSC-derived in vitro models may be lacking in predictivity in terms of physiological immune responses. In such a scenario, it is plausible that hiPSC-derived models may poorly describe immune mediated diseases of the brain tissue and thereby hinder successful drug development.

Using hiPSC-based disease models to model neurodegenerative diseases

Disease modelling using hiPSCs may be highly accurate for monogenic diseases, but the relevance for complex, polygenic and sporadic diseases is debatable. The reports so far compare cellular characteristics of one or a few patient cells to their respective family and gender-matched controls; any differences observed during phenotypic analysis are usually attributed to the mutation under investigation and thereby to the cause of the disease. This approach may prevent researchers from distinguishing disease-related variation from variation attributable to other factors such as epigenetic memory, genetic background and environmental cues. Consequently, and because most neurodegenerative diseases are sporadic in nature, it is necessary to use large numbers of patient-derived hiPSC lines to decrease the signal-to-noise ratio of the studies and increase the accuracy of the results (Doss and Sachinidis, 2019). Moreover, hiPSC derivatives are commonly used in 2D cultures, which lack the multi-faceted interactions in the human body in health and disease; engineered 3D models, on the other hand, provide in vivo-like platforms to investigate how, among others, genetics and environmental cues influence disease phenotypes (Sharma et al., 2020).

Regional identity of hiPSC-derived cells

The complexity of the human CNS arises from the active interactions among multiple neural cellular subtypes. Most neurological diseases, on the other hand, stem from defects in specific cellular subtypes, with the underlying mechanisms remaining largely elusive (Imaizumi et al., 2015). Accordingly, accurate disease modelling requires the use of differentiated cells with specified regionality, thereby providing relevance to the disease of interest. Indeed, researchers worldwide have used hiPSCs from patients to study several neurological diseases, including amyotrophic lateral sclerosis (ALS), Alzheimer’s disease, schizophrenia and Parkinson’s disease (Cooper et al., 2010; Zhang et al., 2014; Ahmad et al., 2018; Fujimori et al., 2018; Henstridge and Spires-Jones, 2018; Li et al., 2018; Mishima et al., 2018, Osaki et al., 2018c; Ishii et al., 2019; Penney et al., 2020). In these studies, hiPSCs were differentiated into the respective cell types, thereby providing cellular disease models with regional characteristics. The conclusions obtained in these studies, however, remain largely ambiguous, owing to the limited capacity of current differentiation protocols to provide highly pure cultures of the various neuronal subtypes. We believe that further development of the current protocols in terms of regional characteristics, combined with large cohorts to overcome patient-specific variability of derived cells, will be of great benefit for in vitro CNS modelling.

The need for a ‘universal medium’

An issue of great importance for iPSC-derived in vitro models based on OoCs is the need for a ‘universal medium’. To effectively mimic the multifaceted nature of brain tissue, it is necessary to co-culture several cell types. As discussed in a remark by CellPress editor Pavlovich, ‘there are biological factors: not every cell thrives in the same medium, responds to the same growth factors or differentiation cues, or adheres to the same matrix’ (Pavlovich, 2018). Therefore, research and industrial laboratories continue to invest enormous efforts into the development of xeno-free, defined protein substrates and media to support iPSC-derived cell expansion and differentiation. Alternative potential solutions include advanced microengineering, elaborate design and compartmentalized devices. It is important to note, however, that the latter approaches require extensive technical expertise.

Engineered in vitro models to mimic CNS physiology

Cellular organization in microfluidic platforms

A basic objective underlying the development of in vitro microfluidic platforms is the creation of an environment in which physiological compartments (e.g. vasculature and parenchymal tissue) are simplified and physically separated yet can still interact with one another-thereby facilitating observations that are effectively impossible to achieve in in vivo models. This simultaneous separation and interaction may be realized either by linking several OoCs or by using one compartmentalized device.

The concept of physically separating the neuronal body and its extending neurites was first proposed by Campenot in 1977 (Campenot, 1977), and since then was extensively applied in the so-called Campenot chambers. In 2003, Taylor et al. incorporated microgrooves in a cell culture chamber, taking advantage of microfluidic, micropatterning and microfabrication technology for the first time; the researchers were able to physically isolate neuronal soma and neurites, and to use the system to study local damage of neurites (Taylor et al., 2003; Neto et al., 2016). Since then, numerous engineered in vitro platforms have relied on single-chip CMDs to support physical isolation of the cellular populations and/or components on the microscale as a basis for mechanistic studies, including studies of axonal biology (Shin et al., 2010) or synapse formation and modulation (Taylor et al., 2010; Coquinco et al., 2014). More than a decade ago, in 2005, Taylor et al. pioneered a microfluidics-based in vitro platform to study axonal regeneration after injury (Taylor et al., 2005). The device was further advanced by Park and collaborators to investigate neuron-glia interactions and axon myelination (Taylor et al., 2005; Park et al., 2006, 2009a, 2012; Higashimori and Yang, 2012; Shi et al., 2013). As a further advancement, intra-system co-cultures of neuronal and glial populations (Park et al., 2009b; Yang et al., 2012; Shi et al., 2013), aiming to elucidate myelination pathways, have been developed with the aim of identifying novel treatment strategies for demyelinating diseases such as multiple sclerosis. Subsequent developments of such systems may include integration of immune cells and particularly hiPSC-derived microglia (Abud et al., 2017; Douvaras et al., 2017; Haenseler et al., 2017; Pandya et al., 2017; Garcia-Reitboeck et al., 2018; McQuade et al., 2018) and hiPSC-derived oligodendrocytes (Ehrlich et al., 2017), towards shedding light on the complex cellular processes occurring in the human demyelinating brain tissue.

In a multi-chip design, several OoCs are linked; thus, the complex cytoarchitecture of the nervous tissue may be largely replicated. For example, a linked three-chip arrangement revealed metabolic coupling between the endothelial and the neuronal cells in an engineered model of the human neurovascular unit (NVU) (Maoz et al., 2018).

Moreover, microfluidics have been used extensively to explore how biochemical cues affect axonal behaviour, outgrowth, pathfinding and synapse function (Wu et al., 2005; Cox et al., 2008; Taylor et al., 2009; Gumy et al., 2011; Park et al., 2014; Kung et al., 2015). Deglincerti et al. (2015) elegantly used the physical separation of axons from the neuronal soma to show that local protein synthesis and degradation are interconnected in growth cones. The study demonstrated that growth cones show elevated levels of ubiquitylation, and that the ubiquitin-proteasome system targets locally translated proteins. Thus, the authors suggested that axonal tuning responses towards guidance signals may incorporate local protein synthesis and degradation (Deglincerti et al., 2015; Neto et al., 2016). Local translation in axons and particularly in the growth cones is now widely accepted; this is a result of the capability to perform studies in isolated axonal fractions with the use of microfluidics.

In addition to providing a means of separating cell populations while enabling interactions to take place between them, OoCs offer the possibility to control a vast array of mechanical properties such as stiffness, geometric confinement, interstitial flow and shear stress, which affect cellular state and differentiation (Sundararaghavan et al., 2009; Peyrin et al., 2011; Song and Munn, 2011; Kim et al., 2013; Galie et al., 2014; Hattori et al., 2014; Asano et al., 2015; Osaki et al., 2018a). Moreover, biochemical gradients on the cellular components can be easily imposed on the devices to recapitulate early neurodevelopmental processes and perform mechanistic studies. This feature is useful, e.g. for manipulating concentration gradients of cytokines, such as BMP4, SHH, FGF, RA, and WNT3, which orchestrate cellular proliferation, differentiation and organogenesis in early brain tissue development (Park et al., 2009b; Demers et al., 2016; Uzel et al., 2016; Osaki et al., 2018a).

The above section briefly describes how OoCs may be used to investigate cellular organization in the brain tissue. Several of the studies mentioned here utilize cells of animal origin; nevertheless, we consider them of tremendous importance for further development of human in vitro models. We envision that incorporation of hiPSC-derived cellular entities in similar platforms may further advance our understanding of how cells organize themselves in the human brain.

