The single most important question when choosing a molecular therapy is what the pathogenic mechanism actually is — because the mechanism dictates the strategy. A loss-of-function disorder (the protein is missing or non-functional, as in recessive disease) calls for adding back the gene or boosting output of a paralog; you cannot fix a deletion by silencing anything. A gain-of-function or dominant-toxic disorder (a misfolded or aggregating protein actively poisons the neuron, as in SOD1-ALS or Huntington disease) calls for lowering the offending transcript. Picking the wrong axis is futile — this is why diagnosis must be precise to the level of gene and variant before a precision therapy is even considered.

The field divides along a second axis: where the change is made and how durable it is. Gene addition delivers a functional copy that sits alongside the genome (episomal, not integrated) — durable in non-dividing neurons but vulnerable to dilution where cells still divide. Antisense oligonucleotides and RNAi act on mRNA, leaving the genome untouched; their effect is potent but transient, demanding repeat dosing. Gene editing alters the DNA sequence itself, offering a potentially one-and-done correction but raising the stakes on specificity and off-target effects. mRNA therapeutics sit at the most transient end — a protein is expressed for days, then gone.

• Reversibility cuts both ways. mRNA and ASO effects fade if dosing stops (a safety advantage if toxicity emerges, a burden because therapy is lifelong). A permanent edit cannot be undone if it misfires.
• Delivery is the recurring bottleneck, not the molecular biology. Every category in this module is rate-limited by getting the agent past the blood-brain barrier and into the right cells at the right dose — the subject of the next two sections.

Adeno-associated virus (AAV) became the workhorse of in vivo neurological gene therapy for reasons rooted in its biology. It is a small (~25 nm), non-enveloped, single-stranded DNA virus that is naturally replication-deficient — wild-type AAV cannot replicate without a helper virus and is not an established human pathogen. That makes it an unusually benign chassis. The older, stronger claim that AAV "has never been linked to human disease" no longer holds: in 2023 three groups reported AAV2 in the livers of most children in the 2022 cluster of severe acute hepatitis of unknown origin, alongside a helper adenovirus and a marked excess of HLA-DRB104:01 ([Ho et al., Nature 2023](https://pubmed.ncbi.nlm.nih.gov/36996873/); Morfopoulou et al., same issue). Causality is unresolved, and the finding concerns wild-type AAV2 rather than recombinant vectors, which retain no viral coding sequence. To build a vector, virtually everything viral is stripped out: recombinant AAV (rAAV) keeps only the two inverted terminal repeats (ITRs)* — short hairpin sequences that flank the cargo and are all that is needed for genome packaging and persistence. The therapeutic transgene plus its promoter sits between them. Removing all viral coding sequence both makes room for cargo and removes the genes whose expression would otherwise mark the cell for immune destruction.

The genome's behavior after entry is what makes AAV well-suited to the brain: it does not meaningfully integrate. Instead the delivered DNA forms stable, circular episomes in the nucleus. In a post-mitotic neuron — a cell that will never divide again — an episome can drive expression for years with no risk of insertional mutagenesis. The same property is a liability anywhere cells still divide, because episomes are not copied at mitosis and are progressively diluted out (the durability problem revisited later).

Serotype is the steering wheel. The capsid protein that forms the outer shell determines which receptors the virus engages, hence which tissues it enters (tropism) and whether it can cross the blood-brain barrier. Swapping the capsid (AAV1, 2, 5, 8, 9, AAVrh10, AAVrh74, and engineered variants) while keeping the same ITR-flanked cargo lets one redirect an identical transgene to muscle, liver, retina, or CNS — which is why capsid choice, not the gene, often defines a product.

The first wave of approved neurological gene therapies share a logic worth making explicit: each treats a recessive loss-of-function disease by restoring the missing protein, and each exploits a delivery route that solves a specific anatomical problem. Onasemnogene abeparvovec uses systemic AAV9 to reach motor neurons in infants whose blood-brain barrier is still permeable; the single-dose AVXS-101 data establishing that the course of motor neuron disease could be altered by one infusion came from Mendell et al. 2017. Because that permeability closes with age, the same transgene now also comes as a concentrated intrathecal formulation (onasemnogene abeparvovec-brve, Itvisma — FDA-approved November 2025 for patients ≥2 years, dosed independent of weight): the delivery route is chosen to fit the patient's biology, not the other way around. Voretigene neparvovec is injected directly under the retina — an immune-privileged, contained space where a tiny dose suffices and systemic antibodies are irrelevant. Delandistrogene moxeparvovec confronts the cargo limit head-on, packaging an engineered micro-dystrophin because the full gene cannot fit.

Four recurring lessons run through these approvals. First, timing is biology, not just policy — presymptomatic or early treatment dramatically outperforms later intervention, because once motor neurons die, no transgene revives them. Second, the immune system is the adversary at every step: pre-existing antibodies can block the vector, and the host response to the capsid drives the immunosuppression and monitoring that surround every infusion. Third, durability is unproven on a human lifespan — these are recent approvals, and episomal dilution, waning expression, and the inability to simply re-dose are open questions, not solved ones. Fourth, the safety profile keeps being written after approval. Elevidys reached the market on an accelerated pathway in 2023 and was broadened in June 2024 to traditional approval for ambulatory patients ≥4 years plus accelerated approval for non-ambulatory patients ≥4 years, but two deaths from acute liver failure in non-ambulatory boys in 2025 led the FDA to add a boxed warning for acute serious liver injury and to withdraw the accelerated non-ambulatory population, leaving the indication at ambulatory patients ≥4 years (FDA safety communication, November 2025); the EMA's CHMP declined authorization altogether after the EMBARK trial missed its primary motor endpoint (Roche, July 2025). Skysona's lentiviral platform carried a theoretical insertional-oncogenesis risk at approval that became concrete — hematologic malignancy in 15% of trial participants by 2025 — prompting restriction to patients without a matched donor. Rare-disease approvals rest on small trials, so the first few hundred treated patients are, in effect, the safety study. The economics follow directly: a one-time AAV infusion priced in the millions is a bet that a single dose lasts decades, a bet still being tested.

