Modeling Cell-Autonomous Motor Neuron Phenotypes in ALS Using iPSCs

Abstract

Amyotrophic lateral sclerosis (ALS) is an aggressive and uniformly fatal degenerative disease of the motor nervous system. In order to understand underlying disease mechanisms, researchers leverage a host of in vivo and in vitro models, including yeast, worms, flies, zebrafish, mice, and more recently, human induced pluripotent stem cells (iPSCs) derived from ALS patients. While mouse models have been the main workhorse of preclinical ALS research, the development of iPSCs provides a new opportunity to explore molecular phenotypes of ALS within human cells. Importantly, this technology enables modeling of both familial and sporadic ALS in the relevant human genetic backgrounds, as well as a personalized or targeted approach to therapy development. Harnessing these powerful tools requires addressing numerous challenges, including different variance components associated with iPSCs and motor neurons as well as concomitant limits of reductionist approaches. In order to overcome these obstacles, optimization of protocols and assays, confirmation of phenotype robustness at scale, and validation of findings in human tissue and genetics will cement the role for iPSC models as a valuable complement to animal models in ALS and more broadly among neurodegenerative diseases.

Introduction

First described by Jean-Martin Charcot in 1869, amyotrophic lateral sclerosis (ALS) is a neurodegenerative disease predominantly characterized by the loss of cortical motor neurons in the brain and lower motor neurons in the brainstem and spinal cord. The progressive dysfunction and degeneration of these neurons lead to initial symptoms that range from focal weakness in a limb to difficulty in swallowing or speaking, depending on the specific neurons first affected. While the presenting symptoms of ALS can be substantially broader, including muscle spasms or even generalized fatigue, they nonetheless invariably progress, generally within two or three years, to the same late stages of paralysis and respiratory failure (Brown and Al-Chalabi, 2017).

The vast majority of ALS cases (90–95%) are apparently sporadic (SALS), occurring in patients with no known familial history of the disease. A minority of cases (5–10%) are familial (FALS) and occur in patients who have a family history of ALS. The list of FALS genes includes more than 40 members and spans a broad range of cellular functions within topologically distinct subcellular compartments; however, the different mutations converge on a surprisingly homogeneous clinical ALS disease phenotype. Indeed, understanding that mechanistic funneling process is one of the central challenges in identifying and developing better treatment targets. The first and best-understood FALS gene, superoxide dismutase 1 (SOD1) (Rosen et al., 1993), accounts for some 20% of FALS cases. Expansion of an intronic hexanucleotide repeat in the C9orf72 gene causes about 30% of FALS (Dejesus-Hernandez et al., 2011; Renton et al., 2011), and TARDBP (also known as TDP-43) (Sreedharan et al., 2008) and FUS (Kwiatkowski et al., 2009; Vance et al., 2009) mutations are each responsible for about 5% of FALS cases (Taylor et al., 2016). Most FALS mutations tend to be inherited in a dominant monogenetic manner, although they can be dominant with incomplete penetrance, recessive, and X-linked (Al-Chalabi et al., 2017). Furthermore, the variation of disease onset and progression – even within families – can be quite variable (Swinnen and Robberecht, 2014). Indeed, the log-linear relationship between age of disease onset and incidence suggests a multi-step disease process (Al-Chalabi et al., 2014), which results from a combination of environmental interactions and genetic risk.

Some of the ALS genes are not unique to ALS and cause the related condition frontotemporal dementia (FTD) or an overlap syndrome of ALS and FTD (FTD-ALS) (Bang et al., 2015). The identification of C9orf72 repeat expansion as a frequent cause of both ALS and FTD established the two as a disease spectrum with shared pathology. While the C9orf72 gene in typical healthy controls has hexanucleotide repeats of less than 30, the gene in ALS and FTD patients can contain repeats in the hundreds or even thousands (Gendron and Petrucelli, 2018). Mutation in several other genes, including SQSTM1, VCP, UBQLN2, and OPTN can also be seen in either disease or the overlap syndrome.

As perhaps expected, FALS and SALS are connected in an intricate and incompletely understood manner. First, de novo single disease-causing mutations occur in a small percentage of SALS-classified patients (Al-Chalabi et al., 2017). Second, some FALS patients can be misclassified as SALS due to incomplete genetic penetrance or late onset. Even the capacity for disease generation can depend on genetic background: a single SOD1^D90A allele causes typical dominant ALS in most genetic backgrounds but no effect in a specific northern Scandinavian population, in which two copies are necessary to cause disease (Pasinelli and Brown, 2006). The relationship between SALS and FALS is particularly complicated in the C9orf72 repeat expansion, where limited familial genetic information as well as repeat heterogeneity within different tissues of an individual patient confound classification, although penetrance seems to be consistent between SALS and FALS patients (Murphy et al., 2017).

For both SALS and FALS, no cure exists, and existing treatments only extend lifetime by several months. To date, the U.S. Food and Drug Administration has approved two medications, riluzole, and edaravone. Despite riluzole being used for more than two decades, its precise mechanism of action remains unclear. Current evidence supports the reduction of presynaptic glutamate release, which may be due to block of sodium or calcium channels, as well as a decrease of persistent sodium currents (Bellingham, 2011). Edaravone possesses neuroprotective features, including reduction of oxidative stress, that appears to ameliorate the disease (Abe et al., 2017). To what extent these approved drugs yield greater benefit in specific disease subgroups remains unclear, particularly as edaravone’s approval was based on an effect in a relatively restricted subgroup of early-disease patients.

The genetic identification of SOD1 as the first FALS gene enabled the generation of the SOD1 mouse model, which has been the primary tool for understanding ALS pathophysiology and developing novel therapeutics for more than two decades (Gurney et al., 1994; Rosen et al., 1993). While the mouse model has been invaluable for identifying the molecular processes contributing to the disease, it has not led to successful treatment of ALS in humans, and there is a disagreement as to its limitations and how it should best be used as a pre-clinical tool (Benatar, 2007). Reasons for this failure of translation to the clinic include insufficiently large effects in the mice, poor design of both mice and human studies, limitations due to feasibility of treatment in mice versus patients, and generalizability from SOD1 to other ALS variants.

By and large, drug trials in the mouse model have not yielded the large-scale benefit seen with SOD1 knockdown (Smith et al., 2006) or growth factor treatments (Kaspar et al., 2003). The limited success with small molecules may be due to the strong overexpression of mutant SOD1, which was estimated to be more than over 25 copies in the original mouse model (Alexander et al., 2004; Gurney et al., 1994). Thus, it remains unknown whether the lack of substantial improvement in the mice reflects an unrealistically and indeed unnecessarily difficult target, on account of which potentially promising clinical targets and compounds are discarded, or whether the lack of human success reflects only the limited and in many cases insufficiently robust benefits in the mouse model. Compounding the issue, insufficient sample size and sex matching have hampered many mouse trials (Ludolph et al., 2010).

Several aspects of human clinical trial design and indeed the human disease itself could limit the success of potentially promising drugs. In humans, the lack of sufficient pharmacokinetic investigation to demonstrate high and stable drug levels in the CSF, pharmacodynamic studies to show target engagement, and biomarkers leaves most clinical studies without a clear understanding of their failure (Zinman and Cudkowicz, 2011). While treatment initiation in mice can occur either prior to disease onset or at an early state of disease progression, the equivalent time of treatment initiation in humans may be substantially later, due to the typical one-year or greater delay between symptom onset and disease diagnosis. Thus, the more advanced corresponding disease state at time of treatment initiation in humans could render a drug that performed well in mice ineffective in humans.

The greatest issue, however, is that mechanisms and targets capable of successfully treating one type of FALS may not generalize to another. Most clinical studies enroll both SALS and FALS patients and have not been powered to detect benefit in small subsets of, for example, SOD1 patients, although transformative effects in SOD1 patients within longer-term treatment studies would presumably have been identified. With improved understanding of ALS genetics, the number of yeast, invertebrate, and mouse models has increased. Genetic and phenotypic conservation across species has led to widespread validation of yeast and invertebrate findings using human post-mortem samples and human genetics (Elden et al., 2010; Freibaum et al., 2015; Jovičić et al., 2015; Kramer et al., 2018; K. Zhang et al., 2015). For mouse models, gene editing has enabled transgenic modifications that avoid overexpression effects (Fratta et al., 2018; Scekic-Zahirovic et al., 2016; White et al., 2018). However, studies using human cells and ALS postmortem samples indicate the critical importance of intronic and other non-coding regions, which are poorly conserved in non-human models (Klim et al., 2019; Ling et al., 2015; Melamed et al., 2019), as well as human genetic background (Pasinelli and Brown, 2006). Most importantly, yeast, invertebrate, and mouse models do not exist for SALS, which represents 90% of ALS cases, and subgroups of SALS may each require their own treatments.

