Innovating spinal muscular atrophy models in the therapeutic era

Spinal muscular atrophy (SMA) is a severe, monogenetic neuromuscular disease characterised by progressive muscle weakness caused by loss or dysfunction of motor neurons (Mercuri et al., 2022). The severity of SMA varies significantly, and it is traditionally classified based on motor milestones achieved and the age of onset (Groen et al., 2018a; Wadman et al., 2017). The most severe type, SMA type 1, is also the most common type of the disease and, when untreated, leads to death around one year after birth, before any motor milestones are achieved. SMA type 4, in contrast, is the mildest and also rarest type of the disease, characterised by adult onset and affected individuals retaining the ability to walk independently (Piepers et al., 2008). Between these extremes, type 2 and type 3 make up a large group of patients that, in addition to limited motor function, can develop any of a range of complications including scoliosis, bulbar symptoms, increased fatiguability and respiratory complications (Wadman et al., 2017). Additionally, it is increasingly recognised that symptoms outside the central nervous system (CNS), including dysfunction of the kidney, liver, heart and bone, may affect some patients to varying degrees (Hamilton and Gillingwater, 2013; Yeo and Darras, 2020). In general, although the life expectancies of patients with SMA types 2, 3 and 4 are close to normal (Wijngaarde et al., 2020), muscle weakness and complications caused by SMA can significantly deteriorate the quality of life (Wohnrade et al., 2023). Without exception, SMA is always progressive when untreated.

SMA is caused by deleterious variants in the survival of motor neuron 1 (SMN1) gene (Lefebvre et al., 1995). The paralogous, but less functional, gene SMN2 is the main modifier of SMA severity. A cytosine-to-thymine variant in exon 7 of SMN2 causes the majority of transcripts to be alternatively spliced, leading to lower levels of the functional, full-length SMN protein (Fig. 1A,B) (Lorson et al., 1999; Monani et al., 1999). However, because the SMN locus on chromosome 5 is complex and unstable, significant copy number variation occurs, with SMN2 copy numbers ranging from one to eight copies (Wirth et al., 2020). In general, a higher SMN2 copy number leads to higher SMN protein levels and therefore a relatively milder phenotype (Wadman et al., 2020). This correlation is especially clear at both ends of the SMA severity spectrum: most patients with SMA type 1 have two SMN2 copies, whereas all patients with SMA type 4 have at least four SMN2 copies. However, the correlation between SMN2 copy number and disease severity for SMA types 2 and 3 is less clear, and the copy number for these subtypes usually ranges between two and four (Calucho et al., 2018; Wadman et al., 2020). In addition to SMN2 copy number variation, other genetic modifiers of SMA have also been described (Hosseinibarkooie et al., 2016; Oprea et al., 2008; Riessland et al., 2017), but up to 40% of variation in disease severity remains genetically unexplained. In a large group of patients with three or four copies of SMN2, prognosis remains difficult to provide based on genotype alone.

Treatment for patients with the most severe types of SMA used to be mostly palliative, and for patients with relatively less severe types, it was mostly supportive (e.g. physical therapy, scoliosis surgery and feeding support). Excitingly, a thorough understanding of the genetic cause of SMA recently led to the approval of three gene-targeting, disease-modifying therapies (Fig. 1C) (Chaytow et al., 2021), representing a milestone in the treatment of neuromuscular diseases. Although treatment response varies from significant clinical improvement to initial stabilisation, the majority of patients with SMA appear to benefit from these therapies.

The success story of SMA therapies was built, in large part, on the availability of multiple, robust disease models ranging from cell-based models to small animals (Table 1) – such as Caenorhabditis elegans (Briese et al., 2008; Doyle et al., 2020; Sleigh et al., 2011a), Drosophila (Chan et al., 2003; Chang et al., 2008; Spring et al., 2019) and zebrafish (Boon et al., 2009; McWhorter et al., 2003; Tay et al., 2021) – and many mouse models (Table 2) (Bowerman et al., 2009, 2012a; Hsieh-Li et al., 2000; Le et al., 2005; Monani et al., 2000). These models allowed key aspects of SMA pathology to be characterised and gene-targeting therapies to be investigated, tested and optimised in vitro and in vivo. Preclinical studies performed in these models generated invaluable data on the efficacy, distribution and other key requirements of gene-targeting therapies (Foust et al., 2009; Hua et al., 2011; Zhao et al., 2021). Importantly, these findings proved to be highly translatable to patients, leading to the relatively efficient and rapid clinical development of gene-targeting therapies for SMA.

