Two routes to replace an oversized retinal gene

What the work claims

This is a preclinical proof-of-concept, and it makes two paired claims. First, the human MYO7A gene can be supplemented in the retina despite exceeding the cargo limit of adeno-associated virus, by splitting it across two viral vectors whose messages are rejoined inside the cell. Second, an alternative that avoids delivering the big gene at all, switching on the related Myo7b gene with a CRISPR activator, can partially substitute for the missing protein. In a mouse model that lacks Myo7a in the eye, the supplementation approach restores the protein to normal levels and corrects the disease-linked cellular defect, while the activation approach helps less and comes with a hint of toxicity.¹

What kind of work this is matters for how much weight the claims carry. These are mechanism-of-delivery results measured against a molecular surrogate, not a demonstration of restored vision. The honest reading is that the study clears an engineering bar, getting a large gene expressed, spliced and localised correctly in the right cells, rather than a clinical bar.²

How it works

Adeno-associated virus is the workhorse of in-vivo gene therapy, but its cargo hold is small, roughly 4.7 kilobases. The MYO7A coding sequence is about 6.7 kilobases, so it does not fit. The authors use a dual-vector trick called REVeRT: each virus carries one half of the gene plus splicing signals, and the two halves are joined at the level of messenger RNA once both reach the same cell. They report that this messenger-RNA trans-splicing reconstitutes the gene about threefold more efficiently than their own DNA-level reassembly construct. That DNA-level approach is the same general strategy now in a first human trial, but the authors are explicit that their comparator is not the identical clinical vector, so the threefold figure is not a head-to-head test against the trial construct.¹

Myosin VIIA, the MYO7A protein, works in the retinal pigment epithelium and in photoreceptors, where it moves pigment granules and helps organise the connecting cilium. To test therapy the authors built a conditional mouse that deletes Myo7a in the retina and its pigment epithelium, avoiding the balance and hearing problems that plague the classic Shaker mouse. That eye-specific knockout reproduces the hallmark measurable defect, misplaced melanosomes in the pigment epithelium, which becomes the readout for rescue. Injecting the dual MYO7A vectors under the retina at three weeks of age restored the protein to normal levels in both the pigment epithelium and the retina, and the melanosome defect became statistically indistinguishable from healthy controls at both four and twenty weeks.²

The second strategy leans on redundancy. MYO7B is a close relative with a similar domain layout, so the authors used a dead-Cas9 activator, delivered again as split halves, to turn up the endogenous Myo7b gene in the same knockout mouse. This improved melanosome placement compared with untreated eyes but did not fully normalise it, and it came with a small reduction in the thickness of the photoreceptor layer and faint markers of retinal stress, signs that switching on a gene where it is not normally expressed at that level is not free.³

Where a skeptic should push

The most load-bearing assumption is that the molecular surrogate stands in for the outcome patients care about. The entire efficacy story rests on melanosome localisation, a cell-biological phenotype, not on any measure of vision or photoreceptor survival. That is a defensible choice, because this mouse, like the Shaker it replaces, has only a weak retinal phenotype, with no detectable defect in photoreceptor protein transport and an unchanged electroretinogram, but it also means the study cannot say the treatment preserves sight. More pointedly, the photoreceptor pathology that actually drives blindness in human Usher 1B is not present in this model to rescue, so the decisive efficacy question is not merely unmeasured, it is unmodelable here. A therapy can normalise a trafficking marker and still fail to protect the photoreceptor over years. The gap between marker and function is where retinal gene therapies have repeatedly disappointed. There is a further problem with the baseline itself: the knockout being rescued is incomplete. The model retains residual myosin VIIA, and the pigment epithelium, the very place the melanosome defect is scored, is only mosaically knocked out, so a rescue that reaches control levels is clearing a bar set by a leaky knockdown within the injected patch rather than restoring a true null.

Second, separate demonstrated from asserted for the two arms. Supplementation is the strong arm: full protein restoration, full correction of the surrogate, no detected toxicity in the window studied. Activation of the cousin gene is the weak arm: partial correction, and measurable costs to layer thickness and glial and immune markers. Reading the two as equivalent options would flatter the weaker one. The authors themselves note that Myo7b may not fully compensate and that a dose-finding study would be needed, which is an admission that the current activation result is not yet therapeutic.

Third, the human evidence is expression, not rescue. The human retinal organoids and the pigs are wild type; they carry a working copy of the gene. Running the therapy through them shows that the human transgene splices and the protein localises in human and large-eye tissue, which is valuable, but it does not show correction of a human disease phenotype, because there is no disease in those samples. The organoids were also transduced at a set maturation stage, and organoids at that stage still lack the fully formed outer-segment and pigment-epithelium interface where myosin VIIA does its most important work. So the human data de-risk delivery and expression, not efficacy.

Where organoids fit in a therapy pipeline

The instructive move in this paper is not the mouse rescue; it is the division of labour between models. Mice answer whether the therapy corrects a phenotype in a living eye. But mice cannot answer a question that is fatal to translation for this class of therapy: does the human gene, delivered as two halves and rejoined by the cell's own splicing machinery, actually reconstitute the correct human transcript and protein in human cells. Human and mouse splicing are not identical, and a trans-splicing construct optimised in one species can misbehave in the other. The human retinal organoid is where that question gets asked, on human sequence in human cells, without a patient.

That reframes what the organoid is for. In much of this field organoids are pitched as disease models that must faithfully reproduce pathology. Here they play a narrower but arguably more decisive role: a human-context validation bridge for the engineering of the therapy itself, checking splicing fidelity, expression level and subcellular localisation of a human construct before it reaches a person. For any large-gene or split-vector programme, and there are many, from Stargardt disease to the muscular dystrophies, that is a concrete, reusable use of a human organ model that does not depend on the organoid perfectly mimicking disease. It only has to be authentically human at the molecular steps that matter.

The threat is over-reading that same bridge. Because a wild-type organoid can show beautiful expression and correct localisation, it is tempting to treat that as evidence of efficacy, and it is not. The organoid says the construct works as a piece of molecular engineering; it is silent on whether restoring the protein rescues a diseased human photoreceptor, and silent on long-term safety. The activation arm is a reminder of the stakes: switching on a substitute gene looked attractive on paper and delivered partial benefit with real costs. A pipeline that uses organoids to confirm human-context expression, then still demands functional and durability evidence in animals and patients, is using them well. A pipeline that lets a clean organoid image substitute for efficacy is fooling itself.

The bottom line

Established here: a large retinal gene can be reconstituted from split AAVs with better efficiency than the DNA-level method now in trials, and doing so restores the protein and corrects a molecular defect in an eye-specific mouse knockout, with the human construct shown to express and localise correctly in human retinal organoids and in pigs. Hypothesis, not result: that this preserves vision, that the effect lasts, and that activating Myo7b is a viable alternative rather than a partial and slightly toxic one. What would confirm the programme is a model with a genuine photoreceptor-degeneration phenotype showing functional rescue that endures, plus organoid work in mutant human lines rather than wild type. What would break it is durable expression failing to translate into preserved photoreceptors, the recurring gap between a corrected marker and saved sight. The delivery engineering is the strong, verifiable core; the therapeutic claim is still ahead of the data.