“Between stimulus and response there is a space. In that space is our power to choose our response. In our response lies our growth and our freedom”—Viktor E Frankl.

In autism spectrum disorder (ASD), a neurodevelopmental disorder affecting ~ 1.5% of the population [1], aetiologically diverse deficits in cognitive plasticity lead to broad impairments in communication and restricted, repetitive behaviours [2]. Comorbidities are common (~ 70% of cases) and include epilepsy; attention, mood and language disorders; sleep disturbance; gastrointestinal problems; and intellectual disability [3].

Despite the great personal and sociological cost of ASD (estimated to be $2 million/patient/year [4]), only the antipsychotics risperidone and aripiprazole are currently FDA-approved to treat ASD, indicated solely in the treatment of irritability symptoms [5]. A fundamental reason for this lack of disease-modifying therapies may relate to ASD’s pathogenesis, which appears to be principally driven by heterogeneous genetic mutations and variants and modulated by diverse gene × environment interactions, to include pregnancy-related factors (e.g. maternal immune activation, maternal toxins) and perinatal trauma [2, 6–10]. Many of the encoded proteins implicated in ASD pathogenesis—such as cytoskeletal proteins, cell adhesion molecules and DNA-binding proteins—may be ‘undruggable’ using conventional small molecule drugs, which principally only modulate the function of receptors and enzymes [11].

In contrast, gene therapy—broadly defined as the delivery of nucleic acid polymers into cells to treat disease—may be used to repair, replace, augment or silence essentially any gene of interest in a target cell, opening up new areas of the proteome for drug targeting [12]. Other advantages of gene therapy versus small molecules include the ability to effect long-lasting clinical benefit with a single treatment and the potential to control cellular targeting via vector modifications [13].

Indeed, gene therapy is already making a clinical impact in the field of neurology, with Nusinersen, an antisense oligonucleotide therapy approved in Spinal muscular atrophy (SMA), and more recently Luxturna, a viral-based gene replacement strategy approved in Leber’s congenital amaurosis, acting as the first disease-modifying therapies in both of these diseases [14, 15]. In addition, a clinical trial in SMA by AveXis using systemic delivery of recombinant adeno-associated virus 9 (rAAV9) carrying a replacement SMN1 gene recently proved safe and efficacious in neonates [16]. On the other hand, gene therapies are clearly expensive in the short-term, with current therapies costing at least $500,000 per treatment, and thus remaining unaffordable in many healthcare systems (see ref [17] for a thorough economic analysis).

This review will highlight key targets for ASD gene therapy, the utility of ASD models, and recent advances in our ability to deliver such therapies to the central nervous system (CNS). It will then move on to discuss recent gene therapy strategies in ASD, concentrating on conditions with available preclinical data, and the roadblocks facing their clinical translation.

More recently, our understanding of the genetic landscape of ASD has been revolutionised by several whole-exome and whole-genome sequencing studies, identifying hundreds of de novo and rare inherited variants influencing sporadic ASD risk [24–32]. Many of these genes appear to be involved, either directly or indirectly, in synaptic morphology and activity, leading to the concept of ASD as a ‘synaptopathy’ [33, 34] (Fig. 1). Certainly, the idea of using gene therapy to increase or decrease the expression of target proteins within this network and ‘retune’ the synapse is a powerful one, which may be applicable to certain ASD cases.

However, in such a heterogeneous condition as ASD, it is important not to become evangelical about a single causative mechanism, especially given recent insights into the apparently critical roles of immune dysfunction and epigenetics in at least certain ASD cases [35, 36]. Furthermore, recent phase II clinical trials which looked to regulate synaptic function via GABA and glutamate receptor modulation failed to demonstrate significant overall benefit, despite strongly positive responses in certain patients [37, 38]. Thus, it is important to consider whether targeting the synapse using gene therapy may be most appropriate for correcting particular ASD endophenotypes in specific patient subsets, rather than seeing it as a panacea for ASD, a topic that is returned to later in this article.

The field of gene therapy is littered with examples of therapies which failed to translate from their preclinical promise. In many cases, blame can be attributed to the predictive validity of the animal model used, which is itself related to its construct validity (i.e. how well the model mimics disease aetiology) and face validity (i.e. how closely the model’s phenotype represents the human disorder) [39].

Given the challenges clinicians have faced in developing diagnostic criteria for ASD [40], and how various social traits often appear to be inherently ‘human’ qualities (although this is itself highly contested [41]), it is little wonder that generating ASD animal models with good face validity has been challenging. Nevertheless, whilst caution should always be taken when ascribing behavioural outcomes in animals to autism, various monogenic ASD rodent models have capably demonstrated quantifiable social and communicative behavioural traits [42–44], laying the ground for preclinical therapeutic studies.

As a caveat, it must be noted that a major limitation of ASD animal models relates to their inability to reflect heterogeneous environmental influences on the ASD phenotype, with various toxic, inflammatory and psychosocial factors difficult to incorporate in robust, reproducible animal models. This deficit in construct validity is especially relevant in modelling polygenic ASD, in which the gene × environment interaction plays a more fundamental role [45], and which therefore bears a particular risk for future translational work.