The availability of human induced pluripotent stem cells (iPSCs) provides a powerful new modeling approach for both FALS and potentially SALS within specific patient genetic backgrounds. As in other neurodegenerative diseases, iPSC-derived neurons may reveal salient early-stage disease-driving mechanisms that may constitute more potent targets than those identified through end-stage post-mortem tissue sample analyses. Thus, a combination of non-human and human iPSC models, confirmed by ALS genetic or postmortem studies, may yield the most benefit in identifying robust disease mechanisms and promising clinical compounds.

The first iPSC line was reprogrammed from mouse somatic cells in 2006 (Takahashi and S. Yamanaka, 2006), followed by the transformative development of human iPSCs in 2007 (Takahashi et al., 2007). Bearing quintessential features of embryonic stem cells, iPSCs derived using virally-delivered transcription factors maintain an unlimited proliferation capability and the capacity to differentiate into specific cellular subtypes of human endoderm, mesoderm, and ectoderm while avoiding the ethical issues surrounding human embryonic stem cells.

The first spinal motor neuron differentiation protocol, using mouse embryonic stem cells, was developed using a combination of neural induction followed by caudal and ventral patterning, as occurs during development (Wichterle et al., 2002). The process was then adapted to human embryonic stem cells (Li et al., 2005) and subsequently to human iPSC cells from ALS patients (Dimos et al., 2008). Since then, many groups have designed and validated more than a dozen varied protocols for differentiating spinal motor neurons from iPSCs (Sances et al., 2016). The three main steps across most of these protocols are neuralization, caudalization, and ventralization. Neuralization is most commonly performed by inhibiting TGF-ß and BMP signaling together, termed dual-SMAD inhibition (Chambers et al., 2009), using SB431542 and either noggin or more recently LDN193189, respectively. Caudalization was initially directed by retinoic acid treatment and then in later studies supplemented by Wnt activation via inhibition of glycogen synthase kinase 3 by CHIR-99021 (Maury et al., 2015). Ventralization occurs through a combination of hedgehog signaling and BMP antagonism, typically using the smoothened agonist purmorphamine and dorsomorphin, respectively. However, the precise order, timing, and combinations of the cocktails differ substantially (see Figure 3 from Sances et al., 2016 for descriptive figure). Separately, strategies using transcription factor overexpression generate spinal motor neurons from pluripotent stem cells (Mazzoni et al., 2013; Shi et al., 2018).

The potential promise of identifying upstream disease phenotypes and their mechanisms in both FALS and SALS has led to an explosion of iPSC-based ALS modeling studies (S. Lee and Huang, 2017). These generally fall into one of three categories: exploratory-based, confirmation-based, or screen-based. Exploratory studies identify phenotypes in predominantly FALS compared to control iPSC-derived motor neurons and then validate those phenotypes using human post-mortem tissue or genetics. Confirmation studies validate a mechanism, gene, or drug candidate identified from other sources, including yeast, invertebrate, and rodent models. Screen-based studies use iPSC motor neurons demonstrating in vitro ALS phenotypes to identify compounds or genetic targets capable of mitigating the phenotypes.

Together, these approaches look to maximize the translatability of in vitro phenotypes to the fundamental in vivo disease process that occurs in patients. Ultimately the strength of this connection will determine the true value of targets and phenotypes identified by iPSC modeling studies. Nevertheless, questions remain as to how well the phenotypes observed in vitro capture the disease pathogenesis in humans. Initial hints may come from a greater understanding of the time course of in vitro phenotype development, correlation with other associated in vitro processes that may reflect ALS pathophysiology, and the use of molecular and genetic perturbations to establish causal connections between the intervention and mitigation of phenotypes. Ultimately, though, complementary information from mouse model, human genetic, and post-mortem studies may still be necessary to resolve which phenotypes, responsible genes, and putative target compounds most likely reflect upstream disease-driving as opposed to epiphenomena-related or even compensatory mechanisms.

Modeling Major FALS Genes and Their Phenotypes

SOD1

The first critical FALS gene, superoxide dismutase 1 (SOD1), was identified in 1993 (Rosen et al., 1993). Over 170 different mutations in SOD1 cause ALS, almost all in a dominant manner (http://alsod.iop.kcl.ac.uk/), with differing and in some cases highly variable age of onset and severity. The disease-causing mutations fall broadly into the 154-amino acid polypeptide. ALS results from a gain-of-function toxicity of mutant SOD1, in which dismutase activity has no correlation with ALS disease severity (Cleveland et al., 1995). Transgenic mouse models for ALS based on the overexpression of mutant SOD1 proteins capture the quintessential behavioral and molecular features of the human disease (Gurney et al., 1994) and have formed the foundation for mechanistic inquiry into ALS.

Mutant SOD1 protein affects many cellular pathways, including oxidative stress and mitochondrial function, sub-cellular transport, and ER stress (Bruijn et al., 2004). SOD1 ALS may be a relatively distinct variant of ALS, as it does not have intracellular TDP-43 aggregates that are present in over 95% of ALS cases (E. B. Lee et al., 2011). Although still controversial, the largest studies have not shown misfolded SOD1 in non-SOD1 variants of ALS (Da Cruz et al., 2017; Paré et al., 2018).

Several stem cell modeling studies have examined the effects of mutant SOD1. Using motor neurons derived from SOD1^A4V ALS patients and healthy and gene-edited controls, Kiskinis and colleagues identified a range of disease phenotypes (Kiskinis et al., 2014). These included a reduction in survival and neurite outgrowth in SOD1^A4V compared to isogenic control motor neurons, but not the formation of typical detergent-resistant SOD1 aggregates under basal conditions. RNA sequencing studies revealed transcriptional differences in genes related to cytoskeletal organization, mitochondrial transport, and oxidative and ER stress. These findings were supported by abnormalities of mitochondrial mobility and structure, as well as increases in catalase activation, supporting oxidative stress, and the unfolded protein response, indicating ER stress. Treatment with salubrinal, which reduces ER stress, improved motor neuron survival. Interestingly, spinal cord motor neurons have inherently higher baseline levels of ER stress and thus may be particularly vulnerable. Another study using an isogenic pair of SOD1^E100G and control iPSCs confirmed the reduction in ALS motor neuron survival and involvement of ER stress pathway activation, and implicated extracellular signal-regulated kinase (ERK) and c-Jun N-terminal kinases (JNK) signaling pathways (Bhinge et al., 2017).

At the same time, Chen and colleagues observed neither SOD1 aggregation nor mitochondrial abnormalities in iPSC-derived ALS motor neurons harboring SOD1^A4V and SOD1^D90A mutations (Chen et al., 2014). They reasoned that the strong overexpression of SOD1 in mouse models may overemphasize these features (Gurney et al., 1994), which they argued are less prominent in human pathology. However, they did document neurofilament aggregates, despite the physiologic SOD1 expression, with early formation only four days after motor neuron plating. Thus, neurofilament accumulation may be an early event in ALS pathogenesis and a reasonable therapeutic target, consistent with its established clinical use as a biomarker (Bowser et al., 2011).

In other studies, mutant SOD1 motor neurons have been used either for screening or to test candidate compounds identified by other means. Imamura and colleagues observed reduced survival in mutant SOD1^L144FVX and used this finding as the basis for a phenotypic screen in which they identified Src/c-Abl inhibitors including bosutinib (Imamura et al., 2017). In one of the most powerful results of the study, the benefit of bosutinib treatment was confirmed independently in TDP-43, C9orf72, and two of three SALS lines. As an interesting aside, they did observe misfolded SOD1 protein both by immunohistochemistry and Western blot in the SOD1 motor neurons, and the accumulation was reduced by bosutinib treatment. Likewise, a recent large study discussed in detail below also observed SOD1 aggregates by immunohistochemistry (Fujimori et al., 2018). In another screen based on trophic factor withdrawal from mouse embryonic stem cell-derived motor neurons, Yang and colleagues found that kenpaullone, a dual inhibitor of GSK3 and HGK kinases, prolonged the survival of human iPSC-derived motor neurons, including wild-type and SOD1^L144F and TDP43^M337V ALS mutations (Yang et al., 2013).

