Organotypic and Microphysiological Human Tissue Models for Drug Discovery and Development—Current State-of-the-Art and Future Perspectives

1. Hepatotoxicity

Drug-induced liver injury (DILI) constitutes the most common cause of acute liver failure, accounting for 52% of acute liver failure cases (Ostapowicz et al., 2002). The majority of cases are attributed to acute acetaminophen toxicity, whereas the remaining DILI events are caused by a multitude of different drugs, including antibiotics and nonsteroidal anti-inflammatory drugs, and manifest mostly after multiple days to months after the initial exposure. Events allotted to those drugs are individually rare with DILI risk estimates per medications ranging from 1 in 10,000 to 1 in 100,000 individuals (de Abajo et al., 2004; Bell and Chalasani, 2009; Björnsson, 2010). As a result, in the general population, annual DILI rates excluding acetaminophen are estimated between 1 in 5000 to 1 in 10,000 (Sgro et al., 2002; Bjornsson et al., 2013). In addition, hepatotoxicity is of major concern in drug development and remains a leading cause of drug failure in preclinical, clinical, and postmarketing stages (Watkins, 2011; Onakpoya et al., 2016). Importantly, however, the risk of failure due to safety is in strong anticorrelation with the confidence in the preclinical safety profile, indicating that thorough preclinical safety testing reduces the risk of safety-related project closures during expensive clinical stages (Cook et al., 2014).

2D monolayer cultures of PHHs have long constituted the cornerstone of in vitro hepatotoxicity testing. However, these models are short-lived and thus can, at best, mimic acute toxicity. To capture the chronic and delayed toxicity associated with most DILI culprit drugs, long-term stable cultures are needed. Given that hepatotoxicity is often tested relatively early in the preclinical pipeline, systems should allow for a relatively high throughput. Consequently, mainly sandwich cultures, spheroids, and MPCCs have been tested and benchmarked for their ability to predict DILI using a large set of drugs. In the following, we provide an updated overview of toxicological screens conducted in 3D human liver models. For other applications, such as mechanistic investigations, comparative studies, or toxicogenomics, we refer the interested reader to recent comprehensive reviews (Zhou et al., 2019; Zhang et al., 2020; Lauschke, 2021; Serras et al., 2021).

Sandwich cultures have primarily been used for investigations of cholestatic hepatotoxicity. Cholestatic toxicity is evaluated by coexposure of cultures with bile acid mixtures (i.e., a compound is classified as cholestatic if bile acids amplify its toxicity). Based on this principle, sandwich cultures could distinguish the cholestatic compounds cyclosporin A, nefazodone, troglitazone, perhexillin, chlorpromazine, bosentan, ticlopidine, and ritonavir and from eight hepatotoxins that exert toxicity by different mechanisms and nontoxic controls (Marion et al., 2007; De Bruyn et al., 2013; Chatterjee et al., 2014; Oorts et al., 2016). Furthermore, sandwich cultures have been established for hepatotoxicity screening irrespective of toxicity mechanisms (Fig. 2A). Evaluation of the acute hepatotoxicity of 344 compounds (200 DILI positives and 144 DILI negatives; 24 hours exposure) in 2D sandwich cultures of PHH using automatic imaging for mitochondrial damage, oxidative stress, and intracellular glutathione resulted in the correct classification of all nontoxic drugs and the identification of 101 of 200 hepatotoxins (sensitivity 51%, specificity 100%) (Xu et al., 2008). Notably, however, these predictions relied on the calibration of the individual cutoffs, and an application of the system to further independent test sets has not been performed.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 2

Organotypic liver models can faithfully predict human drug hepatotoxicity. (A) Representative image of a screen of 300 drugs and chemicals in which hepatotoxicity is assessed using high content imaging of markers for ROS, mitochondrial membrane potential (MMP) and intracellular glutathione (GSH) levels. Cells were treated for 24 hours at 100-fold therapeutic c_max. Nimesulide caused a significant increase in ROS, whereas telithromycin, nefazodone, and perhexiline caused a decrease in MMP and GSH. (B) In MPCCs based on hepatocyte-like cells derived from induced pluripotent stem cells, treatment with the four hepatotoxins acetaminophen, diclofenac, tolcapone, and troglitazone dose-dependently reduced albumin and urea secretion after 8 days repeated exposure, indicative of hepatocyte toxicity. In contrast, no decrease was detected for the nontoxic compounds aspirin, dextromethorphan, propranolol, and rosiglitazone. (C) Summary result of the hepatotoxicity test of 123 drugs (70 hepatotoxic, 53 nonhepatotoxic) in primary human hepatocyte spheroids. The model detected dose-dependent toxicity and, at 20-fold c_max concentration, successfully detected 69% of all hepatotoxic compounds without any false positive hits. Error bars indicate S.D. **, ***, and **** indicate P < 0.01, P < 0.001, and P < 0.0001, respectively. (D) Hepatotoxicity of acetaminophen and methotrexate (MTX) was evaluated using a perfused liver chip system. Increased ROS levels (magenta) were detected at acetaminophen concentrations of 10 mM, whereas MTX-induced toxicity manifested as intracellular lipid droplets (yellow) and αSMA (green). Figure modified with permission from (Jang et al., 2019a; Berger et al., 2015; Vorrink et al., 2018; Xu et al., 2008).

Evaluation of drug hepatotoxicity of 13 compounds after short-term exposure identified most toxic compounds apart from acetaminophen, whereas the nontoxic controls aspirin and caffeine were successfully identified as nontoxic (Khetani and Bhatia, 2008). Using a multiplexed toxicity assay comprising measurements of ATP, urea, albumin, and glutathione, exposure of MPCCs of PHH for 9 days successfully flagged 23 out of 35 compounds as hepatotoxic, whereas 9 out of 10 DILI negative compounds were identified as true negatives (sensitivity 66%, specificity 90%) when using the ratio of IC₅₀ and total plasma C_max for classification (Khetani et al., 2013). Changing the cell model to iPSC-derived HLCs, the four prototypical hepatotoxins acetaminophen, diclofenac, tolcapone, and troglitazone were successfully identified as hepatotoxic after 8 days exposure, whereas the nontoxic controls aspirin, dextromethorphan, propranolol, and rosiglitazone were flagged as nontoxic (Fig. 2B). Further extension of this model resulted in the identification of 24 out of 37 hepatotoxic drugs, whereas all 10 nontoxic drugs were correctly classified (sensitivity 65%, specificity 100%) (Ware et al., 2015).

Spheroids have emerged as the predominant 3D model for the prediction and mechanistic analysis of DILI. To analyze the hepatotoxic liability of compounds, most earlier studies used spheroids generated from cell lines, such as HepG2 (Mueller et al., 2011; Fey and Wrzesinski, 2012; Ramaiahgari et al., 2014; Gaskell et al., 2016) and HepaRG (Leite et al., 2012; Gunness et al., 2013; Mueller et al., 2014; Wang et al., 2015, 2016; Ott et al., 2017; Ramaiahgari et al., 2017; Liu et al., 2018a). However, these studies only evaluated small sets of compounds (<20) and thus do not allow firm conclusion about their predictive accuracy. For spheroids generated from iPSC-derived HLCs, two smaller screens have been published that evaluated the toxicity using single-dose exposures. In a study of 24 compounds (20 DILI-positive and 4 DILI-negative), HLCs were found to perform better than HepG2 cells but remained less sensitive than PHH monolayer cultures (Takayama et al., 2013). In a separate study using 48 (42 DILI-positive and 6 DILI-negative) compounds, evidence of toxicity was identified in 36 of the toxic drugs, whereas the 6 nontoxic compounds remained negative (Sirenko et al., 2016). However, the choice of test compounds was problematic, as the DILI-positive set included various compounds with general cytotoxicity (anthracyclines, taxanes, alkaloids, and platinum compounds), whereas the negative set comprised carbohydrates (sucrose and sorbitol) and commonly used media supplements (penicillin and streptomycin).

For PHH spheroids two larger screens based on simple high-throughput compatible ATP endpoints were published that used compound sets based on regulatory classifications. Two weeks treatment of spheroids cultured in undisclosed media compositions at 100 times the therapeutic C_max of 110 drugs (69 DILI-positive and 41 DILI-negative) resulted in a sensitivity of 59% and a specificity of 80% (Proctor et al., 2017). In contrast, spheroids cultured in conventional ULA plates in chemically defined conditions that were exposed for 2 weeks to 123 compounds (70 DILI-positive and 53 DILI-negative) at 20 times C_max further improved the predictive accuracy to 69% sensitivity and 100% (Vorrink et al., 2018) (Fig. 2C). Notably, spheroids detected toxicity irrespective of DILI pattern and underlying toxicity mechanism, including mitochondrial toxicity and cholestatic toxicity (Hendriks et al., 2016; Hendriks et al., 2019).

Hepatotoxicity has furthermore been evaluated in cocultured spheroids of PHHs and NPCs wherein the latter appear to modulate compound toxicity, likely at least in part by impacting drug metabolism. For instance, NPC coculture resulted in lower expression of CYP2E1 and CYP3A4 and reduced glutathione depletion and acetaminophen hepatotoxicity (Bell et al., 2020). Interestingly, Kupffer cells appear to exert differential effects on compound toxicity. Although Kupffer cell cocultures corroborated the protective effects on acetaminophen, they amplified trovafloxacin toxicity (Li et al., 2020). Thus, these results suggest that liver spheroids can provide useful tools to investigate the effects of specific liver cell types or their subpopulations on hepatocyte function and drug toxicity.

Hepatotoxicity prediction using microphysiological liver systems is making progress, and proof-of-concept studies have demonstrated that such approaches can reveal drug effects on metabolic flux dynamics (Ehrlich et al., 2018) and identify compounds with a propensity for idiosyncratic DILI (Novik et al., 2017; Jang et al., 2019a) (Fig. 2D). However, at present, available small-scale comparative data of microphysiological liver systems compared with spheroids and sandwich culture indicate that the sensitivity of microfluidic devices is similar (Foster et al., 2019) or lower (Rubiano et al., 2021). Thus, in the absence of larger screens, it remains thus unclear whether microfluidic systems have increased or decreased predictive accuracy compared with other organotypic liver models.

Overall, across 3D liver models, most hepatotoxicity data are available for spheroids. Acetaminophen, amiodarone, chlorpromazine, and fialuridine are the most tested compounds and might thus provide suitable reference values for the comparison of platform sensitivities (Table 1). Notably, commercial providers seem to focus predominantly on microfluidic chips, likely at least in part because of the substantial revenue associated with the sales of associated devices and equipment. Although these systems enable the study of gradients and liver zonation, the available data do not seem to indicate that these technically more demanding setups currently translate into improved phenotypes or increased predictive power, and further unbiased studies are required to comprehensively benchmark the available models.

