Neurogenetics

Neurogenetics is two fields that share a name and barely share a method. Established One of them studies single genes whose failure produces a defined neurological disease, and it has produced medicines. Nusinersen for spinal muscular atrophy was approved in December 2016; tofersen for SOD1-ALS in 2023; trofinetide for Rett syndrome on 10 March 2023; onasemnogene abeparvovec, sold as Zolgensma, in May 2019 at a US list price of $2.125 million. Established The other studies the common genetic variation underlying psychiatric diagnoses and cognitive traits, and after twenty years and sample sizes in the millions it has produced no medicine at all — it has produced epidemiology, some genuine biology, and a statistical object that is persistently misread.

Established The framing usually offered is that we are learning the genetics of the brain. The fetched record supports something sharper and less comfortable: we have learned the genetic architecture of brain disorders well enough to see that it consists of thousands of variants of tiny effect plus a short list of very large ones, and that the field's first two decades of candidate-gene results were not merely underpowered but false. Established Border and colleagues tested 18 canonical depression candidate genes in 621,214 people and concluded the prior positives “were false positives.” That episode is the single most instructive thing neurogenetics has to teach any adjacent field.

Established Three numbers carry most of this brief. SNP-based heritability for intelligence is 0.19. The ratio of direct to population effects for the educational-attainment polygenic index is 0.556, meaning about 31% of that index's predictive power is the child's own genotype acting on the child. And in 178,086 siblings the population estimate for education shrinks 47% when you compare brothers and sisters rather than strangers. Frontier A fourth number matters for the therapeutic half: no delivery vehicle crosses the human blood-brain barrier at therapeutic fraction after systemic dosing, which is why every approved CNS genetic medicine is injected into a compartment or a structure by hand.

Established The flagship result is the 2022 schizophrenia GWAS, and its headline is a map rather than a mechanism. Trubetskoy and colleagues assembled 76,755 cases and 243,649 controls and reported 287 distinct genome-wide-significant loci, prioritising 120 genes of which 106 are protein-coding. Established Associations concentrate in genes expressed in excitatory and inhibitory CNS neurons, which is a real biological statement and not a truism — it excludes glia, immune tissue and the liver as the primary site of common-variant risk. Established Only 16 of the 120 carry a credible causal variant. Established The ancestry composition of the primary analysis was 74.3% European, 17.5% East Asian, 5.7% African-American and 2.5% Latino, which is the fact that governs how the resulting score behaves outside Europe.

Established The same paper's effect sizes are the part that gets dropped. SNP heritability on the liability scale is 0.24 with a standard error of 0.007. The polygenic risk score explains a median 7.3% of variance in liability at a p < 0.05 inclusion threshold, and 2.4% if you use only genome-wide-significant SNPs. Established The odds ratio comparing the top centile of the score with the bottom centile is 39, with a 95% confidence interval of 29 to 53. Established The odds ratio comparing the top centile with the remaining 99% is 5.6, interval 4.9 to 6.5. Both are correct; quoting the 39 without the 7.3% and the 5.6 is the standard distortion, and it is the reason a lay reader comes away believing a score identifies individuals.

Established The rare-variant half of the architecture is the more mechanistically informative one. The SCHEMA consortium meta-analysed exomes from 24,248 cases and 97,322 controls plus 3,402 trios and found ten genes in which ultra-rare coding variants confer odds ratios of 3 to 50 at p < 2.14 × 10^-6, including GRIN2A and GRIA3 — glutamate receptor subunits. Established Genes prioritised by the common-variant analysis are enriched for rare-variant risk, so the polygenic background and the near-Mendelian tail point at the same synaptic biology. Established After excluding those ten genes, cases retain what the authors call a substantial excess of ultra-rare damaging variants, so the list is incomplete by the study's own accounting.

Established For cognitive traits the architecture is the same shape and the numbers are smaller. Savage and colleagues ran an intelligence GWAS in 269,867 people, found 205 loci and implicated 1,016 genes, and measured SNP heritability at 0.19 with an out-of-sample polygenic prediction reaching 5.2% of variance in four independent samples. Established The educational-attainment GWAS is the largest behavioural study ever run: 3,037,499 individuals, 3,952 approximately uncorrelated significant SNPs, a polygenic index explaining 15.8% of variance in Add Health and 12.0% in the Health and Retirement Study, sample-size-weighted mean 13.3%. Established The same paper reports a direct-to-population effect ratio of 0.556 with a standard error of 0.020, which implies that 30.9% of the index's R-squared is direct effect. That is the consortium reporting that roughly two-thirds of its own score's predictive power is something other than the measured person's own genes acting on that person.

Established The cleanest way to see this is to stop comparing strangers. Howe and colleagues assembled 178,086 siblings in 77,832 sibships across 19 cohorts and re-estimated genetic associations within families, where the confounds of population structure, assortative mating and parental environment cancel. Educational attainment shrank by 47% (95% CI 41–52%). Cognitive ability shrank by 22% (6–37%). Height shrank by 10% (8–12%). Smoking shrank by 19% (9–30%), and body mass index showed limited evidence of any difference. Established The shrinkage is largest for exactly the traits people want to select on and smallest for the traits nobody does. That ordering is not a coincidence and it is the most important single table in behavioural genetics.

Established The candidate-gene literature that preceded all of this was false. Border and colleagues tested 18 genes, 16 polymorphisms and 4 imputed VNTRs — SLC6A4, DRD4, MAOA, SLC6A3 among them — against depression in 621,214 individuals drawn from UK Biobank subsamples and the Psychiatric Genomics Consortium. Established They found one marginal COMT rs4680 effect at p = 0.002, a DRD2 gene-wise signal in PGC that failed to replicate in UK Biobank, and no significant gene-by-environment interactions at all. Established Their conclusion was that the prior findings “were false positives” and that the field should “abandon” the approach. Frontier 5-HTTLPR, COMT Val158Met, BDNF Val66Met and DRD4 nonetheless remain in commercial genetic-optimisation panels and nootropics marketing, which is a live consumer-protection problem rather than a historical curiosity.

