Genetic Engineering

The framing under test is the one everybody carries: that CRISPR turned genome editing into a general-purpose technology. The measured record says something more specific and more useful. Programmable targeting is close to solved — a sequence can be found and altered in a human cell with reliability that would have been science fiction in 2010. What is not solved is getting the editing machinery into the cells that need it, keeping the edit from damaging the chromosome around it, manufacturing the product at more than boutique volume, and paying for it. Every clinical failure in this brief is a delivery, durability, manufacturing or reimbursement failure. Almost none of them is a targeting failure.

Established Two numbers frame the subject. Casgevy, the first approved CRISPR therapy anywhere, was cleared on 8 December 2023; as of 30 September 2025 its sponsor reported 39 patients infused worldwide against a stated eligible severe US population of about 16,000. Established In the same period, in-body base editing produced dose-dependent, apparently permanent reductions in a human physiological set-point, and in-body Cas9 produced a grade 4 liver injury, two FDA clinical holds and a death. Frontier Both facts are about delivery, not about whether the enzyme cut the right base pair.

Speculative Beyond the clinic sit the genuinely exotic programmes: writing whole chromosomes rather than editing them, synthetic accessory chromosomes that add function without perturbing anything, recoded organisms, and heritable modification of the germ line. Handwave This brief takes those seriously, gives each its strongest form, and then says what the mainstream says back. It also carries the argument that the field's own enthusiasts most dislike — that the delivery problem may be structurally permanent rather than a queue of engineering tasks — because it is the single claim that would most change the subject if it were true.

Established Genome editing did not begin with CRISPR and its clinical history is older than the public one. Zinc-finger nucleases and TALENs reached patients before Cas9 did: a TALEN-edited CD19 CAR-T product was used in acute lymphoblastic leukaemia in 2015, and a zinc-finger nuclease was administered inside the body in 2019. Established What CRISPR changed was the cost and speed of retargeting — a guide RNA instead of a protein redesign — not the fundamental operation. Established It also inherited a hard constraint that is easy to forget: Cas9 can only cut adjacent to a protospacer-adjacent motif, so the addressable positions in a genome are set by where those short motifs happen to sit, and engineered variants exist precisely because the wild-type constraint bites.

Established The first approved CRISPR medicine does not correct the disease-causing mutation. Casgevy treats sickle cell disease by disrupting the erythroid-specific enhancer of BCL11A, which de-represses fetal haemoglobin and lets it substitute for the defective adult protein. Established The sickle mutation itself is left in place. Established That is not a criticism — it is a design choice that avoids a repair template entirely — but it is a precise indicator of where the field's competence actually sits: knocking a regulatory element out is far easier than writing a corrected sequence in.

Established The successors to Cas9 nuclease editing are real, and one of them has now been used in a person. Prime editing, described by Anzalone and colleagues in 2019, fuses a Cas9 nickase to an engineered reverse transcriptase and programmes it with an extended guide; the original paper reported more than 175 edits in human cells covering all twelve point-mutation classes, at PE3 efficiencies of roughly 20–50% in HEK293T cells with 1–10% indels. Established HEK293T is the most permissive cell line in routine use, so those efficiencies are a ceiling, not a typical value. Frontier In December 2025 the first clinical prime-editing results appeared: PM359, for p47phox-deficient chronic granulomatous disease, in two patients, reporting 69% and 83% DHR-positive neutrophils by day 30 against a projected 20% threshold for clinical benefit, with no adverse events attributed to the editor and all toxicity consistent with the busulfan conditioning. Frontier Two patients, one indication, one sponsor, no independent replication.

Frontier In-body base editing works in humans, with a dose–response curve. VERVE-102 pairs an adenine base editor with a guide RNA in a GalNAc-targeted lipid nanoparticle aimed at PCSK9 in hepatocytes. In the Heart-2 phase 1b, data cut 13 March 2025, mean LDL-cholesterol fell 21% / 41% / 53% across 0.3, 0.45 and 0.6 mg/kg cohorts of four, six and four patients, with a maximum individual reduction of 69% and PCSK9 protein down 46% / 53% / 60%; no treatment-related serious adverse events and one grade 2 infusion reaction. Established These are the sponsor's own figures. Frontier The important structural fact is not the effect size but the target organ: the liver filters blood by anatomical design, which is why lipid nanoparticles reach it without being told to.

Established The safety record of in-body CRISPR is no longer clean. Intellia's nex-z, an LNP-delivered Cas9 knockout of transthyretin, is the largest in-body CRISPR programme ever run: more than 450 of about 650 enrolled MAGNITUDE patients dosed, plus 47 in MAGNITUDE-2. A patient dosed on 30 September 2025 was hospitalised on 24 October with grade 4 transaminase elevations and raised bilirubin; the company paused both phase 3 trials on 27 October, the FDA placed clinical holds on 29 October, and the patient died on 5 November 2025. Established Fewer than 1% of MAGNITUDE patients had grade 4 transaminase elevations and none in MAGNITUDE-2, so this is a rare event found only because the trial was large enough to find it. Frontier Read carefully, it localises the risk of systemic LNP-Cas9 to hepatotoxicity at scale — in the one tissue the field has actually solved.

Established The precision claim is wrong in a direction almost nobody expects: the bigger measured problem is on-target, not off-target. Kosicki, Tomberg and Bradley showed in 2018 that ordinary Cas9 double-strand breaks “frequently resolved into deletions extending over many kilobases,” plus complex rearrangements, in mouse embryonic stem cells, mouse haematopoietic progenitors and human RPE1 cells — and that standard genotyping had missed them because it limited exploration to the immediate vicinity of the cut. Established In human embryos specifically, Alanis-Lobato and colleagues edited POU5F1 in 54 embryos and found segmental loss or gain at or adjacent to the locus in about 16% of targeted samples, and open-ended loss-of-heterozygosity patterns in 61.9% against 10% in Cas9-only controls. Established Haapaniemi and colleagues added the uncomfortable corollary: Cas9 editing triggers a p53-mediated DNA-damage response, inhibiting p53 raises homologous-recombination efficiency, and therefore the cells that edit best are enriched for a weakened tumour suppressor.