The use of engineered microfluidic systems to recapitulate distinct brain regions and their connectivity

The human brain consists of more than 250 different brain regions (Ding et al., 2016), which have a highly specialized ECM (Dauth et al., 2016), architecture and functionality (Novak and Kaye, 2000; Lau et al., 2013; Dauth et al., 2016). Nevertheless, higher brain functions, as well as many neuropsychiatric disorders (Quadrato et al., 2016), are regulated by the interaction between multiple brain regions such as the hippocampus, cortex, thalamus, cerebellum, and amygdala (Kato-NegiShi et al., 2013). Therefore, it is of great importance to mimic the different physiologically connected brain regions in vitro. Currently, there are just a few reported in vitro models incorporating cells derived from two or more different brain regions (Kanagasabapathi et al., 2011; Peyrin et al., 2011; Kato-NegiShi et al., 2013; Dauth et al., 2017; Soscia et al., 2017). These include a model developed by Peyrin et al. (2011), who reported the ability to recreate a functional and synchronized cortico-striatal oriented network, using a microfluidic system. The two-chambered microfluidic device was used to culture primary murine cortical and striatal neurons and let them connect through controlled and highly ordered axons (Fig. 2A) (Peyrin et al., 2011); this enabled the researchers to observe that cortical neurons trigger the differentiation of spiny striatal neurons and the formation of dendritic spines. A similar neurofluidic device with microchannels was also used to provide insights into the interactions between the cortex and thalamus (Kanagasabapathi et al., 2011); the system provided easy accessibility and manipulation capabilities, which cannot be easily achieved in in vivo systems (e.g. the ability to investigate cortical-thalamic interactions in isolation, without the influence of other regions). The latter in vitro model allowed for neurite outgrowth, connections between the two different brain tissues, functional readouts (e.g. electrophysiology) and immunohistochemistry (Kanagasabapathi et al., 2012). By using this in vitro system, the authors demonstrated that burst events originate in the cortical region, triggering the cortical-thalamic network, confirming some previous data obtained from in vivo experiments.

The next advance was achieved by Kato-Negishi et al. (2013), who developed a 3D millimetre-sized neural building block, formed from rat hippocampal and cortical cells. This model allowed for Ca²⁺ imaging, gene transduction measurements, and immunohistochemistry, and it was able to show the formation of projections and synaptic contacts between cortical and hippocampal neurons, thereby enabling in vitro investigations of interactions between multiple brain regions. In a similar way, Soscia et al. (2017) developed a platform to co-culture hippocampal and cortical neurons (Fig. 2B), with the possibility of investigating how the different brain regions establish connections, integrate networks and increase their firing rate. Recently, Dauth et al. (2017) were the first to develop an in vitro multiregional CNS model implementing three brain regions, derived by culturing rat prefrontal cortex, hippocampus and amygdala-derived tissues, functionally connected through axons (Fig. 2C). Their model incorporated functional readouts by measuring extracellular field potentials with multielectrode arrays, together with biochemical readouts and immunohistochemistry. Their work demonstrated the significance of connecting different brain regions in vitro, showing that doing so changes the cellular composition, protein expression and electrophysiological properties of the co-cultured cells compared with those observed in monocultures. Moreover, the model was used to mimic the corticolimbic system and to examine the effects of phencyclidine hydrochloride (PCP) on one brain region and to identify how the other regions respond.

Modelling the neurovascular unit (the blood–brain barrier)

The studies discussed above focused on modelling the brain’s complex neuronal architecture and functionality. Yet, recent studies have acknowledged the need for in vitro CNS models to take into account not only the brain’s parenchymal cells but also its unique vasculature, which differs from non-CNS vasculature in that it exhibits continuous tight junctions, is void of fenestrations and has a very low rate of transcytosis (Abbott et al., 2006). Perhaps most importantly, the brain’s vasculature includes a highly specialized endothelium—the blood–brain barrier—which tightly regulates the entry of compounds into the brain. In other words, in vitro models aimed at understanding the brain’s responses to various stimuli [e.g. mechanical perturbations such as traumatic brain injury (TBI), drug development, toxicology, etc.] must model the entirety of the NVU, which is composed of the blood–brain barrier and the perivasculature; brain pericytes, and closely interacting astrocytes and neurons. In general, ensuring that a given molecule can penetrate the blood–brain barrier and thus access the CNS is a major challenge in CNS drug development (Herland et al., 2020). In small molecular approaches, computational pharmaco-distribution and animal models have been fairly successful, but the increasing dominance of human-specific biological pharmaceuticals has led to a greater need for models that are highly predictive of human blood–brain barrier penetrance (Gribkoff and Kaczmarek, 2017).

Several engineered in vitro models have been developed to mimic the NVU (summarized in Table 2 and Supplementary Fig. 2). For example, a recent model by Maoz and colleagues connected three chips, one blood–brain barrier chip connected to a brain chip, which was connected to a second blood–brain barrier chip containing human BMECs, neuronal cells, glial cells and pericytes (Fig. 2D). They used this system to analyse the individual cell types comprising the NVU (Maoz et al., 2018), as well as to mimic the effect of intravascular administration of psychoactive drugs (e.g. methamphetamines) and to identify, for the first time, metabolic ‘crosstalk’ among the cellular components of the blood–brain barrier. Moreover, in silico models of the blood–brain barrier have also been extensively used to predict drug permeability in the blood–brain barrier (Cabrera et al., 2004; Suenderhauf et al., 2012; Miao et al., 2019; Roy et al., 2019).

Table 2.

Summary of in vitro models commonly used in blood–brain barrier research

Model	Shear stress	Cell-cell interactions	High-throughput / cost	Similarity to human physiology
Transwella,b	No	Co-culturing possible, tri-culturing more challenging to evaluate cell populations	Yes / low	Minimal, ECM present only as anchoring points, 2D geometry
Porous-tube modelsc	Yes	Same as Transwell	Minimal / moderate	Improved similarity to human physiology (shear stress, 3D luminal geometry), but minimal ECM present
Microfluidic chips (membrane-based)d,e	Yes	Capability of compartmentalization and studying interactions between cell populations	Yes; however, more time consuming than Transwell / moderate	Same as porous-tube models
Microfluidic chips(ECM-based)f	Yes	Same as membrane-based microfluidic chips	Yes; however, more time consuming than Transwell / moderate	Utmost attempt at in vitro biomimicry (shear stress, 3D geometry, ECM present)

NVC = neurovascular chip.

^aIn this list, we consider studies that use Transwell in static cultures, there are, however, studies that implement flow in Transwell (Hinkel et al., 2019).

^b Zenker et al., 2003; Colgan et al., 2008; Helms et al., 2014; Labus et al., 2014; Canfield et al., 2017; Delsing et al., 2018.

^c Neuhaus et al., 2006; Cucullo et al., 2011;Marino et al., 2018; Moya et al., 2020.

^dIn this list, microfluidic chips with a temporary membrane (i.e. a membrane that degrades over time) are not included, such as the work of Tibbe et al. (2018).

^e Booth and Kim, 2012; Prabhakarpandian et al., 2013; Achyuta et al., 2013; Wang et al., 2017; Maoz et al., 2018.

^f Brown et al., 2015; Herland et al., 2016, Xu et al., 2016a; Adriani et al., 2017; Partyka et al., 2017.

Below, we elaborate on the considerations that must be taken into account when modelling in vitro the properties of the NVU and blood–brain barrier, which is constantly under flow and is characterized by diverse types of cell–cell interactions.

Blood–brain barrier in vivo metrics

Evaluation of barrier properties is done by measuring transendothelial electrical resistance (TEER), passive permeability to small compounds (<1000 g/mol), and the activity of efflux and influx transporters (e.g. by the use of P-glycoprotein substrates and glucose, respectively) (Lippmann et al., 2012; Stebbins et al., 2016). A key component that affects these properties is the source of cells for the model. In particular, as noted above, the use of animal cells can raise concerns regarding inter-species differences. Primary human BMECs retain some blood–brain barrier phenotypes. However, TEER of human primary BMECs (does typically not exceed 200 Ω×cm² (Mackic et al., 1999; Zenker et al., 2003), which accounts for only 10% of in vivo TEER measurements of the blood–brain barrier in rats and frogs (Crone and Olesen, 1982; Butt et al., 1990). In 2012, with the advent of hiPSCs, Lippmann was the first to report hiPSC-derived BMEC-like cells that exhibited TEER values >200 Ω×cm² (Lippmann et al., 2012). The controversy of the identity of these cells were discussed in an earlier section. Since then, a plethora of studies on hiPSC-derived BMEC-like cells emerged, using similar (Lippmann et al., 2014; Hollmann et al., 2017) or conceptually different differentiation strategies (Orlova et al., 2014).

Real-time monitoring of salient features of the NVU-on-chip is a field of expanding interest. To date, researchers have integrated permeability evaluations and TEER measurements in 2D (Booth and Kim, 2012; Walter et al., 2016; Wang et al., 2017) and 3D cultures (Brown et al., 2015, Xu et al., 2016a; Partyka et al., 2017) on chips. However, real-time monitoring of metabolic processes with analytical microfluidic chips has so far been only implemented in conjunction with animal models (Lin et al., 2014).