Established Missing heritability, stated with numbers instead of as a slogan. Twin and family estimates of IQ heritability rise with age — the Wilson effect — from roughly 0.55 in early childhood towards 0.7 in adolescence and higher still in old age, with the highest published figures around 0.9 in Swedish 65-year-olds. Frontier Several of those point estimates were read from a figure rather than a table in the source this brief cites, so treat individual values as approximate; the direction is not in doubt. Established Against that: SNP heritability of 0.19 and out-of-sample prediction of 5.2%. For schizophrenia, a conventional twin heritability near 0.8 against SNP heritability of 0.24 and a polygenic liability R-squared of 7.3%. Established Only the SNP-based, within-family portion is available to selection or editing. The rest may be rare variants, non-additive effects, gene-environment correlation, or measurement artefact of the twin design itself — the candidate resolutions are live and none has won.

Established Meanwhile the monogenic half delivers. The approved neurogenetic drug list is dominated by antisense oligonucleotides delivered intrathecally: nusinersen (SMA, December 2016), eteplirsen (DMD exon 51, September 2016), golodirsen (exon 53, December 2019), viltolarsen (exon 53, August 2020), inotersen (hereditary ATTR, October 2018) and tofersen (SOD1-ALS, 2023). Established Milasen, an antisense oligonucleotide designed for one child's CLN7 Batten mutation and dosed in 2019 under an FDA expanded-access protocol, was explicitly described as “not suited for the treatment of other patients with Batten's disease.” Established Zolgensma delivers an SMN1 transgene in an AAV9 capsid, was approved in May 2019 for children under two, lists at $2.125 million and was negotiated to about $1.53 million in Japan; its adverse reactions include hepatotoxicity and thrombocytopenia. Established An intrathecal formulation, Itvisma, was approved for ages two and up in November 2025.

Established And the record is not clean. Novartis delayed reporting data manipulation by its own executives to the FDA until June 2019, one month after Zolgensma's approval, and withheld animal toxicity findings for seven months. Established Skysona, the gene therapy for cerebral adrenoleukodystrophy, produced myelodysplastic syndrome in six of 67 patients and acute myeloid leukaemia in a seventh, with five clones carrying a vector insertion in MECOM and one in PRDM16, at onsets of 14 to 92 months. Established That is roughly a 10% malignancy rate in a paediatric neurological indication. Frontier uniQure's AMT-130, an AAV-delivered microRNA against huntingtin injected by MRI-guided convection-enhanced stereotactic delivery into the striatum, reported 75% slowing of composite UHDRS progression at 36 months (p = 0.003) and 60% on Total Functional Capacity (p = 0.033), with cerebrospinal-fluid neurofilament light down 8.2% — against a propensity-matched external control rather than the trial's own sham arm, which is the caveat most coverage loses.

Established Two more monogenic results set the ceiling on what genetics currently buys. Trofinetide, approved on 10 March 2023 for Rett syndrome, is a synthetic peptide analogue rather than a genetic therapy at all: it treats a disorder caused by mutations in MECP2 without touching MECP2, met both co-primary endpoints in a trial of 184 females aged 5 to 20, and produced diarrhoea in 80.6% of them. Frontier DB-OTO, an otoferlin gene therapy delivered into the cochlea, improved 14 of 15 treated ears across 12 children aged 10 months to 16 years, with nine reaching 70 decibels hearing level or better at week 24 and three achieving normal hearing sensitivity, stable or improving out to 72 weeks. Established Both required knowing the gene, and neither edits it. That is the shape of the entire delivered record: identify a single gene, then reach its tissue by any means available.

Established Two further established facts set the scale of what is known and unknown. Kaplanis and colleagues sequenced 31,058 parent-offspring trios with developmental disorders, established 285 significantly associated genes including 28 not previously robust, and estimated that more than 1,000 genes associated with developmental disorders remain undescribed. Established Elliott and colleagues genotyped and imaged 8,428 UK Biobank participants across 3,144 imaging-derived phenotypes and found significant SNP heritability in 1,578 of them — highest for structural volumes, lowest for resting-state functional connectivity, where only 235 of 1,771 edges reached significance. Established The substrate underneath all of it is UK Biobank's roughly 500,000 participants and about 96 million testable variants in 487,442 individuals, plus the All of Us programme, which by 30 June 2026 reported over 883,000 enrolled and more than 535,000 whole genomes, 86% of participants from historically underrepresented groups.

Established The structural connectome arrived, on schedule and at the advertised scale, and the denominators are the part that rarely gets printed beside the headline. The FlyWire consortium published a wiring diagram of an entire adult fly brain in 2024: 139,255 neurons and of the order of 54.5 million synapses, reconstructed from volume electron microscopy. A connectome of the male fruit fly central nervous system followed. Established In mammals the unit is still the cubic millimetre: a dense reconstruction of mouse visual cortex published in 2025 covers about 200,000 cells and more than 500 million synapses in roughly one cubic millimetre, and a human temporal-cortex volume of similar size holds about 57,000 cells, 150 million synapses and 1.4 petabytes of image data. Frontier A mouse brain is of the order of 500 such cubic millimetres and a human brain of the order of a million. The fly is finished; the mammal is at roughly a five-hundredth, and the human at a millionth, of one brain.