2. Pharmacokinetic Studies

For approximately 70% of small-molecule drugs, metabolism is the main route of drug elimination, which mainly occurs in the liver (Wienkers and Heath, 2005). Therefore, hepatic intrinsic clearance (CL_int) prediction, identification of metabolites, and delineation of biotransformation routes constitute key elements of drug discovery and development pipelines. Currently, mainly microsomes and primary hepatocytes in suspension culture are used for pharmacokinetic profiling. However, stable metabolic activity in both microsomal preparations and hepatocytes in suspension is limited to several hours, making measurements of CL_int for slowly metabolized compounds difficult. For the analysis of low-clearance compounds, MPCCs and, to a lesser extent, microfluidic chips and spheroid systems have been presented (Table 2).

Using PHHs in MPCCs, CL_int could be predicted for 72%–76% within 3-fold of in vivo, which further increased to 92% if correction for protein binding was only integrated for low-clearance compounds (Qian et al., 2016; Chan et al., 2013). When stratifying by the extent of metabolism, clearance prediction was within 2-fold for 70% of 10 tested low-clearance drugs, whereas 57% of 7 tested intermediate-clearance drugs were within the same interval (Chan et al., 2013). Comparisons of MPCCs with suspension and 2D cultures indicated that MPCCs considerably reduced confidence intervals and exhibited superior clearance predictions except for uptake transporter substrates for which 2D cultures showed the best performance (Lin et al., 2016; Da-silva et al., 2018; Docci et al., 2019, 2020; Umehara et al., 2020). Furthermore, using MPCC PHH data as input for pharmacokinetic modeling using the well stirred model, 75% of compounds were predicted within 3-fold, whereas clearance measurements were considerably lower in HepG2 and iPSC-derived HLCs (Kratochwil et al., 2017). The system moreover allows profiling the responsible drug-metabolizing enzymes of slowly metabolized compounds using P450 inhibition assays based on isoform-specific small molecules (Smith et al., 2021).

Besides MPCCs, intrinsic clearance measurements have been presented using microfluidic chips. Since flow has been integrated in the device, a mathematical modeling using the well stirred model was not necessary, but measured clearance data could be scaled up directly. Clearance predictions of slowly cleared drugs using the HµREL coculture system of liver (HepaRG cells or human hepatocytes) and stromal cells were superior compared with suspension cultures, whereas both models performed similarly for intermediate clearance compounds (Chao et al., 2009; Bonn et al., 2016; Hultman et al., 2016).

For pharmacokinetic studies, liver spheroids have been primarily used to identify nuclear hormone receptor activation and P450 induction (Hendriks et al., 2020) to predict drug-drug interactions and to delineate species-specific induction mechanisms (Nudischer et al., 2020). In contrast, their use for clearance predictions is only emerging in recent years. In an industry-led study using a small panel of four low-to-intermediate clearance compounds, intrinsic clearance could be predicted within 3-fold using as little as 6000 hepatocytes (Kanebratt et al., 2021). In agreement with these findings, clearance of a nonoverlapping set of seven low- and intermediate-clearance compounds was accurately predicted with an average fold error of 0.7 (Riede et al., 2021). Irrespective of the model, the turnover of high-clearance compounds tended to be underpredicted, whereas low-clearance compound predictions were more accurate. Prediction accuracy was also dependent on the protein binding assumption and therefore the prediction method (Table 2). However, clearance studies are highly heterogeneous with regard to medium composition, hepatocyte concentration, and culture time, resulting in large interlaboratory and interexperimental variability. Thus, protocol standardization and in-depth benchmarking, ideally in a multicenter setting, are required to further develop 3D liver models for clearance predictions (Louisse et al., 2020).

In addition to clearance estimations, identification of the route of metabolism and the resulting metabolites is of importance to establish a full safety profile. Conventional methods often fail to predict the full metabolite profile, mainly because they are short-lived and lack complexity (Anderson et al., 2009). For instance, only 22 (39%), 26 (46%), and 31 (55%) of 56 major metabolites corresponding to 27 different drugs could be detected in incubations of liver microsomes, liver S-9 fraction, and human hepatocyte suspensions, respectively (Dalvie et al., 2009). In contrast, MPCCs were able to detect 82% and 75% of the same excretory and circulating metabolites, respectively, after 1 week of exposure (Wang et al., 2010). Furthermore, MPCCs using primary hepatocytes of preclinically relevant animal models could replicate species differences in rate and extent of metabolism and guide preclinical model selection (Ballard et al., 2016; Kamel et al., 2021).

Also, spheroids can be used for metabolite profiling. For instance, PHH spheroids generated the known phase I and phase II metabolites of acetaminophen, diclofenac, lamotrigine, midazolam, propranolol, and salbutamol (Ohkura et al., 2014). Especially the identification of salbutamol-4-O-sulfate and lamotrigine-N-glucuronide were remarkable, as those metabolites had been identified in man in vivo but were not generated in S-9 fractions, microsomal preparations, or hepatocyte suspension cultures. In addition, 14 out of 16 major human phase I and phase II metabolites of carvedilol, imipramine, diazepam, disopyramide, sildenafil, and S-warfarin were detected using spheroid cultures (Kanebratt et al., 2021). Although these data show that organotypic liver models can predict the metabolites for a given compound of interest, they remain qualitative and do not allow to accurately estimate the quantitative metabolite profile in vivo.

In recent years, human liver chimeric mice in which the murine liver is ablated and repopulated by transplanted human PHHs have emerged as additional models for pharmacokinetic analyses (Peltz, 2013; Bissig et al., 2018). These models allow the study of human drug metabolism in an in vivo context. However, their low throughput and high costs in addition to the lack of clear demonstration of added value compared with organotypic and microphysiological liver systems have so far prevented more widespread usage of these models.

3. Liver Disease Modeling

Accurate modeling of hepatic diseases facilitates the identification of novel therapeutic targets and the screening for drug candidates. In recent years, organotypic liver models have been developed for a multitude of liver diseases, including nonalcoholic fatty liver disease (NAFLD), fibrosis, and various infectious diseases, such as malaria, hepatitis B virus (HBV), hepatitis C virus, and SARS-CoV-2 infections.

NAFLD constitutes a spectrum of symptoms and diseases from benign steatosis to NASH, fibrosis, and eventual cirrhosis. Etiology and progression of NAFLD are closely linked to obesity and type 2 diabetes, and it has become the leading cause of chronic liver disease worldwide with a global prevalence rate of 25% (Younossi et al., 2016). Although NAFLD, NASH, and fibrosis have been major targets of drug development efforts, all drug candidates have so far failed to receive regulatory approval, likely at least in part because of a lack of translatability of traditionally used genetic and diet-induced rodent models (Drenth and Schattenberg, 2020). Thus, the development of model systems that can mimic the pathobiology of human NAFLD constitutes a main frontier of organotypic liver model research (Müller and Sturla, 2019; Hundertmark and Tacke, 2020; Soret et al., 2020).

NAFLD and NASH have been predominantly studied in spheroids and microfluidic chips; however, limited MPCC data have been presented. Although the majority of systems use cocultures of PHH and primary NPCs, other approaches using rat hepatocytes (Bulutoglu et al., 2019) or hepatic cell lines (Gori et al., 2016; Suurmond et al., 2019) have been presented.

In vitro, NAFLD can be emulated by supplementing culture media with pathophysiologically relevant levels of carbohydrates and free fatty acids as well as high concentrations of insulin. In PHH spheroids, NAFLD models recapitulate steatosis, insulin resistance, reduced glucose consumption, and the induction of genes involved in de novo lipogenesis and gluconeogenesis (Kozyra et al., 2018; Kemas et al., 2021) (Fig. 3A). Importantly, these parameters are reversible and allow for the evaluation of antisteatotic effects of drug candidates or dietary supplements (Kozyra et al., 2018; Cordero-Herrera et al., 2019). Furthermore, spheroids recapitulate the effects of genetic risk factors, such as variations in TM6SF2 (Prill et al., 2019) or PNPLA3 (Pingitore et al., 2019; Hurrell et al., 2020), allowing for targeted drug development for specific at-risk populations.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 3

Organotypic culture systems for liver disease modeling. (A) Steatosis and insulin resistance can be induced in PHH spheroids by exposure to high insulin (IR) and supplementation with free fatty acids and fructose (“NAFLD”) for 7–14 days. Steatosis was measured as intracellular triglyceride levels, whereas insulin resistance was indicated as elevated expression of key genes involved in gluconeogenesis and de novo lipogenesis and reduced AKT phosphorylation upon insulin stimulation. IS, insulin-sensitive group. Error bars indicate S.E.M. * and **** indicate P < 0.05 and P < 0.0001, respectively. (B) The anti-NASH drug candidates cenicriviroc, elafibranor, and lanifibranor reduce free fatty acid–induced collagen deposition (COL1A1) and αSMA in PHH spheroids cocultured with liver nonparenchymal cells. (C) Exposure to high-fat media (black columns) increased expression of genes associated with fatty liver in a microfluidic liver culture system and reduced activity of CYP3A4 and CYP2C9. Error bars indicate S.D. * indicates P < 0.05. (D) PHH in micropatterned cocultures can be infected with HBV as evidenced by various viral life-circle readouts at 16 days postinfection (dpi). HBV surface antigen (HBsAg) was plotted as mean ± S.E.M. (n = 3). HBV covalently closed circular DNA copies were plotted as average of duplicates ± range. HBV core protein (HBVc) expression was indicated by immunostaining together with isotype-matched control. Figure modified with permission from (Hurrell et al., 2020; Kemas et al., 2021; Kostrzewski et al., 2017; Shlomai et al., 2014).

Coculture spheroids moreover recapitulate interactions between PHH and NPCs and allow the study of disease progression from steatosis to NASH and fibrosis, as evident by stellate cell activation (αSMA) and collagen 1A1 deposition (Hurrell et al., 2020). These features allow the screening for compounds that attenuate endpoints of interest, such as steatosis, inflammation, or fibrogenesis, and facilitate the identification of drug targets with potential therapeutic relevance in a human context (Fig. 3B). For instance, in obesity-induced insulin resistance liver macrophages increase secretion of multiple noninflammatory factors, and experiments in spheroids showed that one of these, IGFBP7, regulates hepatocyte lipogenesis and gluconeogenesis, thus providing an interesting and conceptually novel target for the treatment of hepatic insulin resistance and NAFLD (Morgantini et al., 2019). Furthermore, coculture liver spheroids recapitulated the increased expression of miR-144 seen in obesity, which in turn represses IRG1 (Azzimato et al., 2021). IRG1 catalyzes the production of itaconate, an inhibitor of succinate dehydrogenase and fumarate hydratase, and its downregulation causes depletion of fumarate, reduced activity of the antioxidant master regulator NRF2, and exacerbation of hepatic oxidative stress. Notably, downregulation of miR-144 in either PHH or NPCs could rescue obesity-induced suppression of NRF2, highlighting miR-144 as an interesting target to restore the endogenous antioxidant response in obesity-associated NAFLD (Azzimato et al., 2020). In addition to their utility for the identification of new therapeutic modalities, liver spheroids from specific patient populations also allow testing impacts of disease states or genetic polymorphisms on drug metabolism and toxicity (Ingelman-Sundberg and Lauschke, 2018).