Established Approved gene therapy is expensive, and several products are not one-time cures. Verified US list prices as of 30 November 2025: Lenmeldy $4.25M, Hemgenix $3.5M, Elevidys $3.2M, Skysona $3M, Zynteglo $2.8M, Zolgensma $2.125M. Frontier Casgevy is not on that list and its widely quoted US price could not be confirmed against a primary source for this brief, so no figure is printed here. Established Skysona has caused six myelodysplastic syndromes and one acute myeloid leukaemia in 67 treated patients — five clones carrying a vector insertion in MECOM and one in PRDM16, with onset spread over 14 to 92 months, and six of the seven cases in the arm conditioned with busulfan plus fludarabine. Established Elevidys was approved in June 2023 on gene expression rather than clinical benefit, then missed its primary motor endpoint in the phase III EMBARK trial (n=125), was linked to three deaths from acute liver failure in July 2025 with an FDA clinical hold imposed on 18 July and lifted on 28 July, and was recommended against approval by the EMA. Established Roctavian's factor VIII expression declines after roughly 12 to 18 months, a durability gap that prompted a complete response letter in 2020.

Established And the approved products are barely used. As of 30 September 2025 Casgevy had roughly 300 patients referred, about 165 who had completed a first cell collection, and 39 infused, across 25 authorised treatment centres that had each initiated more than five patients. Established Its sponsor expected total 2025 revenue above $100 million, which for a therapy at this price point is a statement about volume rather than about commercial success. Frontier The gap between 300 referrals and 39 infusions is the shape of the whole subject: patients are found, consented and collected from, and then stall at a step that has nothing to do with genome editing — conditioning capacity, transplant beds, and payers. Speculative Any account of this field that treats approval as the finishing line is measuring the wrong event.

Established The germline record is one man, three children, and a set of edits that were not what he claimed. He Jiankui recruited seven couples; the twins Lulu and Nana were born in October 2018 and a third child in about August 2019. Established Neither twin carries the natural CCR5-delta32 deletion. Lulu is heterozygous, carrying a 15-base-pair in-frame deletion on one chromosome 3 with a wild-type copy on the other, and is mosaic; a 15-base in-frame deletion need not abolish the protein at all, so she retains functional CCR5 and received no plausible protection. Established Nana is homozygous for novel alleles — a 4-base-pair deletion and a single-base insertion — not the 32-base deletion. Established Fyodor Urnov, reading the unpublished manuscript, called the claim to have reproduced the prevalent CCR5 variant “a blatant misrepresentation of the actual data,” and pointed out that the off-target analysis was performed on biopsy cells that were discarded, not on the cells that became the children: “it is technically impossible to determine whether an embryo did not show off-target mutations without destroying that embryo.” Established He was sentenced on 30 December 2019 to three years' imprisonment and a fine of three million yuan, with two colleagues also convicted and permanently barred from assisted reproduction, and was released in April 2022.

Established The law is unusually unambiguous for a frontier technology. Seventy countries explicitly prohibit heritable genome editing for reproduction and five more prohibit it with exceptions — Belgium, Colombia, Italy, Panama and the United Arab Emirates — while no country permits it. Established In-vitro research is treated differently: 19 countries explicitly prohibit it, 4 prohibit with exceptions, and 11 explicitly permit it, including China, India, Japan, the United Kingdom and the United States. Established The 2019 Nature comment calling for a moratorium carried 18 signatories from seven countries, including Feng Zhang and Emmanuelle Charpentier, and notably did not specify a duration — it asked for a governance framework, not a fixed pause.

Established There is a third class of editor that changes no bases at all, and on the same target in the same organ it now competes directly with the base editors. Epigenome editing replaces the nuclease with a catalytically dead Cas9 carrying transcriptional effector domains — KRAB for repression, DNMT3A and DNMT3L for de novo CpG methylation — so the sequence is left untouched and only the chromatin state is rewritten. Established Amabile and colleagues demonstrated in 2016 that a transiently supplied combination of those domains could silence an endogenous gene heritably, and Nuñez and colleagues generalised it in 2021: their CRISPRoff construct, delivered once and then degraded, produced silencing that survived roughly 450 cell divisions and the differentiation of induced pluripotent stem cells into neurons, and was reversible on demand with a TET1-based counterpart. Established The operative word is transient: the effector arrives as mRNA, works over days and is cleared, so the window in which it can act anywhere it should not is measured in days rather than in the weeks or years of vector-driven nuclease expression.

Frontier The result that makes this a clinical proposition rather than a cell-culture trick is a single dose of lipid-nanoparticle mRNA silencing Pcsk9 in mouse liver for close to a year. Cappelluti and colleagues fused the three effector domains into one hit-and-run repressor, delivered it by lipid nanoparticle, and reported near-complete silencing of the target in hepatocytes from a single administration, maintained across the observation period, transmitted through cell division, and erasable afterwards by a targeted demethylator. Established The target and the tissue are the ones VERVE-102 uses, which is what makes the comparison legible: a repressor and a base editor are being run against the identical gene, in the identical organ, by the identical delivery route. Frontier The only thing that differs is what is left behind — a methylation pattern rather than an A-to-G substitution — and that difference cuts in both directions at once.

Frontier Reversibility is the safety case and the regulatory problem in the same sentence. A DNA edit is permanent by construction, so an error is permanent; a methylation mark can be taken off again by a second editor, so an error is in principle correctable. Speculative That argument holds only if the second editor can be delivered to the same cells as the first, which is the unsolved problem of this whole brief restated — inside the liver it is plausible, and outside it the correction route is blocked in exactly the way the original route was. Frontier The mirror image is durability. A silenced locus can drift back, mark maintenance depends on a maintenance methyltransferase the therapy does not control, and there is no agreed regulatory endpoint for “durable” short of measuring it. Speculative For a one-dose product that means years of surrogate follow-up before anyone can say whether what was sold as a cure is a very long-acting drug.

Established Off-target regulation is a different failure mode from off-target cutting, and it is far less well instrumented. A dead Cas9 that binds briefly at a mismatched site leaves no permanent lesion, which is the optimistic reading; the pessimistic one is that the DNMT3A domain carries documented guide-independent activity, and Galonska and colleagues showed that dCas9-methyltransferase constructs deposit methylation across the genome whether or not a guide directs them there. Frontier That failure does not appear in an off-target sequencing panel, because no sequence has changed; finding it requires whole-genome bisulphite or enzymatic methylation sequencing against a matched baseline, and the assay burden is the reason the field’s off-target claims are weaker than they look. Frontier Multiplexing compounds it: the guide-independent component scales with effector dose rather than with the number of guides, so editing six loci at once is not six times one experiment.