Established The control experiment for “wiring predicts function” has been running for forty years, and its result is negative. The nematode Caenorhabditis elegans has had a complete synaptic connectome — 302 neurons in the hermaphrodite — since 1986, and connectomes across development for multiple individuals since 2021. Frontier Nobody can predict the animal's behaviour from that diagram. Forward simulation of the full circuit does not reproduce locomotion without fitted parameters that the connectome does not supply, and the parameters are where the behaviour lives. Frontier The developmental series also showed that a large minority of connections — on the order of 40% — are not shared between individuals of the same genotype raised in the same conditions. Speculative A wiring diagram of 302 neurons, complete for four decades, is the strongest single piece of evidence against the claim that a wiring diagram of 139,255 or of eighty-six billion will explain anything by itself.

Established The reason is specific and it is not a lack of resolution: the electron micrograph does not contain the variables that set the dynamics. A reconstructed synapse carries a location and an apposition area. It does not carry a neurotransmitter identity — in the fly volume those were assigned by a trained classifier, not measured — nor a synaptic weight, nor a short-term plasticity rule, nor the neuromodulatory state of the animal. Established It also omits a second signalling system entirely. A neuropeptide signalling network mapped in the nematode from receptor-ligand expression describes a communication graph that is largely distinct from the synaptic one: neuropeptides diffuse, act on cells they never touch, and change the sign and gain of the wired circuit. Speculative For this brief the implication is exact rather than rhetorical: if the neurogenetic variants that matter act by tuning transmitter identity, receptor expression and peptidergic signalling, they act on precisely the layer that connectomics does not image.

Frontier The convergence of rare and common architectures is the decade's real structural result. SCHEMA's finding that common-variant-prioritised genes are enriched for rare-variant risk is what makes drug-target inference from GWAS defensible at all; without it, 287 loci is a list of correlations. Frontier Whether that convergence extends beyond schizophrenia to depression, ADHD and autism is not settled in the sources this brief can cite, and the cross-disorder papers that would answer it were record-verified but not readable here.

Frontier The direct-effect fraction is now measurable, and most secondary literature has not caught up. The 0.556 direct-to-population ratio and the 47% within-sibship shrinkage for education both date from 2022. Textbooks, press coverage and commercial marketing overwhelmingly still quote uncorrected between-family figures. Frontier This is the rare case where the correction is not contested by anyone who works on it — it is simply not yet distributed.

Frontier Polygenic embryo selection is being sold ahead of its evidence, and the numbers on both sides are specific. Herasight states that its CogPGT 1.0 predictor explains 16.4% of fluid-intelligence variance among unrelated UK Biobank individuals and correlates 0.456 with latent general cognitive ability within 4,642 UK Biobank sibling pairs. Orchid states that it reads 99% of an embryo's genome across more than 400 genetic conditions including autism, Alzheimer's, ALS and schizophrenia. Frontier Both are interested parties, neither has published an independently replicated within-family randomised outcome, and Orchid publishes no sensitivity or specificity. Frontier The sharpest technical critique comes from Sasha Gusev, who puts within-family heritability ceilings at roughly 4% for educational attainment and 14% for IQ, estimates an upper bound of about 6 IQ points from selecting the top of ten embryos, and reports that selecting on educational attainment raises bipolar-disorder relative risk by about 1.4 times at ten embryos. Frontier That last figure is the most useful thing in the debate and it comes from a single critic; it needs replication.

Established The ancestry gap has moved from projected to measured. Polygenic scores are several times more accurate in individuals of European ancestry, and the applied consequence is quantified: the expected educational-attainment gain from selecting the best of ten embryos is 0.53 years for European ancestry, 0.40 for admixed American, 0.35 for East Asian and 0.23 for African — the same product delivering roughly 43% of its benefit to African-ancestry users, with non-European groups 25–65% worse across all conditions studied. Frontier All of Us is the first substrate at a scale that could fix this, with more than 275 million previously unreported variants including over 3.9 million with coding consequences in its first 245,388 clinical-grade genomes.

Frontier Bespoke, one-patient genetic medicine is now a demonstrated turnaround time rather than a proposal. KJ Muldoon, born 1 August 2024 with CPS1 deficiency, received a patient-specific adenine base editor in a lipid nanoparticle first dosed on 25 February 2025, three infusions over seven weeks, discharged 2 June 2025 — design to manufacture in about six months against an estimated eighteen. Frontier The disease is metabolic rather than neurological, but its damage is neurotoxic hyperammonaemia, and the regulatory precedent of an n-of-1 protocol is the template for the long tail of ultra-rare neurogenetic disease that will never support a trial.

Frontier The hardest open question is the one nobody markets: getting from a locus to a mechanism. Twenty years of GWAS have produced 287 schizophrenia loci and 16 credible causal variants. The gap between an associated haplotype and a molecule that does something in a neuron is filled by expression quantitative trait loci, chromatin conformation and heroic single-locus functional work, and it is crossed a few times a year. Speculative Whether that rate can be industrialised — base-editing candidate variants into isogenic human iPSC-derived neurons and reading out electrophysiology at scale — is the question that decides whether psychiatric GWAS ever becomes pharmacology.

Frontier One live neurobiological question this brief declines to resolve. Whether neurogenesis persists in the adult human hippocampus has been contested since two 2018 papers reached opposite conclusions, and a 2025 report of proliferating neural progenitors bears directly on it. Those papers were record-verified but could not be read for this brief, so it states no resolution and prints no figure from them. Speculative The answer matters because it determines whether a genetic intervention in an adult brain has any dividing target population at all.