NAFLD can furthermore be mimicked in microfluidic devices using similar cocultures of PHHs, stellate cells, and macrophages in high-nutrient, high-insulin conditions. Specifically, these culture conditions resulted in hepatic lipid accumulation and lipidomic signatures similar to NASH biopsies as well as increased TGFβ levels and hepatic stellate cell (HSC) activation (Feaver et al., 2016). Furthermore, the model corroborated in vivo observations that BMP8B might be an important contributor to stellate cell activation and NASH progression (Vacca et al., 2020). In addition to steatosis, exposure to elevated levels of free fatty acids resulted in altered gene expression and metabolic activity (Fig. 3C). Treatment of steatotic cultures with metformin resulted in reduced free fatty acid uptake and lower cellular fat content (Kostrzewski et al., 2017). Hepatic fat accumulation resulted in the generation of an inflammatory milieu with elevated expression of profibrotic markers, which could be ameliorated by treatment with the FXR agonist obeticholic acid (Kostrzewski et al., 2019). Furthermore, as in spheroids, the PNPLA3 mutation I148M enhanced the overall NASH phenotype. Using a different three-channel fluidic device, exposure to lipotoxic stress induced steatosis, hepatocellular ballooning, HSC activation, and elevation of inflammatory and profibrotic markers, effects that could be blunted using the dual PPARα/δ agonist elafibranor (Freag et al., 2020).

Although fewer studies have been described about the modeling of NASH in MPCCs, the available data indicate that steatosis and insulin resistance can be induced by culture in hyperglycemic media, similar to spheroids and microphysiological liver models (Davidson et al., 2016). Moreover, coculturing conventional MPCCs with activated stellate cells resulted in increased levels of IL6 and hepatocyte dysfunction, which could be alleviated by treatment with obeticholic acid or the NADPH oxidase inhibitor setanaxib (Davidson et al., 2017).

In addition to NAFLD, NASH, and fibrosis, liver models have been used in drug development for hepatotropic protozoan infections. The life cycle of Plasmodium falciparum, berghei, or vivax malaria parasites from infection to the release of hepatic merozoites can be emulated in both MPCCs (March et al., 2013; Gural et al., 2018) and spheroids (Arez et al., 2019; Chua et al., 2019). Although various cell models are compatible with malaria infections, high and stable expression of CD81 and SRB1 is required for efficient sporozoite progression, which favors PHHs over hepatoma cell lines and iPSC-derived HLCs (March et al., 2013; Ng et al., 2015). Importantly, spheroids provide a scalable translational tool for the development of antiplasmodial therapies as demonstrated by close agreement of in vitro and in vivo efficacy data of the Plasmodium Elongation Factor 2 Inhibitor M5717 (Arez et al., 2019). Similarly, MPCCs allowed for the testing of antihypnozoite drugs for prophylactic or cidal therapies (Gural et al., 2018). In addition to malaria, a sandwich culture of liver sinusoidal endothelial cells and Huh-7 cells in a 3D collagen matrix sandwich recapitulates infection with the protozoan parasite Entamoeba histolytica (Petropolis et al., 2014).

Organotypic 3D models have been applied to a multitude of viral liver diseases. HBV infection relies on expression of the viral receptor NTCP (SLC10A1). HBV infection models have been established in PHH MPCCs that recapitulate productive viral infection, including viral gene expression, and enable the detection of HBV core protein throughout the 3-week culture period (Fig. 3D) (Shlomai et al., 2014). In addition, HBV infection can be successfully modeled in microfluidic chips (Ortega-Prieto et al., 2018) and iPSC-derived HLC spheroids (Fu et al., 2019), and the authors show that the model can replicate antiviral effects of entecavir, a reverse-transcriptase inhibitor that interferes with viral replication, and the NTCP substrate tauroursodeoxycholic acid that blocks HBV entry. Similarly, Huh-7 and PHH spheroids (Murakami et al., 2006; Ananthanarayanan et al., 2014) as well as PHH MPCCs (Ploss et al., 2010) are permissive to hepatitis C virus infection and allow for miR-122–dependent viral replication. Recently, liver spheroids were furthermore used to evaluate hepatotoxicity of antiviral candidate drugs (Monteil et al., 2021) and to study extrapulmonary effects of SARS-CoV-2 infections (Stebbing et al., 2020). In this model, increased exposure to class I interferons, as typically observed during excessive inflammation in COVID-19, induced expression of the viral receptor ACE2, which in turn resulted in increased infectivity and altered expression of genes involved in coagulation and platelet activation (Stebbing et al., 2021). Importantly, these effects were blunted by therapeutic concentrations of the JAK/STAT inhibitor baricitinib that had been suggested by artificial intelligence-based knowledge graphs to inhibit both cytokine signaling and viral endocytosis (Richardson et al., 2020) and contributed to the decision by FDA and Japanese Pharmaceuticals and Medical Devices Agency to grant emergency use authorization of baricitinib in combination with remdesivir for hospitalized adults with severe COVID-19.

1. Drug Transporter Studies

Regulators in Europe, the United States, and Asia mandate that interactions of new compounds with key renal transporters are tested during preclinical drug development, but current systems do not meet the needs of pharmaceutical companies. Glomerular filtration, tubular secretion, and reabsorption must be considered when investigating the renal clearance of drugs and their metabolites. Using preclinical animal models, prediction of renal excretion works well when glomerular filtration is the major way of clearance. For tubular secretion and reabsorption, however, in vivo models are often inaccurate because of species differences in transporters and drug-metabolizing enzymes (Zou et al., 2018; Bueters et al., 2020). For kidney models to be of use for ADME applications, they should have polarized activity of OAT1/3 and OCT2 on the basolateral side and MDR1, BCRP, MRP2, and MATE on the apical side, show barrier function, and feature appropriate reuptake transport and kidney-relevant metabolism (Fowler et al., 2020). Furthermore, the system should recapitulate known drug-drug interactions that influence drug transport or metabolism.

Common assays to evaluate renal solute carrier (SLC)-mediated transport are based on PAH or creatinin, whereas ATP-binding cassette transporters (ABC) transport is measured using acetoxymethyl-calcein or 5-chloromethylfluorescein diacetate as substrates of P-gp and MRPs, respectively. These dyes become fluorescent and membrane-impermeable only after intracellular removal of the acetomethoxy or chloromethyl groups by esterases. Renal metabolism is mostly tested by vitamin D biotransformation. Evaluation of at least some of those endpoints constitutes an essential element of the functional characterization of organotypic or microphysiological renal models. Importantly, however, although molecular phenotypes and cellular morphology are increasingly benchmarked against in vivo tissue or primary freshly isolated cells, most characterizations of transporter function or metabolism remain qualitative. For instance, OCT2, P-gp, and MRP activity was demonstrated in immortalized PTECs using the fluorescent substrates 4-(4-(dimethylamino)styryl)-N-methylpyridinium iodide, acetoxymethyl-calcein, and 5-chloromethylfluorescein diacetate, respectively (Jansen et al., 2015; Vormann et al., 2018; Vriend et al., 2018). However, although the authors demonstrate that uptake is specific and can be blocked by inhibitors of the respective transporter, whether the transport kinetics quantitatively correspond to in vivo function remained unclear. Similarly, PTECs in perfusion cultures were able to bioactivate vitamin D at rates around 0.01 µl/h, but again, it was not analyzed whether these kinetics resemble the activity of proximal tubules in vivo (Weber et al., 2016).

In an interesting recent study, renal reabsorption of five compounds was determined in a perfused PTEC chip, and the resulting data were used to predict overall in vivo renal clearance using a physiologically based parallel tube model (Sakolish et al., 2020). For creatinine the model correctly predicted no reabsorption, which aligns with in vivo data that creatinine excretion is driven by the glomerular filtration rate. Furthermore, in vivo and chip data aligned for perfluorooctanoic acid, cisplatin, and cadmium clearance, whereas gentamicin reabsorption was underpredicted (20% in vivo, 0.5% in PTEC chip). Similar attempts for alignments with physiologic data were performed for a microfluidic model of iPSC-derived glomerular cells in which glomerular filtration rates of albumin and inulin were benchmarked against in vivo values and shown to be indistinguishable (Musah et al., 2017). Overall, these data indicate that the currently available systems give promising results, but further characterization using liquid chromatography coupled with tandem mass spectrometry–based measurements of several compounds and benchmarking to in vivo renal clearance is needed.

2. Nephrotoxicity

Kidneys take part in the elimination of 32% of the 200 most prescribed drugs (Morrissey et al., 2013). Given these essential roles in drug clearance, drug-induced kidney injury (DIKI) is relatively common with frequencies of 14%–26% in adults and even higher numbers in pediatric populations (Awdishu and Mehta, 2017). Mechanisms and patterns of injury differ between drugs with aminoglycosides causing acute tubular necrosis, whereas puromycin and adriamycin are associated with glomerular injury, nonsteroidal anti-inflammatory drugs cause injury to the loop of Henle, and sulfonamides and amphotericin B result in predominant injury of the distal tubule (Pannu and Nadim, 2008). Current models used for toxicity studies often fail to predict nephrotoxicity (Tiong et al., 2014). Especially interspecies differences in renal transporter expression limit the use of animal models to study adverse drug effects (Chu et al., 2013).

The aforementioned organotypic kidney models using human cells aspire to improve the accuracy of nephrotoxicity predictions. To be of relevance for the pharmaceutical industry, novel microphysiological kidney models should incorporate the tissue mechanical environment (e.g., fluid shear and pressure) and allow the evaluation of different endpoints, such as filtration rate, barrier function, transport activity, viability, or release of kidney tubular injury biomarkers, such as KIM-1 (Phillips et al., 2020). However, as mentioned above, a complete and phenotypically accurate model that contains all segments of the kidney does not exist yet.

Since the proximal tubule is the major site of nephrotoxicity, most studies have used proximal tubule models. Using rat kidney scaffolds recellularized with conditionally immortalized human PTECs overexpressing OAT1, concentration-dependent nephrotoxicity of cyclosporine A, tenofovir, and the OCT2 substrate cisplatin was observed (Fedecostante et al., 2018). Furthermore, cells were more sensitive to tenofovir and cisplatin than in 2D culture, which aligned with increased expression of the respective transporters OAT1 and OCT2, respectively. However, sensitivity remained very low. For instance, the tested cisplatin (300 µM) and cyclosporine A concentrations (1mM) exceeded therapeutic levels (200 nM and 300 nM for cisplatin and cyclosporine A, respectively) by more than 1000 times. By contrast, using the same cell line in a plate design of two channels separated by a collagen gel, cisplatin, and cyclosporine A toxicity was reproducibly detected in a multicenter study at 5–15 µM and 5–30 µM, respectively (Vormann et al., 2021). This is, however, in contrast to other studies culturing non–OAT1-overexpressing immortalized human PTECs in the same device in which cyclosporine A toxicity was not detected at 30 µM (Vriend et al., 2020).