Frontier Inheritance here means mitotic memory, not transgenerational transmission, and conflating the two is the most common category error made about this technology. The marks propagate through DNA replication because a maintenance methyltransferase copies them onto the daughter strand; they do not survive the two rounds of genome-wide reprogramming that separate a somatic cell from a gamete and an early embryo. Established Imprinted loci and a small set of repeat elements escape that erasure, and they are the entire empirical basis for mammalian transgenerational epigenetic inheritance, which is a narrower foundation than the popular literature implies. Speculative The practical consequence is favourable and under-argued: a somatically applied epigenome editor is not a germline intervention on any reasonable reading, which places it outside the prohibition regime described above and inside ordinary somatic gene-therapy regulation.

Frontier The clinical record is early, small, and concentrated in precisely the two places the delivery argument predicts. The programmes reported as reaching first-in-human dosing during 2025 target the liver — epigenetic silencing of hepatitis B viral DNA, where the therapeutic object is covalently closed circular DNA that no nuclease can safely cut — and locally injected skeletal muscle, where a repressor aimed at DUX4 addresses facioscapulohumeral muscular dystrophy, a disease caused by derepression rather than by a coding mutation. Frontier Neither has published peer-reviewed efficacy data, and first-in-human dosing is a statement about a regulatory filing, not about a result. Established The indication logic is nonetheless the strongest thing the epigenome-editing case has: where the pathology is unwanted expression, a nuclease offers nothing that a repressor does not offer without a break.

Frontier The most informative live result is a central-nervous-system gene therapy with a caveat that is easy to lose. uniQure's AMT-130 delivers an AAV-borne miRNA against huntingtin by MRI-guided convection-enhanced stereotactic injection into the striatum. Pooled US (n=26, 16 treated and 10 sham) and European (n=13) phase I/II data reported 75% slowing of cUHDRS progression at 36 months (mean change -0.38 versus -1.52, p = 0.003), 60% slowing on Total Functional Capacity (p = 0.033), cerebrospinal-fluid neurofilament light down 8.2% from baseline, and no new drug-related serious adverse events since December 2022, with a BLA planned for Q1 2026. Frontier The comparison is against a propensity-matched external control cohort, not the trial's own sham arm. That is the single most important thing about the result and it is routinely dropped in summary. Established It is also, notably, delivered by a neurosurgeon through a burr hole — the vector did not find the striatum, a stereotactic frame did.

Frontier Bespoke editing for one patient is now a six-month process. KJ Muldoon, born 1 August 2024 with CPS1 deficiency, received a custom adenine base editor delivered as mRNA in a lipid nanoparticle, first dosed on 25 February 2025, with three infusions over seven weeks. Design to manufacture took about six months against an estimated eighteen. Frontier The regulatory precedent matters as much as the chemistry: the FDA permitted an escalating, patient-specific dosing protocol for a therapeutic that will never be given to anyone else. Speculative If that becomes routine rather than exceptional, the economics of ultra-rare disease change more than any nuclease improvement would change them.

Frontier Removing the manufacturing step is being attempted directly. Kelonia's KLN-1010 generates CAR-T cells inside the body: in four patients with relapsed or refractory myeloma, all four reached partial response by month 1 and all four were minimal-residual-disease negative at 10^-5 to 10^-6, with no lymphodepletion, no apheresis and no ex vivo manufacturing, and a consent-to-infusion interval of 13 to 18 days. Frontier Four patients, first dosed 19 August 2025. Speculative If in-body generation holds up, it removes exactly the constraint that limits Casgevy — the cell-collection and transplant-centre pipeline — without touching the editing chemistry at all.

Frontier Sensory restoration is the cleanest non-liver delivery success, and it is also an injection. The DB-OTO otoferlin gene therapy treated 12 children aged ten months to sixteen years, with 14 of 15 treated ears improved, nine reaching thresholds of 70 dBHL or better at week 24 and three achieving normal hearing sensitivity, responses appearing within weeks and stable or improving out to 72 weeks. Established The cochlea is a small, enclosed, immune-privileged fluid compartment that a surgeon can inject directly. Frontier Every non-liver clinical success in this brief shares that property, which is the observation the whole delivery argument turns on.

Frontier Gene drives have a biology that works and a field pathway that is closed. No gene-drive organism has ever been released into the wild anywhere. What was released in Burkina Faso on 11 August 2025 was a non-drive, self-limiting male-bias construct, after favourable biosafety and environmental approvals in July and community agreement. National authorities requested suspension on 18 August and the ministry issued formal termination on 22 August — eleven days after a single release of the deliberately conservative version. Established The canonical demonstration that a CRISPR drive can crash a caged Anopheles gambiae population by targeting doublesex was published in 2018 and is not disputed. Frontier So the binding constraint on gene drives is now governance, running ahead of the biology rather than behind it.

Frontier Bottom-up genome writing has crossed the megabase line, in yeast. The Sc2.0 consortium completed all sixteen synthetic yeast chromosomes, from the first synthetic arm in 2011 through nine further chromosomes published simultaneously on 8 November 2023, with a strain carrying more than half of its genetic information in synthetic form and more than 250 researchers involved. Frontier Separately, a 1.024 Mb de novo synthetic accessory chromosome carrying 542 exogenous genes was built and moved into six phylogenetically diverse yeasts by CRISPR-mediated programmed haploidisation, conferring thermotolerance, glutarate utilisation and novel metabolite production. Speculative That is the first credible demonstration that you can add a designed chromosome rather than edit an existing one — in an organism that mates, using a delivery route with no mammalian analogue.

Frontier The most consequential recent advance in writing DNA is not an editor at all: it is a recombinase bolted onto a prime editor. Serine integrases such as Bxb1 perform conservative site-specific recombination between a short attachment site in the genome and a matching site on a circular donor, joining the two without ever leaving a free double-strand break for the cell to repair. The obstacle was always that mammalian genomes contain no attachment site. Established Prime editing removes the obstacle by writing one in: an attachment sequence of roughly 38–50 bp is installed by the prime editor, and the integrase then loads multi-kilobase cargo onto it. Frontier In the PASSIGE work, directed evolution of the recombinase itself lifted integration of whole genes in human cells from low single-digit percentages to roughly 30% at favourable loci — more than an order of magnitude, obtained by evolving the protein rather than by improving the guide, which is where nearly all editing effort has gone since 2019.