Looking at the SOD1-based ALS studies together, thematic questions - both specific to SOD1 and also relevant to the broad evaluation of iPSC modeling studies - emerge. One relates to the intersection of post-mortem human pathology, animal models, and iPSC-derived motor neurons. In SOD1 ALS post-mortem tissue and in mouse models, the presence of SOD1 granules is a defining feature; in contrast, they have been largely absent from iPSC SOD1 motor neurons, with the exception of the reports by Imamura (Imamura et al., 2017) and Fujimori (Fujimori et al., 2018). Whether these differences result from variation in lines, differentiation protocol, or culture conditions remains unclear. However, as discussed by Chen and colleagues (Chen et al., 2014), both the size and abundance of SOD1 granules are notably higher in the mouse models than in humans. Thus, iPSC modeling studies raise the concern that mouse models, on account of the strong overexpression of mutant SOD1, may be overly focused on SOD1 granules in SOD1 ALS and consequently mask potential target pathways like ER stress and neurofilament accumulation.

TDP-43

Tar DNA binding protein (TDP-43) was identified as the main component of ubiquitinated inclusions in both ALS and FTD patients (Arai et al., 2006; Neumann et al., 2006), and TDP-43 mutations are responsible for about 5% of FALS (Sreedharan et al., 2008). TDP-43 is a highly conserved ribonuclear protein that normally shuttles between the cytoplasm and nucleus and mediates a range of cellular functions including RNA regulation, splicing control, intron suppression, and mRNA transport and stability (Cohen et al., 2011). As cytoplasmic aggregates of phospho-TDP-43 are found in almost all FALS and SALS post-mortem samples (with the exception of SOD1 and probably FUS variants), understanding the events regulating TDP-43 dysfunction may yield powerful opportunities for therapeutic interventions. In the case of TDP-43 FALS, mutant TDP-43 likely exerts both gain and loss of function contributions, with the added caveat of an autoregulatory mechanism that maintains tight control over TDP-43 levels (Ayala et al., 2011; Polymenidou et al., 2011).

The gain of function, which may reflect altered transcript regulation in the nucleus as well as established toxicity of aggregates in the cytoplasm (Igaz et al., 2009; Pesiridis et al., 2011), is perhaps most clearly illustrated by homozygous overexpression of wild-type TDP-43 in mice causing rapid death with many features of motor neuron disease (Becker et al., 2017). Furthermore, White and colleagues generated TDP-43^Q331K knock-in mice that showed increased TDP-43 transcript levels, impaired cognitive function, and loss of parvalbumin-positive cortical interneurons in male mice, albeit without an obvious spinal motor neuron phenotype (White et al., 2018).

In contrast, loss of function mechanisms include splicing defects that lead to the depletion of genes with long introns (Polymenidou et al., 2011) and emergence of cryptic exons (Ling et al., 2015) including an abnormal short isoform of stathmin-2 (Klim et al., 2019; Melamed et al., 2019). Adding to the complexity, the gain and loss of function phenotypes may share similar mechanisms, in particular abnormal splicing: gain of TDP-43 function results in skipping of constitutive exons while loss of function yields intron retention (Fratta et al., 2018).

In an important iPSC modeling paper, Egawa and colleagues examined seven control and nine mutant TDP-43 ALS lines, including three clones each from patients harboring Q343R, M337V, and G298S mutations (Egawa et al., 2012). Key findings included decreased protein levels of neurofilament medium and light chain RNAs, increased TDP-43 RNA, and detergent-insoluble TDP-43 accumulation with cytoplasmic granules linked to the spliceosomal factor SNRPB2 and exacerbated by stress granule induction. Using candidate modifiers of transcription and spliceosome activity, the authors showed that the histone deacetylase inhibitor anacardic acid mitigated many of the disease phenotypes observed in their model.

In a similar study, Bilican and colleagues examined motor neurons from two iPSC clones of TDP-43^M337V and found increased soluble and insoluble TDP-43 protein levels compared to two wild-type controls, without differences in TDP-43 RNA levels (Bilican et al., 2012). TDP-43 ALS motor neurons showed nearly a three-fold increased risk of death using longitudinal Kaplan-Meier survival analysis of individual cells. Evaluation of different cellular stressors demonstrated an increased susceptibility to inhibition of phosphoinositide 3-kinase but not mitogen-activated protein kinase or ER stress induction in the mutant motor neurons.

Other studies did not show TDP-43 insoluble granules but nonetheless did observe disease phenotypes, including impaired motor neuron survival, neurofilament accumulation, and impaired axonal transport in TDP-43^S393L and TDP-43^G294V iPSC motor neurons (Kreiter et al., 2018). This result provides additional support for the idea that granule accumulation may be a downstream event in the disease, as suggested by the aforementioned SOD1 iPSC studies (Chen et al., 2014; Kiskinis et al., 2014), and is consistent with the idea that upregulation of TDP-43 and loss of nuclear TDP-43 may underlie the key disease events (Klim et al., 2019; Melamed et al., 2019). Surprisingly, sorbitol treatment, which has been used to induce stress granule and promote the cytoplasmic localization of TDP-43 (Dewey et al., 2011), relieved the transport deficits (Kreiter et al., 2018).

Stem cell modeling offers the opportunity to examine the effects of mutations that cause a range of disease severity. The TDP-43^A90V mutation, which is located in the nuclear localization signal, has been observed in ALS/FTD families (Winton et al., 2008) but also controls (Guerreiro et al., 2008; Kabashi et al., 2008), indicating that it may be an ALS risk factor or exert a milder phenotype. Zhang and colleagues evaluated TDP-43^A90V motor neurons using iPSC modeling and found that the baseline percentage of cells with mislocalized TDP-43 in the A90V mutant was similar to control lines (Z. Zhang et al., 2013); however, following treatment with staurosporine, a non-selective protein kinase inhibitor, a larger percentage of TDP-43^A90V mutant neurons showed cytoplasmic mislocalization of TDP-43 compared to control neurons. These findings were consistent with in vitro overexpression studies that showed a more modest capacity of TDP-43^A90V to generate insoluble TDP-43 aggregates than other typical mutations such as M377V (Wobst et al., 2017). These results support the idea that stressors may be useful or required to identify phenotypes, particularly in the case of less aggressive mutations.

Motor neurons from mutant TDP-43 patient iPSCs have also been used to test and implicate specific disease mechanisms including axonal transport and autophagy. Alami and colleagues used drosophila and mouse neurons to show that TDP-43 is trafficked to distal neurites in ribonucleoprotein granules, which were defined in several ways: using labeled TDP-43; by co-localization with the RNA trafficking protein Stafuen as well as a tagged oligonucleotide that labels neurofilament-L mRNA within the granules (Alami et al., 2014). Consistent with these observations, axonal transport of the neurofilament-L-labeled ribonucleoprotein granules was impaired by TDP-43^M337V, TDP-43^A315T, and TDP-43^G298S mutations.

Investigating the potential mechanism of cytoplasmic TDP-43 accumulation, Barmada and colleagues showed that TDP-43 mutation results in shorter protein half-lives, and that stimulating autophagy increases TDP-43 clearance and improves in vitro survival (Barmada et al., 2014). TDP-43^M337V iPSC-derived motor neurons had impaired survival at baseline compared to control lines. Fluphenazine and methotrimeprazine, which had both been identified in their screen as promoters of autophagy, each reduced cytoplasmic TDP-43 and improved neuronal survival.

As discussed below, nuclear cytoplasmic transport disruption is a rapidly expanding focus for ALS mechanistic investigation, particularly in the C9orf72 repeat cases, but likely also as a broader link between nuclear pathological processes and cytoplasmic inclusions in neurodegenerative diseases (H. J. Kim and Taylor, 2017). Proteomic analysis of overexpressed mutant TDP-43 C-terminal fragment using proximity labeling detected an abundance of bound nuclear cytoplasmic transport components (Chou et al., 2018). Staining of patient fibroblasts from TDP-43, C9orf72-expansion, and SALS patients, as well as iPSC-derived motor neurons from TDP-43 FALS patients confirmed abnormalities in the nuclear pore complex and laminin morphology. Staining of human post-mortem tissue from subsets of TDP-43 and sporadic ALS patients showed disrupted Nup-205 immunoreactivity (Chou et al., 2018).