For cisplatin, disruption of barrier integrity, lactate dehydrogenase release, and apoptosis were observed at lower levels of 25–100 µM when PTECs were cultured in a two-channel microfluidic device (Jang et al., 2013; Nieskens et al., 2020). By contrast, a bioprinted proximal tubule model comprised of fibroblasts, HUVEC cells, and primary human PTECs detected cisplatin toxicity at concentrations below 5 µM (King et al., 2017).

Besides cisplatin, the P-gp substrate cyclosporine A has been repeatedly used to evaluate function-dependent nephrotoxicity. In a bioprinted 3D proximal tubule on chip, reduced viability and significant disruption of barrier function were detected at 50 µM (Homan et al., 2016). Sensitivity to cisplatin and cyclosporine A was even lower in perfused proximal tubule models based on HK-2 cells (Sakolish et al., 2019). Here, changes in viability were only detected at concentrations of 1.4 mM cisplatin and 270 µM cyclosporine A, and even at those levels, only mild elevations of the secreted toxicity markers NGAL and action potential (AP) were observed. In addition to cisplatin and cyclosporine A, few studies have evaluated the nephrotoxicity of other prototypic toxins, such as polymyxin B (Weber et al., 2018) and aristolochic acid (Chang et al., 2017) in perfused proximal tubule models, and of tacrolimus in kidney organoids (Kim et al., 2020b). Although these studies are useful to elucidate the toxicity mechanisms of the respective compounds, larger screens are needed to conclude about the overall accuracy of organotypic kidney systems for the prediction of nephrotoxicity.

Only few studies have been presented that evaluated the nephrotoxicity of a larger set of compounds. In an important seminal study, the predictivity of 2D cultured PTECs was compared with HK-2 and LLC-PK1 cells using a panel of 41 drugs and chemicals consisting of 22 nephrotoxins that injure the proximal tubule, 11 nephrotoxins that injure other parts of the nephron, and 8 drugs that are not nephrotoxic (Li et al., 2013). Using IL6 and IL8 expression as endpoints after 16 hours of treatment, the model achieved sensitivity and specificity between 60% and 80%. The area under the receiver operating characteristic (AUC_ROC) across three PTEC donors was 0.85 compared with 0.71 and 0.73 for HK-2 and LLC-PK1 cells, respectively. When ESC-derived instead of primary proximal tubule cells were used for the same 41 compounds, AUC_ROC was only slightly lower with 0.8 (Li et al., 2014). By contrast, evaluation of a subset of 30 compounds using iPSC-derived proximal tubule cells and random forest-based model training, the model predicted proximal tubular toxicity in humans with 99.8% in the training set and 87% in the test set using crossvalidation using changes in IL6 and IL8 expression as readout (Kandasamy et al., 2015).

Besides IL6 and IL8 expression, other studies used different biomarkers for nephrotoxicity detection. A small study using six nephrotoxins showed that protein levels of KIM-1 and NGAL in primary PTECs correlated well with toxicity, whereas viability or the same endpoints in HK-2 cells was not suitable (Huang et al., 2015). Similarly, clusterin levels in primary PTECs achieved 65% sensitivity and 94% specificity in a set of 20 DIKI-positive and 16 DIKI-negative test compounds (Bajaj et al., 2020). Multiplexed gene expression profiling of primary PTECs moreover identified heme oxygenase-1 (HO-1) protein deregulation as a significant biomarker for all nine tested nephrotoxic compounds (Adler et al., 2016). Importantly, using an additional panel of 39 mechanistically distinct nephrotoxins, HO-1 achieved higher AUC_ROC than conventional measures of cell viability (0.89 compared with 0.78). Although these in vitro findings align with earlier results showing that HO-1 was increased in urine of patients with acute kidney injury (Zager et al., 2012) and tubulointerstitial damage (Yokoyama et al., 2011), it remains to be determined whether HO-1 is indeed superior in vivo compared with conventional biomarkers.

These data indicate that available test systems based on primary or stem cell–derived proximal tubule cells provide useful tools for preclinical nephrotoxicity assessments. However, the tested endpoints remain diverse, and integration of the best-performing markers promises to further increase the predictive power. Notably, to the best of our knowledge, all current test systems used conventional 2D cultures of proximal tubule cells. It will be interesting to see whether the increased complexity of 3D renal models, such as fluidic chips or organoids, will translate into increased predictive accuracy. Furthermore, whether models of other sites of kidney injury, such as the glomerulus, can be useful in reliably identifying nephrotoxic compounds remains to be determined.

1. Undirected Cerebral Organoids

In 2013, neuroectoderm of embryoid bodies was first maintained in 3D culture embedded in matrigel (Lancaster et al., 2013; Lancaster and Knoblich, 2014). The droplets were then transferred to a bioreactor to ensure adequate nutritional supply and kept in culture for up to 10 months. Immunohistochemistry revealed that these 3D cultures usually referred to as cerebral organoids or undirected whole-brain organoids displayed a cellular diversity and architecture that partly recapitulate the structure of the brain as early as 20–30 days after the start of culture. The self-organizing capacity was later confirmed by single-cell transcriptomic analyses that revealed a mixture of region specific ectodermal and mesodermal cells. Upon graftment of such undirected human cerebral organoids into the adult mouse brain, the models established graft-to-host functional synaptic connectivity and became vascularized (Mansour et al., 2018). Due to significant organoid-to-organoid variability, modifications of the original protocol have been proposed to support the formation of neuroectoderm (e.g., by using a microscale internal scaffold or small molecules or by decreasing the size of the embryoid bodies) (Lancaster et al., 2017; Sivitilli et al., 2020). On the contrary, a recent study increased the percentage of mesodermal progenitors, by lowering the heparin concentration and delaying matrigel embedding, and observed innately developing microglia-like cells (Ormel et al., 2018). However, this strategy likely entails increased organoid-to-organoid variability, and single-cell transcriptomic analyses are warranted to evaluate the effects on cell composition.

2. Directed Cerebral Organoids

In contrast to undirected methods, directed protocols rely to a higher degree on the addition of external patterning factors, such as small molecules and morphogens, to achieve the development of patterned organoids with region-specific neural cell types from ESCs and iPSCs (Eiraku et al., 2008; Kadoshima et al., 2013). Subsequently, a number of different protocols have been reported to generate region-specific patterned organoids of the cortex (Mariani et al., 2015; Paşca et al., 2015; Lancaster et al., 2017; Qian et al., 2017a, 2018), midbrain (Jo et al., 2016; Qian et al., 2016), cerebellum (Muguruma et al., 2015), pituitary gland (Ozone et al., 2016), and hypothalamus (Qian et al., 2016). In an interesting strategy, these region-specific organoids can be fused into so called “assembloids” that to some extent enable modeling of functional connectivity (Birey et al., 2017).

3. Neural Spheroids

In addition to complex and heterogeneous stem cell–derived organoids, spheroid cultures have been presented in which differentiated cells are aggregated in engineered scaffolds or in concave microwell plates. Although these models have clear advantages over conventional 2D culture, they lack the cellular complexity that can be achieved with organoids and do not recapitulate the tissue architecture of the human brain. Most 3D neural spheroids have used alginate or collagen scaffolds (Li et al., 2007; Xu et al., 2009; Frampton et al., 2011; Odawara et al., 2013; Tang-Schomer et al., 2014). However, more recently scaffold-free methods have emerged (Choi et al., 2013; Kato-Negishi et al., 2013). One promising example is the formation of spheroids using primary human astrocytes, pericytes, and endothelial cells, in which these cell types form layered structures with a core of astrocytes surrounded by pericytes and endothelial cells, thus recapitulating important structural BBB features (Urich et al., 2013).

4. Microfluidic Brain-on-a-Chip Models

Brain‐on‐a‐chip models comprise microengineered systems that create opportunities for spatiotemporal control of cell growth and differentiation cues by combining surfaces mimicking the ECM with microfluidic channels regulating the mass transport of fluids and soluble factors. These chips can be manufactured in a variety of material, of which PDMS remains most commonly used for brain-on-a-chip devices (Mofazzal Jahromi et al., 2019). BBB microchip models frequently based on conventional transwell cultures in which a porous barrier membrane coated with endothelial cells, pericytes, and astrocytes separates two adjacent compartments. Notably, the two different compartments are commonly perfused using different media resembling blood and cerebrospinal fluid. Such perfusion devices improve the stability of pH and oxygen in the media compared with static cultures in which media were exchanged periodically. Seminal studies presented a PDMS device with a vascular compartment containing primary endothelial cells separated from a neural chamber containing primary human astrocytes and pericytes as well as iPSC-derived glutamatergic neurons by a porous polycarbonate membrane (Brown et al., 2015; van der Helm et al., 2016). Cells remained viable for up to 21 days and displayed relevant expression of the tight junction marker ZO-1.

The two compartment chips were recently extended to include an additional compartment so that BBB influx, brain parenchyma, and BBB efflux could be modeled simultaneously (Maoz et al., 2018). The brain compartment consists of stem cell–derived glutamatergic, GABAergic, dopaminergic, and serotonergic neurons as well as glial cells, whereas the BBB compartments each consist of a basal channel lined with primary endothelial cells separated from an apical, perivascular channel by a porous membrane coated with astrocytes and pericytes. Shear stress on astrocytes and pericytes was >25-fold lower than in the endothelial compartment (0.0007 vs. 0.02 dyne/cm²), closely mimicking the native microenvironment. By restricting active flow to the apical channel of the brain chip, flow velocity neural compartment approached zero, causing diffusion-mediated molecular transport to dominate, as previously reported for the brain in vivo (Ballabh et al., 2004). The model could mimic the effect of intravascular administration of the psychoactive drug methamphetamine on the reversible breakdown of the BBB observed in vivo (Turowski and Kenny, 2015) and revealed previously unknown metabolic coupling between the BBB and neurons, offering an in vitro platform for probing transport, efficacy, mechanism of action, and toxicity of neuroactive drugs.

Other BBB models are designed so that the endothelium forms microvessels, which are in direct contact with astrocytes and pericytes without an artificial membrane (Herland et al., 2016). Stimulation with tumor necrosis factor-α resulted in significantly increased endothelial secretion of G-CSF and IL6 compared with static transwell cultures, emphasizing the importance of direct cell-cell crosstalk and perfusion for proper mimicking of human BBB biology and neuroinflammation.

Several microfluidic brain-on-a-chip as well as BBB-on-a-chip devices have been applied to study neural toxicity and BBB permeability as well as to identify novel drug target candidates for the treatment of brain diseases. Although these systems have obvious advantages in terms of scalability and control of the microenvironment, it is important to consider that the reversed-engineering approach remains hampered by our incomplete understanding of brain ontogenesis and physiology.