Established The comparison that matters is against the two older ways of inserting DNA, and both carry documented genotoxicity that site-specific recombination is designed to avoid. Homology-directed repair needs a double-strand break and a template, works poorly in cells that are not dividing, and inherits every on-target rearrangement problem set out in section 2. Established DNA transposases — piggyBac and Sleeping Beauty, both already used to manufacture CAR-T products — cut and paste at large numbers of pseudo-sites, and a piggyBac-generated CD19 CAR-T trial produced T-cell lymphomas in two of ten treated patients, in products carrying high transgene copy numbers. Frontier Integrase insertion is arbitrary only in the sense that a human chose where the attachment site went; after that the landing is one place, which converts an insertional-mutagenesis problem into a site-selection problem that can be settled before dosing rather than discovered years after it.

Frontier The all-in-one version, in which a CRISPR effector carries the transposase with it, works beautifully in bacteria and has only just begun to work at all in human cells. CRISPR-associated transposons were described in 2019 in two independent systems and reach high RNA-guided insertion efficiencies in Escherichia coli with no host repair involvement whatsoever; moved into human cells they were, for several years, effectively inert. Frontier Directed evolution of the type I-F system reported during 2025 recovered useful efficiency in human cells for the first time, and a second route — bridge RNAs from IS110 insertion sequences, which specify target and donor in one structured RNA — sits at the same stage: a clean mechanism, a convincing bacterial demonstration, and a mammalian question mark. Speculative If either matures, the two-component prime-editor-plus-integrase architecture is an interim design rather than a destination, and the field will have spent a decade optimising a scaffold it then discards.

Established Every one of these schemes then hits the same wall, and it is a delivery wall of a new kind. Inserting a ten-kilobase gene means delivering a ten-kilobase donor, and the vector that reaches hepatocytes reliably — the lipid nanoparticle — is an mRNA vehicle; double-stranded DNA loose in the cytosol is an agonist of the cGAS–STING pathway and is precisely what innate immunity evolved to notice. Frontier Every workaround is a compromise. An adeno-associated virus as the donor reimposes the packaging ceiling and pre-existing capsid immunity; doggybone or minicircle DNA moves the problem into manufacturing; engineered virus-like particles carry protein and RNA well and large DNA badly. Speculative So the honest summary of gene writing as of 2026 is that the writing chemistry has travelled considerably further than anything capable of carrying it into a patient, which is the same sentence this brief has already had to write about base and prime editing.

Frontier Manufacturability is the quiet difference between a paper and a product. A prime-editor-plus-integrase insertion is a four-component drug — prime-editor mRNA, extended guide, integrase mRNA, circular donor — each with its own release specification, its own stability profile and its own contribution to a combination-potency matrix, against a base editor’s two. Established Regulators treat each component as a separate item for chemistry, manufacturing and controls purposes, so the component count sets the size of the filing and the cost of every subsequent change to any part of it. Speculative That arithmetic, rather than integration efficiency, is the most likely reason the first approved insertion product will be an ex vivo cell therapy, where the four components meet in a bioreactor and only the finished cells are given to a person.

Frontier Mosaicism is the metric that survives all of this, and it is the one most often left out. A population-average integration rate of 30% means most cells received nothing, and whether that matters is a property of the disease rather than of the technology: a secreted factor made by a minority of hepatocytes can be therapeutic, a structural protein required in every myofibre cannot, and a dominant-negative allele is entirely untouched by adding a correct copy elsewhere. Established The chronic granulomatous disease result earlier in this brief is the instructive case, because it names the second number as well as the first — a projected 20% corrected-neutrophil threshold for clinical benefit. Frontier Almost no insertion paper states what fraction of cells its target disease would actually require, which is the difference between a capability and an indication.

Established The chain to genome engineering as a general-purpose technology has eight links, and the one the public treats as the whole problem is the least binding. Ordered: (1) editing chemistry without double-strand breaks or bystander damage; (2) delivery beyond the liver; (3) non-genotoxic conditioning for ex vivo products; (4) manufacturing throughput; (5) a long-horizon safety readout; (6) reimbursement at $2–4M a dose; and, for germline only, (7) epigenomic correction in laboratory-derived gametes and (8) a jurisdiction that permits reproductive use.

Frontier Link 1 is closest to solved. Base and prime editors already avoid the double-strand break; the target specification is efficiencies above 50% with indels below 0.5% in primary human cells rather than in HEK293T, where the published 20–50% with 1–10% indels was measured. Established Link 2 is binding for somatic therapy. The whole in-body clinical record is hepatic; nothing has achieved above 30% editing in muscle, central nervous system or haematopoietic stem cells in vivo. Established Link 3 is where the harm is. Six of Skysona's seven malignancies came from one conditioning arm, and Casgevy's throughput is limited by busulfan conditioning and transplant capacity, not by the nuclease. Established Link 4 is industrial. Thirty-nine infusions in 22 months has to become thousands a year, and in-body generation is the only visible route that does not simply build more transplant centres.

Established Link 5 is time plus institutions. Skysona's leukaemias appeared between 14 and 92 months, so nothing shorter than a decade of clonal tracking with long-read sequencing across a thousand-plus patient registry is a safety readout at all. Frontier Link 6 is a market failure, not a science one. Approved products go unused at $2–4M a dose because no reimbursement instrument matches a one-time payment against a lifetime of avoided cost. Speculative Links 7 and 8 are germline-only and are of different kinds. Epigenomic correction in gametes derived in vitro has not been achieved in human cells at all, and no permissive jurisdiction exists anywhere. Established The binding link is 2 for somatic and 8 for germline. Nothing in the somatic chain is waiting on a better nuclease. Frontier That ordering is the most useful thing this brief has to offer, because it predicts where progress will and will not come from: a laboratory announcing a cleaner editor is announcing an improvement to the least binding link in the chain, while a laboratory announcing 30% editing in skeletal muscle without surgery would have moved the field more than every editor published since 2019 combined.

Frontier Epigenome editing does not add a ninth link to that chain; it reshuffles two of them and leaves the binding one untouched. It improves link 1 by removing the double-strand break and the base substitution alike, and it changes the character of link 5, because the long-horizon readout becomes duration of effect rather than clonal safety — a cheaper question to ask and a slower one to answer. Established It does nothing whatever for link 2: every clinical epigenetic silencer so far is hepatic or locally injected, for the same anatomical reasons every other programme in this brief is. Speculative A reader who takes a durable mouse-liver result as a delivery advance has misread it. What it is, is a chemistry advance in the one tissue where chemistry was never the constraint.