Now that three gene-targeting therapies are used extensively in clinics around the world, follow-up data of treated SMA patients are becoming available for large, real-world patient cohorts (Antonaci et al., 2023; Scheijmans et al., 2022). These studies highlight some important limitations of current therapies with regards to variability in therapy response, and illustrate that the cause of variability in disease severity and treatment outcomes remains incompletely understood. Existing SMA animal models present a disease that is genetically and phenotypically much more homogeneous than in patients, and thus their use to study the underlying causes of this variability may be limited. In addition, current gene-targeting therapies lead to a shift towards milder phenotypes in patients, highlighting the need to develop models that also recapitulate less severe disease. Now is therefore an important time to reflect on the use of current disease models in SMA research and discuss further possibilities to develop and innovate existing and new models, focusing on key features of SMA in the therapeutic era. This will enable the optimisation of gene-targeting therapies to maximise their benefits and the development of novel therapies that combat the remaining symptoms.

The approval of SMN-targeting therapies has significantly changed disease course and outcomes of SMA. Although drug approval and the start of widespread use of therapies in real-world settings is relatively recent, the majority of patients who started treatment after symptom onset experienced a degree of benefit from receiving these therapies (Antonaci et al., 2023). This ranges from significant clinical improvement in young children who started treatment early to stabilisation of the disease in the majority of older children and adults. The biggest treatment effects have been observed in children for whom treatment was started as early as possible and at a time when motor function was still well-maintained.

Each treatment uses a different approach to increase SMN expression (Fig. 1C): (1) nusinersen (Spinraza) is an antisense oligonucleotide (ASO; see Glossary, Box 1) SMN2-splicing modifier (Finkel et al., 2017; Hua et al., 2008; Mercuri et al., 2018); (2) onasemnogene abeparvovec-xioi (Zolgensma) is an adenoviral vector-based SMN1 gene-replacement therapy (Mendell et al., 2017; Valori et al., 2010); and (3) risdiplam (Evrysdi) is a small-molecule SMN2-splicing modifier (Baranello et al., 2021; Naryshkin et al., 2014). As these drugs require different delivery routes – intrathecal injection, one-off intravenous infusion and oral delivery, respectively – the choice of treatment for patients is often pragmatic and based on, for example, age, weight or the possibility to access the spinal column for intrathecal injection. As a consequence, the different characteristics and delivery routes lead to CNS-specific, partially systemic or systemic increases of SMN. This currently complicates interpretation of the clinical progression of SMA after the start of treatment in symptomatic patients. Treatment efficacy may appear more limited for some drugs than for others as more severely affected patients are more likely to be treated with a certain drug. For example, adult patients with severe scoliosis may not be able to receive nusinersen owing to limited access to the spinal column for intrathecal injection, whereas less severely affected patients who receive nusinersen may be more likely to respond to the treatment (Erdos and Wild, 2022). Next, as some of the therapies lead to CNS-specific or partially systemic increases of SMN levels in specific groups of patients (Ramos et al., 2019; Thomsen et al., 2021) systemic phenotypes of SMA that were previously considered subclinical may become more pronounced in groups of patients that receive those treatments. Additionally, the wide range in severity and disease duration before patients start treatment complicates direct comparisons between treatments and patient groups. These points illustrate the challenges in discussing prognosis and expected treatment benefit with patients, as many features of SMA treatment remain uncertain. Long-term follow-up with patients receiving treatment is therefore of the utmost priority.

Excitingly, the availability of multiple treatments and the recognition that early start of treatment leads to significantly better outcomes is now leading to the inclusion of SMA in newborn screening programmes around the world (Dangouloff et al., 2021). Early genetic diagnosis of SMA (generally within two weeks after birth) allows the start of treatment before symptom onset in most patients. Recent phase III studies on onasemnogene abeparvovec in presymptomatic children with two or three copies of SMN2 showed that, although varying degrees of delayed motor development remained, many children developed in line with or close to the normal developmental window for newborns (Strauss et al., 2022a,b). For this specific group of patients, it is still unclear if and to what extent they will develop features of SMA in later life, and what will be the best approach to monitor putative disease progression in the future. In addition, and especially for children with two SMN2 copies, it is unclear to what extent they will develop new, unexpected symptoms in the long term that may be different from those that we have previously associated with SMA. For example, in Pompe disease – a rare metabolic disease leading to progressive muscle weakness – enzyme replacement therapy has been very effective in improving development and survival, but can unexpectedly lead to progressive white-matter abnormalities later in life (Ebbink et al., 2018). As for symptomatic SMA patients, careful long-term monitoring in real-world populations will be vital to better inform patients and clinicians for expected prognosis of disease severity and long-term treatment effects.