1. Drug Screening Platforms

Organotypic and microphysiological human brain models are extensively used for studies of BBB drug permeability. Using spheroids of primary human astrocytes, pericytes, and endothelial cells, drug penetration could be studied by confocal microscopy or mass spectrometry. Specifically, analyses revealed that the BBB-permeable compounds angiopep-2 and BKM120 passed the BBB, whereas the non–BBB-permeable tyrosine kinase inhibitor dabrafenib was not detected within spheroids (Bergmann et al., 2018), rendering this model suitable for cost-effective high-throughput drug permeability testing (Cho et al., 2017). The model system was recently extended to also include iPSC-derived microglia, oligodendrocytes, and neurons (Nzou et al., 2018). Although spheroids expressed cell-specific markers and functional barrier assessments using MPTP, MPP+, and mercury chloride indicated charge selectivity, further characterizations of ECM structure and composition as well as functional analyses of the individual cell types in the model are necessary to better understand its physiologic relevance.

Such 3D systems have obvious advantages in comparison with 2D models regarding drug screening. However, limitations concerning scalability and intraexperimental and interexperimental reproducibility have stalled their adoption as high-throughput assays for drug development, particularly for models that use primary human cells. To overcome these limitations, a fully automated workflow was developed that provides an integrated solution for the generation, maintenance, and high-content imaging of human midbrain organoids in 96-well format (Renner et al., 2020). The authors demonstrate that the formed organoids recapitulate midbrain structures and display reduced morphologic, structural, and functional heterogeneity, thus providing a promising platform for future drug screening efforts. Similarly, high-throughput generation of homogeneous cerebral organoids derived from iPSCs of patients with spontaneous Alzheimer disease facilitated the testing of FDA-approved drugs as repurposing candidates to reduce amyloid β (Aβ) or τ deposition and enhance neuronal cell viability (Park et al., 2021).

In an interesting approach, ESC-derived neural progenitor cells, endothelial cells, mesenchymal stem cells, and microglia precursor cells were integrated in hydrogels (Schwartz et al., 2015). The resulting 3D neural constructs exhibited reproducible RNA sequencing gene expression signatures, and training with 34 toxic and 26 nontoxic chemicals yielded a linear support vector model that correctly classified 9 out of 10 compounds from an additional blinded test set as toxic or nontoxic, respectively, with one false positive.

A further interesting application is the use of brain organoids with clinically relevant oncogenic mutations introduced by CRISPR/Cas for the development of targeted treatments for brain tumors (Bian et al., 2018). As a proof of concept, the authors found that only those organoids with mutations resulting in EGFR overactivation were responsive to the EGFR inhibitors. Furthermore, human cerebral organoids and brain-on-a-chip systems have been used in several studies to evaluate effects of environmental exposures, such as ethanol (Zhu et al., 2017), nicotine (Wang et al., 2018b), and cocaine (Lee et al., 2017a), on prenatal neurogenesis.

2. Disease Models

Organotypic and microfluidic 3D models of the human brain have in recent years been widely used to model CNS pathologies using patient-derived iPSCs or genomic editing in which the effect of genetic risk variants can be directly compared with isogenic control lines. Both undirected and directed differentiation protocols as well as spheroids and microfluidic systems have been used to assess a growing number of diverse neural pathologies, including monogenic diseases, complex neurodegenerative disorders, brain tumors, and neurotropic infectious diseases, which facilitates the understanding of underlying disease mechanisms and enables subsequent drug discovery studies.

Undirected organoids were successfully used to recapitulate autosomal recessive primary microcephaly due to nonsense mutations in CDK5RAP2, which resulted in premature neuronal differentiation and smaller overall organoid size as compared with controls (Lancaster et al., 2013). Similarly, undirected cerebral organoids could mimic phenotypes associated with Sandhoff disease, an infantile G_M2 gangliosidosis caused by mutations in HEXB that results in a deficiency of the lysosomal enzyme β hexosaminidase and early ganglioside accumulation in the brain (Allende et al., 2018). Specifically, organoids generated from patient-derived iPSCs recapitulated GM2 ganglioside accumulation, which could be rescued by CRISPR/Cas of the underlying mutation. Furthermore, directed midbrain organoids with the Parkinson disease–associated LRRK2 mutation recapitulated several pathologic hallmarks of LRRK2-associated sporadic Parkinson disease (Kim et al., 2019).

Brain organoids have moreover been used to model polygenic diseases. Transcriptomic comparison of dorsal/ventral forebrain organoids from members of four families with severe idiopathic autism spectrum disorder (ASD) or unaffected, first-degree family members revealed an upregulation of genes involved in cell proliferation, neuronal differentiation, and synaptic functions as well as an overproduction of GABAergic inhibitory neurons in affected organoids compared with controls (Mariani et al., 2015). Importantly, by downregulating FOXG1, which was upregulated in ASD organoids, the overproduction of GABAergic neurons could be reversed, suggesting the FOXG1 axis as an interesting target for the treatment of ASD.

Organotypic or microfluidic 3D models have also been used to study a variety of neurodegenerative disorders, such as Alzheimer disease, Parkinson disease, and amyotrophic lateral sclerosis. Using a microfluidic 3D triculture system consisting of neurons and astrocytes derived from human neural progenitor cells together with SV40 immortalized microglia, overexpression of a human β-amyloid precursor variant resulted in Aβ aggregation and increased microglial recruitment, which entailed secretion of proinflammatory cytokines as well as neuron loss (Park et al., 2018). For further information, we refer the interested reader to a recent comprehensive review on this topic (Chang et al., 2020).

3D culturing of tumor cells from biopsies or tumor resections has been used during the past decades to study brain tumor biology (Klein et al., 2020). In contrast to cultures of stem cell–derived or primary neural cells, tumor cells can be cultured in chemically defined conditions, which facilitates mechanistic analyses (Hubert et al., 2016). Furthermore, they can be studied in vivo after implantation into the brain of immunodeficient mice to generate patient-derived orthotopic xenografts (Huszthy et al., 2012). However, these models are less suitable to study oncogenesis, and their use for pharmacological applications is limited by the absence of interaction between healthy and tumor tissue (in vitro models) or the lack of scalability (patient-derived orthotopic xenografts). To overcome these issues, two recent studies have used genome-editing to introduce oncogenic mutations into human cerebral organoids. CRISPR/Cas-mediated introduction of oncogenic HRas^G12V into a subset of cells in undirected cerebral organoids after 4 months in culture resulted in the transfected cells displaying invasive phenotypes within the organoid and a gene expression profile consistent with mesenchymal subtype human glioblastoma (Ogawa et al., 2018). Furthermore, stereotactic injection of organoid-derived tumor cells into the brain of immunodeficient mice allows the study of disease pathology and invasive growth in vivo. In a similar approach, brain tumorigenesis was recapitulated by introducing different oncogenic mutations in undirected cerebral organoids at the end of the neural induction phase by transposon- and CRISPR/Cas9-mediated mutagenesis (Bian et al., 2018). This strategy allowed to define combinations of mutations that result in glioblastoma-like and primitive neuroectodermal tumor-like neoplasms.

During the Zika virus (ZIKV) epidemic in the Americas, a dramatic increase in newborns with microcephaly as well as other neurologic impairments was observed (Teixeira et al., 2016). An initial study showed that ZIKV could infect human iPSC-derived neural progenitors in 2D culture and cause dysregulation of cell cycle, attenuation of proliferation, and cell death (Tang et al., 2016). In agreement with these findings, infection of undirected cerebral organoids with ZIKV in 3D resulted in an average growth reduction by 40% (Garcez et al., 2016). Infection in directed forebrain organoids revealed tropism toward more immature SOX2⁺ radial glia cells as compared with TBR2⁺ intermediate neural progenitors and immature neurons (Qian et al., 2017b), a finding that was confirmed in studies of fetal human brain slices (Onorati et al., 2016). This established cerebral organoids as a useful platform for studies of neuroinvasive viruses. With the growing number of reported neurologic symptoms in patients with COVID-19, the number of studies in which SARS-CoV-2 is used to infect cerebral organoids is steadily increasing (Jacob et al., 2020). However, as compared with ZIKV, findings regarding infectivity and viral tropism are less clear. A general limitation in most cerebral organoid models used to study infectious disease is the lack of nonectodermal microglia (i.e., the resident immune cells of the brain parenchyma). Thus, more recent protocols, which either modify undirected organoids to innately develop microglia-like cells (Ormel et al., 2018) or add iPSC-derived microglia-like cells grown in 2D to directed organoids (Abud et al., 2017), are likely to play a prominent role in future modeling of diseases with pronounced immune components.

1. Electrophysiology Assays for Single Cardiomyocytes

The patch clamp is the gold standard for measuring single-cell APs and ionic currents. Manual patch clamping of isolated single cardiomyocytes has facilitated the development of treatments for channelopathies, such as the long QT syndrome, and the technique was used to study the distribution of ion channels in engineered cell pairs (Aratyn-Schaus et al., 2016; van der Voorn et al., 2020). Automated patch-clamp systems trade control over cell environment for the ability to patch hundreds of cells at a time (Danker and Möller, 2014). The most well established application of automated patch clamp in cardiotoxicity assessment is the ERG activity assay, which measures the effect of a compound on the potassium current flowing through ERG channels genetically expressed in easy-to-patch noncardiomyocytes, such as oocytes. Even if these cells do not have an AP, changes to the ERG current are considered proarrhythmic, since mutations to ERG have been linked to arrhythmias in vivo. Unfortunately, this approach is prone to many false positives as proarrhythmic signals in the expression system do not necessarily translate into human cardiac toxicity, wherein tissue- and organ-level compensatory mechanisms exist. To combat this trend, an international public-private partnership known as comprehensive in vitro proarrhythmia assay was tasked with developing computational models to interpret the ERG results in the context of human heart pathophysiology and whose predictions will be validated against suitable in vitro assays (Ridder et al., 2020).

2. Excitation-Contraction Coupling in Single Cardiomyocytes

Calcium transients can be measured optically using fluorescent dyes. Conventional calcium-sensitive fluorophores are used for comparing relative changes in the same cell, whereas ratio-metric dyes, such as Fura-red, are characterized by two distinct calcium-sensitive signals, wherein their ratio creates a signal proportional to the absolute level of calcium in the cell (Aratyn-Schaus et al., 2016).

Single-cell contractility can be assessed with miniaturized force transducers, but such analyses are low-throughput and invasive (Wang and Lin, 2007). Instead, optical analyses of cellular shortening offer minimal interference and high throughput. For instance, the raw displacement and beat rate of cells may be obtained using open-source software, such as MUSCLEMOTION (Sala et al., 2018), which is also applicable for tissue-level preparations (Ribeiro et al., 2017). Alternatively, cell-generated forces can be quantified by tracking cell-induced deformation of soft substrates of known material properties (Wang and Lin, 2007).

In aggregate, single-cell methods are powerful, scalable, and reproducible. Yet, isolated cells lack the important physiologic signals in a native tissue, which is electrically and mechanically coupled in aligned arrangements (Kojima et al., 2006). In turn, these signals have further proven critical for maturating human iPSC-derived cardiomyocytes (hiPSC-CMs) (Ronaldson-Bouchard et al., 2018).