Established Payload size is the constraint that silently decides which diseases are addressable at all, and it is a packaging number rather than a biological one. An adeno-associated virus carries about 4.7 kb; the dystrophin coding sequence is roughly 11 kb, which is why every Duchenne gene therapy in the clinic delivers a truncated micro-dystrophin rather than the protein the patient is missing. Established Base editing changes one base; prime editing reliably writes tens of bases; neither can supply a gene that is absent. Frontier So for the large share of genetic disease caused by hundreds of different mutations scattered through one gene, the relevant technology is whole-cassette insertion, and insertion has a different bottleneck list from the eight-link chain above — donor manufacture and donor delivery sit where editing chemistry sits in the somatic chain, and neither is a nuclease problem either.

Established This subject consumes results it does not produce. Lipid chemistry and capsid engineering belong to materials science and virology; the field's tropism problem is being solved, where it is being solved at all, by people optimising ionisable lipids and evolving AAV capsid libraries, not by people improving nucleases. Established Pre-existing neutralising immunity to AAV serotypes excludes a substantial fraction of candidate patients before any question about the transgene arises, and that is an immunology problem this field can state but not solve. Frontier Redosing an AAV product is generally impossible for the same reason, which is why durability failures such as declining factor VIII expression are not recoverable by giving the therapy again.

Established Haematology owns the conditioning problem. Busulfan is the reason ex vivo gene therapy needs a transplant centre, and a non-genotoxic conditioning regimen with equivalent engraftment would unlock more clinical benefit than any editor improvement now in development. Frontier Long-read sequencing and clonal-tracking analytics own the safety readout; the MECOM and PRDM16 insertions in the Skysona cases were found by integration-site analysis, and the kilobase deletions that ordinary genotyping missed are only visible to assays designed to look far from the cut site.

Frontier Two dependencies sit outside biology's usual suppliers. High-fidelity synthesis of very long DNA is a chemistry and instrumentation problem that the genome-writing programmes cannot solve themselves — and centromere biology, specifically which sequences are necessary and sufficient to specify a functional human centromere de novo, is a basic-science question that the human-artificial-chromosome literature has named as its own limiting unknown for three decades. Speculative Neither is being funded at a scale proportionate to what depends on it. Frontier A third dependency is actuarial rather than scientific: nobody can price a one-time therapy without a durability estimate, and durability estimates require follow-up longer than any commercial trial, so the reimbursement question is downstream of a registry that does not exist. See Synthetic Biology for the construction side and Precision Medicine for the patient-stratification side.

Established The decisive somatic experiment is a delivery measurement, and it has a number attached. Demonstrate an LNP or capsid achieving above 30% editing in vivo in muscle, central nervous system or haematopoietic stem cells, without stereotactic placement or surgical injection, at a dose below the hepatotoxicity threshold the nex-z programme has now established. Frontier Nothing in the published clinical record does this. Until something does, every non-hepatic gene therapy is a surgical procedure that happens to involve nucleic acid.

Frontier Second: conditioning. A randomised comparison of busulfan-free conditioning against busulfan in an autologous gene-therapy setting, powered on engraftment rather than on a disease endpoint. Speculative A positive result addresses the Skysona MECOM signal and the Casgevy throughput ceiling simultaneously, which no other single experiment in this brief does. Established Third: durability. Ten-year clonal tracking with long-read sequencing across at least a thousand treated patients, structured as a registry rather than as a set of sponsor commitments, because the Skysona latency distribution runs to 92 months and every sponsor obligation is shorter than that.

Established Fourth, and decisive for germline: repeat the human-embryo experiment with an editor that makes no double-strand break. The 61.9% loss-of-heterozygosity result was obtained with Cas9 nuclease. Base and prime editors do not cut both strands, so the honest question is whether the chromosomal damage is a property of editing embryos or a property of cutting them — and that has not been tested at scale in human embryos. Speculative A clean negative would remove the strongest technical objection to heritable editing and leave only the legal and ethical ones; a positive would close the route on evidence rather than on assertion. Handwave Fifth, far out: demonstrate de novo specification of a functional human centromere on a synthetic construct, which is the gating experiment for every artificial-chromosome proposal in section 14 and which has been the named blocker for three decades without a serious funded attempt. Speculative Sixth, and cheapest of all: publish a cost-per-base series for long high-fidelity DNA synthesis with error rates attached. It is a measurement, not an experiment, and its absence is why nobody can say whether genome writing is five years or fifty years from being affordable.

Established The manufacturing chain for approved gene editing is a transplant chain. Casgevy requires autologous cell collection, ex vivo editing, busulfan conditioning and reinfusion at an authorised centre; as of late 2025 only 25 centres had initiated more than five patients each, about 300 patients had been referred, roughly 165 had completed a first cell collection, and 39 had been infused. Established The attrition between those numbers is the engineering problem, and none of it is molecular biology.

Frontier Genome writing has a different and starker constraint. Andrew Hessel, of the Human Genome Project-write and now the Synthetic Human Genome project, states the requirement directly: “We need much better DNA synthesis — pieces that are hundreds of kilobases long, error-free, and affordable.” Established That is a statement about instrumentation, not about design. Frontier The scale gap is arithmetic: human chromosome 21 is about 47 Mb with 200–300 genes and chromosome 1 is about 249 Mb with more than 2,000, against a completed yeast genome of roughly 12 Mb that took from 2011 to 2023 and more than 250 people. Frontier A published, verifiable cost-per-base curve for long high-fidelity fragments could not be sourced for this brief, so none is quoted — which is itself informative, because a field whose central industrial constraint has no public price series is not yet an industry.

Speculative The engineering ask is therefore unusually concrete. Not a better editor: a synthesis platform producing error-free fragments in the hundreds of kilobases at a price that makes a chromosome-scale build a project rather than a career, plus an assembly and delivery route for megabase episomes into mammalian cells. Handwave The yeast route used mating and programmed haploidisation. Mammalian cells do not mate, so the entire delivery step for a designed chromosome would have to be invented from nothing. Frontier Meanwhile the near-term engineering that would help most patients is unglamorous: more transplant-capable centres, a conditioning regimen that does not require one, and closed-system automated cell processing.