In summary, despite its monogenetic origin, SMA is a complex disease. SMA prognosis was never straightforward because of an incomplete correlation between genotype and disease severity. The approval of different types of therapies administered at various stages of the disease further increases the uncertainty around disease severity and therapeutic outcomes for clinicians, and patients with SMA and their families.

To study SMA pathogenesis and support drug development, a wide range of SMA model systems are available. A detailed discussion of all models would go beyond the scope of this Review, but an overview of the most commonly used models has been included in Tables 1 and 2.

Humans and several closely related Homininae species are the only organisms with SMN duplication, and the SMN2 gene is unique to humans (Rochette et al., 2001). All other organisms carry one SMN gene only and therefore cannot develop SMA, as a complete loss of SMN is embryonically lethal (Schrank et al., 1997). The ability to model SMA therefore depends on the use of a variety of genetic approaches, including the use of SMN2 transgenes, partial-loss-of-function mutations or knockdown of SMN1, and hybrid genes (Box 1). However, the SMN protein is evolutionarily and functionally well conserved (Osman et al., 2019), which means that overlapping mechanisms and phenotypes are present in many SMA models. The most common and widely investigated phenotypes in SMA are related to defects in the neuromuscular system, such as loss of motor neurons (Box 1), axonal defects (Box 1), neuromuscular junction (NMJ; Box 1) dysfunction or muscle-specific pathology. The specifics of these phenotypes, their severity and how they have been studied depend on the possibilities and limitations of the model that is used.

Many species and approaches have been used to study SMA and range from small invertebrate models to large animal models, such as pigs. The possibilities for genetic manipulation to induce SMA phenotypes differ between models and species. For instance, Smn knockout in Drosophila, C. elegans and zebrafish is not embryonically lethal owing to maternal Smn contribution (Box 1) (Boon et al., 2009; Briese et al., 2008; Chan et al., 2003), providing the opportunity to develop models in which no functional Smn is present at later developmental stages. Moreover, knockdown of Smn using genetic tools, such as morpholinos, can be used to transiently reduce Smn levels in zebrafish (Boon et al., 2009; McWhorter et al., 2003). Finally, in animal models in which modification of the genome is challenging, such as pigs, transient knockdown of Smn1 has been induced by adeno-associated virus (AAV; Box 1) 9-mediated delivery of specific short harpin RNAs (shRNAs) (Duque et al., 2015). In invertebrate models of SMA, synaptic defects, studied in C. elegans (Briese et al., 2008; Doyle et al., 2020), and altered NMJ development, studied in Drosophila (Chan et al., 2003), lead to defective locomotion and reduced lifespan as the most prominent disease phenotype. Moreover, Drosophila are relatively suitable for studying intermediate SMA phenotypes, as Smn loss-of-function mutations can be used to reliably generate animals with varying degrees of phenotypic severity and survival (Spring et al., 2019). Similarly, in zebrafish, knockdown of Smn1 leads to varying degrees of defects in axon and NMJ development (Boon et al., 2009; McWhorter et al., 2003) and locomotion behaviour (Tay et al., 2021). As zebrafish are optically transparent during development, the use of this model contributed significantly to our understanding of the effect of low levels of Smn1 on nervous system development (Boon et al., 2009; McWhorter et al., 2003). In addition to these phenotypes, pigs provided an opportunity to study neurophysiological characteristics upon SMN reduction, showing similar changes to those observed in patients (Duque et al., 2015).