Frontier The effectome is the proposed answer, and its contribution is a measurement strategy rather than a measurement. The proposal, published on the fly connectome in 2024, is to treat the wiring diagram as a prior over a dynamical system: fit a linear dynamical model whose structure is given by the connectome, derive the matrix of causal effects one neuron has on another, and use the predicted dynamical modes to choose which perturbations are worth running. Established The combinatorial argument for why this is necessary is not in dispute. 139,255 neurons make roughly 1.9 × 10¹⁰ ordered pairs; at one perturbation per second the exhaustive experiment takes about six hundred years. Frontier So the all-pairs effectome is not underfunded, it is impossible, and the live claim is the narrower one: that structure constrains dynamics tightly enough for selective perturbation to be informative. Speculative That claim has not been validated against a measured effectome in any animal, because no measured effectome exists.

Frontier The perturbation tools are real, and they measure a different thing from what the model predicts. Two-photon holographic optogenetics can stimulate specified small ensembles in a behaving mouse and drive behaviour, and pan-neuronal calcium imaging can read a whole nematode or larval zebrafish brain at cellular resolution. Frontier What neither delivers is a causal effect coefficient between identified cells in the same individual whose wiring has been reconstructed, because the electron-microscopy volume is acquired from a dead animal and the physiology from a live one. Speculative The decisive experiment is therefore a co-registration problem, not a stimulation problem: functional recording and perturbation in an individual, followed by dense reconstruction of that same individual, with the predicted and measured effect matrices compared cell by cell. Speculative The mouse cortical volume published in 2025 is the closest approach to that design at mammalian scale, and it is one animal.

Established For the therapeutic half, the binding bottleneck is a membrane. The entire approved and late-stage in vivo genetic-medicine record for the CNS is delivery into a compartment or a structure by hand: intrathecal injection for the antisense oligonucleotides, MRI-guided convection-enhanced stereotactic injection into the striatum for AMT-130, intracochlear injection for otoferlin gene therapy. Established Nothing in this record achieves therapeutic transduction of a defined neuronal population after systemic dosing in an adult human. Frontier The chemistry works: in vivo base editing has produced dose-dependent 21%, 41% and 53% mean LDL-C reductions in humans by editing the liver, which is the organ a lipid nanoparticle goes to whether you want it to or not. Every non-liver programme is a surgery.

Established The workback chain for a delivered monogenic CNS therapy, in order. First, a vector or formulation with greater than 30% transduction of a named neuronal population in a large animal after systemic administration — this does not exist and is the binding link. Second, redosing: AAV capsids provoke neutralising immunity, so a single shot must last, and Roctavian's declining expression after 12 to 18 months in a non-CNS indication is the warning. Third, an on-target damage readout: ordinary Cas9 cuts resolve into deletions extending over many kilobases that short-range genotyping misses, and post-mitotic neurons cannot dilute a bad clone. Established Fourth, an endpoint a regulator will accept for a slow degeneration in a few dozen patients. Fifth, a price that a health system will pay more than a few dozen times.

Speculative For the polygenic half the chain is different and the binding link is not where the public argument is. Link one is a within-family-validated cognitive predictor replicated by someone who does not sell it. Link two is evidence that the within-family signal is causal rather than assortment-driven. Link three is a quantified pleiotropic cost across bipolar, autism and schizophrenia liability. Speculative Link four is enough embryos, which requires human in vitro oogenesis producing meiotically competent, correctly imprinted oocytes — not achieved for human cells at all. Speculative Links one to three are what the debate is about; link four is what stops it. Even a perfect predictor runs into the embryo-count ceiling, because a typical IVF cycle yields far fewer than ten euploid embryos. Speculative Naming the binding link on each side: delivery across the blood-brain barrier for the therapeutic programme, and the supply of embryos for the selection programme. Neither is a genetics problem, and neither is helped by a better polygenic score.

Frontier The binding constraint on causal brain mapping is industrial and institutional, and it is worth stating plainly because the field is usually described as if it were waiting on a discovery. Volume electron microscopy is an imaging and proofreading pipeline: petabyte-scale acquisition, automated segmentation, and then human correction of the segmentation, which for the fly brain consumed a large distributed proofreading effort over years. Speculative Segmentation error rates are improving and the correction step has not disappeared, so cost scales with volume rather than falling with it, which is the opposite of the sequencing analogy the field borrows when it asks for money. Frontier The funding position is the second constraint and it moved the wrong way: multisite connectome programmes lost federal support in 2025 as a consequence of institutional funding disputes rather than of any scientific judgement about the work. Speculative A field whose unit of progress is one cubic millimetre and whose funding arrives in political weather is not on a technology curve.

Established This brief waits on results that Genetic Engineering produces, not on more genetics. The editing chemistry, the delivery vehicles, the on-target damage record and the manufacturing economics all arrive from there. Established Neurogenetics contributes the target list and inherits the toolkit, and every failure mode in the fetched therapeutic record — insertional oncogenesis, declining expression, hepatotoxicity, unaffordable price — is a delivery or durability failure rather than a targeting failure.

Established It depends second on biobank scale and biobank diversity. Every number in section two is a function of sample size, and the two results that most changed the field's self-understanding — the 621,214-person candidate-gene refutation and the 178,086-sibling shrinkage estimate — only became possible when cohorts crossed six figures. Frontier The next such result requires within-family designs at ten times current scale, and non-European cohorts at a scale that only All of Us currently approaches.

Speculative It depends third on a lever that is not the sequence. If the genetic ceiling on cognition is real, the remaining route is developmental timing rather than allele frequency, which is Neuroplasticity Engineering's subject. Established The mechanism is characterised: critical periods open with the maturation of parvalbumin-positive interneurons and close behind perineuronal nets, whose removal with chondroitinase ABC restores plasticity in adult rodent visual cortex. Speculative Whether any of that transfers to a human adult is unproven.