Established The nearest adjacent technology is the one that competes with editing entirely. Antisense oligonucleotides, small molecules and RNA-targeting agents treat genetic disease without touching the genome, and they are frequently the faster route. Frontier In-body CAR-T is adjacent in a different way: it applies the delivery insight to cell therapy, removing apheresis, lymphodepletion and manufacturing in one move, and if it works it changes the economics of ex vivo genome editing more than any change to the editor would.

Established Embryo selection is the genuinely competing germline technology, and it is already commercial. It requires no editing at all, produces no chromosomal damage, and is legal in most of the jurisdictions that ban heritable editing outright. Established Its measured power is modest: selecting the top-scoring of ten embryos on the best published education polygenic index yields an expected gain of about 0.53 years of education with a 95% prediction interval running from -3.2 to +4.2 years, about 2.5 cm for height, and roughly 0.23 years in African-ancestry populations where the scores transfer badly. Frontier A prediction interval spanning zero means that for any individual family the procedure is close to a coin flip.

Speculative Further out, three adjacencies matter. Synthetic Biology and Designer Organisms own the construction methods that genome writing depends on; De-Extinction Technologies is the same toolkit pointed at a different target and inherits every delivery constraint in this brief; and Artificial Wombs plus in-vitro gametogenesis are what would make iterated germline schemes physically possible at all. Speculative A fourth adjacency is defensive rather than enabling: the same synthesis and assembly capability that would let a consortium write a human chromosome is the capability that makes synthesis screening a biosecurity requirement rather than a courtesy, and the two roadmaps are the same roadmap read in opposite directions. Frontier See also Biological Enhancement and Neurogenetics for the trait-side arguments this brief deliberately does not make.

Established Heritable editing is the rare frontier technology with a settled global answer. Seventy countries prohibit reproductive use outright, five prohibit with exceptions, none permits it — and the prohibition survived the one attempt to defy it, which ended in criminal convictions in the country with the largest permissive research regime. Established The 2019 moratorium call asked for an international governance framework and set no end date; the field's periodic international summits have since become the venue where that framework is argued rather than a body that can bind anyone.

Established The gene-drive case shows what consent actually requires. Target Malaria obtained national biosafety and environmental approvals in July 2025 and community agreement in the release village, released a non-drive construct on 11 August, and was terminated by ministerial decision on 22 August. Frontier Nothing technical intervened. The lesson institutions should draw is that for an organism designed to spread, village-level consent and national regulatory approval are not the same instrument as national political consent, and the third one is the one that binds.

Established Two institutional facts about the approved products deserve more weight than they get. 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 Separately, the FDA has now approved a patient-specific editing protocol for a single infant, which is a new regulatory category rather than a new science. Speculative The institution that does not yet exist is a reimbursement instrument for a one-time $2–4M therapy; outcome-based contracting is discussed everywhere and operates nowhere at national scale, and it is the reason approved therapies sit unused. Frontier A second missing institution is a durable, publicly funded clonal-tracking registry with a mandate measured in decades rather than in trial periods; the Skysona malignancies were detected because integration-site analysis happened to be performed, not because any body was obliged to look. Speculative A third is a funding route for germline-adjacent research that does not depend on private philanthropy: the in-vitro gametogenesis work now being advanced toward human application is largely privately backed, which means the research agenda for the most consequential branch of this technology is being set outside any public deliberation. See Precision Medicine.

Established The germline case has a factual core that most ethical commentary skips. The children were not given the protective variant their creator claimed; one is mosaic and retains functional CCR5; the off-target analysis was performed on cells that were discarded rather than on the cells that became the children, which Urnov noted is the only analysis that is physically possible. Established So the case is not a trade-off between benefit and risk. It is a case of risk with no benefit at all, imposed on people who cannot consent and who will be studied for the rest of their lives.

Frontier A second ethical problem is buried in the biology. If editing selects for cells with a blunted p53 response, then the population of successfully edited cells is systematically enriched for weakened tumour suppression — a harm that is invisible at the moment of treatment and appears, if the Skysona latencies generalise, somewhere between one and eight years later. Established Informed consent for a therapy whose principal risk manifests after most trial follow-up has ended is a genuinely unsolved problem, not a paperwork one.

Established Distribution is the third. A $4.25M therapy that 39 people have received in 22 months, against 16,000 eligible in one country, is not a technology with an access problem at the margin; it is a technology whose access pattern is its dominant clinical fact. Speculative The standard reply — that prices fall — has to be argued rather than assumed here, because the cost driver is a transplant admission and a conditioning regimen, not a molecule whose synthesis gets cheaper. Frontier And for gene drives the consent unit is not a patient or even a village but an ecosystem shared across borders, for which no consent mechanism exists in any body of law. Speculative The uncomfortable symmetry is that the strongest ethical objection to gene drives — that nobody can consent on behalf of an ecosystem — is also the reason the intervention with the largest plausible reduction in human mortality of anything in this brief has never been tried.

Speculative The civilizational case for this technology is real and it is almost entirely about infectious disease, not enhancement. A gene drive that suppresses Anopheles gambiae is the only proposed intervention that could end malaria transmission rather than manage it, and the caged demonstration exists. Established The pathway to using it was closed by a political decision in a country with among the highest malaria burdens in the world, eleven days after a release of the deliberately non-spreading version.

Speculative The second civilizational claim is about the species rather than a disease. If heritable modification ever became routine and effective, it would be the first technology to alter the substrate that transmits every other one, and the resulting divergence would be measured in generations rather than product cycles. Handwave That argument is at least fifty years old and has never had a working technology attached to it; on the present record it remains a philosophical position, and this brief flags it as such.

Established The honest counterweight is scale. The entire global output of approved CRISPR therapy in its first 22 months was 39 people. Frontier A technology at that throughput is not yet civilizational in any sense, and the interesting civilizational question is therefore not what genome writing will do to humanity but whether it will ever escape the transplant ward. Speculative There is a third implication that neither camp likes. If somatic editing can permanently reset a lifelong physiological set-point in an adult with one dose, then the case for heritable modification collapses from a capability argument to a convenience argument — you are no longer buying something otherwise unobtainable, you are only avoiding a treatment. Frontier That would be a decisive reframing, and it is currently true for exactly one organ. See Longevity Therapies and Cellular Rejuvenation for the same throughput argument in adjacent fields.