1. Cardiac Cell Aggregates as Scaffold-Free 3D Models

Spheroid cultures have been developed for the study of 3D cardiomyocyte constructs, which are methodologically similar to spheroids of other tissues. One of the early examples of cardiomyocyte spheroid formation has been performed in 96-well ULA plates in static culture (Koike et al., 2005). The method relies on the spontaneous differentiation of embryonic stem cells in the spherical embryoid body toward different progenitors, including cardiac. Improved methods to aggregate more homogeneous cardiomyocyte spheroids include centrifugation (Burridge et al., 2007) and the use of bioreactors (Lü et al., 2008). A versatile and popular alternative is the hanging drop method, wherein aggregation and cardiomyocyte differentiation of a predefined number of cells occurs in droplets of specified size (Chen et al., 2011). The beat rate of the resulting spheroids can be assayed in response to drug applications either optically (Zuppinger, 2017) or using microelectrodes (Schmid et al., 2016). Embryoid bodies can also be combined with electrical and mechanical stimulation to synchronize and promote iPSC-cardiomyocyte differentiation (Richards et al., 2016; LaBarge et al., 2019).

However, both 2D monolayers and cardiac spheroids fail to recapitulate the structural architecture of native cardiac cells and tissues. To overcome this problem 3D cardiomyocyte aggregates can be cultured in elongated microfluidic channels (Mathur et al., 2015; Marsano et al., 2016). In this system, the tissues remain viable for many weeks, and their response to drugs is more consistent with ex vivo data compared with single-cell studies. Importantly, such systems can be parallelized to increase throughput (Schneider et al., 2019). Despite these advantages, cardiomyocytes in scaffold-free culture have yet to show similarly elongated morphology, aligned and organized intracellular structure, and directionality in action potential propagation, as found in the scaffolded alternatives.

2. Scaffold-Based Cardiac Tissue Models

Scaffold-based approaches have been established for replicating the structure and function of the human myocardium. The most prominent strategies include the formation of monolayered or multilayered laminar tissues on deformable substrates or 3D strips derived from deformable scaffolds and external anchoring pillars.

Laminar cardiac tissue models are organized into aligned, anisotropic architectures that display a native-like organization of the contractile cytoskeleton, especially in the alignment and organization of sarcomeres (Fig. 7B). Similarly, the expression and longitudinal localization of gap junctions in laminar tissue enable physiologic AP propagation and can therefore provide more relevant tissues for drug screening and disease modeling. Substrate cues or patterns that induce laminar tissue formation include microcontact printed ECM proteins, micromolded hydrogels, 3D-printed soft substrates, and nanofiber coatings (Grosberg et al., 2011; Wang et al., 2014; Li et al., 2016a; Lind et al., 2017a; Ahn et al., 2018).

Like conventional monolayer cultures, laminar tissues are thin enough to be compatible with conventional epifluorescent or confocal imaging techniques. Laminar tissue can also be readily integrated with noninvasive electrophysiological assays, such as MEAs (Kujala et al., 2016; Li et al., 2016a; Alassaf et al., 2019). This can serve to overcome the common drawbacks of MEA studies, which include unphysiologic tissue architectures, limited culture time, and tissue delamination.

By formulating the tissue-guiding substrates into deflectable cantilevers, the laminar tissues can be purposed into simple tissue-scale contractility assays. As the tissue contracts, the cantilever deflects proportionally to the contractile force (Grosberg et al., 2011; Wang et al., 2014). Initially cantilevers were made of silicones, but softer formulations derived from gelatin hydrogels were later introduced to extend tissue culture time. More recently, soft, multilayer cantilevers embedding flexible strain gauges were introduced, thus replacing optical tracking of cantilever motions with noninvasive electrical readouts (Lind et al., 2017a,b; Kim et al., 2020a). This enables scalable drug dose-response studies in user-friendly multiwell formats as well as noninvasive monitoring of the long-term maturation of laminar human iPSC-derived cardiomyocyte tissues (Lind et al., 2017a,b).

Engineered tissue strips, also referred to as engineered heart tissues (EHTs), are the most established 3D cardiac tissue models (Fig. 7C). Although multiple variations have been reported, they are based on the same key principles. First, a mold containing a set of micropillars or beams that serve as anchoring points for the EHT is fabricated. Then, cardiomyocytes are seeded within a deformable hydrogel matrix, which over time is compacted around the anchoring points by cellular contractions and enzymatic activity, generating an assayable, dense 3D tissue.

The most common formulations rely on collagen combined with laminin or fibronectin introduced directly or by addition of matrigel, which may further be modulated by addition of fibrin or by regulating cell-induced matrix turnover (Ronaldson-Bouchard et al., 2018). Dependent on the mold design, EHT can be formulated as linear or circular strips at a range of sizes, and the anchoring pillars provide either isometric or auxotonic load dependent on their stiffness. To enable miniaturization, molds are often fabricated through PDMS-based soft lithography (Hansen et al., 2010; Boudou et al., 2012; Ronaldson-Bouchard et al., 2018) but have also been made from stereolithographically 3D-printed poly(ethylene glycol) diacrylate hydrogels (Ma et al., 2019) or in multitool approaches that avoid PDMS (Zhao et al., 2019).

EHTs are particularly relevant for contractility studies, which can be performed with force transducers (in the case of large tissue constructs) or by optically tracking the movement of the anchoring pillars. The contractility readouts are often combined with optical electrophysiological assays, such as voltage- and calcium-sensitive sensors based on dyes and proteins.

Importantly, several studies have demonstrated improved maturation of stem cell–derived cardiomyocytes when integrated into EHTs, leading to more adult, tissue-level drug responses. Improved tissue maturation is in part a result of the passive mechanical stretch provided by the anchoring points (Abilez et al., 2018) and can be accelerated drastically by electrically induced contractile training (Hirt et al., 2014) or active mechanical loading (Liaw and Zimmermann, 2016). Moreover, multicellular constructs can be applied to generate EHTs with coupled atrial/ventricular phenotypes (Zhao et al., 2019).

3D EHTs have provided the most adult-like models of the cardiac muscle derived from human iPSC-CMs. Miniaturized assays that combine EHTs with automated video analysis of multiwell plates are by now commercially available and constitute promising systems for medium- to high-throughput drug screening and disease modeling purposes.

1. Cardiotoxicity

As illustrated above, patch clamping, MEAs, and various imaging methods have long been used to monitor cardiac function and constitute an important pillar of high-throughput cardiotoxicity screening. The potential to increase predictivity by incorporating human genetics with iPSC-CMs is further enhanced when engineered culture conditions can provide native-like, mature tissue-level performance. As the use of these engineered culture platforms is becoming more widespread, their predictive power was benchmarked against a gold standard provided by the hERG assay and in vivo data. Commonly used compounds include isoproterenol and epinephrine that increase (positive inotropy) or verapamil and nifedipine that decrease (negative inotropy) contractile force, respectively.

As a simplistic and easy-to-scale 3D platform, cardiac spheroids in multiwell plates for calcium imaging are currently the preferred platform for cardiotoxicity screening (Sirenko et al., 2017). For instance, spheroids formed of cocultured hiPSC-CMs and fibroblasts (Verheijen et al., 2018) or endothelial cells (Polonchuk et al., 2017) in physiologic ratios have successfully identified the dose-dependent toxicity of doxorubicin, a cancer therapeutic with known cardiac side effects, which complemented earlier findings derived from traditional monolayer cultures (Burridge et al., 2016). Furthermore, spheroid testing of 29 and 13 compounds with known inotroptic and noninotroptic effects in vivo in 384-well format under field stimulation in an automated calcium imaging system resulted in 80% sensitivity and 91% specificity (Pointon et al., 2017).

Laminar-tissue cantilever assays provide force readouts and allow long-term culturing, rendering them suitable for nonacute in vitro studies (Grosberg et al., 2011; Ahn et al., 2018). Instrumented cantilevers with embedded sensors in multiwell formats have allowed noninvasive monitoring of the contractile maturation of hiPSC-CM tissues over a 4-week period and 12 dose-response studies of positive and negative inotropic compounds (Lind et al., 2017a,b).

Although high-throughput drug screening using EHTs was initially challenging because of their size, several approaches achieved suitably small tissues (Hansen et al., 2010; Boudou et al., 2012), and EHT constructs have by now been established in 96-well format (Mills et al., 2017; Thavandiran et al., 2020). Known compounds such as quinidine, chromanol, erythromycin, and doxorubicin have been tested on EHTs in dose-dependent manner (Hansen et al., 2010). In a recent intersectoral, multicenter drug screening study EHTs achieved 93% accuracy for drug-induced cardiac response (Saleem et al., 2020). Responses aligned most closely to the free therapeutic plasma concentration, and contraction amplitude was found to be a good predictor of negative and positive inotropes in the 3D EHT culture, underlining the value of iPSC-CM-based EHTs for drug safety evaluation.

2. Cardiac Disease Models for Drug Development

Although confidence in engineered cell culture platforms increases with the addition of more and more compounds into validation libraries, these platforms are also being rapidly adopted for discovery and efficacy studies, especially in those cases wherein an animal model of the disease of interest is not available. Current state-of-the-art engineered cardiac tissue models were initially built from primary rodent cell sources and were validated by exposure to several compounds with known effects (Grosberg et al., 2011). However, there is a lack of studies validating engineered cardiac tissue models based on stem cells, in particular with cells originating from larger animals applied in preclinical cardiac research, such as dog. The importance of cell source is underlined by recent research showing that engineered tissues based on primary rodent and stem cell–derived human tissues display markedly different response to potential emerging exosome therapies against myocardial infarct (Yadid et al., 2020).

Long QT syndrome is an inherited arrhythmogenic syndrome characterized by aberrant ion channel function leading to sudden heart failure. Using patient-derived iPSC from diseased individuals, in vivo phenotypes can be recapitulated in vitro, as validated by electrophysiological studies that reveal action potential prolongation through patch-clamp and MEA assays (Itzhaki et al., 2011; Malan et al., 2016). The diseased phenotypes could be pharmacologically rescued in vitro as proof of successful tissue modeling. Congenital long QT syndrome can furthermore be modeled in 2D using fully automated 384-well assays based on voltage sensitive dyes. This may enable testing of patient cells at early stages of drug development and predict clinical outcomes of patient-specific differences (McKeithan et al., 2017).

Patient-specific iPSC BRAF mutant cells (Cashman et al., 2016) and genome-edited MYH7 mutant lines (Mosqueira et al., 2018) can be used to model dysfunction of the left ventricle. EHT constructs generated from these cells recapitulate in vivo tissue morphology with altered contractility and typical electrophysiology with accelerated twitch dynamics as well as altered signaling and metabolism. Likewise, dilated cardiomyopathy can be modeled in vitro using genetically edited iPSC-derived cardiomyocytes with mutations in TNNT2 (Sun et al., 2012) or TTN (Hinson et al., 2015). Cells from TNNT2 variant carriers showed reduced beating rates, compromised contraction, and abnormal sarcomeric α-actinin distribution, which could be partially rescued by treatment with blockers of β-adrenergic signaling, whereas TTN mutations manifested in aberrant sarcomeric bands, sarcomere insufficiency, and impaired response to adrenergic and mechanical stress in miniaturized EHTs.