Identifying specific cells and circuits dysregulated in ASD is crucial in gene therapy design, as vectors may be targeted to specific brain regions or cell types (discussed in detail later). In particular, nascent data suggest the hippocampus, cerebellum and corpus callosum contain key pathogenic circuits [34, 46–48], whilst influential cell types include pyramidal cells, Purkinje cells and glial cells [34, 49].

Recent technological advances have begun to elucidate the relationship between cell type-specific function and particular ASD endophenotypes. For example, in a RS mouse model, cre/lox-mediated deletion of MECP2 specifically from forebrain glutamatergic neurons led to a partial disease phenotype, with deficits in social behaviour and motor coordination, but preserved locomotor activity and fear-conditioned learning [50]. Meanwhile, in a TSC mouse model, chemogenetic excitation of the Right Crus I (RCrusI) region of the cerebellum—an area consistently noted to be altered in the ASD brain in neuroanatomical and neuroimaging studies [51, 52]—was sufficient to specifically rescue social impairments, without rescuing repetitive or inflexible behaviours [53].

Within individual circuits, it appears that different ASD mutations may have opposing effects on synaptic function. For example, TSC2^+/− and FMR1^−/y knockouts appear to have opposite effects on mGluR-dependent long-term depression (LTD) in the hippocampus, whilst mice bred with both mutations balance each other out at the synaptic and behavioural levels [54]. Such data not only exemplify the heterogeneic nature of ASD but also highlight the necessity of optimal synaptic control, and are the first hint that a successful gene therapy must walk a narrow therapeutic tightrope between over- and under-stimulating synaptic transmissions.

A crucial further question is whether autistic phenotypes can be reversed or are neurodevelopmentally fixed. Remarkably, an array of studies in different monogenic ASD animal models have consistently demonstrated the potential for reversal of established neuronal dysfunction, either after pharmacological intervention or genetic reactivation of silenced alleles [55–60]. These findings imply that postnatally, indeed post-symptomatically, the genetic horse may not yet have bolted, and genetic correction via a delivered vector might be useful in treating cognitive dysfunction.

If the ASD phenotype is reversible rather than neurodevelopmentally fixed, as implied by studies in monogenic animal models, then it follows that continuous genetic correction will be necessary for a sustained therapeutic effect. Currently, only viral packaging systems have combined efficient transduction with long-term gene expression in vivo [61] (although, as will be discussed later, certain ASD conditions may be amenable to non-virally delivered antisense nucleic acid therapies).

Of the viruses which can transduce post-mitotic cells, rAAVs have emerged as the principal CNS delivery candidate [62]. This is based upon their relatively low immunogenicity (compared to adenovirus and herpes simplex virus), limiting the likelihood of an encephalitic immune response, their ability to persist in episomal form, reducing their oncogenic potential (compared to retroviruses [63, 64]), and their high production titres [12]. Indeed, rAAV vectors have already been used safely in a number of early clinical studies in CNS gene therapy, in disorders ranging from SMA to idiopathic Alzheimer’s disease [65].

Optimising cell-specific targeting is critical in maximising the number of transduced target cells/dose and limiting off-target toxicity. There are two major ways in which the spatial dynamics of rAAV vectors can be adjusted. Firstly, the properties of the vector itself can be modified, to include a cell type-selective capsid (e.g. AAV9 is particularly neurotropic [66, 67]) and/or promoter [68].

Secondly, the mode of delivery can be adjusted. Historically, rAAV vectors have been delivered intraparenchymally via stereotactic CNS injection, leading to high local concentrations with limited vector spread [69]. Although invasive (each injection requiring a craniotomy), such a localising method of delivery might have utility in correcting specific dysregulated ASD circuits linked to particular clinical endophenotypes, analogous to the recent improvement in motor scores seen after lentiviral vector delivery of a dopamine-producing gene therapy to the nigrostriatal pathway in Parkinson’s disease [70].

However, ASD appears to involve global synaptic dysregulation and thus will require global CNS gene correction to fully reverse cognitive phenotypes [71, 72]. The discovery that rAAV9 crosses the blood-brain barrier (BBB) and globally transduces CNS neurons and glial cells [73], and the recent derivation of more efficient BBB-traversing rAAVs by targeted evolution [74], has opened up the possibility of using intravenous injection in ASD gene therapy.

It remains to be seen, however, whether side effects relating to peripheral tissue transduction, as well as the presence of neutralising circulating antibodies (anti-AAV9 antibodies are present in 47% of humans), will preclude intravenous administration in various ASDs [75–77]. Changes to the viral vector nucleic acid sequence outside of the transgene—such as the inclusion of ‘detargeting’ sequences recognised by micro RNAs (miRNAs) expressed specifically in off-target cells [78]—might circumvent the former issue, but use up highly limited space (rAAV’s packaging capacity is limited to ~ 5 kb [79]). The latter issue may be negotiated by the use of engineered rAAV capsids, which may have lower neutralising antibody seropravalences [80].