These horizons track the two clocks that actually run at different speeds: the editing chemistry, which improves steadily and is already ahead of what can be delivered, and the delivery-plus-manufacturing chain, which has moved very little in a decade.

• 10 yr: Frontier In-body base and prime editing become standard of care for at least one hepatic indication, on the strength of the dose–response data already published. Frontier A second solid organ is reached, or is not, and that single fact determines whether this subject is a platform or a liver speciality. Speculative Non-genotoxic conditioning reaches phase III and, if it works, does more for patient numbers than any editor released in the same window. Frontier The nex-z holds resolve one way or the other and set the regulatory tolerance for systemic LNP-Cas9 for a decade. Speculative Bespoke n-of-1 editing becomes a defined regulatory pathway rather than a series of exceptions. Frontier An epigenetic silencer reports multi-year durability in humans, or fails to, and that single readout decides whether reversible regulation is a therapy class or a research tool. Speculative Prime-editor-plus-integrase insertion enters the clinic ex vivo before it enters anyone in vivo, for manufacturing reasons rather than efficiency ones.
• 25 yr: Speculative A synthetic human chromosome exists as a laboratory object, most plausibly a small one built as proof of principle rather than a functional replacement, and the SynHG programme's five-year proof-of-concept scale suggests it will be built by a successor programme rather than by SynHG. Speculative Human artificial chromosomes either acquire a reliable de novo centromere or are abandoned as a therapeutic route; there is no third outcome and the question has been open for thirty years. Handwave A jurisdiction somewhere permits heritable editing for a narrow monogenic indication, or none does; both are defensible predictions and the honest position is that nobody knows which. Speculative Gene drives are field-tested somewhere, or the technology is quietly retired without ever having been tried. Speculative Targeted insertion of whole genes displaces micro-dystrophin-style truncations for the large-gene diseases, or the 4.7 kb packaging ceiling remains the de facto definition of what is treatable.
• 50 yr: Handwave Whole-genome writing for mammalian cells becomes an engineering discipline only if long-fragment synthesis improves by something like two orders of magnitude in fidelity-per-cost, for which no published roadmap exists. Speculative If it does, adding designed chromosomes plausibly displaces editing for anything requiring more than a handful of changes, because it perturbs nothing and can in principle be removed. Handwave If it does not, this subject in 2076 looks like this subject now with better vectors: a set of expensive, effective, narrow therapies for monogenic disease.
• 100 / 250+ yr: Handwave Recoded organisms with genetic codes incompatible with natural horizontal transfer are the coherent long-horizon endpoint of genome writing, and the argument for them is biosafety rather than capability. Handwave Whether anything in this brief scales to the polygenic architecture of ordinary human traits is genuinely open, and the strongest available answer is that it does not, because the per-edit damage rate multiplies with the number of edits and nobody has proposed a mechanism that escapes that arithmetic. Handwave On this horizon the constraint is no longer the technology but whether any institution can hold a safety commitment across a century, which is the same problem every long-latency technology has and none has solved.

• Depends on Depends on results this subject does not generate: ionisable-lipid and capsid engineering capable of tissue tropism outside the liver, which belongs to materials chemistry and virology; non-genotoxic conditioning regimens, which belong to transplant haematology; long-read clonal-tracking analytics adequate to see a MECOM insertion years before it becomes a leukaemia; and de novo human centromere specification, which is a basic chromosome-biology question that the artificial-chromosome literature has named as its own blocker for three decades. None of the four is nuclease work, none of the four is being funded in proportion to what waits on it, and no brief on this map produces any of them. That is the honest reason this subject has no typed dependency on another brief: what it waits on is chemistry, surgery, transplant medicine and instrumentation, none of which is a frontier hypothesis.
• Requires (not on this map) Two constraints sit outside this map and neither is a scientific unknown. The first is industrial: error-free synthesis of DNA fragments hundreds of kilobases long at a price that makes a chromosome-scale build affordable — the requirement stated in those words by the Synthetic Human Genome project's own participants, and the reason a completed 12 Mb yeast genome took twelve years and 250 people while a human chromosome runs from 47 to 249 Mb. No public cost-per-base series for long high-fidelity fragments could be verified for this brief. The second is institutional: a consent mechanism for releasing a self-spreading organism into a shared ecosystem. Village agreement and national biosafety approval were both obtained in Burkina Faso in 2025 and neither survived contact with national politics, which terminated a non-drive trial eleven days after a single release. Until an instrument exists that can carry that decision, gene-drive biology has nowhere to go. A third constraint is regulatory rather than scientific, and it arrives with epigenome editing: there is no agreed endpoint for the durability of a silencing effect, so a one-dose therapy whose mark could in principle drift has no defined way to demonstrate that it has not, and the first sponsor to reach that question will be negotiating a standard rather than meeting one. A fourth is supply. Closed-circular donor DNA at good-manufacturing-practice grade and clinical dose scale is the physical input every integrase-based insertion needs, and no established plasmid supply chain was built to make it; that is a factory question, not a laboratory one, and nothing on this map answers it.
• Enables Enables Precision Medicine in its only genuinely curative form, supplies De-Extinction Technologies with the entire editing toolkit it assumes, and provides Designer Organisms and Synthetic Biology with the construction step they specify but do not perform. It also gates Biological Enhancement and Neurogenetics in a way those subjects should state more plainly than they usually do: without multiplex editing that does not accumulate chromosomal damage per cut, the polygenic version of enhancement has no delivery mechanism, only a scoring function.
• Adjacent Adjacent to Synthetic Ecosystems, which owns the release-and-consent problem that terminated the only field trial in this brief; to Artificial Wombs and in-vitro gametogenesis, which are the physical prerequisites for every iterated germline scheme; to Longevity Therapies and Cellular Rejuvenation, which share the delivery constraint exactly and rarely say so; and to Biological Computing, which shares the DNA-synthesis cost curve and is the other subject whose progress is hostage to it.

Established “CRISPR can edit anything.” This is the enthusiast's error and it has an exact correction: CRISPR cannot edit what it cannot reach. The complete approved and late-stage in-body record is liver — Verve's base editor, Intellia's nuclease — or cells taken out of the body first, which is Casgevy, Skysona, Zynteglo and PM359. Established Every non-hepatic in-body programme in this brief is placed by hand: stereotactic injection into the striatum for AMT-130, injection into the cochlea for DB-OTO. Frontier The correct statement is that CRISPR can be programmed to edit anything and can be delivered to roughly two places.