Established It depends fourth on diagnostic infrastructure that already exists but is unevenly distributed. Every approved neurogenetic therapy requires a molecular diagnosis before a patient can be treated, which makes trio sequencing and newborn screening the actual gate on uptake for the monogenic half. Established With 285 established developmental-disorder genes against more than a thousand estimated to be undescribed, the limiting factor for most affected families is still whether anyone can name their condition. Frontier See Precision Medicine for the machinery around that.

Established These are the specific measurements that would move the field, in rough order of cost.

Frontier One. A pre-registered sibling-pair validation of a published cognitive polygenic score in a cohort not used for training, reporting within-family R-squared with a confidence interval. The only claimed within-family figure — a correlation of 0.456 in 4,642 pairs — is a company's own. This experiment is cheap, needs no new data collection, and nobody has run it. Frontier Two. A multi-trait selection simulation on real sibling genomes, reporting the induced change in bipolar, autism-spectrum and schizophrenia liability from selecting on education or cognition. Gusev's approximately 1.4-times bipolar figure at ten embryos is the only published number and it needs a second one.

Speculative Three. Base- or prime-edit a credible causal schizophrenia variant into isogenic human iPSC-derived neurons and measure a synaptic or electrophysiological phenotype. Sixteen of the 120 prioritised genes have a credible causal variant identified; this is the experiment that converts one of them into a mechanism, and it is the rate-limiting step for every drug-discovery claim made for psychiatric GWAS. Frontier Four. Dose-ranging for any systemically administered vector achieving greater than 30% transduction of a defined neuronal population in a non-human primate, with long-read sequencing for kilobase-scale deletions in the transduced tissue.

Frontier Five. Repeat AMT-130 against a concurrent randomised sham arm rather than a propensity-matched external control. The result is the largest CNS genetic-therapy signal on record and its control is its weakest feature. Speculative Six. Electronic-health-record-linked treatment-response phenotyping at All of Us scale: predict who responds to clozapine rather than who develops schizophrenia. Pharmacological response plausibly has fewer contributing loci than the disorder, the sample sizes have never existed because response is not collected at biobank scale, and for the first time they plausibly could. Established Seven. Long-read whole-genome sequencing of edited human embryos, since the only published figure is 61.9% loss of heterozygosity at a single targeted locus against 10% in controls, and nobody has measured it genome-wide. Speculative Eight. A registry with ten-year clonal tracking of CNS gene-therapy recipients, since the malignancies in the one product that produced them appeared between 14 and 92 months after dosing and no standard follow-up window would have caught the late ones.

Established The engineering requirement that dominates everything else is crossing or bypassing the blood-brain barrier. Bypassing is what currently works, and it is expensive per patient: intrathecal dosing requires repeated lumbar punctures for life in the antisense products, and convection-enhanced striatal delivery requires an MRI-guided neurosurgical procedure per patient. Frontier Neither scales to a common disease, and both put the cost floor above what any health system pays for a chronic neurological condition.

Established The vector constraints are hard and well characterised. AAV9 carries roughly 4.7 kilobases, which excludes most large CNS genes outright; systemic dosing at CNS-relevant titres drives hepatotoxicity and thrombocytopenia, both listed for Zolgensma; and pre-existing or induced neutralising antibodies make redosing difficult. Speculative Human artificial chromosomes are the only demonstrated route for a payload the size of dystrophin at 2.4 megabases, and thirty years of proof-of-concept has produced no therapy, with the required centromeric sequences still unidentified.

Established Manufacturing and price are engineering problems too. Approved neurogenetic and adjacent gene therapies list at $4.25 million (Lenmeldy), $3.2 million (Elevidys), $3 million (Skysona) and $2.125 million (Zolgensma). Established The comparison case for throughput is the first approved CRISPR therapy, where about 165 patients had completed a first cell collection and 39 had been infused roughly 22 months after approval, against a stated eligible US population of about 16,000. Frontier The n-of-1 route is the one piece of manufacturing that got dramatically faster: six months from sequence to dosing, against an eighteen-month estimate.

Established The one CNS delivery route that genuinely scales is chemical rather than viral. Antisense oligonucleotides reach the neuraxis by intrathecal injection, are redosable indefinitely, carry no integration risk, and are the reason the approved neurogenetic drug list is mostly oligonucleotides rather than gene therapies. Frontier Their limitation is equally structural: they modulate transcripts and do not repair genes, so they work for exactly as long as they are given, which converts a genetic disease into a lifelong procedure schedule and a permanent revenue stream rather than a cure.

Established Neuroplasticity engineering is the nearest neighbour and the most plausible substitute. Critical periods are opened by parvalbumin-positive basket-cell maturation and closed by perineuronal nets — chondroitin sulphate proteoglycan lattices whose removal restores adult plasticity in rodents. Established Restoring chondroitin 6-sulphate in 20-month-old mice, by chondroitinase ABC or by AAV-chst3, rescued memory deficits and restored cortical long-term potentiation. Speculative If the genetic ceiling on cognition is real, this is where the remaining headroom would have to be.

Frontier Neural interfaces are the competing route to the same clinical endpoints, and they scale differently. Adaptive deep brain stimulation that self-adjusts to individual brain activity received US FDA approval on 24 February 2025 — a closed-loop device fitted to one person in a clinic visit, which is a different scaling law from anything that has to be manufactured per patient from their own cells. Frontier Decoding bandwidth has meanwhile reached 62 words per minute from intracortical arrays in a single participant with ALS, at a 23.8% word error rate over a 125,000-word vocabulary. Frontier See Brain-Computer Interfaces for where that is going.