Looking at results from the papers together, fundamental questions remain unresolved. Some studies have recapitulated cytoplasmic TDP-43 localization and detergent-insoluble aggregates, as seen in post-mortem human tissue, while others have not (Bilican et al., 2012; Egawa et al., 2012; Fujimori et al., 2018; Kreiter et al., 2018; Osaki et al., 2018). Reports also differ as to the dysregulation of TDP-43 and whether it occurs on the RNA or protein levels (Bilican et al., 2012; Egawa et al., 2012). As in the case of SOD1 mutation, several reasons such as differentiation protocol, motor neuron purity, and maturation time may all contribute to the variation among studies. Most, but not all, studies that found basal TDP-43 aggregation also observed reduced neuronal survival; however, the relationship and specific effects of TDP-43 dysregulation, loss of nuclear TDP-43, increase in cytoplasmic TDP-43, granule formation, and downstream effects of the granules require further investigation.

FUS

The fused in sarcoma (FUS) gene was identified as a FALS gene in 2009 (Kwiatkowski et al., 2009; Vance et al., 2009). FUS mutations causing ALS largely fall within two regions, the glycine-rich prion-like domain and the nuclear localization signal (NLS) located at the extreme C-terminus of the protein (Lagier-Tourenne et al., 2010), and lead to large, ubiquitinated, and TDP-43-negative neuronal cytoplasmic inclusions. Different mutations in the FUS gene lead to a broad range of ALS severity and disease onset. Several mutations either severely disrupt (FUS^P525L) or delete (FUS^R495X; FUS^M511Nfs*⁶) the nuclear localization signal and cause juvenile-onset ALS (Schwartz et al., 2014). In contrast, FUS^R514S and FUS^R521C are both missence mutations that lead to more typical ALS disease onset. These mutations allow researchers to correlate the strength of in vitro phenotypes with degree of disease severity.

FUS plays key roles as an RNA-binding protein that participates in RNA transcription, splicing, transportation, and translation (Lagier-Tourenne et al., 2010). Similar to TDP-43, FUS actively exports mRNA by shuttling between the nucleus and the cytoplasm. In addition to these roles, the FUS protein can affect DNA damage repair by interacting with histone deacetylase 1 (Wang et al., 2013) and stress granules (Dormann et al., 2010). Finally, FUS serves as a quintessential example of mediating liquid-liquid phase separation, by which interactions among low-complexity domain regions promote droplet and potentially aggregate formation (Shin and Brangwynne, 2017). These roles define several areas of mutant FUS dysfunction and potential contribution to ALS pathology.

Using iPSC-derived motor neurons harboring FUS nuclear localization signal mutations, Guo and colleagues demonstrated cytoplasmic mislocalization of FUS, as well as transport deficits of both mitochondria and ER vesicles (Guo et al., 2017). Isogenic correction of the FUS^R521H mutation resolved the transport deficits and thus demonstrated that the mutation was necessary for the phenotype. In addition, overexpression of mutant FUS in control motor neurons was sufficient to induce the disease phenotype. As previous experiments using Charcot-Marie-Tooth type 2 and ALS mouse models demonstrated the benefit of HDAC6 inhibition, the group tested HDAC6 blockers as well as HDAC6 knockdown, both of which reversed the transport deficits but interestingly did not correct FUS mislocalization.

Building on similar themes, Naumann and colleagues elegantly showed stepwise degenerative phenotypes using microfluidic chambers: first, impaired organelle trafficking; second, distal axonal degradation; and third, motor neuron cell death. While the trafficking deficits were present as early as 21 days in culture, cell death did not occur until after 100 days in culture (Naumann et al., 2018). Interestingly, double strand DNA breaks were present as early as day 14 culture, which precedes FUS mislocalization and suggests an upstream role in FUS pathology. Induction of DNA double strand breaks in control motor neurons mimicked the FUS phenotype of distal axonal impairment of mitochondrial and lysosomal motility, in addition to compromise of mitochondrial membrane potential and structural integrity. Whereas laser microirradiation induced the rapid recruitment of FUS to DNA damage sites in control motor neurons, the presence of FUS mutations impaired this response, suggesting that DNA damage exacerbates FUS mislocalization. Reciprocally, impairment of FUS nucleocytoplasmic shuttling also induced DNA damage. Finally, promoting DNA damage repair via poly(ADP-ribose) glycohydrolase inhibition reduced FUS cytoplasmic accumulation and distal axon pathology. Thus, the triad of DNA damage repair, ribonucleoprotein localization, and axonal transport emerge as integral and self-promoting components of FUS toxicity. Where and how other mechanisms such as autophagy, which decreases FUS-dependent stress granules (Marrone et al., 2018), fits into this network remains to be seen.

In contrast with SOD1 and TDP-43, FUS studies have uniformly confirmed pathogenic FUS mislocalization to the cytoplasm and accumulation within stress granules (Guo et al., 2017; Higelin et al., 2016; Ichiyanagi et al., 2016; Lenzi et al., 2015; Marrone et al., 2018; Naumann et al., 2018). Many phenotypes are present without the addition of specific cellular and DNA damage stressors, supporting the physiological saliency of iPSC models for FUS ALS mutations. iPSC modeling using FUS may benefit from the unfortunate existence of particularly aggressive frameshift and deletion mutations that severely perturb or eliminate the nuclear localization signal and cause juvenile-onset ALS. These mutations as well as FUS^P525L, which markedly disrupts the nuclear localization domain, led to more mislocalization and extreme cellular phenotypes compared to FUS^R514S and FUS^R521C, which cause a less aggressive form of the disease (Guo et al., 2017; Higelin et al., 2016; Lenzi et al., 2015).

The combined picture of early DNA damage with later structural deficits that begin and remain for a substantial period localized to the distal axon support a “dying back” process (Fischer et al., 2004), albeit with signalling from the cell body. These early axonal phenotypes – combined with the importance of axonal biology in ALS demonstrated by FALS genes such as PFN1, DCTN1, and TUB4A – support the development of improved ways to use iPSCs to study axonal pathology. Achieving neuromuscular junctions in co-cultures of iPSC-derived motor neurons and predominantly primary muscle may be particularly useful for such investigations (Santhanam et al., 2018; Steinbeck et al., 2016). Developing these co-cultures has proven difficitult using iPSC-derived muscle; only one group has been successful and used commercially-derived muscle precursors from an iPSC line with a custom-fabricated microfluidic device (Osaki et al., 2018).

In ALS, FUS aggregation has generally been limited to patients harboring FUS mutations. In contrast, some FTD subtypes (Urwin et al., 2010) and two rare potentially related conditions, neuronal intermediate filament inclusion disease (Neumann et al., 2009) and basophilic inclusion body disease (Munoz et al., 2009), have FUS accumulation without known FUS gene mutation. How FUS mislocalization occurs in these cases remains unclear. While iPSC modeling studies using motor neurons in FUS ALS have been quite successful, comparisons across neuronal subtypes may yield further clues into the mechansims of cell-type specificity for these different diseases.

C9orf72

Discovering the chromosome 9 open reading frame 72 (C9orf72) hexanucleotide repeat expansion was a pivotal step in elucidating the pathophysiology of ALS and its relationship with frontal temporal dementia (FTD) (Dejesus-Hernandez et al., 2011; Renton et al., 2011). C9orf72-expansion is responsible for about 30% of FALS and 10% of SALS in Europeans, although substantial variation exists among ethnicities and geographical regions (Al-Chalabi et al., 2017; X. Liu et al., 2018). The presence of ALS, FTD, and ALS/FTD patients within the same C9orf72 repeat expansion families established that the two diseases share common mechanisms and should be considered as a continuous spectrum. There are three major pathological features frequently seen in post-mortem tissue from C9orf72 repeat expansion patients: predominantly nuclear RNA foci; five distinct dipeptide repeats, Gly-Ala, Gly-Pro, Gly-Arg, Pro-Ala, and Pro-Arg, generated by non-canonical repeat-associated non-ATG (RAN) translation of the sense and antisense strands in 6 total reading frames; and TDP-43 inclusions (Gendron and Petrucelli, 2018).

A blitz of investigations have sought to determine how RNA foci, different dipeptides, TDP-43 aggregate accumulation, and loss of function of the native C9orf72 gene each contribute to pathology. Intriguingly, animal model studies exploring the interactions among repeat length, RNA foci, dipeptides, TDP-43 pathology, and motor neuron death have led to potentially conflicting results. On the one hand, the combination of RNA foci and dipeptide repeats may not be sufficient to cause the disease, as C9orf72 ~500-repeat transgenic mice with both RNA foci and several dipeptide repeats had only modest pathological findings (Jiang et al., 2016; O’Rourke et al., 2015; Peters et al., 2015), although more pronounced in one model for reasons that remain unclear (Y. Liu et al., 2016). On the other hand, strong overexpression of a 66-copy repeat in an AAV vector was sufficient to cause both RNA foci and dipeptide repeats, as well as TDP-43 pathology, neuronal loss, and behavioral and motor deficits (Chew et al., 2015).