Cell–cell interactions

Co-culturing cells of the NVU adds another level of complexity to the in vitro model, enabling the model to capture in vivo conditions more faithfully. In particular, perhaps unsurprisingly, co-culturing BMECs with CNS cells contributes to the blood–brain barrier-like properties of endothelial cells, e.g. through fortification of tight junctions and expression of polarized transporters (Kasa et al., 1991; Megard et al., 2002; Didier et al., 2002, 2003; Haseloff et al., 2005; Lippmann et al., 2012; Herland et al., 2016; Hollmann et al., 2017). Moreover, studies reveal that astrocytes and BMECs secrete factors that confer each other’s maturity (Janzer and Raff, 1987; Fukushima et al., 2009; Blanchette and Daneman, 2015).

The Transwell model (Table 2 and Supplementary Fig. 2) has been used extensively for co-culturing BMECs with CNS and non-CNS cells; this model allows for non-invasive TEER measurement, permeability assays, and evaluation of efflux pumps (Zenker et al., 2003; Colgan et al., 2008; Helms et al., 2014; Labus et al., 2014; Canfield et al., 2017; Delsing et al., 2018). However, this approach reflects a static environment with a non-continuous cell monolayer, and thus does not capture the in vivo blood–brain barrier setting.

Flow

Flow is a significant parameter that should be considered in the development of NVU models. Endothelial cells in the capillaries experience a force (shear stress) parallel to the endothelium, exerted from the blood flow. Jiang et al. (2019) recently discussed the impact of flow on cellular functionality in a comprehensive review of microfluidic models of the blood–brain barrier.

Siddharthan et al. (2007) showed a correlation between shear stress and the upregulation of the tight junction protein ZO-1 in BMECs. In 2011, Cucullo et al. reported a thorough evaluation of the impact of shear stress on the transcriptome of BMECs; shear stress upregulated tight and adherens junctions as well as multidrug resistance transporters (Cucullo et al., 2011), which is in accordance with the blood–brain barrier transcriptome footprint.

The emergence of OoCs has further enabled researchers to observe the functionality of vascular and other cells in vitro in the presence of flow stimulation. Huh et al. (2010) were the first to demonstrate flow-induced organ-level functions, in an OoC model, specifically a lung-on-a-chip model (Huh et al., 2010). Booth and Kim (2012) first reported an NVU-on-a-chip using brain endothelial and astrocytic cell lines, documenting that the NVU-on-a-chip resulted in higher TEER than the static conditions as well as permeability to tracers that resembled in vivo levels. However, another study using hiPSC-derived BMECs suggested that there was no difference in tight and adherens junction expression between static and dynamic conditions; this lack of difference may have been attributable to immaturity of the differentiated BMECs (DeStefano et al., 2017). Other studies have suggested that the permeability levels of tracers and compounds in NVU-on-chip systems are on par with in vivo data (Wang et al., 2017), nevertheless, these studies did not carry out direct comparisons between chips and Transwells. More thorough studies suggest that NVU-on-chip systems exhibit lower permeability than the Transwell model (Prabhakarpandian et al., 2013; Walter et al., 2016; Partyka et al., 2017). Recently, Vatine et al. (2019) demonstrated that flow significantly decreases the blood–brain barrier permeability in an isogenic NVU-on-chip. In addition, they perfused the isogenic NVU-on-chip with blood from the same donor and could detect inter-individual blood–brain barrier characteristics. The various discrepancies between different studies highlight the fact that variables such as cell sources, materials, media volumes as well as cell-to-medium ratios are important aspects when comparing various in vitro models.

Flow might also influence the alignment of BMECs; notably, however, in vitro studies investigating this alignment process have produced different results. A recent study by Moya et al. (2020), which used a brain endothelial cell line, suggested that BMECs align in the direction of the flow. Another study on primary BMECs also suggested that primary cells align in the direction of the flow (Garcia-Polite et al., 2017). However, several studies using primary BMECs (Ye et al., 2014; Reinitz et al., 2015) and iPSC-BMECs (DeStefano et al., 2017) have suggested that BMECs do not align in the direction of the flow. None of these in vitro studies addressed whether the BMECs (or human BMECs) have a venous or arterial phenotype. This omission is notable, given that, in vivo, the capacity of endothelial cells to align in the direction of the flow is dependent on variables such as the nature of the endothelium (arterial or venous) (dela Paz and D’Amore, 2009) as well as the level of shear stress (Masumura et al., 2009). These variables may have a direct implication on endothelial cells’ capacity to align in the direction of the flow.

Extracellular matrix

Mimicking the brain ECM in vitro is a great challenge (Rauti et al., 2019); the ECM components, discussed in the ‘Overview of recent developments in in vitro CNS models’ section, differ throughout the brain (i.e. different brain regions have unique ECM composition, and the brain vasculature ECM is different from the brain ECM) while that is also amenable to the developmental stage i.e. BMECs in the brain vasculature swift their signalling from fibronectin in development to laminin in adulthood (Herland et al., 2016; Adriani et al., 2017; Linville et al., 2019). To recapitulate the NVU in vitro, the ECM of both the brain vasculature and the rest of the brain should be considered to ensure the accurate replication and effectiveness of the NVU in vitro model (Rauti et al., 2019).

The majority of iPSC-derived BMECs protocols use a combination of collagen IV and fibronectin as a purification step during BMEC differentiation (Lippmann et al., 2012, 2014; Hollmann et al., 2017). Nevertheless, other types of ECM components have been used when replicating the NVU in vitro, such as type I collagen. Albeit type I collagen is not naturally present in the brain, the gelation properties of type I collagen have made it quite favourable in 3D in vitro rendering of the NVU (Herland et al., 2016; Partyka et al., 2017; Wevers et al., 2018; Grifno et al., 2019; Linville et al., 2019).

ECM-derived gels have been incorporated into microfluidic devices, adding yet another layer of complexity to the existing fluidic models (Herland et al., 2016; Adriani et al., 2017; Linville et al., 2019). Currently, microfluidic systems with incorporated ECM gels constitute the most comprehensive attempts at achieving biomimicry in NVU-on-chip models. Yet, the process of setting up systems that incorporate ECM gels is time-consuming, costly, complex, and the throughput of these systems is lower than that of traditional Transwells. Given the physiological importance, the recapitulation human relevant ECM in vitro is a growing research field, however with many challenges unsolved.

Engineered in vitro models mimicking CNS disease

Cell death or alterations in the CNS cellular microenvironment may lead to network disruption and pathologies, including neurodegenerative diseases, TBI and cancer (Osaki et al., 2018a).

The brain ECM is altered in some pathological conditions and ageing (Bonneh-Barkay and Wiley, 2009; Burnside and Bradbury, 2014; Caldeira et al., 2018). There are alterations that directly contribute to the progression of certain pathological conditions diseases (Baeten and Akassoglou, 2011) such as autism (Mercier et al., 2012), epilepsy (Dityatev, 2010; McRae and Porter, 2012), Alzheimer’s disease and schizophrenia (Lu et al., 2011; Berretta, 2012; Pantazopoulos et al., 2015; Reed et al., 2019). Moreover, recently it was shown that the brain ECM have a significant role in the development of neurodegenerative diseases such as Alzheimer’s disease, schizophrenia and bipolar disorders (Barcelona and Saragovi, 2015).

To elucidate the underlying mechanisms of brain tissue pathologies, it is necessary to emulate the occurring processes accurately in vitro. In CNS pathology, animal models have shown particularly low predictive capacity; Alzheimer’s is a striking example of how animal trials have shown promising results, whereas, one after the other, clinical trials have failed (De Felice and Munoz, 2016; Mofazzal Jahromi et al., 2019). Below, we discuss advanced engineered 2D and 3D in vitro models that have recently been developed to investigate brain pathologies.

Neurodegenerative disease

Neurodegenerative diseases are escalating in prevalence and have devastating effects on individual and societal well-being (Marras et al., 2018; Patterson, 2018; Fisher and Bannerman, 2019). Most neurodegenerative diseases are incurable, and the neurobiological mechanisms governing disease initiation, progression and therapy remain elusive (Centeno et al., 2018). Some of these diseases are monogenic, whereas for others, e.g. Alzheimer’s disease, Parkinson’s disease and ALS, the vast majority of cases (90% or more) have not been linked to a genetic cause, making it almost impossible to generate relevant animal models (Centeno et al., 2018). The use of hiPSC-based 3D engineered in vitro models offer a potential alternative to animal testing and provide human-specific mechanistic insights in both monogenic and sporadic disease pathology.

Alzheimer’s disease

Alzheimer’s is currently the most prevalent neurodegenerative disease (WHO, 2019), and extensive efforts have been devoted to the development of predictive in vitro models for this disease. Patients suffer from progressive cognitive dysfunction characterized by excessive amyloid-β accumulation and neurofibrillary tangles (Hardy and Selkoe, 2002; Tanzi and Bertram, 2005). Transgenic mouse models of Alzheimer’s disease mutations mimic aspects of the disease such as memory loss; however, they fail to recapitulate key characteristics of the condition such as the neurofibrillary tangle pathology (Chin, 2011).