Mouse models have been especially critical in the development of current therapies and to study the roles of the SMN protein (Table 2) (Lauria et al., 2020; Sleigh et al., 2011b). Genetically, most mouse models for SMA are based on genetic knockout of the Smn1 gene and the introduction of a second SMN2-like allele. Commonly used models include those based on SMN2 transgenes (Burghes and Taiwanese models; Monani et al., 2000; Hsieh-Li et al., 2000), SMN2 with the addition of an SMNΔ7 allele (Δ7 model; Le et al., 2005), and models based on the mouse Smn2B allele that drives alternative splicing of Smn1 exon 7 (Bowerman et al., 2012a). Phenotypically, most mouse models of SMA are characterised by lower body weight, reduced survival and changes in locomotor behaviour, such as changes in the righting reflex (Box 1) in neonatal mice (Butchbach et al., 2007). Pathologically, they are characterised by varying degrees of motor neuron loss, NMJ pathology and muscle defects (Le et al., 2005; Monani et al., 2000; Murray et al., 2008). In addition to varying motor pathologies, systemic pathologies affecting almost all organs have also been described in mice (Allardyce et al., 2020; Deguise et al., 2019; Šoltić et al., 2019; Thomson et al., 2017). The severity of such systemic, non-CNS pathology, however, varies extensively between mouse models and depends, at least partially, on the genetic background of the mouse model (Buettner et al., 2021). The extensive genetic toolbox available to modify the mouse genome, including models based on the bacterial recombinase Cre and models with inducible alleles, has been particularly useful in improving our understanding of the developmental and tissue requirements for SMN expression. For example, studies using these tools have shown that SMN expression is required throughout the neuromuscular system (Fletcher et al., 2017; Hunter et al., 2016; Kim et al., 2020; Park et al., 2010) and that sufficient levels of SMN expression are critical during a narrow time window in postnatal development (Kariya et al., 2014; Lutz et al., 2011).

The central contribution of SMA models to the development of SMA therapies was instigated by the fact that they reflected key features of severe SMA: a short lifespan and severe motor neuron pathology that could be (partially) rescued by restoring SMN expression. However, the current shift in SMA phenotypes caused by the treatment of many patients at various symptomatic and presymptomatic stages calls for disease models that accurately reflect this development.

To allow us to assess the characteristics that novel SMA disease models should ideally have, we will first discuss some specific limitations of current models. Although some of these points are general limitations of using other species to model human disease, we highlight them because of relevance in the light of recent developments in clinical SMA research.

The functional domains of SMN are highly conserved across species (the amino acid homology ranges from 25% identity in C. elegans to 53% in zebrafish) and Smn depletion leads to locomotor defects and neuromuscular phenotypes in most species. However, when studying the conservation of SMN function across species, the expression of zebrafish smn1 but not Drosophila Smn or C. elegans smn-1 was able to improve the phenotype of SMNΔ7 mice (Osman et al., 2019). This indicates that despite similarities, evolutionarily and functionally, not all SMN genes are equal. When using small animal models, it is also important to consider the widely different anatomy and structure of their neuromuscular systems compared to those of humans. In particular, the metamorphosis process that Drosophila undergoes is not in any way comparable to the development of neuromuscular systems in vertebrates. Finally, using common genetic tools, it is challenging to generate these models, as discussed in the previous section, to draw comparisons to the inverted duplication that makes up the SMN locus in the human genome, with limitations comparable to those in mouse models, as discussed in the next section.

Because only humans possess SMN2, several innovative genetic approaches have been used to generate SMA animal models that survive past embryogenesis. However, the transcription and translation of SMN in models that rely on human transgenes (e.g. SMN2 and SMN2-SMN2Δ7) may not be physiologically regulated, owing to differences between mouse and human genetic elements that regulate expression, and to variation in insertion sites of transgenes between models. Similarly, regulatory elements, such as the long non-coding RNA SMN-AS1 (Box 1) (d'Ydewalle et al., 2017), are thought to be human-specific, and studying their function in mouse models may therefore be challenging. Although subtle, these differences may complicate comprehensively answering important questions, such as how much SMN protein is required by different tissues at varying developmental time points.

As has already been observed in mouse models of other neuromuscular diseases, including amyotrophic lateral sclerosis (Heiman-Patterson et al., 2005) and Duchenne muscular dystrophy (Coley et al., 2016), the genetic background of mouse strains can be an important determinant of the observed phenotype. In SMA, the Taiwanese and 2B mouse models have a less severe phenotype, with longer survival and reduced peripheral pathology, when the mice are on a C57BL/6 background compared to an FVB background (Buettner et al., 2021; Hosseinibarkooie et al., 2016) (Table 2). A possible explanation for this difference is the altered expression of modifier genes (Box 1); for instance, in the spinal cord of 2B mice on a C57BL/6 background, the expression of plastin-3 (Pls3) – a positive modifier gene in patients of specific families – was significantly increased compared to the same model on an FVB background (Eshraghi et al., 2016). This highlights the importance of careful reporting of the genetic background for correct interpretation of the results from studies using such inbred strains.