An alternative to systemic delivery is intrathecal administration, which potentially combines (relatively) safe administration and global CNS transduction with fewer peripheral complications, and a higher spatial resolution limiting the dose requirements. However, there are conflicting data regarding the ability of intrathecally delivered rAAV to efficiently transduce areas outside the spinal cord [81], as well its own ability to avoid both peripheral leakage [82] and a neutralising antibody response [83, 84]. Of note, an intrathecal AAV9 approach has been corrective in a model of giant axonal neuropathy [85] and has progressed to a clinical trial.

As previously mentioned, in comparison with monogenic ASD, polygenic ASD has a greater environmental component driving its phenotype [45]. In this respect, damaging, nonsynonymous postzygotic mutations in whole-exome sequences from the largest collection of trios with ASD were recently identified, with some of these genes being particularly enriched for expression in the amygdala, a key brain region for social conditioning and learning [114]. Such factors, combined with the current paucity and constructional limitations of animal models in polygenic ASD, make it a less obvious target for gene therapy.

Nonetheless, despite the bewildering array of rare genetic mutations linked to polygenic ASD, an important focus of these mutations appears to be in the regulation of synaptic function, with diverse ASD mutations potentially connecting aberrant translational inputs and outputs (Fig. 1) [34]. For example, in mice, deletion of the translational repressor Eukaryotic translation initiation factor 4E-binding protein 2 (4E-BP2) led to overexpression of the NLGN class of cell adhesion molecules [115], mutations of which have been causally linked to ASD [46, 116, 117]. Furthermore, such deletion resulted in disruption of the ratio of excitatory to inhibitory synaptic inputs, as well as an ASD behavioural phenotype, which was corrected by NLGN1 knockdown [115].

This leads to the idea that many ASD mutations might be treated by fine-tuning the expression of influential proteins acting within dynamic translational loops. The apparently fundamental role of the PI3K-AKT-mTOR pathway in various causes of monogenic ASD [118], as well as the phenotypic reversal seen using small molecule inhibitors of mTORC1 preclinically [119], suggests that this pathway may be a critical target for gene therapy in certain cases of ASD.

However, given ASD’s heterogeneity, it is once again important not to focus myopically on a single pathway. Rather, instead of embarking on a ‘one size fits all’ therapeutic approach, the effect of any particular ASD mutation on translational output and synaptic function should be categorised, before deciding whether and how to target a particular gene or pathway. For instance, NLGN3 knockout mice demonstrate a FXS-like disruption of mGluR-dependent synaptic plasticity [120], suggesting that either FMRP overexpression or PI3K-AKT-mTOR pathway knockdown (given the aforementioned opposition of these two pathways) might correct this phenotype.

Finally, recent evidence has emerged of ASD behaviours caused by amino acid deficits [121, 122]. For example, homozygous dysfunction of the BBB solute carrier transporter 7a5 (SLC7A5) and corresponding CNS loss of branched chain amino acids (BCAAs) has been linked to ASD, which is reversible in mouse models upon intra-CNS administration of BCAAs [122]. Thus, direct protein replacement therapy might provide an important additional therapeutic avenue in certain ASD cases. It is also possible to imagine using gene therapy as an adjunct here: for example, combining systemic BCAA replacement with vector-delivered SLC7A5 targeting BBB cells.

Given the heritable component of ASD, gene therapy offers a promising alternative to conventional small molecule therapies. Preclinical studies over the last 5 years using animal models displaying autism-like traits have demonstrated that directly altering gene expression using rAAV-delivered transgenes can reverse the behavioural phenotype, either via gene replacement or RNA knockdown. Such studies establish proof-of-concept and set up a platform for clinical translation in various monogenic ASDs, with RS being a frontrunner in this regard.

However, major hurdles remain, not least the fact that the majority of ASD disorders, even monogenic ones, show variable penetrance, with epistatic and gene × environment interactions determining phenotype. Not only is such genetic and environmental heterogeneity inherently difficult to model, hindering clinical translation, but also in clinical trials that do go ahead, ASD subgroups that benefit from a particular treatment may be lost amongst other unsuitable subgroups. Furthermore, we still do not know whether, or in which cases, epigenetic factors may preclude reversibility in humans. Cyclically, this brings us back to the question of animal models and whether these have sufficient construct validity to actually begin to answer such questions in the first place.

A number of additional questions remain: Firstly, can vector design be optimised to the extent that intravenous delivery achieves sufficient CNS transduction without peripheral toxicity? Secondly, where is the optimum balance between CNS transduction and the risk of on-target transgene-related toxicity for each ASD syndrome? Thirdly, will demonstrations of acceptable levels of CNS toxicity hold when studies commence in larger animal models? Fourthly, is there a time point beyond which some or all autistic features lose their reversibility? Answering these questions will be key to moving ASD gene therapy into clinical trials, and perhaps one day generating a genetic treatment for ASD.