In 2014, Tanzi and his team were the first to report a human 3D in vitro model that resembled the pathophysiology of the disease (Choi et al., 2014). Even though this was one of the greatest advancements in the field an important piece to the puzzle was still missing; the contribution of the blood–brain barrier breakdown to the disease phenotype could not be investigated. To fill the gap, in 2019 the team proposed an advanced, physiologically relevant 3D human microfluidic-based platform, which incorporated a tubular BMEC layer with barrier-like properties into their Alzheimer’s culture system (Shin et al., 2019). Scientists may now use this platform to model the progressive accumulation of amyloid-β peptides in the ECM and the sequential transport via the NVU microenvironment.

Incorporation of hiPSC-derived cells in the 3D culture platforms may bring us as a step closer to understanding the mechanics of this devastating disease and design effective human-oriented therapeutics (Choi and Tanzi, 2012). To this end, Zhang et al. (2014) generated a 3D culture system in which hiPSC-neurons were cultured in a commercially available soft hydrogel composed of laminin and a synthetic peptide (i.e. RADA-16), mimicking the 3D neural microenvironment, especially the tissue stiffness. Park et al. (2018) proposed a novel 3D organotypic tri-culture system, which incorporated neurons, astrocytes, and microglia in a microfluidic platform to study the pathogenesis of Alzheimer’s with respect to neuroinflammatory stimuli (Fig. 3A).

Figure 3.

3D engineered in vitro models for neurodegeneration and TBI. (A) A 3D organotypic human triculture model for Alzheimer’s disease (AD) (Park et al., 2014). (i) Neural progenitor cells (NPCs) were differentiated to Alzheimer’s disease neurons and astrocytes, while monitoring microglia recruitment. (ii–iii) Schematic of the multicellular interactions in the in vitro microfluidic AD model (ii) and in the AD brain (iii). (iv) Image i: Fluorescent image of the microfluidic platform. Alzheimer’s disease neurons (Neu)/astrocytes (AC) (green) are in the central chamber and microglia (red) are in the angular chambers. Scale bar = 250 μm. ii: Microglial recruitment across the angular microchannels. Scale bar = 250 μm. iii and iv: Confocal imaging confirms the 3D physiological intercellular communication among neurons (green), astrocytes (green) and microglia (red) in the central chamber. Nuclei are shown in white. Scale bars = 100 μm in iii; 40 μm in iv. (v) Comparison of microglial recruitment (red) by the control Neu + AC (green) and the AD Neu + AC (green). Scale bars = 250 μm (top) and 200 μm (bottom). (vi) Microglial recruitment by hiPSC AD Neu + AC. Scale bar = 10 μm. (B) 3D model of Parkinson’s disease (PD) dopaminergic (DA) neurons for high content phenotyping and drug screening (Bolognin et al., 2019). (i) Schematic illustration of the experimental procedure. The setup allows for automated image acquisition, segmentation, feature extraction and data analysis. (ii) A clear clustering of the lines according to genetic background is shown in the heat map. (C) A 3D ALS motor unit microfluidic model (Osaki et al., 2018c). The ALS motor unit (right) exhibits fewer thick neural fibres and decreased neuromuscular junction (NMJ) formation compared with the embryonic stem (ES) cell-derived motor unit (left). Motor neurons are stained with TUJ1 (green), actin filaments with F-actin (purple) and nuclei with DAPI (blue). Scale bars = 100 μm. (D) A brain-on-a-chip to model TBI. Schematic sketch of the uniaxial axonal strain device (i) an example of axonal beading observed before and after the strain injury (ii) and a bar plot representing the correlation between the diameter of the axon/bundle and the number of beads used to injure the cells (iii). Figure components are modified from Dollé et al. (2014), Osaki et al. (2018c), Park et al. (2018) and Bolognin et al. (2019) with permission.

Nevertheless, both in vivo and in vitro experimental platforms have this far failed to mimic tissue maturity. Recent studies have shown that epigenetic modifications underlie the link between ageing and disease progression (Fyfe, 2018; Nativio et al., 2018). Future incorporation of mature neurons and immune cells in in vitro platforms may substantially enhance predictivity and bring us closer to effective diagnostics and therapeutics for patients with Alzheimer’s disease.

Parkinson’s disease

Parkinson’s disease is characterized by selective loss of dopaminergic neurons in the substantia nigra (Antony et al., 2013). Since 2010, when hiPSC-derived dopaminergic neurons (Cooper et al., 2010; Hargus et al., 2010) were developed, researchers have used either organoids (Monzel et al., 2017) or OoC methods (Moreno et al., 2015; Bolognin et al., 2019) to mimic the parkinsonian brain in vitro. The G2019S mutation in the leucine-rich-repeat-kinase-2 (LRRK2) is frequently associated with the pathophysiology of both familial and sporadic forms of Parkinson’s disease (Paisán-Ruı´z et al., 2004; Zimprich et al., 2004; Healy et al., 2008; Simón-Sánchez et al., 2009; Abud et al., 2017; Islam and Moore, 2017). Animal models with core LRRK2 mutations are widely used in Parkinson’s research since LRRK2 has a druggable kinase domain and it is therefore considered a potential therapeutic target (Bolognin et al., 2019). Nevertheless, Bolognin et al. recently exploited microfluidics combined with high-content imaging technology to develop an advanced 3D in vitro model enabling pharmacogenomics in pathophysiological conditions (Fig. 3B). Intriguingly, the most penetrant disease phenotypes were a result of patients’ overall genetic background and were not solely dependent on the LRRK2-G2019S mutation (Bolognin et al., 2019).

Amyotrophic lateral sclerosis

ALS is a fatal neurodegenerative disease characterized by progressive degeneration of motor neurons, disturbed neuromuscular junction (NMJ) and muscle atrophy; in most cases death occurs within 3 years after ALS diagnosis due to respiratory failure (Centeno et al., 2018; Ionescu and Perlson, 2019). Osaki et al. (2018c) proposed a novel ALS-on-chip model, a motor unit, comprising 3D skeletal muscle bundles and optogenetic motor neurons from an ALS patient (Fig. 3C). The authors further enhanced their platform with an hiPSC-derived blood–brain barrier to study the CNS penetration of putative therapeutic agents (Osaki et al., 2018b). Future incorporation of mature neurons will mimic later stages of the disease; this may be accomplished either via genetic manipulation (Miller et al., 2013; Osaki et al., 2018c) or via direct differentiation of fibroblasts of ALS patients (Tang et al., 2017). Altman et al. (2019) developed a compartmentalized device to mimic the NMJ; their data depict the importance of mitochondrial accumulation for NMJ functionality (Altman et al., 2019). Replacement of the animal cells with hiPSC-derived populations might further enhance the translatability of the model and provide a platform to study the implication of mitochondria in the NMJ vulnerability in ALS.

We suggest that similar setups may be used to study neurodegenerative diseases that are associated with deterioration in muscle strength and motor skills, such as Alzheimer’s and Parkinson’s diseases (Boyle et al., 2009; Cano-de-la-Cuerda et al., 2010; Antony et al., 2013).

Traumatic brain injury

TBI is a severe health and socioeconomic problem (LaPlaca et al., 2005), and is also considered to be a risk factor for neurodegenerative diseases such as Alzheimer’s disease, Parkinson’s disease and ALS (Sivanandam and Thakur, 2012). In vivo models of TBI have mainly focused on its behavioural and systemic effects, whereas in vitro studies provide a powerful tool to deeply investigate the cellular mechanisms. Currently, the ability to accurately model TBI in engineered in vitro models is limited because of the low physiological resemblance of standard neural in vitro models and incompletely understood pathophysiology manifested in mechanical shear, twist and compression forces as well as subsequent hypoxia (Kumaria, 2017). LaPlaca et al. (2005) pioneered a reproducible model of in vitro TBI. By using an electro-mechanical cell shearing device, they were able to mechanically perturb a 3D neuronal-glial model, with deformation rates and magnitudes comparable to those that occur in inertial human head injuries (LaPlaca et al., 2005). Dollé et al. (2013, 2014) developed a brain-on-a-chip microsystem investigating neuronal response to a mechanical injury (Fig. 3D). The device allowed easy manipulation of the dimensions of the microchannels, which influenced the strain on either individual axons or bundles of axons, thereby suggesting that axonal diameter plays a significant role in strain injury, and thus in TBI. While there has been a great development in new in vitro models for TBI (Morrison et al., 2011), the combination of the advantages of engineered platforms with in vitro capabilities of monitoring cellular mechanisms and functional response to mechanical injury is a powerful tool for studying TBI responses and developing potential therapeutics (Shrirao et al., 2018).