Although SMN2 is often suggested to produce ∼10% of functional SMN protein when compared to SMN1, animal and post-mortem studies suggest that this is substantially more nuanced and dependent on many factors, including age (or developmental stage) and the tissue or cell type analysed (Groen et al., 2018b; Wadman et al., 2016). The only study to robustly determine SMN protein and RNA levels in a large number of post-mortem spinal cord samples, from unaffected individuals and patients with and without SMA, found that SMN levels were especially high in prenatal controls, followed by a quick decline after birth (Ramos et al., 2019). They also found that SMN levels were substantially lower in patients with SMA both prenatally and neonatally, but that later in life, the expression levels were often comparable to those in unaffected individuals. However, the stability of SMN protein levels throughout life is challenging to study as repeated sampling of disease-relevant tissues is impossible. Studies in more accessible tissues suggest that SMN levels may decline further throughout life in patients with SMA (Ramos et al., 2019; Wadman et al., 2016) or remain relatively stable – especially in adults – depending on the cell type studied (Groen et al., 2018b; Kobayashi et al., 2011; Zaworski et al., 2016).

The most severe mouse models of SMA are characterised by a very progressive phenotype and low levels of the SMN protein. The development of models that mimic milder forms of SMA, and are thus based on higher levels of SMN expression, is therefore especially important. An allelic series of mouse models provided some interesting insights into the challenges associated with developing such models (Bogdanik et al., 2015; Osborne et al., 2012). In these studies, varying combinations of mouse Smn1 mutants and human SMN2 transgenes led to phenotypes that were either extremely severe (embryonically lethal) or relatively mild (long survival) and showed that the absence or presence of one extra, partially functional copy of an SMN gene was sufficient for this sudden switch in phenotype severity. This is different from the broad range of phenotypes in patients in which normal development can occur with symptom onset as late as >30 years of age, for example, in higher SMN2 copy number carriers (Piepers et al., 2008). In general, the development of disease models that accurately reflect relatively milder forms of SMA has not been straightforward, suggesting that in mice, the range of variation in SMN protein expression may be narrower than in humans (O'Hern et al., 2017), complicating the generation of models with intermediate phenotypes required for studying milder forms of SMA.

Before the start of clinical trials, the efficacy of SMN-targeting treatments was well established in animal models (Hua et al., 2011; Poirier et al., 2018; Valori et al., 2010). A major difference between currently approved treatments and those developed at preclinical stages are the delivery routes used for these medicines, which are difficult to replicate in cellular and animal models. For example, in humans, the blood-brain barrier is more permeable after birth than later in life and this permeability possibly persists up to 6 months after birth (Lam et al., 2015). In mice, this is the case for several days (possibly up to postnatal day 10) (Coelho–Santos and Shih, 2020), which overlaps with the critical time window for NMJ development, thus complicating a reliable comparison of insights from animal models and patients. In general, it is difficult to compare the biodistribution of different AAV serotypes between species and care should be taken in extrapolating observations between species. Studies in non-human primates were therefore instrumental in improving our understanding of the possibilities and limitations of using different AAV serotypes for therapy delivery (Davidoff et al., 2005; Foust et al., 2009) and may remain critical for the development of future AAV-based treatment, including base-editing strategies (Arbab et al., 2023). In a single post-mortem study of two cases, onasemnogene abeparvovec distribution was observed systemically, although viral copy number per cell varied extensively (up to 1000-fold) between tissues and cell types (Thomsen et al., 2021). For ASO treatment, although intracerebroventricular injection in mice mimics the delivery of ASOs into the cerebrospinal fluid of patients (Chiriboga et al., 2016), the final distribution of the treatment may still differ. Indeed, in limited post-mortem analyses of patients receiving nusinersen treatment, ASO distribution appeared to be restricted to mostly the lower segments of the spinal cord, near the lumbar injection site (Ramos et al., 2019). It should be noted, however, that in post-mortem cases, treatment effects may have been especially limited and therefore do not completely reflect common treatment effects. If and how it is feasible to better recapitulate treatment delivery in disease models remains questionable, but should be considered when translating model findings into the clinic.