Established Precision medicine supplies the honest comparison class. Pharmacogenomics is the one part of genomic medicine with a randomised outcome trial behind it, and what it reliably predicts is individual harm — which drug will injure this patient — from a small panel of well-characterised metabolic genes. Frontier Polygenic prediction of individual benefit has no comparable randomised result in any indication, neurological or otherwise; the numbers behind that asymmetry are in Precision Medicine. Speculative Germline routes — in vitro gametogenesis, chromosome selection, multiplex editing — sit adjacent to all of this and are treated in Genetic Engineering; the neurogenetic question they raise is whether there is anything worth selecting for. Frontier Biological Enhancement reaches the same verdict from the other direction: the strongest replicated result in healthy cognitive enhancement is the absence of transfer, and the strongest result in cognitive genetics is the absence of a large-effect enhancing allele.

Established The institutional constraint this brief actually waits on is an endpoint. Neurogenetic diseases are rare and slow, which means trials are small and long, and the regulatory system has no settled answer for either. Frontier AMT-130's pivotal readout used a propensity-matched external control because a sham craniotomy arm large enough to power a 36-month readout in Huntington's disease is close to impossible to recruit. Established Elevidys took the other route and was approved on gene expression rather than clinical benefit; its phase III trial with 125 participants then missed its primary motor endpoint, three deaths from acute liver failure followed in July 2025, the FDA imposed and lifted a clinical hold within ten days, and the EMA recommended against approval. Established Those are the two available strategies and both have now failed publicly.

Established The n-of-1 pathway exists and is being used, which is genuinely new. Milasen was made for one child under an expanded-access protocol in 2019 and explicitly disclaimed for anyone else; a patient-specific base editor was dosed three times over seven weeks in 2025 under a patient-specific protocol. Frontier What does not exist is a reimbursement mechanism for a therapy with a denominator of one, or a manufacturing standard that scales the practice past a handful of academic centres. Established The wider access problem is measured elsewhere in this cluster: the first approved CRISPR therapy had infused 39 people worldwide roughly 22 months after approval, against about 16,000 eligible US patients, which is a reimbursement and transplant-capacity outcome rather than a scientific one.

Established The heritable end is closed by law, not by capability. Seventy countries explicitly prohibit heritable genome editing for reproduction, five more prohibit it with exceptions, and no country permits it; for in vitro research 19 prohibit, 4 prohibit with exceptions and 11 explicitly permit. Established An eighteen-signatory call for a moratorium on heritable editing was published in 2019 without specifying a duration. Frontier Embryo scoring, by contrast, is barely regulated anywhere: companies sell cognitive and psychiatric polygenic reports on embryos with no published validation and no premarket review, which is the widest gap between practice and oversight in this subject.

Established The first ethical problem is a measurement error with consequences. A polygenic score with a liability R-squared of 7.3% is an excellent epidemiological instrument and a poor individual test, and it is routinely presented as the second. Established The top centile of the schizophrenia score carries a 5.6-fold risk against the other 99%, which still leaves the great majority of top-centile individuals unaffected for life. Frontier Deploying such a score in education, insurance, employment or family planning treats a population statistic as a personal fact, and the sibling-comparison results say the between-family version of that statistic is substantially not about the person's own genome at all.

Established The second is distributional and quantified. Predictive accuracy falls 25 to 65% outside European ancestry and the modelled selection benefit falls from 0.53 years of education to 0.23. Frontier A technology sold as individually empowering therefore delivers less than half its claimed benefit to a large fraction of the world's population, and does so by construction rather than by accident, because the discovery cohorts were 74% European.

Speculative The third is what embryo screening for psychiatric and neurodevelopmental conditions means. A commercial product currently screens embryos for autism, schizophrenia and ALS liability without publishing sensitivity or specificity. Speculative Even setting aside the disability-rights objection, the liability-threshold model those scores rely on is contested when applied to spectrum conditions, and selecting hard on one trait moves others: the only published estimate puts the bipolar-disorder relative-risk increase from selecting on educational attainment at roughly 1.4 times with ten embryos. Handwave Nobody has computed the full multi-trait consequence, and no regulator has asked.

Speculative There is a fourth problem that cuts the other way and deserves recording. The within-family polygenic estimate is the closest thing genetics has to a causal claim about an individual, which makes it simultaneously the most defensible number in the field and the most dangerous one to hand to an institution. Frontier A field that succeeds at its own stated goal produces a better instrument for exactly the uses it says should be off-limits, and no version of the ethics discussion currently addresses that.

Speculative The most likely civilizational consequence of neurogenetics is not enhancement; it is a slow, expensive, unevenly distributed retreat of a class of childhood neurological disease. Spinal muscular atrophy, otoferlin deafness, several leukodystrophies and a growing list of ultra-rare metabolic-neurological conditions are becoming treatable one at a time at prices between $2 million and $4.25 million. Frontier That is a real change in what a paediatric neurology clinic does, and it arrives as a health-financing problem rather than a scientific one. Established The comparison that sets the stakes is diagnostic rather than therapeutic: 285 developmental-disorder genes are established and more than a thousand are estimated to remain undescribed, so most affected families still have no molecular diagnosis and therefore no route to any of this.

Speculative The second consequence is epidemiological rather than clinical. Polygenic scores work well on populations and poorly on people, which makes them a serious instrument for questions about environments, interventions and trial stratification — and a poor one for the individual predictions they are being sold for. Frontier A field that made that distinction load-bearing in public would be more useful than one that keeps relitigating heritability.

Speculative The third is a negative and it deserves to be said plainly. No genetic intervention in the fetched literature has moved a healthy adult human cognitive capability. Speculative If cognitive genetic architecture really is a plateau maintained by stabilising selection — thousands of loci of tiny effect, no identified large-effect enhancing allele in 269,867 people, and every known large-effect neural variant deleterious — then the enhancement debate is arguing about a resource that is not there, and the interesting civilizational question moves to development, education and devices. Frontier That is a conclusion the field's own data support and its own funding structure discourages anyone from stating plainly.