Cancer

Despite tremendous efforts to identify putative treatments for brain cancer, most drug candidates fail in human clinical trials (Huszthy et al., 2012; Caragher et al., 2019; Sontheimer-Phelps et al., 2019). The complex tumour microenvironment in neural tissue is difficult to recapitulate; thus, most proposed compounds are inadequate for treating brain tumours (Sontheimer-Phelps et al., 2019). To increase the predictivity of current experimental platforms, it is necessary to mimic tumour dissemination, reduction and metastasis, cancer stem cell proliferation and differentiation, drug penetration across the blood–brain barrier and immune responses.

Animal modelling of primary brain tumours has been the gold standard in cancer research; however, translatability of the results tends to be questionable (Denayer et al., 2014; Mak et al., 2014). Patient-derived xenograft animal models maintain most of the biological characteristics of the original tumour, and they hold promise for translation to humans (Choi et al., 2018; Yada et al., 2018). Nevertheless, the mice used in these models are usually immunocompromised to prevent possible rejection of the xenograft, and immunity is neglected (Choi et al., 2018). Two-dimensional in vitro models, on the other hand, fail to recapitulate the tumour microenvironment, intercellular communication and tumour cell metastasis. Recent developments of vascularized brains-on-chips that incorporate tumorigenic cells may offer the possibility of improved translatability and effective drug discovery (Saliba et al., 2018; Wang et al., 2018; Sontheimer-Phelps et al., 2019).

OoCs may also provide a means of investigating the cascade of events that contributes to tumor metastasis from other organs to the brain (Fig. 4A), towards identifying novel therapeutics (Caballero et al., 2017). Lei et al. (2016) developed an OoC to study interactions between cortical neurons and cancer cells. The authors showed that functional neurites promoted cancer migration to the neuronal compartment, while perturbed neurites inhibited neuronal signalling cascades, cancer progression and metastasis (Lei et al., 2016). As a future development, advanced 3D microscopy and machine learning (Kingston et al., 2019) may soon augment the predictivity of similar tumour extravasation models.

Figure 4.

3D engineered in vitro models for brain cancer. (A) Brain cancer development. Cancerous tumours are classified into two main categories: primary tumours, which begin within the brain tissue (i) and secondary tumours, which arise due to metastasis from other organs, such as the breast, following a series of events as illustrated in ii. Servier Medical Art (SMART) was used for the illustration. (B) Primary tumours: GBM. (i) Construction of a bioprinted GBM-on-a-chip (Yi et al., 2019); (ii) Photographs of the GBM-on-a-chip from above (top) and the corner (bottom). Scale bar = 2 cm. The brain decellularized extracellular matrix (BdECM) bioink includes human umbilical vein endothelial cells (HUVECs; magenta) or GBM cells (blue). (iii) Phase-contrast (left) and fluorescent image (right) of the GBM-on-a-chip. GBM cells are stained with DiI (red) and HUVECs with DiO (green). Scale bar = 200 μm. (C) OoCs to model cancer metastasis to the brain (i–iii) A physiologically relevant blood–brain barrier device (Xu et al., 2016a). The device consists of 16 independent functional units connected via microchannels (i, left). Detailed view of each functional unit (i, right). Magnified view (ii) and side view (iii) of the blood–brain barrier region composed of BMECs, astrocytes and ECM. The red arrow indicates the flow direction. (iv–vi) A multi-organ microfluidic chip to model lung cancer metastasis (Xu et al., 2016c). (iv) Schematic of lung cancer metastasis to distant organs including the brain. (v and vi) 3D cell cultures of different organs in distinct chambers. Lung cancer cells (A549) flow through the media in the microvascular channel (red) to mimic cancer metastasis via the blood vessels. (D) Blood–brain barrier-on-chip device to investigate metastatic brain tumours (Xu et al., 2016a). (i) Time-lapse imaging of different cancer cell types (green) across the blood–brain barrier via the vascular compartment. Cell extravasation to the brain was monitored for 72 h. Lung cancer cells (A549), breast cancer cells (MDA-MB-231) and melanoma cells (M624) disrupted the integrity of the blood–brain barrier and migrated to the brain, whereas liver cancer cells (BEL-7402) did not. (ii) Functional responses of the blood–brain barrier to therapeutic agents. The glioblastoma cells (U87) showed a dose-dependent response to the lipophilic and blood–barrier-permeable medication temozolomide, which was added to the vascular compartment of the chip. Green = live cells; red = dead cells. Scale bar = 25 μm. Figure components are modified from Xu et al. (2016a, c) and Yi et al. (2019) with permission.

Xu et al. (2016a) used a blood–brain barrier-on-a-chip to model tumour invasion of the brain via a disrupted barrier. Their study confirmed the synergic role of astrocytes and endothelial cells in maintaining barrier integrity, as well as the prohibitive role of astrocytes in cancer metastasis. The authors later extended their setup to incorporate a multi-chamber device accommodating organ-specific cell types. They used this system to study lung cancer metastasis to several organs, including the brain [Fig. 4C(iv–vi)] (Xu et al., 2016c). In a subsequent study, Xu and colleagues further addressed effects of inflammatory microvasculature on tumour extravasation, and they showed that TNFa-induced inflammation increases adhesion of adenocarcinoma cells to the inflammatory endothelium (Xu et al., 2017).

Additional in vitro models have sought to model solid brain tumours, which are characterized by high heterogeneity, obstruction of solute production, high waste accumulation, and a hypoxic inner microenvironment (Sleeboom et al., 2018; Wang et al., 2018). 3D scaffolds (Gomez-Roman et al., 2017), organoid-based 3D models (Lancaster et al., 2013; Hubert et al., 2016; Ogawa et al., 2018), bioprinting (Heinrich et al., 2019), glioblastoma multiforme (GBM)-on-chip (Ayuso et al., 2017), and combinatorial methods have been used for this purpose. A recent effort by Yi et al. (2019) describes a bioprinted human GBM-on-chip, which demonstrated patient-specific sensitivity to putative therapeutic agents (Fig. 4B). Shen et al. (2017) used a 3D in vitro platform for quantitative high-throughput screening and showed that several antiparasitic agents may have therapeutic potential for paediatric solid tumours. Incorporation of immune cells in these models might further elucidate the complex interplay among microglia and/or macrophages and tumour cells (Roesch et al., 2018; Sevenich, 2018).

High scalability and ease of use are crucial for effective and accurate cancer disease models in vitro. Phan et al., 2017 presented a novel and versatile OoC drug screening platform, requiring minimal equipment and no external pumps for flow generation. Custom fitted vascularized microtissues on a standard 96-well plate have been used for large-scale drug screenings (Fig. 4D), enabling anti-angiogenic and anti-cancer drugs to be detected (Phan et al., 2017). We envision that hiPSC-based, vascularized cancer-on-chips combined with advanced 3D imaging may pave the way for patient-specific therapeutics.

Host–pathogen interactions

Various pathogens such as bacteria, viruses, and fungi may cause life-threatening infections of the human CNS (Giovane and Lavender, 2018). The increased morbidity and mortality rate of these cases demand accurate disease models for infectious diseases. Upon CNS infection, multiple cells and tissues engage in a vivid interplay while innate and adaptive immune responses lead to cytokine overproduction widely known as ‘cytokine storm’ (Tisoncik et al., 2012; Koyuncu et al., 2013). This process exposes the neural tissue to a cascade of biological phenomena with tremendous long-term sequelae.

Scientists use animal models to mimic the complexity of infectious disease phenotypes (Swearengen, 2018). Even though these models have shed light on various aspects of disease initiation, progression, and manifestation, they often lack translatability to humans (Mestas and Hughes, 2004; Seok et al., 2013; Costamagna et al., 2019). In the case of neurotropic viruses, for example, animal models are challenged by the versatility and the complexity of the intruders. Their translatability is limited by (i) low reproducibility of subtle neurological clinical phenotypes; (ii) the use of transgenic mice to induce severe clinical phenotypes; (iii) inability to reproduce indirect disease mechanisms, which often enhance CNS virulence (e.g. inflammation); and (iv) differential viral infection kinetics due to animal-specific characteristics that result in faster virus clearance (Natoli et al., 2020). On the other hand, hiPSCs are susceptible to human pathogen infections (Costamagna et al., 2019); thus, hiPSC-based infectious disease models may have higher translational value.