Especially in mouse models of SMA, many non-CNS phenotypes have been described (see Table 2), including heart, gastrointestinal, liver, pancreas, kidney and vascular defects (Hamilton and Gillingwater, 2013; Yeo and Darras, 2020). Many of these phenotypes have also been reported in patients, although often analyses were restricted to post-mortem tissues of untreated patients (Hamilton and Gillingwater, 2013; Yeo and Darras, 2020), as the ability to study organ function in patients is substantially more limited than in animal models. To what extent the non-CNS phenotypes observed in disease models are relevant to patient symptoms remains to be further investigated, although differences are likely. For example, in mice with mild and severe SMA phenotypes, skin necrosis of tails and ears is commonly observed, whereas in patients, necrosis has been reported only in a limited number of very severe cases with one SMN2 copy (Weissman et al., 2021). Moreover, SMA mice often do not die from progressive neuromuscular decline but rather die abruptly, usually from heart defects (Bogdanik et al., 2015). In contrast, in patients, major heart impairments have only been reported in limited numbers of patients that are generally characterised by a severe phenotype and neonatal onset (Palladino et al., 2011; Rudnik-Schöneborn et al., 2008; Wijngaarde et al., 2017). These and other discrepancies can be due to varying reasons, including different roles or requirements of SMN, or alternative expression patterns of SMN in transgenic mice that remain to be identified. However, it is important to remember that these non-CNS phenotypes may appear during long-term treatment follow-up of patients with SMA, especially in patients that receive CNS-specific treatment of SMA. The extent and type of non-CNS pathology may again lead to a more variable phenotype in patients, depending on initial disease severity and type and start of treatment.

In summary, many factors including genetics, protein expression regulation, treatment delivery routes and non-CNS phenotypes may limit the use of current animal models to study important aspects of SMA in the therapeutic era. However, discussing these points may lead to insights that can facilitate the development and improvement of novel and existing model systems going forward.

Drosophila and C. elegans have been particularly useful for large-scale screening of drugs or genetic knockdown libraries in vivo (Aquilina and Cauchi, 2018; De Carlos Cáceres et al., 2018) and to study the effect of specific modifier genes (Hensel et al., 2021), and are a useful part of the SMA research toolbox that can be harnessed further. The transparency of C. elegans, combined with a detailed knowledge of number, location and connections of all (motor) neurons, allows for detailed analyses that are not always possible in other species.

In more complex small vertebrates, such as zebrafish, optimisation of screening and imaging tools increasingly enables high-throughput screening. For example, zebrafish were used to screen the effect of a library of synapse stabilisation compounds for SMA using a semi-automated imaging platform (Oprişoreanu et al., 2021). A similar platform might be used in the future for screening across a broad range of non-neuronal phenotypes by transgenic labelling of specific cell types. Moreover, owing to their high-throughput capabilities and genetic tractability, zebrafish have been and can be used to identify and investigate (genetic) modifiers and their role in influencing disease severity (Akten et al., 2011; Boyd et al., 2017; Lotti et al., 2012; Oprea et al., 2008; Powis et al., 2016; Riessland et al., 2017; Wishart et al., 2014). The use of tools, such as precise CRISPR/Cas9-mediated genome editing, in zebrafish will further extend and enhance the possibilities to perform these kinds of analyses (Ding et al., 2023).

Despite numerous challenges in using mouse models to mimic important features of SMA, the genetic toolbox available for mouse model development is extensive and has not yet been exhausted for SMA research. A perhaps obvious but important addition to the current palette of SMA models would be the generation of models based on outbred lines, such as the JAX diversity outbred mouse population (Lutz, 2018). The current use of inbred strains prohibits the possibility to model the genetic variation that occurs in real-world patient populations and is therefore more like studying a small group of comparable patients. Even though the use of outbred models may initially require an increased number of animals and concurrent ethical discussions, observations about SMA pathogenesis and treatment effects may ultimately prove to be more translatable when population genetics have been considered at the modelling stage.

Important insights into the spatial and temporal requirements of SMN, especially in the neuromuscular system, have been obtained by crossing existing SMA models with cell-type-specific or inducible-Cre mouse lines (Kariya et al., 2014; Kim et al., 2020; Lutz et al., 2011; Sleigh et al., 2011b). This approach will still be useful in the future, as the role of SMN in different organs and during development remains incompletely understood, and specific Cre lines for many cell types are readily available to investigate this further. This would also allow the study of the effect of rescuing SMN expression specifically in the CNS, as is currently the case with nusinersen treatment, compared to systemic approaches. Cre-based models also provide powerful tools to investigate the role of the SMN protein in specific tissues and cell types, at various time points, and independent of SMA pathology.