Other studies have dissected out the distinct effects of the RNA foci and dipeptides and their associated mechanisms. In an elegant investigation in flies, stop codons introduced into the repeats prevented dipeptide generation and yielded reduced toxicity compared to a dipeptide-competent construct (Mizielinska et al., 2014). Additional studies evaluated the interactome of the RNA foci and dipeptides, leading to an intense investigation of nuclear pore dysfunction (Freibaum et al., 2015; Haeusler et al., 2014; Jovičić et al., 2015; K.-H. Lee et al., 2016; K. Zhang et al., 2015).

Finally, loss of endogenous protein function seems to play a role in the disease. Initial support for this hypothesis came from the observation that C9orf72 transcripts are decreased in repeat expansion patient post-mortem samples (Gendron and Petrucelli, 2018). While knockout of C9orf72 in mouse models does not cause motor neuron disease, it does yield myeloid dysregulation and neuroinflammatory activation (Burberry et al., 2016; Jiang et al., 2016; Koppers et al., 2015; O’Rourke et al., 2016). Interestingly, Shao and colleagues investigated the consequences of reducing C9orf72 expression in the background of a C9-BAC mouse (gain-of-function) (Shao et al., 2019). Results showed that haploinsufficiency of C9orf72 exacerbated motor behavior deficits. Ultimately, mouse C9orf72 animal models have confirmed the causative role of the repeat expansion in generating RNA foci and dipeptide repeats. However, what then determines the progression to cell death, potentially via TDP-43 accumulation, remains unclear. One interesting possibility is that loss of C9orf72 endogenous gene function contributes to the sensitivity to RNA foci and dipeptide toxicity (Shao et al., 2019).

iPSC-based studies have proved important for understanding and comparing different potential mechanisms of C9orf72 pathology. Most studies confirmed the expected neuropathological disease features, including RNA foci and several dipeptide repeats (Almeida et al., 2013; Dafinca et al., 2016; Donnelly et al., 2013; Lopez-Gonzalez et al., 2016). Although Sareen and colleagues found increased splicing to favor the repeat-containing isoform and RNA foci co-localized with the RNA binding proteins hnRNPA1 (heterogeneous nuclear ribonucleoprotein A1) and Pur-∝, thus supporting the mechanism of RNA binding protein sequestration, they did not observe dipeptides (Sareen et al., 2013).

Several of the same cellular mechanisms at play in other FALS variants extend to C9orf72. These include susceptibility to autophagy inhibition (Almeida et al., 2013) and ER stress (Dafinca et al., 2016), with both increased presence of apoptotic markers and cell death in C9orf72 repeat expansion than control iPSC-derived motor neurons. Lopez-Gonzalez and collaborators connected C9orf72 repeat expansion and the GR dipeptide repeat with three key downstream mechanisms: DNA damage, oxidative stress, and mitochondrial dysfunction (Lopez-Gonzalez et al., 2016). Notably, C9orf72 iPSC-derived motor neurons showed increased basal levels of DNA damage, as detected by the DNA double-strand break marker γH2AX and by comet assay. After four months in culture, the authors detected increased levels of the p53 tumor suppressor gene, consistent with expected signaling after DNA damage. Expression of a (GR)₈₀ construct amplified these features and also revealed increased oxidative stress with impaired mitochondrial function.

Three major studies have confirmed a sensitivity of C9orf72 motor neurons to glutamate stress, but with different potential molecular mechanisms. Donnelly and colleagues confirmed that antisense oligonucleotides targeting the C9orf72 repeat sequence mitigated the increased glutamate sensitivity (Donnelly et al., 2013). They then showed that RNA foci bound the ADARB2 protein, which normally functions to Gln/Arg edit GluR2 AMPA receptors and render them calcium impermeable. Sequestration of this protein would therefore increase the calcium permeability of AMPA receptors. To support this mechanism, knockdown of ADARB2 independently increased the glutamate sensitivity of control-derived motor neurons.

In a second study evaluating isogenic pairs of C9orf72 repeat expansion iPSCs and edited iPSC lines with short repeats, RNA-Seq of motor neurons revealed a more than two-fold increase in the AMPA-subtype of glutamate receptor GluR1 subunit (Selvaraj et al., 2018). As the levels and editing of GluR2 subunits did not differ between C9orf72 and isogenic controls, the authors hypothesized that the increase in GluR1 subunit levels would be sufficient to increase calcium entry. Indeed, there was a progressive decrease over time in AMPA calcium permeability in control but not C9orf72 repeat motor neurons, consistent with mutant motor neurons becoming more sensitive to AMPA-induced excitotoxicity. Finally, they validated their findings by showing increased GluR1 in the anterior horn but not the prefrontal cortex of C9orf72 expansion patient post-mortem samples compared to controls.

A third parallel link between C9orf72 repeat expansion and glutamate sensitivity came from examining a loss of function mechanism (Shi et al., 2018). C9orf72 expansion-derived motor neurons showed reduced survival compared to control-derived motor neurons in a glutamate stress assay; surprisingly, knockdown of the native C9orf72 gene impaired survival to a similar degree in control-derived motor neurons. Further investigation using isogenic constructs with either one or both C9orf72 alleles deleted implicated a role for the native C9orf72 gene in lysosomal dysfunction, and reduced C9orf72 expression yielded increases in glutamate receptors at the cell membrane. In an interesting twist, loss of C9orf72 gene function and subsequent lysosomal impairment enhanced dipeptide toxicity. Moving to a small molecule phenotypic screen based on motor neuron survival following glutamate stress, the authors identified inhibitors of the PIKFYVE enzyme, which converts phosphatidylinositol 3-phosphate to phosphatidylinositol 3,5-bisphosphate, the former being critical for endosome maturation and lysosome formation. PIKFYVE inhibitors, such as apilimod, improved lysosome function and motor neuron survival in their assay.

Several important efforts have used C9orf72 iPSC-derived motor neurons to validate specific mechanistic findings. Zhang and colleagues used iPSCs to confirm previously discovered suppressors of repeat toxicity in flies (K. Zhang et al., 2015), and highlight the connection between repeat expansion and nucleocytoplasmic pore and transport dysfunction. Using a fluorescent reporter downstream of nuclear export and import signals, the authors showed that the C9orf72 repeat was sufficient to induce nucleocytoplasmic transport defects. C9orf72 iPSC-derived motor neurons revealed impaired localization of Ran GTPase, which was corrected by antisense oligonucleotide-mediated knockdown of the C9orf72 repeat.

Finally, KPT-276, an exportin 1 inhibitor, rescued the transport deficits and repeat-associated neurodegeneration in the fly eye. Strikingly, independent unbiased screens in yeast (Jovičić et al., 2015) and flies (Freibaum et al., 2015) implicated the same nucleocytoplasmic transport process, although there were differences in the involved genes.

Based on the ability of Spt4/SUPT4H1 knockdown to reduce the transcription of the trinucleotide repeat in Huntington’s disease, Kramer and colleagues evaluated its role in C9orf72 ALS (Kramer et al., 2016). Remarkably, reducing Spt4 led to a decrease in dipeptides and RNA foci, while improving survival phenotypes across yeast, flies, and human C9orf72 iPSC-derived motor neuron models. In human post-mortem tissue, the level of SUPT4H1 correlated with the C9orf72 variant 3 RNA (which contains the hexanucleotide repeat) as well as the GP dipeptide.

A screen in mammalian neurons for modifiers of dipeptide repeat toxicity highlighted ER stress and identified the TMX2 gene as a potential target (Kramer et al., 2018). Decreasing TMX2 mitigated PR-toxicity and increased longitudinal survival of C9orf72 iPSC-derived motor neurons. Separately, immunoprecipitation of influenza hemagglutinin-tagged C9orf72 protein in mouse neuroblastoma cells identified cofilin and additional actin binding proteins, which had been shown to be dysregulated in both iPSC-derived motor neurons and post-mortem samples (Sivadasan et al., 2016).