In the past, hiPSC-derived brain organoids have been used extensively to model host-virus interactions in the human CNS. Zika virus (ZIKV) is associated with Guillain-Barré syndrome as well as congenital infection and microcephaly (Aragao et al., 2016; Qian et al., 2017; Costamagna et al., 2019). Several efforts have now established hiPSC-based brain organoids as an effective method to investigate ZIKV-induced microcephaly (Cugola et al., 2016; Garcez et al., 2016; Qian et al., 2016). ZIKV impaired the overall growth of the organoids; specifically, the infection resulted in neural precursor cell death and disruption in neurosphere formation. Moreover, Xu et al. (2016b) used forebrain-specific hiPSC-derived organoids to perform a big screen with the prospect of drug repurposing against ZIKV. This work led to the identification of compounds that inhibited either the ZIKV replication or the neural precursor cell death (Xu et al., 2016b). Herpesviruses are often responsible for CNS infections in humans, which mostly result in encephalitis, meningitis, or myelitis (Bulakbasi and Kocaoglu, 2008). Herpesvirus CNS infections exhibit dramatic sequelae especially in immunocompromised or elderly patients and newborns (Costamagna et al., 2019). Unfortunately, and despite early antiviral drug administration, infection of newborns results in encephalitis and high mortality (Kimberlin, 2004). Researchers have used hiPSC-derived cerebral organoids to mimic herpes simplex virus 1 (HSV-1) and cytomegalovirus (CMV) infection in vitro (D’Aiuto et al., 2018, 2019; Brown et al., 2019; Sun et al., 2020). These efforts have greatly advanced our understanding of disease development after infection and manifest the value of using human 3D organoid cultures to recapitulate virus-induced pathologies in the developing brain.

Meningococcal meningitis, caused by Neisseria meningitidis bacteria is a devastating CNS infection associated with increased mortality and frequent severe sequelae (WHO, 2018). N. meningitidis is a human-specific pathogen (Hodeib et al., 2020), thus animal models fail to replicate the pathophysiology of the disease. Martins Gomes et al. (2019) used hiPSC-derived BMECs as a cellular model of N. meningitidis infection. The authors showed barrier permeability, tight junction disruption, and bacterial transmigration into the CNS via the hiPSC-BMECs after N. meningitidis infection. We hypothesize that engineered in vitro models of the human NVU coupled with hiPSC technology might further elucidate the mechanisms underlying the pathophysiology of meningococcal meningitis.

Accumulating clinical evidence shows that respiratory viruses can escape the immune responses of the human body and not only cause severe respiratory issues but also migrate to other organs including the CNS (Vareille et al., 2011; Desforges et al., 2019). The infection of the resident neural populations may then lead to several pathologies including encephalitis (Bohmwald et al., 2018). Like other neuroinvasive viruses, human coronaviruses (hCoVs) can enter the CNS via the hematogenous route or the neuronal retrograde route (Desforges et al., 2019). After studies in both animals and microfluidic devices, Desforges et al. (2019) proposed a putative mechanism that the human coronavirus OC43 (HCoV-OC43) invades the CNS. Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) causes the coronavirus disease 2019 (COVID-19) and has led to the ongoing global pandemic (WHO, 2020). The COVID-19 outbreak is at a very early stage; nevertheless, clinical data support the neurotropic character of the virus as several patients suffer from neurological symptoms (Alvin et al., 2020; Mao et al., 2020; Natoli et al., 2020). Both neurons and endothelial cells express the angiotensin-converting enzyme 2 (ACE2) receptor, which hCoVs utilize to enter the cells (Li et al., 2003). SARS-CoV-2 binds the cell membrane via the CoV spike glycoprotein, which in turn binds ACE2 with high affinity. According to Natoli et al. (2020) this might explain the higher neuroinvasive potential of SARS-CoV2 compared to the previous SARS-CoV.

Effective mimicking of the complex CNS cytoarchitecture and cell responses due to SARS-CoV infection is pivotal to perform mechanistic studies, investigate the time course of infection, and disease progression. Beyond any doubt, patient-oriented disease models set the basis for the development of effective therapeutics. The existing mouse models fail to recapitulate human CNS infection and disease progression sufficiently; therefore, the development of humanized in vitro models is critical. Structural differences between ACE2 mouse and human proteins result in the poor tropism of SARS-CoV for mouse tissue; hence, the virus cannot infect mice efficiently (Cleary et al., 2020; Wan et al., 2020). The K18-hACE2 transgenic mice, first proposed by McCray et al. (2007), and now offered by Jackson laboratory, express the human ACE2 receptor and provide us with a valuable tool to study the course of SARS-CoV infection in humans. In these mice, however, the expression of ACE2, driven by the human keratin 18 (KRT18) promoter, is limited to the epithelia. Thus, K18-hACE2 mice are not useful to investigate the effects of the infection on the human brain endothelium. A recent study by Varga et al. (2020) highlights the implication of the endothelial cell layer in various tissues from COVID-19 patients. Moreover, Monteil et al. (2020) have recently shown with the use of human organoids as a model system that human recombinant soluble ACE2 (hrsACE2) but not mouse recombinant soluble ACE2 (mrsACE2) can significantly reduce SARS-CoV-2 infection. Certainly, SARS-CoV-2 research is at an immature state and the involvement of the CNS in disease progression is largely ambiguous (Li et al., 2020a, b; Turtle, 2020). Nevertheless, the severity of the disease and the detrimental impact on our society necessitates the development of accurate, patient-oriented disease models, which emulate the overall disease phenotype including the neurological symptoms (Zhou et al., 2020).

We believe that OoCs offer limitless opportunities for such studies. We envision that linked microdevices might serve as putative platforms to study how systemic inflammation due to respiratory infection contributes to neurological deficits and vice versa. Researchers can use OoCs to unravel how viruses use immune cells as ‘trojan horses’ to surpass the blood–brain barrier and transmigrate to the neural tissue. We foresee that the use of hiPSC-based OoCs may help the scientific society to surpass the translational limits of the animal disease models and pave the way towards improved diagnostics and patient-oriented therapeutics.

Neurotoxicity

For the pharma industry and regulatory authorities there is a growing need to increase the output and efficacy of current toxicity testing and to minimize the use of animals (Krewski et al., 2010; Crofton et al., 2011). In particular, the scientific community is striving to minimize animal use by using in silico studies, such as pharmacokinetic and pharmacodynamic models and quantitative structure activity relationship (QSAR) models (Raies and Bajic, 2016), coupled with advanced in vitro models.

Neurotoxicity testing is multifaceted, focusing not only on whether a potential drug can penetrate the blood–brain barrier but also the vulnerability of the CNS. The developing CNS is particularly sensitive and goes through several highly orchestrated processes such as cell migration, proliferation, and differentiation, and subsequent cell populations develop processes such as synaptogenesis, cell pruning, and myelination. Hence, the conventional dose-response battery of tests needs to factor in the developmental phase when evaluating a potentially harmful chemical (Rice and Barone, 2000; Giordano and Costa, 2012). Alarmingly, many drugs that are on the market have not undergone developmental neurotoxicity testing (Grandjean and Landrigan, 2014; Judson et al., 2014; Schwartz et al., 2015; Tohyama, 2016).

Specific OoC demonstrations of neurotoxicity have not been reported; instead, organoid approaches have so far been used. In 2012, Eiraku and colleagueswere the first group to consider 3D brain organoids (Eiraku and Sasai, 2012) as potential tools for investigating neurotoxicity, even if lacking vasculature and microglia. Schwartz et al. (2015) reported neural constructs encompassing vasculature, albeit not functional, and microglial cells that, in combination with machine learning approaches, resulted in a model that could identify 9 of 10 toxic chemicals in a blinded trial. Future advances for developing organoids for neurotoxicity testing steer towards generating organoids with a functional blood–brain barrier, possibly combined with OoC technology to increase bio-fidelity. Organoid culturing can indeed be a stepping stone towards recapitulating brain development in vitro that exhibits not only distinct developmental phases but also higher brain functions.

Applicability of engineered in vitro models in neuroscience settings

The discussion above points to the vast potential of engineered in vitro systems to enhance our understanding of the CNS in health and in disease. Notably, adoption of such systems in commercial and academic labs may ultimately also be cost-effective: for example, continuing to develop OoC devices for drug screening is expected to reduce research and development costs for each new drug by 10–26%, within a timescale of 5 years (Franzen et al., 2019). Below, we elaborate on some of the considerations that labs should take into account when considering the adoption of an engineered in vitro model system.

There are several channels through which a commercial or academic neuroscience lab can adopt new engineered CNS in vitro models: (i) collaborating with model developers; (ii) fabricating models in-house; or (iii) purchasing commercial systems. When deciding which of these approaches to take, it is important to consider the complexity and portability of the system in relation to the biological question that is to be addressed.

Model fabrication in-house or with the aid of collaborators

Biological considerations

The constraints of biological entities such as cells, tissue slices and native structure of proteins should guide the design of a specific engineered device. Technically, geometry, porosity, mechanical parameters (such as stiffness of the surface), topography, transparency, thickness and material chemistry are just a few of the parameters to be considered in fabricating a microfluidic in vitro device (Kilic et al., 2016; Vernetti et al., 2017).