Engineered cell culture platforms also facilitated the study of congenital arrhythmic disease, such as catecholaminergic polymorphic tachycardia (Goldfracht et al., 2019). A recent study with laminar 2D tissues using optogenetics and optical mapping revealed calcium wave re-entry as the substrate for the life-threatening arrhythmias (Park et al., 2019). By using genome editing the authors traced the phenotype to a single catecholamine-driven phosphorylation event of the ryanodine receptor, which helps in the development of treatment strategy.

Barth syndrome is a monogenic mitochondrial myopathy resulting in global ATP depletion, disrupted sarcomeric architecture, and decreased contractile stress generation. The in vitro model exhibited abnormalities in sarcomere assembly and myocardial contraction typical of the disease (Wang et al., 2014). Interestingly, using genetic and pharmacological rescue experiments, the model revealed that these aberrations were not linked to the lack of ATP per se as previously believed but rather with the loss of mitochondrial reactive oxygen species buffering capacity. Familial hypertrophic cardiomyopathy is a hereditary cardiac disorder characterized by cardiac hypertrophy associated with mutations in structural genes, such as MYH7 (Lan et al., 2013), MYBPC3, and ACTN2 (Prondzynski et al., 2019). EHT cultures of cardiomyocytes generated from patient-derived iPSCs replicate myocyte hypertrophy and myofibrillar disarray, leading to hypercontractility, impaired relaxation, and electrophysiological abnormalities with higher calcium sensitivity and prolonged action potential duration (Prondzynski et al., 2019).

Although patient-specific and genome-edited iPSCs enable the in vitro investigation of genetic diseases, most cardiovascular disease cases are associated with environmental conditions, such as arterial blockages and myocardial infarction. One of the early examples to model such factors in EHTs focused on effects of increased afterload on cardiac hypertrophy (Hirt et al., 2012). After 2 weeks, the soft PDMS posts were reinforced with metal braces leading to EHTs enlargement by 28.4% together with metabolic signs of hypertrophy and reduced contractile force after release. Cardiac fibrosis is another nonhereditary condition leading to reduced contractility and, in severe cases, heart failure. TGFβ-activated fibroblasts were mixed with cardiomyocytes, and the resulting EHTs recapitulated the distinctive features of cardiac fibrosis, including excessive collagen deposition, increased tissue stiffness, and passive tension (Mastikhina et al., 2020). Notably, treatment with the antifibrotic drug pirfenidone reduced tissue stiffness and reversed fibrosis-associated transcriptomic alterations, suggesting the utility of this platform for drug discovery. To recapitulate diseases that manifest in specific areas of the heart, a recent EHT platform was populated with iPSC-derievd cardiomyocytes differentiated to recapitulate the characteristics of atrial and ventricular cardiomyocytes. The platform enabled the characterization of normal electrophysiological and contractile parameters as well as the pathophysiological and pharmacological characteristics of the atrio-ventricular muscle (Zhao et al., 2019).

Ischemia-reperfusion injury (IRI) was also mimicked in vitro in EHTs by first culturing cells for 6 hours in low-medium volumes in a hypoxic chamber perfused by anoxic gas to emulate the nutrient deprivation, hyperkalemia, high lactate concentration, and low extracellular pH observed in cardiac ischemia (Chen and Vunjak-Novakovic, 2019). Subsequently, ambient oxygen concentrations were restored, and fresh medium was added to replenish nutrients and re-establish pH. IRI resulted in elevated ROS, increased lactate dehydrogenase release, and caspase activity as well as decreased contractility, which could be ameliorated using ischemic preconditioning in alignment with in vivo data. In addition, engineered human laminar tissues with embedded contractility sensors were used to monitor the reduction in contractility during IRI in real time (Yadid et al., 2020). Furthermore, the authors showed the therapeutic potential of human extracellular vesicles derived from endothelial cells and could trace the protective effects to protein modules involved in AMPK signaling, redox homeostasis, and unfolded protein response. Combined, these results suggest that the presented models might constitute interesting platforms for the development and translation of novel therapeutic approaches for IRI.

The number of examples using hiPSC-CMs on engineered in vitro models emphasizes the versatility of the system for both genetic and nongenetic heart pathologies. Faithful disease modeling of these conditions carries a high translational potential by pinpointing potential targets, screening for novel drug candidates, and facilitating the identification of patient-tailored solutions, particularly for rare conditions.

1. Muscle-Specific Toxicity

Myotoxicity of drug candidates is primarily evaluated using cell viability endpoints, which can furthermore be complemented with measurements of contractile force, oxygen consumption rate, calcium ion transient, and fatigue resistance. Studies presented so far have evaluated the myotoxicity of few selected compounds, whereas comprehensive screens have not yet been presented. For instance, cerivastatin and lovastatin caused progressive weakness and lipid accumulation in 3D myoblast-derived microphysiological myobundle cultures (Juhas and Bursac, 2014; Madden et al., 2015). Notably, reduced force production was observed at therapeutically relevant concentrations of the myotoxic statin cerivastatin (5 nM), whereas reduced force generation due to lovastatin treatment was only observed at 40-fold higher concentrations (200 nM). Mechanistically, in vitro experiments demonstrated that cerivastatin myotoxicity was mediated by OATP2B1-mediated statin uptake, whereas MRP1 overexpression on the sarcolemmal membrane ameliorated statin-induced myopathy (Knauer et al., 2010).

Besides statins, 3D muscle bundles successfully recapitulated chloroquine-induced autophagic myopathy as well as hypertrophy and increase in contractile strength upon exposure to the β-adrenergic agonist clenbuterol, mimicking effects observed in vivo (Madden et al., 2015). Moreover, live imaging of acetylcholine-induced calcium release using myogenic cells engineered to express the calcium reporter GCaMP6 revealed inhibition of calcium release upon treatment with the muscle relaxant tubocurarine. Overall, these results provide proof of principle that microphysiological muscle models can recapitulate the functional responses to pharmacological stimuli of native human muscles, suggesting that such systems might pose promising preclinical screening platforms for myotoxicity testing.

2. Mitochondrial Toxicity

Drug-induced mitochondrial dysfunction constitutes a common off-target toxicity not always identified during early drug development stages that commonly manifests in muscle injury (Dykens and Will, 2007). The difficulty to detect mitochondrial toxicity prior to clinical trials is mainly due to the lack of physiologically relevant cellular oxidative stress indicators in existing screening platforms and the relatively short duration of many preclinical studies. To this end, functional human 3D muscle models constitute an ideal system to evaluate mitochondria toxicity because of their high energy demands. Specifically, primary myoblast-derived 3D models allow for measurements of respiration rate using polarographic high-resolution respirometry and muscle fatigue (Davis et al., 2017). Furthermore, by using several mechanistically distinct respiration inhibitors, such as oligomycin (ATP synthase inhibitor), FCCP (electron transport uncoupler), and antimycin A (complex III inhibitor), the authors revealed that previously used models for the evaluation of mitochondrial toxicity, such as isolated mitochondria and myotube monolayer cultures, had higher basal respiration rate than bioengineered 3D muscle bundles, possibly due to mitochondrial stress and non–transport-limited oxygen consumption. Importantly, the model was able to recapitulate dose-dependent decrease in respiration and ATP generation upon treatment with the complex I inhibitor rotenone, which propagated into reduced twitch and tetanus force, thus providing proof of principle that 3D skeletal muscle models can be useful for bioenergetic profiling.

3. Muscle Disease Models

Duschenne muscular dystrophy (DMD) is a severe muscle-wasting disease caused by mutation in the DMD gene. Although few antisense oligonucleotide-based treatments have been approved by EMA and FDA in recent years (eteplirsen, golodirsen, viltolarsen, and casimersen), these drugs are specific for certain mutations. As such, development of novel drugs for DMD therapy remains an area of ongoing research. Drug development is furthermore complicated by the scarcity of models that recapitulate the complexity of the pathology, especially with intricate interplays between skeletal muscle tissues, vascular networks, and fibroblasts responsible for deposition of extracellular matrices, such as collagen I, fibronectin, and α-smooth muscle actin.

To study DMD in vitro, 3D myobundle models generated from differentiated immortalized myoblasts were cocultured with endothelial cells to recapitulate CD36+ and CD31+ vascularized endothelial networks (Bersini et al., 2018). The myobundles were also positive for myosin heavy chain marker, suggesting full maturation of the skeletal muscle fibers. Furthermore, introduction of muscle-derived human fibroblasts resulted in the formation of a sheath structure, resembling the connective endomysium found in vivo. Interestingly, incorporation of patient-derived fibroblasts phenocopied the elevated expression of collagen I and fibronectin seen in DMD in vivo, whereas the corresponding 2D cultures failed to recapitulate the pathophysiology. Notably, however, results are based on cells from a single patient and thus warrant validations.

Amyotropic lateral sclerosis (ALS) is a neuromuscular disease leading to progressive loss of motorneurons, muscular atrophy, and lack of voluntary muscle movements. Ten percent of ALS cases are attributed to SOD1 mutations, whereas the remainder is idiopathic. Two drugs have been approved for symptomatic treatments (riluzole and edaravone); however, both drugs only moderately delay disease progression, and response rates remain low.

To develop high fidelity and robust testing platforms for ALS drug discovery and development, patient-derived iPSCs were used to generate isogenic 3D skeletal muscle fiber models expressing muscle markers (myogenin) and neural spheroids expression markers for motor neurons, such as choline acetyltransferase and TUJ1 (Osaki et al., 2018). The models were cultured in two separate chambers of a microfluidic device, and, after 14 days, neurite elongation from the motor neuron spheroids toward the skeletal muscle fibers resulted in the formation of a mature NMJ (Fig. 8G), which could not be mimicked using conventional 2D cultures (Dimos et al., 2008). Notably, the model showed reduced skeletal muscle innervation with ALS patient-derived motorneuron spheroids (Fig. 8H). To allow for spatiotemporal control of neural activity and muscle contraction, the authors used an elegant optogenetic system that allowed for light-inducible cation influx into neurons. Upon 1-Hz optical stimulation, muscle contraction was observed in muscle fibers connected to either control or ALS-derived motor neurons; however, coupling was 32% less efficient in ALS neuromuscular junctions. Importantly, treatment with rapamycin and bosutinib reduced dysfunctional signal transmission and increased autophagy, demonstrating the neuroprotective effect of this combination therapy and highlighting the power of the platform for ALS drug development (Fig. 8I).