Taking the C9orf72 iPSC modeling studies together, most iPSC studies – like the transgenic mouse models – captured both RNA foci and some dipeptides. However, the iPSC models, also like the transgenic mouse models, have struggled to capture both TDP-43 pathology and survival differences. The link between more aggressive mouse phenotypes and TDP-43 pathology in the viral overexpression approach (Chew et al., 2015) compared to transgenic mouse models suggests that the processes by which TDP-43 pathology occurs may be closely related to toxicity. To what extent TDP-43 pathology requires additional environmental or genetic insult beyond the fundamental C9orf72-related mechanisms of RNA foci, dipeptides, and reduced RNA expression, or whether there is just insufficient burden of RNA foci or specific dipeptides to effect TDP-43 pathology in the iPSC motor neurons, remains unclear. The pooled C9orf72 studies establish strong cases for nuclear pore disruption and DNA damage, mechanistic themes that have echoed through other FALS cases including FUS and TDP-43 mutations (H. J. Kim and Taylor, 2017). Stress granule assembly has also been linked to nucleocytoplasmic transport dysfunction (K. Zhang et al., 2018), suggesting that nucleocytoplasmic transport disruption may occur from either the nuclear or cytoplasmic side. Furthermore, C9orf72 dipeptides are connected with liquid phase separation and consequent demixing (K.-H. Lee et al., 2016). Differences in the ability to reach critical thresholds for positive feedback and downstream pathological cascade may help explain why RNA foci and at least some dipeptides can exist without cell death in mouse models, iPSC motor neurons, and indeed in human brain and spinal cord; similarly, such a dynamic equilibrium hypothesis may help explain the pronounced sensitivities to of C9orf72 motor neurons to cellular stressors such as glutamate. Finally, the role of C9orf72 endogenous gene loss of function remains to be clarified: whether it acts via non-cell autonomous effects on immune function or via a predominantly cell-autonomous role to sensitize motor neurons, as suggested by the iPSC modeling study (Shi et al., 2018).

Motor Neuron Excitability

A separate group of studies sought to use iPSC-based modeling to identify mechanisms underlying motor neuron hyperexcitability. This feature has been strongly supported by human neurophysiological studies (Bae et al., 2013), but has been controversial in SOD1 mice (Delestrée et al., 2014; J. Kim et al., 2017; Kuo et al., 2004).

At early time points in culture, typically two to four weeks after differentiation, SOD1^A4V iPSC-derived motor neurons fired more than controls, as measured by both whole cell patch clamp and microelectrode array recording (Wainger et al., 2014). Similar results were also seen for TDP43 and C9orf72 repeat expansion motor neurons (Devlin et al., 2015). For SOD1 ALS neurons, patch clamp showed reduced delayed-rectifier potassium current amplitudes compared to healthy control and gene-edited isogenic control motor neurons, thus offering a mechanistic explanation for the observed hyperexcitability (Wainger et al., 2014); however, C9orf72 and TDP-43 motor neurons showed a more broad and progressive decrement in both potassium and sodium current amplitudes, with less differences evident at early time points (Devlin et al., 2015). Nonetheless, the deficit of delayed-rectifier potassium channels in the SOD1 neurons led to investigation of the Kv7 agonist retigabine (ezogabine), and treatment with the drug reduced motor neuron firing of SOD1, TDP43, and C9orf72 repeat expansion iPSC-derived motor neurons and improved in vitro survival of SOD1^A4V motor neurons (Wainger et al., 2014). As an aside, a similar survival benefit of retigabine was seen in C9orf72 repeat expansion-derived motor neurons (Shi et al., 2018), providing broader support for the mechanism.

Generally in contrast to these reports, studies performed at later timepoints in culture obtained seemingly opposite results: decreased excitability of motor neurons from C9orf72 repeat expansion (Sareen et al., 2013) and several mutant SOD1 and FUS (Guo et al., 2017; Naujock et al., 2016; Naumann et al., 2018) iPSC lines. Indeed, the study by Naujock and colleagues showed that the K-channel blocker 4-aminopyridine (4-AP) reduced the percentage of SMI32-positive neurons that co-labeled with cleaved caspase-3 using SOD1 (D90A and R115G) and FUS (R521C, R521L, and R495Q) mutant motor neurons.

To integrate these conflicting results, Devlin and colleagues showed a progression from early hyperexcitability after two to six weeks in culture to reduced firing capacity later in culture after 7–10 weeks in mutant C9orf72 and TDP-43 neurons compared to control motor neurons (Devlin et al., 2015). Although lactate dehydrogenase staining did not show a difference between TDP-43 and control neurons, suggesting that survival was not affected, the more depolarized membrane potential could have reflected worsening cellular health after longer culture periods, consistent with modeling studies of an ATP deficiency contributing to impaired energetic homeostasis (Le Masson et al., 2014).

No doubt, fully understanding the differences among these studies will require employing a more granular time course of evaluation, investigating the different mechanisms of cell death, as well as controlling for differentiation protocol and density-influenced network effects of neurons activated by depolarization. In comparing the contrasting effects of activating Kv7 potassium channels with retigabine versus blocking voltage-gated potassium channels with 4-AP, differences in the specific gene mutations studied could also add to the disagreement. Despite the controversies, the observation of early hyperexcitability and a positive-feedback vicious cycle between hyperexcitability and ER stress (Kiskinis et al., 2014), in the setting of the clinical picture of cortical and spinal motor neuron hyperexcitability, led to a clinical trial in ALS patients (ct.gov ) for which successful results in reducing motor neuron excitability in patients were presented at the Motor Neuron Disease Association meeting in 2018 (Wainger et al., 2018). The fact that insights from iPSC-derived motor neurons identified a mechanism and drug that exerted profound effects toward normalizing ALS neurophysiological phenotypes, stronger perhaps than riluzole (Kovalchuk et al., 2018), represents an initial demonstration of the value of iPSC-based modeling for rapid translation to clinical studies.

The hyperexcitability studies illustrate several issues and controversies in iPSC modeling. iPSC studies showed a broad spectrum of altered excitability, ranging from hyperexcitable to nearly inexcitable motor neurons. A time-dependent progression from increased to decreased firing explained some but not all of the discrepancy. Importantly, much of the contradictory-appearing findings may reflect a continuum from hyperexcitability to impaired firing as depolarization block occurs, instead of hyperexcitability and apparent hypoexcitability – resulting directly from prolonged hyperexcitability – being opposites (Le Masson et al., 2014; Wainger et al., 2014)).

In order to maximize the utility of iPSC modeling, one needs to identify and understand which in vitro conditions best reflect the disease. From there, iPSC studies can be of additional value in helping to guide future clinical studies, for example, by supporting mechanistic targets or treatment candidates. Care must be taken to critically evaluate connections made between in vivo and in vitro phenotypes. For example, the main measure of upper motor neuron hyperexcitability in ALS, short-interval intracortical facilitation, is thought to reflect impairment of interneuron inputs onto cortical motor neurons (Bae et al., 2013; Hallett, 2007). Thus, assuming that increased firing of isolated cortical motor neurons in vitro reflects circuit-dependent mechanisms of hyperexcitability in vivo could lead to incorrect findings. However, knowing the importance of interneuron modulation could help focus attention on mouse and potentially in vitro organoid models. Similarly, dynamic changes in motor neuron excitability in iPSC motor neurons could underscore the need to clarify in greater detail how neurophysiological metrics of lower and upper motor neuron excitability change during disease course in patients.

Finally, the excitability studies highlight the issue of generalizability. Although the results of different investigations differ in terms of the propensity for increased or decreased firing after distinct culture durations, the individual reports generally find similar effects across FALS genes. One may posit that different upstream FALS mechanistic pathways converge to produce the disease. However, whether the mechanisms of abnormal excitability are convergent across ALS variants remains unclear. Furthermore, in targeting increased excitability, as for other target mechanisms, one needs to be confident that correcting the phenotypes will be sufficiently upstream and potent to impede disease progression.

Modeling Sporadic ALS

While most studies have focused on approaching ALS through highly penetrant monogenic FALS disease models, only a few have examined SALS, despite the fact that SALS accounts for 90% of all ALS cases. The largest paper to date evaluated the pathological features in 32 sporadic lines, using an established multiplex of phenotypes identified in TDP-43, FUS, and SOD1 FALS iPSC motor neurons (Fujimori et al., 2018). The FALS iPSC-derived neurons developed a reliable set of features including reduced neurite outgrowth, increased lactate dehydrogenase release after approximately 40 days in vitro, and granule formation. In contrast, the SALS-derived lines were much more variable, with some showing little outgrowth loss even after 70 days in vitro. The time course of pathology development in vitro did not appear to reflect the age of disease onset in the patients; however, the data suggested a correlation between the time of in vitro phenotype and clinical progression course. The study clearly made inroads into the heterogeneity of SALS, particularly as their multiplex analysis clustered individual SALS lines with specific FALS variant features. Motor neurons from 16 of 22 SALS lines responded to ropinirole, which was the lead hit compound identified in the screen using TDP-43 and FUS FALS lines. However, motor neurons from more than a quarter of the SALS lines did not respond, underscoring the challenge that disease heterogeneity presents for both drug discovery and clinical trials in SALS.