Established “Prime editing can correct 89% of genetic diseases.” The source sentence says prime editing “in principle could correct up to 89% of known genetic variants associated with human diseases,” and the 89% is the fraction of catalogued pathogenic variants whose type — point substitution, small insertion, small deletion — falls inside prime editing's addressable class. Established It is arithmetic about mutation categories. It says nothing about delivery, nothing about efficiency in the relevant tissue, and nothing about whether the corrected fraction of cells is therapeutically sufficient. Frontier Converting it into a clinical claim is the single most common overstatement in this field.

Established “CRISPR's problem is off-target edits.” The larger measured problem is on-target. Cas9 breaks resolve into kilobase-scale deletions and complex rearrangements that conventional short-range genotyping does not see; edited human embryos showed loss-of-heterozygosity in 61.9% of targeted samples against 10% of Cas9-only controls; and editing itself selects for cells with a blunted p53 response. Frontier A field that spent a decade improving guide specificity was optimising the smaller of its two error terms.

Established “The CRISPR babies got the Berlin Patient's mutation.” They did not. Neither child carries CCR5-delta32. One twin has a 15-base-pair in-frame deletion on one chromosome with a wild-type copy on the other and is mosaic; the other is homozygous for a 4-base-pair deletion plus a single-base insertion. Established A related claim in wide circulation — that CCR5-delta32 homozygotes die younger, which was used as the cautionary tale about unintended consequences — rests on a 2019 Nature Medicine paper that overlooked a genotype-calling bias and was retracted four months after publication. Frontier It should not be stated as fact, and the correction cuts against the safety argument that used it.

Established “Gene drives have been released in Africa.” No gene-drive organism has been released into the wild anywhere on Earth. One release of non-drive, self-limiting male-bias mosquitoes took place in Burkina Faso on 11 August 2025 and the national programme was formally terminated on 22 August. Established “Approved gene therapies are one-time cures.” Skysona caused a haematologic malignancy in roughly one in ten treated patients with latencies out to 92 months; Roctavian's expression declines after 12–18 months; Elevidys missed its phase III primary endpoint and was linked to three acute-liver-failure deaths. Established “Approval means access.” Thirty-nine infusions against sixteen thousand eligible patients in one country.

Frontier “A synthetic human genome is imminent.” The Synthetic Human Genome project is a £10 million, five-year, proof-of-concept programme across five UK universities that has not committed to a first chromosome. Established The completed eukaryotic precedent, a 12 Mb yeast genome, ran from 2011 to 2023 and involved more than 250 researchers. Speculative A human chromosome is between four and twenty times larger and carries regulatory architecture that yeast does not.

Speculative The strongest heterodox position in this field is that the delivery problem is permanent, not temporary. Its case has four parts. First, targeting had a single unifying principle — base pairing — and delivery has none; it is not one problem but one problem per tissue, each with a different barrier. Second, the liver is not a success, it is an anatomical gift: it filters blood by design, so lipid nanoparticles arrive there without being told to, and no second organ has that property. Third, the therapeutic window may close before efficacy is reached, because the systemic dose required for a low-uptake tissue can exceed the hepatotoxicity threshold that the largest in-body trial has now measured. Fourth, every non-hepatic clinical success is a surgeon placing the vector by hand, which is a confession that the vector did not do the work. Frontier What the mainstream says back is that the liver was also once impossible, that engineered capsid libraries and targeted lipid formulations are improving on a visible curve, and that the argument mistakes an unfinished programme for a proof of impossibility. Speculative The two positions make a testable divergence: if no non-hepatic tissue reaches 30% in-body editing without surgical placement within a decade, the permanence argument has won the empirical round.

Speculative The second heterodox position is that editing does not scale to polygenic traits at all. The arithmetic is unfriendly in both directions. On the selection side, the top-of-ten-embryos gain on the best published education index is about half a year with a prediction interval spanning zero. Frontier On the editing side, each cut carries an independent probability of kilobase deletion or loss of heterozygosity, so an intervention requiring dozens or hundreds of edits multiplies a per-cut damage rate that is already measurable at one. Speculative Notably, the most systematic published advocacy for strong germline engineering concedes this in its own terms, arguing that the bottleneck is not polygenic scores but epigenomic correction — and that in-vitro gametogenesis has not been achieved in human cells at all, in either the maternal or paternal direction. Frontier That is an interest-against admission from an organisation founded to advance the technology, and it is worth more than any critic's version of the same point.

Speculative The most interesting minority proposal is to stop editing and start adding. If editing perturbs the existing genome and produces kilobase deletions, loss of heterozygosity and p53 selection, then delivering an engineered chromosome perturbs nothing: the payload carries its own regulatory context and can in principle be removed. Frontier The precedents are real — a 1.024 Mb designer accessory chromosome carrying 542 genes moved into six yeast species, and a 2.4 Mb dystrophin gene carried on a human artificial chromosome in 2009 when no viral vector could hold it. Speculative Three links are missing: reliable de novo human centromere specification, faithful segregation through many mitoses in the target lineage, and any mammalian route for delivering a megabase episome, since the yeast method used mating. Handwave Further out sits whole-genome recoding — rebuilding an organism's genetic code so that it cannot exchange sequence with anything natural, which would be the strongest biocontainment argument ever made. The bacterial recoding results were not verified for this brief and no numbers from them are stated here; what can be said is that recoding is a genome-writing problem, and genome writing is gated on the synthesis constraint in section 13.

Handwave Finally, the idea that keeps being reinvented. Iterated embryo selection — deriving gametes from selected embryos, crossing, selecting again, and compressing many generations of selection into laboratory time — has been independently proposed at least five times since 1991, under the names velogenetics, whizzogenetics, iterated embryo selection and in vitro eugenics, and reinvented again at least three times in 2018. Established The livestock literature's own efficiency claim is that in-vitro breeding could be 30–40 times more efficient than conventional breeding. Established What has been demonstrated is IVF, preimplantation diagnosis and stem-cell derivation. Handwave What has never been demonstrated, in any mammal, is a single complete iterative cycle producing viable offspring across multiple rounds. The proposal is nearly thirty-five years old and has not advanced past the first step in any species.