1. Elastomers

Elastomers are characterized by macromolecular structures with linear chains of high flexibility and few crosslinks, resulting in elastic and rubbery properties. Notably, chain elongation in elastomeric networks is reversible. The rather low cost of fabrication in combination with its facile use have made PDMS favorable for rapid prototyping. Moreover, PDMS offers a biocompatible, gas-permeable, and transparent surface, rendering this material suitable for cell-based applications. In addition to its common use as a bulk microfluidic material, the gas permeability of PDMS has also enabled its use for other innovative applications, such as the integration of embedded pumps or valves. In these applications, a thin layer of PDMS membrane can be embedded into a system to control and manipulate fluid inside microchannels (Eddings and Gale, 2006). Such approaches reduce the need for the addition of external materials that are prone to introduce contaminations into the system. PDMS membranes can further be harnessed for filtering applications as a more facile and cost-effective alternative to silicon membranes (Nam et al., 2019).

Despite such advantages, however, PDMS suffers from considerable drawbacks. PDMS strongly absorbs hydrophobic molecules, which hinders the use of PDMS devices for pharmacological and toxicological studies. In addition, leached monomers from bulk to the surface and medium can affect cellular phenotypes. Furthermore, the hydrophobic surface of PDMS often needs to be modified to allow for a smoother flow in microfluidic devices, particularly for narrow channels. Such surface modifications are often not durable, which in turn leads to restoring the original surface energy of this material.

To overcome the aforementioned limitations, the development of alternative elastomeric materials, such as poly(itaconate-co-citrate-co-octanediol) (Davenport Huyer et al., 2019), petrafluoroethylene-propylene (Sano et al., 2019), and poly(octamethylene maleate (anhydride) citrate) (Zhao et al., 2019), is ongoing. These efforts aspire to provide a more diverse range of flexible materials. However, further research and characterization efforts are required before these polymers can replace PDMS in organ-on-a-chip and tissue engineering applications.

2. Thermoplastics

Thermoplastics are a class of polymers that can be reversibly shifted from a solid to a liquid state and vice versa by altering the temperature below melting temperature (to melt) or above the glass transition temperature (to stiffen). Thermoplastics are linear or branched and not covalently crosslinked, resulting in stiff mechanical properties that enable solid support for microfluidic devices. Notably, thermoplastic materials often feature hydrophobic surfaces and are prone to swelling when subject to organic solvents. Although the use of thermoplastic materials is currently not as prevalent as PDMS, partly due to the need for complex processing and lack of oxygen permeability, the paradigm is slowly shifting in their favor, particularly due to the high degree of drug absorption correlated of PDMS. For instance, a multilayer PMMA chip with an integrated electrospun membrane in poly-l-lactide was developed for the culture of human endothelial cells (Chuchuy et al., 2021). The study showed a considerably lower absorption of hydrophobic molecules into PMMA chips compared with their PDMS counterparts. Similarly, a multilayer PMMA microfluidic chip containing a thermoplastic polyurethane membrane provided enough flexibility to form on-chip microvalves and micropumps with increased chemical resistance.

Polycarbonate (PC) is another well known low-cost thermoplastic material that has raised the interest in its off-the-shelf use in microfluidic devices. The ease of fabrication is an enabler for the rapid prototyping of PC-based microfluidic devices. Additionally, the gas impermeability of PC can leverage its application when precise control over oxygen concentration is needed. For instance, PC-based microfluidic devices have been used to enable oxygen sensing in a gastrointestinal human microphysiological system (Shah et al., 2016). Furthermore, PC is commonly used as material for integrated porous membranes (Novo et al., 2017; Wang et al., 2017a; Bai et al., 2021).

Polystyrene (PS) is another thermoplastic material with well established knowledge base in biology that is routinely used for standard 2D cell-culture formats. Furthermore, PS plates are more recently used for 3D cell culture after ULA coating and, because of the extensive knowledge available with respect to the cell culture, PS has also been leveraged for numerous applications in perfused microfluidic devices. Cyclic olefin polymers and copolymers are other examples of thermoplastic materials that have been used for microphysiological systems. Excellent optical properties, suitability for molding and very low rates of impurity have rendered them ideal candidates for organ-on-a-chip applications (Nunes et al., 2010). Nevertheless, the high costs of these materials and their availability mostly in the form of polymer granules have mostly limited their use to commercial applications. In addition to these prominent thermoplastics, multiple alternatives have been proposed for microfluidic devices, including polyether ether ketone (Song et al., 2004; Ildefonso et al., 2012) and polysulfone (Jiang et al., 2015). These materials are cheaper and universally compatible with organic solvents and, in the case of polysulfone, offer excellent hemocompatibility for applications involving perfusion of blood.

3. Thermosets

Thermosets are a class of polymers that undergo irreversible polymerization. Prior to polymerization, thermosets are in the form of one- or often two-component liquid resins. After monomer mixing and polymerization, thermosets form covalent crosslinks, which in turn imparts chemical and thermal stability to the material. SU-8, polyurethane, and polyimide are some examples of thermosetting polymers frequently used for microphysiological systems. SU-8 has established itself as a standard resin for microelectromechanical system applications because of its uniquely high thickness (up to hundreds of microns) while also offering a high lateral resolution that enables the formation of high-aspect-ratio features. Moreover, the high crosslink density and stability when exposed to high temperature and harsh chemicals have rendered it favorable for microfabrication processing. Although SU-8 is commonly used for the fabrication of master molds in soft lithography, its biocompatibility and high mechanical stability have also opened directed uses in medical device fabrication (Nemani et al., 2013).

A more recent generation of thermosetting materials based on thiol-ene chemistry has also been presented. Off-stoichiometry thiol-ene (Carlborg et al., 2011) and thiol-ene-epoxy (Carlborg et al., 2014) are formed via click-chemistry reactions occurring in one and two polymerization stages, respectively. This class of polymers is specifically tuned for microfluidic and laboratory-on-a-chip applications, offering unique features, such as robust covalent bonding (Sandström et al., 2015; Zandi Shafagh et al., 2017), in situ surface energy patterning (Zandi Shafagh et al., 2019), and direct biofunctionalization (Zandi Shafagh et al., 2018). Given these advantages, the use of thermosetting materials for the fabrication of microfluidic devices is on the rise.

1. Photolithography

Photolithography is a standard process through which a pattern, with microfeatures and nanofeatures is transferred from a photomask onto a substrate, often by means of exposure to UV light (Chen et al., 2012). In this process, polymers, referred to as photoresists, are often used as sacrificial layers for the pattern transfer into a solid substrate, such as silicon or glass. In contrast, for some specific applications, such as fabrication of master molds for PDMS replication, the polymer remains intact after UV exposure. For instance, SU-8 is a negative resist that covalently crosslinks upon exposure to UV light, often without subsequent etching. Because of the high thickness of SU-8 resist reaching hundreds of microns while having micron and submicron lateral resolution, high-aspect-ratio structuring of this material is feasible (Dervisevic et al., 2020). The use of photolithography has been commonly used for organ-on-a-chip applications, for instance for the photopatterning of silicon substrates and SU-8 to form 3D structured membranes mimicking intestinal villi (Esch et al., 2012).

Photolithography is by far the most common microfabrication method for creating molds and stamps for organ-on-a-chip systems. Nevertheless, photolithography is a complex and labor-intensive process that needs to be conducted inside standard cleanroom facilities. This often leads to the lack of scalability and the high cost of fabrication that is not favorable for biomedical applications.

2. Soft Lithography

Soft lithography is a well established method for the facile fabrication of microfluidic devices in academic settings (Xia and Whitesides, 1998). This is mostly originated from the low tooling costs required for the processing that allows for easy prototyping. Soft lithography is mainly based on casting PDMS onto a mold followed by subsequent heat curing and demolding, which results in the high-resolution replication of features onto a PDMS replica. Most microfluidic organ-on-a-chip devices are at least partially produced by soft lithography (Kim et al., 2012; Lee et al., 2013a, 2017b; Ahmad et al., 2014; Rigat-Brugarolas et al., 2014; Zhang et al., 2016b). Although soft lithography is the most common replication method in academic research laboratories, it is not commonly used in commercial settings because of the lack of scalability and long cycle time.

3. Microinjection Molding

Injection molding (IM) has been recognized as one of the most versatile techniques for the fabrication of polymer parts and devices. Microinjection molding (µIM) was further established for the fabrication of microfluidic devices (Yang et al., 2013). The scalability of µIM renders it suitable for the large-scale production of polymer devices for biomedical applications, which is of prime importance for applications, such as point-of-care diagnostics requiring disposable and cost-effective chips. Although µIM is a well established fabrication method for industrial production, its use in the academic domain is still limited mainly due to its complex processing conditions.

4. Reaction Injection Molding

Reaction injection molding (RIM) is based on the injection of two-component thermosetting resins in an enclosed molding setting (Heckele and Schomburg, 2003). Due to the low viscosity of resins, there is a higher replication fidelity associated with RIM compared with that of IM. UV curing systems have facilitated the use of RIM for the fabrication of microcomponents and nanocomponents as well as microfluidic devices (Sandström et al., 2015). In addition, RIM of microwell arrays with defined surface energy for applications in digital assays was reported (Zandi Shafagh et al., 2019). RIM has great potential for biomedical applications both in academic research and on industrial scales.

5. Nanoimprint Lithography

Nanoimprint lithography (NIL) is a stamp-based approach for the replication of high-resolution features from a stamp onto a substrate (Guo, 2007). NIL can be used either for direct pattern transfer implemented for end applications or followed by subsequent etching through the substrate material. NIL enables high-resolution and cost-effective manufacturing of micropatterned and nanopatterned surfaces and has found great potential for applications in life sciences. For instance, we recently developed a method for facile nanoimprinting of high-aspect-ratio nanotopographies, which modulate the mechanosensing of human hepatic cells and provide precise spatial control of cell attachment (Zandi Shafagh et al., 2020). NIL has also been implemented for the fabrication of microfluidic and nanofluidic devices with nanochannels (Sun et al., 2016). Although the capability of NIL for manufacturing of organ-on-a-chip devices is still not fully realized, the cost effectiveness and scalability of roll-to-roll nanoimprinting can pave the way for high-throughput replication of such devices (Stuart and Chen, 2009).

6. Electron Beam Lithography

Electron beam lithography (EBL) is a maskless nanopatterning technique that can be used for the direct writing of nanotopologies into an e-beam resist (Chen, 2015). EBL enables ultra-high-resolution fabrication of nanofeatures with high precision and low structural collapse and has been used for high-resolution patterning of biomolecules on nanopatterned surfaces that allows for spatially refined biofunctionalization of polymer substrates (Zandi Shafagh et al., 2018). Furthermore, direct protein inactivation through exposure to electron beams can be used to guide cell attachment, resulting in cell shapes constrained to the EBL-written pattern (Glezos et al., 2002; Rundqvist et al., 2007). In addition to surface patterning, EBL has been used for the fabrication of various microfluidic units, including microchannels and micromixers (Mali et al., 2006). However, because of the costs and low availability of e-beam infrastructure, the use of EBL for organ-on-a-chip devices remains in its infancy.