While the Fujimori study was transformative in several ways, particularly in terms of scale and robustness of modeling phenotypes, the results of the screen raised some specific issues (Wainger and Lagier-Tourenne, 2018). First, drugs that have failed in large ALS clinical trials, such as dexpramipexole and ceftriaxone, performed better in the screen than the approved drugs riluzole and edaravone. Second, although the authors demonstrated why ropinirole may be more potent than dexpramipexole (on account of both increased dopamine receptor affinity as well as a greater capacity to target lipid peroxidation), the fact that the drug has not been successful in Parkinson’s disease – where presumably these considerations would apply as well – is concerning.

Two earlier smaller studies also examined sporadic lines. In the first, nuclear TDP-43 and phospho-TDP-43 granules were observed in three of 16 SALS iPSC-derived motor neurons but not controls (Burkhardt et al., 2013). The second used microarray and pathway analysis in motor neurons derived from two SALS patients compared to two controls and implicated mitochondrial and other pathways (Alves et al., 2015).

While SALS holds the greatest potential for iPSC modeling, researchers have appropriately first focused on FALS, specifically on techniques to capture robust molecular phenotypes and neurodegeneration within FALS motor neurons. As suggested by Fujimori and colleagues, distinct genetic FALS variants may share significant mechanisms and targets with subgroups of SALS (Fujimori et al., 2018). Interestingly, bosutinib, a Src/c-Abl inhibitor identified in the aforementioned SOD1 FALS iPSC-derived motor neuron survival screen, showed benefit in several FALS and three of three SALS lines (Imamura et al., 2017). While three lines is not sufficient to address SALS heterogeneity, these results indicate that the Src/c-Abl pathways may be more broadly-relevant targets, particularly if the generalizability from SOD1 to more SALS lines is confirmed; in contrast, ropinirole, the lead hit in the Fujimori paper, did not prove useful in SOD1 cases. These findings are, in fact, surprising: based on the unique pathological features in SOD1 ALS and the ubiquitous presence of TDP-43 pathology in SALS, one may have expected greater generalization to SALS from a screen-based on TDP-43 FALS than from one based on SOD1 FALS lines. In the following section, we discuss achievements of iPSC disease modeling approaches in ALS as well as the challenges, with particular considerations for maximizing success in targeting SALS.

Achievements and Challenges in Using iPSCs to Model ALS

We have highlighted a number of successes, particularly with regard to modeling FALS, in which iPSC-based motor neuron models have captured quintessential disease phenotypes. In many cases, investigations using iPSC-derived motor led to conclusions that were subsequently validated in human post-mortem tissue from both FALS and SALS patients. While some phenotypes – such as SOD1 inclusions, the presence and toxicity of TDP-43 inclusions, and the different manifestations of abnormal motor neuron excitability – remain controversial, iPSC modeling has nonetheless contributed substantially to the understanding of FALS mechanisms. Encouragingly, several implicated pathways span multiple FALS mutations and thus are particularly promising, such as DNA damage repair and nucleocytoplasmic transport, as well as the conserved involvement of stress granule processing and autophagy.

Nonetheless, major challenges remain in the iPSC modeling approach, any of which could derail the translation of targets and blockers to human clinical trials. We will briefly discuss the need to address non-cell autonomous effects in ALS, improve the ability to capture late-onset phenotypes, acquire better resolution of variance components inherent to the iPSC modeling process, and apply solutions to the formidable task of targeting SALS.

Substantial non-cell autonomous effects contribute to ALS (Ilieva et al., 2009), including from microglia (Boillée et al., 2006), astrocytes (K. Yamanaka et al., 2008), oligodendrocytes (Kang et al., 2013), and potentially muscle (Loeffler et al., 2016). Indeed, the contributing cell types may themselves vary among the different ALS subgroups (Serio et al., 2013). For this review, we have focused on cell-autonomous effects of motor neurons, and thus accepted that additional influences from other cell types may affect the generalizability of conclusions based on motor neuron-autonomous findings. Thus, targets and compounds that prove successful in cell autonomous environment could be evaluated orthogonally using models that offer at least some non-cell autonomous component. For example, co-cultures of motor neurons and astrocytes can help address simple interactions (Di Giorgio et al., 2008). However, because such non-cell autonomous contributions may reflect complex interactions among multiple cell types (Liddelow et al., 2017), more complicated models may be necessary to capture such effects. One possibility is that organoid models, which suffer from greater variability but provide exposure to a larger range of additional cell types, can help model non-cell autonomous effects in future studies (Quadrato et al., 2016).

Given that ALS only becomes clinically apparent in mid or late life, the ability to obtain such phenotypes using motor neurons matured for weeks to months in vitro is in itself quite remarkable, particularly as age estimates for motor neurons suggest an embryonic range (Arbab et al., 2014; Ho et al., 2016). It would seem reasonable – if not necessary – that hastened cellular aging would be an important way to add value to in vitro phenotypes. Refined molecular manipulation, such as progerin overexpression, may be used to enhance aging and augment disease phenotypes (Miller et al., 2013). However, as with other stressors described in the studies we have discussed, a careful balance must be weighed between the use of stressors – genetic and otherwise – to enhance phenotypes and confidence that such stressors do not betray physiological relevance.

Another major area of concern results from a lack of understanding of multiple different sources of variation. Protocols for making motor neurons differ substantially among labs and may explain potential disagreement among studies (see Figure 2 in Sances et al., 2016). Methods vary in the length of neural induction in 2-D or 3-D format, motor neuron specification, and maturation steps. Different protocols produce a wide range of purities of motor neurons. While some aim to reduce contaminants using FACS purification based on Hb9 or another marker, even these methods can yield substantial remaining populations of other neuronal or expanding non-neuronal cell types. Unfortunately, the non-motor neuron populations in the culture and their effects on motor neurons are often poorly defined.

Even more concerning, sources of variation within protocols are also not well understood. For example, motor neuron identity and consequentially the assessment of batch variation are typically based only on a small number of immunocytochemical markers. Accurate distribution estimates of variance due to reprogrammed tissue (blood versus skin samples), its collection and handling, reprogramming method, clone, iPSC passage, and differentiation batch have all been largely inadequate. What is clear is that an accumulation of genetic mutations occur throughout reprogramming and iPSC processing (Gore et al., 2011). Perhaps not surprisingly in working with passaged cell lines, mutations in the p53 tumor suppressor and other genes tend to increase with passage number (Merkle et al., 2017). For C9orf72 expansion studies in particular, Southern blots revealed instability of the repeat during the differentiation process, adding an additional potential confounder and arguing for monitoring of repeat length (Almeida et al., 2013). The impacts of these effects are poorly defined and likely depend on the specific assays and experimental questions being addressed.

Efforts to unravel these complications are limited by the high cost and labor necessary for reprogramming primary blood or fibroblast samples, culturing iPSCs, and differentiating and maintaining motor neurons. Dedicating substantial resources, through ambitious programs such as Answer ALS (https://www.answerals.org/), which will generate iPSCs from 1,000 ALS patients as well as controls, will be necessary to provide rigorous assessment of the capacity for SALS modeling. Infrastructural changes allowing experiments to be performed at sufficiently large scale will almost surely be required.

All of these challenges converge in modeling SALS, which embodies the greatest potential of the iPSC approach, due to both the large percentage of SALS patients and the limited capability of other models to address it. Because epigenetic changes are likely erased during the reprogramming process (Takahashi and S. Yamanaka, 2016), iPSC models may not sufficiently capture environmental effects and their contributions to disease. Whether genetic predisposition to such environmental or epigenetic modifications will be sufficient to recapitulate quintessential disease features remains to be seen. Alternative direct lineage reprogramming approaches, which avoid a pluripotent cell stage and thus more faithfully preserve epigenetic information, may be necessary (Abernathy et al., 2017; Mertens et al., 2015).