An important challenge for further mouse model development is the refinement of SMN protein expression to enable reliable modelling of milder forms of SMA. Some approaches for this already exist but have had comparatively little investigation, such as models based on milder mutations in SMN1 and treatment-based approaches (Monani et al., 2003; Sahashi et al., 2013; Workman et al., 2009; Zhou et al., 2015). The latter would, for example, include approaches that use suboptimal dosing of an SMN-targeting ASO to partially restore SMN levels and improve mouse lifespan. Although this is a valuable approach, in practice, it is not straightforward to achieve consistent expression across mice and between litters. The former would include the generation of mouse lines based on mutations in SMN1 that are known to be associated with relatively milder forms of SMA, such as p.S230L (Nölle et al., 2011) and p.G279C (Sharifi et al., 2021). The advantage of this approach would be that it may be possible to generate mouse models without the addition of a human transgene, which is comparable to the current 2B/− model (Bowerman et al., 2012a), so that SMN expression is more physiologically regulated.

By considering the remaining possibilities for mouse models, they are likely to keep their central role in SMA research for years to come.

The NMJ of sheep and pigs resembles the human NMJ much more closely than that of rats or mice (Boehm et al., 2020). However, the development of SMA models based on these larger animals still remains relatively new. A pig model of SMA, based on SMN knockdown using a virus expressing an shRNA targeting pig SMN1, has previously been published (Duque et al., 2015). This model allowed the characterisation of the postnatal role of SMN in motor neurons and the evaluation of clinically relevant biomarkers, such as compound muscle action potential (CMAP) and motor unit number estimation (MUNE), in response to SMN restoration. This represents an advantage of working with large animal models, as neurophysiological biomarkers may be more difficult to investigate in smaller animal models. Additionally, pigs possess several characteristics that are considerably more comparable to those of humans, such as their cardiovascular system, digestive tract and cytochrome P450 enzymes, (Ribitsch et al., 2020; Schelstraete et al., 2019), which imply that the outcomes of drug metabolism and toxicity studies from pigs would be more translatable to humans. In addition to obvious logistical challenges around working with large animal models, genetic manipulation is generally not straightforward. In other diseases, spontaneous mutations in disease-relevant genes have been used successfully to study diseases such as Batten's disease in sheep (Frugier et al., 2008). However, due to the absence of SMN2 in other species, spontaneous SMN1 mutations would most likely also be embryonically lethal. Should it become possible to genetically edit pigs or sheep to mimic SMA, it would provide an important, disease-relevant readout for further analysis and translatability to the clinic.

A remaining challenge when using animal models for SMA is the fact that SMN2 is specific to humans, which will always limit the specificity of animal models in SMA research. Advances in in vitro disease model systems in recent years have, however, been significant and provide important steps forward. An additional advantage of using patient-derived cellular models of SMA is the ability to include a diverse genetic population in analyses, thereby bypassing an important limitation of having to use many inbred mouse models as discussed above. Moreover, human models would allow us to better study the physiological regulation of SMN2 expression in more detail. We will discuss several options that may be of particular use, depending on the type of experimental design required and the goal of specific analyses.

One of the most basic patient-derived cellular models are primary fibroblasts. Fibroblasts are readily obtained from skin biopsies from patients and can be collected, cultured and analysed from large numbers of patients. In addition, many researchers in the SMA field have obtained fibroblast lines from public biobanks, such as the Coriell repository (Adami and Bottai, 2019). Although the morphological and functional difference between fibroblasts and motor neurons is substantial, fibroblasts do reflect key features of SMA, such as reduced SMN2 RNA and SMN protein levels, and respond well to in vitro treatments (Garbes et al., 2013; Singh et al., 2006; Wadman et al., 2016). Moreover, primary patient-derived fibroblasts can be used without further modifications, therefore preserving (epi)genetic variants that may be putative modifiers of SMA outcomes. For example, within families, SMN2 copy number is often equal, whereas SMA severity often is not – one possible explanation for this discrepancy may be alternative methylation of the SMN2 copies between siblings (Hauke et al., 2009; Wadman et al., 2020). However, when aiming to preserve epigenetic variants, fast expansion of low-passage cultures into the number of cells required for the experiment is critical, as, in high-passage cultures, DNA methylation changes over time (Franzen et al., 2021; Koch et al., 2011).