These horizons track specific results that would have to land, not announcements.

• 10 yr: Frontier An independently replicated within-family validation of a cognitive polygenic score, which is cheap, decisive and currently unrun. Frontier AMT-130 or a successor reports against a concurrent randomised control, settling whether the 75% slowing survives a proper comparator. Frontier Systemically administered CNS gene delivery demonstrated in a non-human primate at therapeutic fraction — the single result that would change the shape of this field. Established More approvals in the monogenic tail, all of them injected by hand, all of them priced above $2 million.
• 25 yr: Speculative A handful of the 16 credible causal schizophrenia variants converted into mechanisms with a drug-discovery programme attached, which is the only route by which psychiatric GWAS becomes pharmacology. Speculative Treatment-response prediction at biobank scale replacing diagnosis prediction as the field's clinical product. Frontier Non-European cohorts large enough that the portability gap closes for at least the best-studied disorders. Speculative A regulator-accepted endpoint framework for slow degenerations in cohorts of a few dozen, most plausibly borrowed from oncology's single-arm precedents rather than invented.
• 50 yr: Speculative If systemic CNS delivery is solved, a somatic genetic therapy for a common-variant-dominated neuropsychiatric condition becomes conceivable for the first time — and it would still require a target, which the polygenic architecture does not currently supply. Speculative Alternatively, and just as plausibly, a settled negative: common-variant psychiatric genetics retired to epidemiology and trial design, with the therapeutic action entirely in monogenic disease and in devices. Handwave Human in vitro gametogenesis with correct imprinting, which would remove the embryo-count ceiling that currently makes selection arguments moot.
• 100 / 250+ yr: Handwave Multiplex germline editing at the scale that would be required for a meaningful polygenic shift — hundreds of simultaneous edits with viable development, no karyotypic abnormality and demonstrated additivity, none of which has been shown beyond a handful of edits in any mammal. Handwave Whole-chromosome selection or transplantation as a bypass, which requires genotyping a chromosome in situ and moving it intact into a viable cell; neither method exists and the proposal's central step is asserted rather than demonstrated. Handwave A description of brain function at a level where genetic intervention has a well-posed target at all, which may be the actual prerequisite and which nobody can currently specify.

Frontier One causal-mapping line belongs in the ten-year window, and it is a specific comparison rather than a scale milestone. A single animal in which the same individual is functionally recorded, perturbed and then densely reconstructed, with a connectome-predicted effect matrix compared against a measured one, would settle whether wiring constrains controllable function or merely describes it. Speculative Absent that comparison, further volume buys resolution and not inference, and the neurogenetic question this brief cares about — how a common variant of small effect reaches a behaviour — stays exactly where the polygenic architecture left it.

• Depends on Depends on Genetic Engineering for everything on the delivery side: the editing chemistry, the vectors, the on-target damage record and the manufacturing economics all arrive from there, and every failure in the therapeutic record assembled here is a delivery or durability failure rather than a targeting failure. Depends second on biobank scale and biobank diversity, because every quantitative claim in this brief is a function of sample size and the two results that most changed the field's self-understanding only became possible above six figures: a 621,214-person refutation of the candidate-gene literature, and a 178,086-sibling estimate of how much of a between-family association survives a within-family comparison. Depends third on a functional-genomics throughput that does not yet exist, since 287 associated loci have yielded 16 credible causal variants and the conversion rate from locus to mechanism is the rate-limiting step for every drug-discovery claim the field makes.
• Requires (not on this map) Neurogenetic diseases are rare and slow, so their trials are small and long, and no agreed way exists to demonstrate benefit in that setting. The two strategies actually used have both failed publicly: the largest CNS gene-therapy result on record was measured against a propensity-matched external control rather than a concurrent sham arm, and the best-known accelerated approval on a molecular surrogate was followed by a missed motor endpoint, three deaths and a European rejection. What is missing is not a laboratory result but an endpoint framework a regulator will accept for a few dozen slowly declining patients. Two further requirements arrive with causal brain mapping, and neither is a discovery. The first is volume electron-microscopy and proofreading throughput: dense reconstruction is an imaging and human-correction pipeline whose cost scales with volume, and the mammalian unit of progress is still the cubic millimetre — roughly a five-hundredth of a mouse brain. Established The second is the functional-genomics throughput this brief already names as missing on the depends-on side, typed here as the requirement it is: without it a locus does not become a mechanism, and a mechanism is the only thing a circuit map can be tested against. Speculative Both are capacity that a funder could buy.
• Enables Enables precision medicine in neurology first and most concretely: a monogenic diagnosis is the entry ticket to every approved neurogenetic therapy, and the 285 established developmental-disorder genes against more than a thousand still undescribed sets the size of the prize. Enables, more usefully than its reputation suggests, better trials rather than better predictions — polygenic stratification is a serious instrument for enrichment, randomisation balance and post-hoc heterogeneity analysis even where it is a poor individual test. Enables neuroplasticity engineering to be posed as a competing hypothesis rather than a separate subject, since the strength of the genetic ceiling is exactly what determines how much is left for developmental timing to do. Enables, if systemic CNS delivery is ever solved elsewhere, a target list that is already written. Enables, finally, a clearer public argument, because most of what is contested about this subject turns out to be a confusion between a population statistic and a personal fact, and the sibling designs settle that confusion without needing anyone's opinion.
• Adjacent Sits alongside neural interfaces and neuromodulation, which pursue the same clinical endpoints on a device scaling law rather than a biological one, and alongside pharmacogenomics, which is the only part of genomic medicine with a randomised outcome trial behind it. Sits alongside the germline programme — in vitro gametogenesis, iterated embryo selection, multiplex editing — where neurogenetics supplies the traits that motivate the whole enterprise and, on the present evidence, also supplies the strongest reason to doubt it. Sits alongside precision medicine, whose central asymmetry this subject inherits without exception: genomic medicine predicts individual harm reliably and individual benefit poorly, and neurogenetics has not yet produced a counterexample.