An additional problem is simply the unknown heterogeneity of SALS, independent of iPSC modeling. How different are the driving mechanisms, non-cell autonomous cellular effects, and environmental influences? The capability of iPSCs and derived motor neurons to capture a critical portion of the SALS disease phenotype remains unknown. All of the issues we have just discussed – non-cell autonomous effects, multiple unknown variation components, and resulting uncertainty in choosing sample sizes – are augmented for SALS.

Despite these concerns, several promising candidate therapeutics have emerged directly from iPSC-modeling in vitro studies (Table 1; see also Table 1 from S. Lee and Huang, 2017 for tabulated listing of ALS modeling studies). Preliminary results from the study of retigabine (ezogabine) on hyperexcitability as assessed by transcranial magnetic stimulation and threshold tracking nerve conduction studies have been encouraging (Wainger et al., 2018). Although the drug was taken off the market as an anti-epileptic, interest remains regarding the mechanism of Kv7 activation, and several companies have prior and ongoing programs in this area.

Table 1.

Screening assay and candidate evaluation studies based on iPSC-derived motor neurons

Reference	Type of study	Disease subgroups # includes isogenics	iPSC to motor neuron differentiation protocol	Analysis date after end of differentiation	Measurement (stressor?)	Compound	Mechanism: pathway or effect
Burkhardt et al. 2013	Screen of 1757 compounds	SALS*, FUS, TDP-43, and SOD1 (34 lines)  FUS (G1566A)  TDP-43 (A315T)  SOD1 (A4V, N139K)	18 day small molecule mediated	Days 7*, 25, 32, 97, 104	TDP-43 aggregates (none)	Digoxin, lanatoside c, and proscillardin a (cardiac glycosides)	Na+/K+ ATPase pump inhibitors: may alter Ca++ influx and affect Ras, IP3, and NF-kB signaling pathways
Fujimori et al. 2018	Screen of 1232 compounds	FUS*, TDP-43*, SOD1, and SALS (58 lines)  FUS (H517D)  TDP-43 (M337V, Q343R)  SOD1 (H46R, H43R)	19 day small molecule mediated	Days 5–50 Day 31*	Neurite regression (none) cytotoxicity (none) aggregates (none) stress granules (none)	Ropinirole	Dopaminergic agonist: improves neuronal health and mitigates lipid peroxidation that could alter the fatty acid-related pathway
Imamura et al. 2017	Screen of 1416 compounds	SOD1*, TDP, C9orf72, and SALS (16 lines)  SOD1 (L144FVX, G93S)  TDP-43 (M337V, Q343R, G298S)	7 day Ngn2, Isl1, and Lhx3 vectors (NIL cassette) doxycycline mediated	Day 7	Neuronal survival (none) p62 (none) autophagy (none) misfolded SOD1 (none)	Bosutinib	Inhibitor: inhibition of the Scr/c-Abl pathway leads to autophagy
Marrone et al. 2018	Screen of ~1000 compounds	FUS# (2 lines)  FUS (P525L)	6 day small molecule mediated	Day 21	Stress granules (arsenite) autophagy (arsenite)	Rapamycin, torkinib, paroxetine, promethazine, trimipramine, and chlorpromazine	Inhibitors: rapamycin and torkinib directly inhibits mTOR pathway while the other compounds are reported to induce autophagy
Shi et al. 2018	Screen of 800 compounds	C9orf72# (10 lines)	15 day Ngn2, Isl1, and Lhx3 vectors (NIL cassette) doxycycline mediated	Days 0–10 longitudinal*	Neuronal survival (glutamate treatment) Neuronal survival (neurotrophic factor withdrawal)	YM201636	PIKFYVE kinase inhibitor: promotes autophagy
Bhinge et al. 2017	Candidate evaluation based on pathways identified by RNA-seq	SOD1# (2 lines)  SOD1 (E100G)	10 day small molecule mediated	Day 20*, 34	Neuronal survival (none)	FR180204, SB203580, SP600125, XAV939, and pifithrin-a hydrobromide	Pathway inhibitors: inhibiting the ERK1/2, p38/MAPK, JNK, WNT, and TP53 pathways respectively
Egawa et al., 2012	Candidate evaluation based on pathways identified by RNA-seq	TDP-43 (16 lines)  TDP-43 (Q343R, M337V, G298S)	35 day small molecule mediated	Day 22	Neurofilament (none) neurite outgrowth (none) neuronal survival (arsenite)	Anacardic acid	Histone acetyltransferase inhibitor: may downregulate TDP-43 mRNA expression
Guo et al. 2017	Candidate evaluation based on Charcot Marie Tooth and SOD1 mouse studies	FUS (9 lines)  FUS (P521H, P525L)	9 day small molecule mediated	Days 15, 22, 29	ER transport (none) mitochondrial transport (none)	ACY-738 and tubastatin a	HDAC6 inhibition: prevent deacetylation of a-tubulin, increasing motor molecule binding to microtubules
Kiskinis et al. 2014	Candidate evaluation based on pathways identified by RNA-seq	SOD1 (4 lines)  SOD1 (A4V)	24 day small molecule mediated	Days 3, 15*, 21, 30	ER stress (none) motor neuron firing (none) neuronal survival (none)	Salubrinal	Inhibitor: inhibits ER stress
Lopez-Gonzalez et al. 2016	Candidate evaluation based on pathways identified by oxidative stress and DNA damage	C9orf72 (7 lines)	31 day small molecule mediated	Days 14, 28, 56, 84, 120	Mitochondrial oxidation (none) comet assays (none)	Trolox	Antioxidant: inhibits oxidation and reduce related damage
Naujock et al. 2016	Candidate evaluation based on mechanism of hypoexcitability	FUS and SOD1 (9 lines)  FUS (R521C, R521L, R495QfsX527)  SOD1 (D90A, R115G)	52 day small molecule mediated	Day 11–18, 22, 32–39*, 81–88	Excitability (none) ER stress (none) caspase activation (none)	4-Aminopyridine	Potassium channel blocker: restores excitability in motor neurons
Naumann et al. 2018	Candidate evaluation based on pathways identified by DNA damage	FUS (9 lines)  FUS (R521C, P525L, R521L, R495QfsX527)	9 day small molecule mediated	Days 0, 5*, 12*, 21*, 26, 47, 51, 101	FUS recruitment (DNA damage) FUS aggregation (none) FUS trafficking (none)	Gallotannin	Poly(ADP-ribose) glycohydrolase (PARG) inhibitor: increases PAR expression and DNA repair
Wainger et al. 2014	Candidate evaluation based on mechanism of hyperexcitability	SOD1*, FUS, and C9orf72 (13 lines)  FUS (fs511)  SOD1 (A4V)	24 day small molecule mediated	Days 4, 8, 12, 16, 20, 24*, 28, 30*	Excitability (none) ER stress (none) neuronal survival (none)	Retigabine	Activation of subthreshold Kv7 currents: reduces excitability

^*for primary focus

Plans have been announced for clinical trials in ALS of both bosutinub and ropinirole based on in vitro studies (Fujimori et al., 2018; Imamura et al., 2017). Additional encouraging mechanisms include PIKFYVE and HDAC6-inhibiting drugs as well (Guo et al., 2017; Shi et al., 2018). Notably, for most of these drugs, the pathway to the clinic has not involved animal models, demonstrating a growing trend toward the acceptance of pathways and targets identified in stem cell models. Ultimately, the true validation of the iPSC approach will lie in the success or failure of these and other clinical studies.

Conclusion and Future Perspectives

iPSC-based modeling in ALS has made rapid and substantial advances, fueled by improvements in reprogramming, motor neuron differentiation, and gene editing techniques. The success in recapitulating disease features in FALS is particularly encouraging and has led to the mobilization of substantial resources aimed at similar achievement for SALS. Although studies of SALS have been limited, it seems encouraging that compounds identified using FALS models may exert benefit in at least subgroups of SALS iPSC-derived motor neurons.

Despite these successes, continued vigilance will be necessary to maximize the chance that targets and compounds capable of mitigating disease phenotypes in vitro will be able to do so in ALS patients. The importance of validating phenotypes, particularly with evidence from human genetics and post-mortem samples, cannot be overemphasized. Robustness of findings across lines, differentiation techniques, and more complicated models for incorporating non-cell autonomous effects will all be critical.

With the availability of screening equipment and compound libraries, an increasingly large number of promising compounds are and will continue to result from phenotypic screens based on features such as aggregate formation, reduced motor neuron outgrowth, and survival. Ranking such targets and hit compounds using robust methods will be a priority in determining how best to choose and advance compounds through drug development and into human ALS subjects.