Since the Yamanaka paper in 2006 (Takahashi and Yamanaka, 2006), many steps have been taken to refine the differentiation of patient-derived primary cells, such as fibroblasts, into a pluripotent stem cell state. The generation of such induced pluripotent stem cells (iPSCs) provides an exceptionally powerful tool, as it allows scientists to subsequently differentiate these into practically any cell type of their choice. Indeed, many groups in the SMA field have used this approach to study the effect of low SMN levels on iPSC-derived motor neuron function. Currently published studies, however, have some limitations. Many groups are using one or several of a limited number of Coriell repository cell lines as a starting point for their analyses (Adami and Bottai, 2019), largely bypassing the opportunity to capture as much genetic variability as possible and with the risk of missing relevant variation in phenotype or treatment effect (Garbes et al., 2013). Extensive characterisation of primary patient-derived fibroblasts from various biobanks will support optimal selection of representative patient cell lines, based on SMA type, genetics and therapy response, when generating iPSC lines for further study. Moreover, the SMN protein has been suggested to play an important role in the differentiation process from primary fibroblast to an iPSC state (Chang et al., 2022; Varderidou-Minasian et al., 2021), which may cause additional variation after differentiation. Alternatively, for SMA research, direct conversion approaches (see e.g. Ladewig et al., 2012) may be particularly promising, as this allows fibroblasts to be differentiated into neurons without the need for a pluripotent intermediate state. Here, the main limitation will be the lack of a cell expansion step comparable to iPSC culture, limiting the amount of material that can be generated in single experiments. Still, the use of iPSC-derived motor neurons, for example, in co-culture with iPSC-derived myotubes to generate in vitro NMJ systems (Faustino Martins et al., 2020; de Jongh et al., 2021; Southam et al., 2013), holds great future promise for detailed analyses of treatment effects and molecular analyses at various important cellular sites. Comparable to how Cre-based mouse models have been used to characterise the requirements for SMN expression in different parts of the neuromuscular system, these in vitro NMJ systems allow the study of these requirements in different patient-derived cell types. Moreover, they can be used to gain valuable insights on the local mechanistic effects of reduced SMN expression. Lastly, they can be used to study the effect of treatment when targeted to the entire neuromuscular system or, in contrast, when restricted to specific parts only.

Finally, many research groups around the globe are developing increasingly complex organoid systems to model a range of complex diseases. The use of organoids in the SMA field has so far been relatively limited, as spinal organoids do not as reliably reflect spinal cord anatomy and development as, for example, brain organoids mimic cortical development and anatomy (Hor et al., 2018). However, current advances in culturing multi-modal organoids, for example, systems that connect cortical, spinal and muscle organoids, essentially mimicking the entire neuromuscular system in a dish (Andersen et al., 2020), illustrate the increasing versatility of complex culture systems. This opens up many new opportunities for research into local treatment effects and tissue-specific pathological changes that may occur in SMA, in line with in vitro NMJ systems as discussed above. Additionally, multi-organoid systems that have been developed to study, for example, the interaction between the kidney and liver (Nguyen et al., 2022) may provide a great possibility to study non-CNS phenotypes of SMA in detail, including targeting treatments to such phenotypes. So far, most of our knowledge on non-CNS phenotypes in patients with SMA is based on limited post-mortem studies. Therefore, the opportunity to use patient-derived organoid cultures to gain a more in-depth understanding of the consequences of low levels of SMN on organ function is highly promising. Although challenging to set up, concurrently culturing different organoids in separate compartments of the same culture system provides a fascinating opportunity to study the interaction and signalling between organoids. This also enables the comparison of treatment effects when it is either targeted to one organoid or the whole system.

In summary, even though the SMA field has already benefitted from the availability of many useful disease models, several routes for further optimisation and development are still available. Efficiently combining multiple model systems – including small animal and cellular models for high-throughput screening, mouse models to study molecular mechanisms, large animal models for biomarker discovery and drug development, and patient-derived models to better study patient heterogeneity and SMN regulation – holds great promise, both in terms of scalability and inclusion of genetic variation, as well as the detailed study of local pathological changes and treatment effects.

The SMA research field is changing rapidly. The approval of multiple gene-targeting treatments and their optimal use through early, often presymptomatic treatment provides a significant step forward for patients with SMA and their families. The observed heterogeneity in both disease severity and treatment outcomes, however, highlights that the current ‘one-size-fits-all’ SMN-targeting treatment approach may, in the future, need to be replaced by more individualised approaches (Fig. 2). Without the continuous further development and investigation of current and new SMA models, it will not be possible to develop such individualised treatment approaches. Although this Review has been speculative at times, we have highlighted several features to consider when contemplating the use and further development of various SMA models in the future. We hope that this will help the field to determine the important next steps that need to be taken to advance our understanding of SMA further and keep improving the lives of all those affected by SMA.

The authors would like to thank Maria Zwartkruis, Dr Kiterie Faller (University of Edinburgh) and three anonymous reviewers for critically reading and evaluating our manuscript.