Microfabrication techniques and materials

Currently, microfabrication techniques that are suitable for fabricating engineered devices for a tissue-specific environment at the microscale, and that can easily be implemented in an academic lab, include soft lithography, photolithography, direct laser patterning and microcontact printing and 3D printing (Campisi et al., 2018). All these techniques allow for the development of devices at reasonable costs, and can also serve to reduce the amounts of chemicals and biological materials used compared to large-scale plate formats (Wang et al., 2018).

One of the most easily implemented materials in a non-engineering lab is poly(dimethylsiloxane) (PDMS). Templates or moulds defining the channels and shapes of the device can be generated to order by central foundries or by companies at low cost. PDMS is then simply cast in the mould and cured at ∼80°C. Additional treatments, such as oxygen plasma, may be needed for assembling separate parts into a complete device. An oxygen plasma chamber can be acquired by a non-engineering lab at relatively low cost. The reason to choose PDMS for small-scale (e.g. academic) chip fabrication is that it is the most widely used material to fabricate microfluidic devices, and it allows for fabrication easily adopted by non-experts, transparency, biocompatibility and gas permeability, essential requirements in developing engineered in vitro models. Specifically, gas permeability enables cells or tissues to be kept alive even for long-term experiments, while transparency allows cellular morphology to be monitored using various high-quality microscopy techniques (Park et al., 2015; Kilic et al., 2016). Yet, PDMS-based devices also suffer from several shortcomings. Primary limitations include absorption of small molecules and diffusion of water vapour, which affect concentrations of components in the cell culture, such as certain amino acids, vitamins, neurotransmitters and growth factors (Toepke and Beebe, 2006), as well as drugs, such as verapamil and nifedipine (Berthier et al., 2012; van Meer et al., 2017).

To overcome this limitation, Ingber and colleagues began to use polyurethane in fabricating polymer-based devices, which preserve PDMS properties, such as flexibility and transparency, without absorbing small hydrophobic molecules (Zhang et al., 2017). Another potential candidate is styrene-ethylene-butylene-styrene (SEBS) copolymers, whose use in constructing devices was patented by Emulate (Huh et al., 2011; Zhang et al., 2017). Polyetherimide (PEI), polycarbonate, silicon, glass, silk protein, and agarose are additional materials that might potentially be used. Importantly, many of these materials can be processed in non-engineering settings because of the relatively large dimensions (>100 mm) of OoCs. More detailed discussion about the production methods for engineered in vitro brain can be found in other relevant papers (Haring et al., 2017; Sosa-Hernández et al., 2018; Yu and Choudhury, 2019), and filmed chip fabrication protocols can be found in some recently published studies (van der Helm et al., 2017; Novak et al., 2018; Jagadeesan et al., 2020).

System setup

When an academic lab has chosen the chip design and material based on a specific biological question, the next step is to introduce a sterilization method, surface coatings, cells or tissue as well as a fluidic interface. For the latter, depending on budget and model design, there are a variety of solutions using external pumps and tubing or pumpless, gravity and diffusion driven solutions. Indeed, when developing and designing a highly accurate OoC system, fluid control (directing, monitoring, controlling) should be a key focus, particularly since the high surface-to-volume ratio necessitates frequent renewal of the cell culture medium. Different types of systems have been developed to accurately deliver and control flow in microfluidic devices (e.g. peristaltic and recirculating pumps (Skafte-Pedersen et al., 2009); pressure-control systems (Heo et al., 2016) or syringe pumps (Kuczenski et al., 2007; Chen et al., 2018). Furthermore, several recent studies have focused on reproducing physiological blood flow changes, implementing OoCs to generate laminar (Zheng et al., 2012), pulsatile (Shao et al., 2009) or interstitial flow (Kingsmore et al., 2016; Kaarj and Yoon, 2019).

Some of the parameters that need to be taken into consideration when developing new OoC devices are summarized in Box 1.

Box 1.

Overview of some important parameters when developing a new OoC device

ADVANTAGES	LIMITATIONS
HUMAN TISSUE SOURCE
Embryonic stem cells (ESCs) a , b
Unlimited differentiation potential  More consistent phenotype  Easier to obtain and last longer in culture  Potential to recreate multiple organ-like structures	Ethically controversial (they derived from human embryos) Difficult to create large numbers of genetically diverse cell lines Variability in efficiency of differentiation protocols Difficult to differentiate into distinct, mature cell phenotypes Low efficiency in generating neuronal subtypes Lack of native 3D tissue structure High time and cost when designing OoC devices
Human induced pluripotent stem cells (hiPSCs) a , b , c , d , e , f
No ethical concerns (they derive from adult tissue)  Defined disease phenotypes  Ideal and unlimited source of cells  Patient-specific  Possibility to expand and differentiate into multiple lineages  Genetic homogeneity  Ideal for target-specific drug development  Low preclinical research time	Low efficiency in generating specific neuronal subtypes Lack of native 3D tissue structure High time and cost associated when designing OoC devices Difficult to develop and achieve complete maturation Lack of robust protocols for their differentiation and maturation Availability of patient-specific human cells Limitation in accurate mimicking of human organs Limitation in reproducing cell-cell interactions
Tissue biopsies b , f , g
Derived directly from adult tissue  Maintaining some of the natural ECM and 3D tissue structures	Do not survive more than 48 h Lack of cell proliferation and of human tissue sources
Cell lines a , b , f , h
Widely available and facile handling  Easy to culture and economical  High proliferation under simple culture conditions  Useful in optimizing parameters during OoC development	Lack of natural extracellular matrix Lack the patient-specificity Not accurately recapitulate tissue function Lack the phenotypic function characteristic of the organ they intend to represent
FLOW MANIPULATION
Microfluidic systems f , h , I , j , k , l , m
High reproducibility and sophisticated fluid manipulation  Ideal in mimicking the dynamic cellular environment  Able to sustain complex microfluidic gradients for long time  Can replicate the complexity and interconnectivity of real organs  High throughput and low reagent consumption  Spatial control of liquid composition at subcellular resolution	Presence of air bubbles Laminar flow only produces relative slow diffuse mixing Difficulty in fluid handling
MATERIALS: BIOCOMPATIBLE POLYMERS
Polydimethylsiloxane (PDMS) e , f , h
Transparent and excellent flexibility  Biocompatibility, oxygen permeability and low cytotoxicity  Low cost and easy of processing	Drug adsorption and highly hydrophobic Not degradable Not scalable, due to its softness and elasticity
Poly(methyl methacrylate) (PMMA) n , o , p
Reduce drug, protein or small molecule absorption/adsorption  Can improve the robustness of the OoC during long operations  Low cost, easy to fabricate and manipulate  Low auto-fluorescence and excellent transparency	Affected by important solvents used in microfabrication and sterilization Barely permeable to gas
Polycarbonate (PC) n , p
Transparent and Low cost  High heat resistance and high stiffness and strength	Barely permeable to gas Poor resistance to certain organic solvents
FABRICATION TECHNIQUES
Photolithography e , h , q
Cells can be cultured directly on the patterned materials  Hydrogels can be incorporated, to promote cell seeding and include a physiological ECM environment	Pattern resolution is limited by the light diffraction Expensive and time-consuming Not possible the direct insertion of specific materials (e.g. ECM)
3D printing h , p , r
Cells can be printed continuously and accurately  Controllable resolution, high printing speed, rapid technique and low material costs  Can incorporate proliferation and differentiation cues  Versatile technique able to reproduce 3D geometry  Able to integrate mechanical and electrical sensors	Sometimes, slow printing speeds, not useful for larger tissues or organ printing Low spatial resolution and cellular perturbation Cross-linking: potentially cytotoxic factors, High viscosity of some biomaterials Multiple treatment session with limited micro size precision
Microcontact printing s
Low cost and rapid prototyping	Difficulty in controlling the ink and the surface robustness
Laser-based patterning s
Cells and any particles can be manipulated	Large instrumentation, complex setup
Injection moulding g
Mass production  Low cycle time and highly automated	Restricted to thermoplastic High costs for moulds and complex moulding equipment
Casting u
Process, equipment setup and replication accuracy	Long process time (e.g. labour and lab costs)
CHIP DESIGN
2D system v , w
Study of cell behaviour using simple technologies  Universally known and several protocols available  Simple realization and low cost	Does not adequately represent the natural 3D environment Does not properly reproduce in vivo conditions
3D system v , w
3D architecture very close to in vivo model	Very complex and expensive to build and to control

^a Ronaldson-Bouchard and Vunjak-Novakovic, 2018;

^b Wnorowski et al., 2019;

^c Takahashi et al., 2007 b;

^d Burridge et al., 2016;

^e Cavero et al., 2019;

^f Jodat et al., 2018;

^g Luni et al., 2014;

^h Ahadian et al., 2018;

ⁱ Andersson et al., 2004;

^j Dittrich and Manz, 2006;

^k Kang et al., 2008;

^l Velve-Casquillas et al., 2010;

^m Sivagnanam and Gijs, 2013;

ⁿ Ren et al., 2013;

^o Gencturk et al., 2017;

^p Rodrigues et al., 2017;

^q Chapanian and Amsden, 2010;

^r Nahmias et al., 2005;

^s Martinez-Rivas et al., 2017;

^t Fiorini and Chiu, 2005;

^u Becker and Gärtner, 2008;

^v Osaki et al., 2018b;

^w Coluccio et al., 2019.