Antisense oligonucleotides (ASOs) are short (18–25 nucleotide) single strands of modified DNA that find their RNA target by ordinary Watson-Crick base pairing. The elegance is that a single chemical platform yields two opposite mechanisms, selected by where on the transcript the ASO binds and how it is chemically built.

• Degradation (RNase H1 pathway). When an ASO binds its complementary mRNA, the DNA:RNA duplex is recognized by the cell's own enzyme RNase H1, which cleaves the RNA strand and leaves the ASO intact to bind again. This lowers the target protein — exactly what a gain-of-function disease needs. Tofersen (SOD1) and inotersen (TTR) work this way.
• Splice-switching (steric block). Here the ASO is chemically designed not to recruit RNase H1, so it does not destroy the transcript. Instead it sits on the pre-mRNA and physically blocks a regulatory site, redirecting the spliceosome. Nusinersen masks the ISS-N1 silencer in SMN2 intron 7, forcing inclusion of exon 7 and raising full-length SMN. Eteplirsen does the inverse — it forces skipping of a DMD exon to restore the reading frame. Same molecular class, opposite effect on protein quantity. Note that risdiplam, which reaches the same exon 7 decision, is not an ASO and does not act at ISS-N1: it is a small molecule that stabilizes U1 snRNP on the weak 5′ splice site at the end of exon 7. The Duchenne exon-skipping drugs also remain on accelerated approval granted on dystrophin as a surrogate: the confirmatory ESSENCE trial of golodirsen and casimersen missed its primary endpoint in November 2025 (4-step ascend velocity at 96 weeks, p=0.31) while significantly raising dystrophin, and the sNDAs seeking conversion to traditional approval were accepted in June 2026 with a PDUFA date of 28 February 2027; eteplirsen’s own confirmatory study (MIS51ON) is still running.

Chemistry is what makes any of this survivable in the body. A naked oligonucleotide is degraded within minutes by serum nucleases. The phosphorothioate backbone (a sulfur swapped into the phosphate linkage) resists nucleases and, importantly, promotes protein binding that keeps the drug in tissue; 2'-O-methyl / 2'-O-methoxyethyl sugars and locked nucleic acids (LNA) boost target affinity and stability; phosphorodiamidate morpholino oligomers (PMO) replace the sugar-phosphate backbone entirely. Notably, the splice-switching chemistries (2'-MOE, PMO) are the ones that do not trigger RNase H1 — the modification simultaneously protects the drug and dictates its mechanism.

Because ASOs are small synthetic molecules rather than viruses, they sidestep the entire problem of anti-capsid immunity and can be re-dosed. The trade-off is the mirror image of gene therapy: effect is transient, so intrathecal delivery must be repeated for life. Injected into CSF, an ASO distributes across the neuraxis and is taken up by neurons and glia without needing to cross the blood-brain barrier from the blood side. The pivotal ENDEAR trial that proved this approach in infantile SMA was reported by Finkel et al. 2017. A decade of use has since refined the regimen: a higher-dose schedule (two 50 mg loading doses, then 28 mg every four months) was FDA-approved in March 2026 on the DEVOTE trial. The platform also keeps reaching new neurological targets — zilganersen (Zanvastro), an RNase H ASO that lowers GFAP, became the first disease-modifying therapy for Alexander disease in September 2026: a leukodystrophy treated by silencing a toxic gain-of-function transcript rather than by replacing a missing enzyme.

Most of the hard problems in gene therapy are not molecular — they are immunological, developmental, and economic — and they cluster around two facts: the body fights the vector, and a single dose has to last.

The immune system attacks AAV at two distinct stages, and each demands a different countermeasure. Before infusion, pre-existing neutralizing antibodies (from common natural AAV exposure) coat the capsid and block transduction outright — hence mandatory antibody screening, and the reason a seropositive patient may be deferred and retested as titers fall. After infusion, a cellular (T-cell) response to the viral capsid can kill the very cells that were just transduced, erasing the benefit weeks later; this is why prophylactic corticosteroids are started around dosing and liver enzymes are watched closely. The two are not interchangeable — antibodies are a gate at the front door, T-cells are a threat after entry.

Durability is the quietest but deepest problem. Because rAAV stays episomal, any cell that divides dilutes it. In an infant treated systemically, the liver is still growing, so hepatocyte division steadily lowers transgene copies per cell over years — a structural reason expression may wane as a child reaches adolescence, and part of the rationale for considering an ASO alongside gene therapy to hold protein levels up. Critically, you usually cannot just give a second dose: the immune memory raised by the first exposure neutralizes the same serotype, and cross-reactivity limits simply switching capsids. Re-dosing is therefore an unsolved engineering challenge, driving work on deimmunized and computationally designed capsids.

• The therapeutic window is unforgiving. For diseases like SMA, neurons lost before treatment do not return, so newborn screening and presymptomatic treatment are not conveniences but determinants of outcome.
• Lowering a transcript is not always benefit. The halt of the HTT-lowering ASO tominersen in Huntington disease is a reminder that engaging the target (reducing the protein) does not guarantee clinical efficacy and can even harm — mechanism is necessary but not sufficient.
• Access and ethics scale with price. Multimillion-dollar one-time therapies force questions about equity, long-term value when durability is unproven, and how to fairly allocate a treatment whose benefit may depend on being given before symptoms — i.e., before a family may even know the disease is coming.