Established “IQ is eighty per cent heritable, so most of it is selectable.” Two different quantities are being merged. Twin-based heritability for adult IQ is high and rises with age. SNP-based heritability is 0.19, and the best out-of-sample polygenic prediction is 5.2% of variance. Established The gap between those is what missing heritability names, and only the SNP-based, within-family portion is available to selection or editing. Frontier The candidate resolutions — rare variants, non-additive effects, gene-environment correlation, and inflation in the twin design itself — are all live, and which one wins changes what the ceiling is.

Established “A polygenic score tells you about this person.” It does not, and the cleanest demonstration is the sibling comparison. Compare 178,086 siblings rather than strangers and the educational-attainment signal shrinks 47%, cognitive ability 22%, smoking 19%, height 10%. Established Height barely moves because a between-family height association really is mostly the person's own genome; education halves because a between-family education association is substantially parental environment, assortative mating and population structure travelling under a genetic label. Established The educational-attainment consortium reports the same thing from a different direction: a direct-to-population effect ratio of 0.556, so 30.9% of its index's predictive power is direct effect. Quoting 13.3% for education without the 0.556 overstates the heritable, selectable component by roughly a factor of three.

Established “The score identifies who will develop schizophrenia.” The odds ratio between the top and bottom centiles of the schizophrenia score is 39, which sounds decisive and is the number that circulates. The odds ratio between the top centile and everyone else is 5.6, and the score explains 7.3% of variance in liability. Established The great majority of people in the top centile never develop the disorder. Frontier Area under the curve makes the same point: about 0.71 for schizophrenia, 0.64 for coronary disease, 0.63 for breast cancer — useful for ranking a population, not for adjudicating a person.

Established “Embryo selection can buy 8.5 IQ points.” The marketed figure is a spread, not a gain: the average maximum-minus-minimum difference across three embryos on a predictor, which is not the expected benefit of choosing the top one and is not a phenotypic measurement. Established The peer-reviewed estimate of gain from selecting the top of ten embryos is 0.53 years of education for European ancestry, with a 95% prediction interval running from −3.2 to +4.2 years. That interval spans zero. Frontier An independent critic's upper bound is about 6 IQ points and about ten months of extra schooling. For any individual family this is close to a coin flip.

Established “5-HTTLPR, COMT, BDNF and DRD4 explain personality and cognition.” They do not. Eighteen such genes were tested against depression in 621,214 people, producing one marginal COMT effect and no gene-by-environment interactions, and the authors concluded the prior positives were false positives. Established The lesson generalises past this subject: thousands of papers with plausible biology and consistent narratives were produced by small samples, flexible analysis and publication bias.

Established “GWAS found the schizophrenia gene.” It found 287 loci and prioritised 120 genes; a separate exome study found ten genes with rare-variant odds ratios of 3 to 50. Neither is a gene for schizophrenia, and the interesting result is that the two sets converge on the same synaptic biology rather than competing. Established “Brain imaging phenotypes are too noisy to be genetic.” 1,578 of 3,144 UK Biobank imaging-derived phenotypes show significant SNP heritability. Established “Most developmental disorders now have a genetic diagnosis.” 285 genes are established and the same study estimates more than 1,000 remain undescribed.

Established “Neurogenetic diseases are untreatable.” Nusinersen, tofersen, trofinetide, otoferlin gene therapy and AMT-130 are approved or in pivotal trials. Established “Gene therapy is a one-time cure.” Skysona produced haematologic malignancy in seven of 67 patients at onsets of 14 to 92 months; the best-known Duchenne product missed its phase III motor endpoint and was followed by three acute-liver-failure deaths; a haemophilia product's expression declines after 12 to 18 months. Established “CRISPR will fix brain disease.” No approved or late-stage in vivo editing programme reaches the CNS; the successful in vivo record is the liver, and every CNS programme in this brief is injected into a compartment or a structure surgically.

Speculative The strongest heterodox claim, stated properly. Psychiatric genetics may have been looking at the wrong level of description all along: it predicts diagnostic categories that were defined by committee for clinical convenience, and if those categories are not natural kinds then a GWAS of one is a GWAS of a mixture. Speculative On this reading the tiny per-locus effects are an artefact of a badly specified phenotype rather than a fact about the brain, and the right units would be circuit-level or computational parameters — something like gain, precision or learning rate — on which effects might be larger and fewer. Established What the mainstream says back is specific and strong: SCHEMA and the flagship GWAS converge on the same glutamatergic synaptic biology from opposite ends of the frequency spectrum, which is not what a mixture artefact produces, and 120 prioritised genes concentrated in excitatory and inhibitory neurons is a mechanistic result about brains and not about committees. Handwave No one has produced a circuit-level phenotype measurable at biobank scale, which is what the heterodox position would need, and until someone does it remains an argument rather than a programme.

Speculative And the honest terminal position on headroom. With SNP heritability of 0.19, thousands of loci of tiny effect, within-family shrinkage of 22 to 47%, and no identified common variant with an effect greater than about a tenth of a standard deviation on general cognitive ability in 269,867 people, the architecture looks like a plateau held by stabilising selection rather than a resource waiting to be mined. Speculative The rare-variant record points the same way: large effects on cognition-adjacent phenotypes exist, with odds ratios up to 50, and they are all deleterious. Speculative The counterargument is fair and should be recorded — GWAS was never meant to predict, it was meant to nominate biology, and by that standard it worked. Frontier Both statements are true, and declaring the tie is the correct terminal position rather than a failure to conclude.