Importantly, a new OoC user should be informed of simple strategies to avoid air bubble formation, contamination and flow disturbances in their specific systems. We recommend visiting an OoC lab or participating in workshops to acquire the expertise needed to set up an in-house microfluidics system.

Commercial engineered tissue culture devices

Because of the promising results of novel in vitro models, several approaches to commercializing engineered tissue culture devices are ongoing (Zhang et al., 2017). Some of the leading developments in commercial microfluidic devices used for recapitulating brain tissues are summarized in Table 3 and Supplementary Fig. 3. GlaxoSmithKline, BASF, Sanofi and AbbVie funded MIMETAS B.V. (https://mimetas.com), an OoC company with a multiple-chip solution integrated on one plate, and which has recently begun to develop a microfluidic 3D cell culture system for neurotoxicity screening (Wevers et al., 2016; Kane et al., 2019). The Wyss Institute for Biologically Inspired Engineering at Harvard University is collaborating with the start-up Emulate, Inc. (Wyss Institute, 2014) for commercializing its OoC technology (Sances et al., 2018). TissUse GmbH is another company that, together with the HiPSTAR consortium, began to develop a new in vitro model of the human blood–brain barrier (https://www.tissuse.com/en/news/press-releases/); their work focuses on identifying new drugs and therapies targeting dementia. Hesperos (https://hesperosinc.com/technology/) commercialized a new microfluidic hiPSC-derived blood–brain barrier model capable of mimicking in vivo blood–brain barrier characteristics; their platform includes reliable in vitro transport mechanisms and accommodates rate measurements for drug permeability screening (Wang et al., 2018; Ramme et al., 2019). In 2019, the company AxoSim, already a cutting-edge facility for modelling human physiology in vitro, acquired Organome (https://www.axosim.com) and subsequently developed and patented two novel platforms (the nerve-on-a-chip and the mini-brain organoid), expanding the possibility to address the growing burden of neurodegenerative diseases, both at central and peripheral nervous level (Huval et al., 2015). Additional commercial options include a 3D-cell culture chip developed by AimBiotech (https://www.aimbiotech.com/about-us.html), which gave rise to the development of a new 3D blood–brain barrier model replicating the in vivo neurovascular organization (Campisi et al., 2018), offering a new platform for drug discovery. Additional manufacturers that warrant attention include Synvivo (https://www.synvivobio.com), Xona microfluidics (https://xonamicrofluidics.com/) and Ananda (https://anandadevices.com/application/extension-of-axons-isolated-using-anandas-neuro-device/); these companies have successfully created novel microfluidic devices to direct neuronal growth and axonal extension (Deosarkar et al., 2015; Magdesian et al., 2016; Paranjape et al., 2019).

Table 3.

Commercial OoC or chip providers

Developer	Engineered devices	Strengths	Limitations
MIMETAS the organ-on-a-chip company	OrganoPlate®, a microfluidic 3D culture plate, made of 96 independent microfluidic chips Each culture cells contain a perfusion channel It can be used as culture system for different cell types, including neurons, endothelial cells and organoids	Cell can be embedded within the hydrogel, resembling the parenchymal space and mimicking the vascular interface Versatile Possibility to grow 3D-culture system Highly compact and higher throughput than competitors	Only operational with bi-directional flow
EMULATE	Dual-channel microfluidic chip able to recreate the body’s dynamic cellular microenvironment (e.g. tissue-tissue interaction, blood flow and mechanical forces)	Presence of two microfluidic channels with the possibility to culture two different types of cells Possibility to modulate and mimic various tissue specific fluid conditions	Made of PDMS Cell-to-liquid and surface-to-volume ratio Not easily adapted to high-throughput assays
TissUse Emulating Human Biology	HUMIMIC Chip 4, microfluidic four-organ-chip devices, includes two separate microfluidic process, designed to host intestinal, liver, renal and brain cultures	Built-in micropump driven by an external pneumatic controller Constructed of thermoplastic while PDMS is restricted to a thin membrane Open tissue chamber and separated from the fluid channels Possibility to combine with different tissue assembly method	Tissue volume scaling and cell-to-liquid ratio Not easily adapted to high-throughput assays
AxoSim	Microengineered nerve-on-a-chip device enabling the growth of 3D neural fibres bundles for peripheral neurotoxicity and physiological testing	Ideal for clinical nerve compound action potential (CAP) and nerve fibre density (NFD) tests 3D in vitro system Successfully adapted for electrophysiological recordings	Only tested on rat tissue explants Not easily adapted to high-throughput assays
SynVivo	SynBBB, 3D OoC model, allowing real-time studies of cellular behaviour, drug delivery and drug discovery, closely mimicking in vivo cellular microenvironment	Possibility to maintain and image the micro-vessel for long periods of time Tissue compartment and microvascular channels that mimic the 3D-morphology of in vivo microvessels Porous interface that replace the use of membranes	Size of the micro-channels Not easily adapted to high-throughput assays
Xona® microfluidics	XonaChips®, multicompartment microfluidic chip, allowing neuron cell culture It offers the ability to isolate and grow axons, for specifically studying neuronal response to axonal damage, in an isolated fluidic environment	Made of cyclic olefin copolymer and no autofluorescent Ideal hydrophilic surface for attachment and growth of stem cells with the possibility of co-cultures	Gas impermeable Not easily adapted to high-throughput assays
anandaTM	The Neuro Device enables to pattern neurons and direct axonal extension It enables the growth of 100 axons with more than 1 mm in length	Possibility to be removed any time for direct manipulation of neurons Good for axonal extension measurements with the possibility of co-cultures	Made of PDMS and not easily adapted to high-throughput assays

In this context, it is also important to mention Ibidi (https://ibidi.com/content/34-ibidi-at-a-glance) and the microfluidic ChipShop (https://www.microfluidic-chipshop.com/). These two companies have become some of the leading providers of innovative functional microfluidic chips currently utilized from various academic research and pharma. Many of these commercial models are interesting options for industrial researchers, and some are also affordable for academic researchers.

Concluding remarks and outlook

Advanced in vitro platforms comprising human cells, and that can recreate integrated human physiological functions, have vast potential to contribute to our understanding of cellular mechanisms and the pathogenesis of neurological disorders. Engineered platforms based on microfluidics, such as OoCs, have emerged as a particularly promising in vitro technology, one that is also versatile and flexible enough to be integrated in a wide range of experimental settings (Huang et al., 2012). The main advantage of OoC systems is the possibility to mimic human organ physiology while observing individual cell systems in isolation, and in a manner that is highly reproducible as well as cost effective. Though in vitro techniques are unlikely to replace in vivo models as the gold standard for cancer research (at least in the foreseeable future), engineered in vitro platforms such as those discussed herein hold substantial potential for overcoming many of the limitations of animal models (e.g. questionable translatability to humans). Importantly, as emphasized herein, OoC technologies have matured to the point where CNS researchers in both academic and commercial research labs should be able to easily adopt them—either alone or in conjunction with alternative methodologies—to study basic physiological mechanisms, disease, as well as drug pharmacokinetic and pharmacodynamic models.

Funding

This research was supported by the Wallenberg foundation, Grant 2015-0178, KTH Tenure Track and Forska Utan Djurforsok (A.H.); Azrieli Foundation (B.M.M), The Israel Science Foundation Grant 2248/19 (B.M.M.) and The Brainboost grant (B.M.M).

Competing interests

The authors report no competing interests.

Supplementary material

Supplementary material is available at Brain online.

Glossary

ALS =

amyotrophic lateral sclerosis

BMEC =

brain microvascular endothelial cells

COVID-19 =

coronavirus disease 2019

ECM =

extracellular matrix

GBM =

glioblastoma multiforme

(h)iPSC =

(human) induced pluripotent stem cell

NVU =

neurovascular unit

OoC =

organ-on-chip

PDMS =

poly(dimethylsiloxane)

SARS-CoV-2 =

severe acute respiratory syndrome coronavirus 2

TBI =

traumatic brain injury

TEER =

transendothelial electrical resistance