To edit a genome better, first switch off the genes that slow delivery down

For ten years, the effort has gone almost entirely into the tool: more precise molecular scissors, finer base editors, better-tolerated vectors. A team at the University of Wisconsin–Madison approached the question from the other end — not "how do we improve the editor?" but "what does the human cell put in the way of its entry?". By screening 19,114 genes, they identify six that act as brakes, and show that lifting a single one is enough to increase eightfold the correction of a mutation responsible for a childhood blindness. All of it in cultured cells: this is a methodological advance, not yet a treatment.

Source: nature.com

In Plain Terms

Correcting a faulty gene requires two distinct things: having a good correction tool, and getting it into the cell nucleus. Much attention has gone to the first point. The second remains the bottleneck: most injected molecules never reach their destination — they are swallowed into vesicles, then degraded, like a parcel that never reaches the right room in the house.

So the team methodically switched off, one by one, nearly every gene in the human genome, then looked at which ones improved delivery when absent. Six stand out: the cell does indeed make proteins that hinder the tool's entry. Neutralising the main one, GJB2, raises correction of a mutation that blinds children, in retinal cells derived from a patient, from 5.8% to 46.6% — and a fraction of the edited cells starts conducting current again, which is the sign that the repaired protein works. Each gene was tested separately: the study does not attempt to switch off two at once, it mentions this as a lead to explore.

Two reservations. The first, which the authors do not frame as a limitation but which is obvious: all of this takes place in a culture dish, no animal, no patient. The second, which they raise themselves: the main one of these brakes, GJB2, is essential to hearing — so it is out of the question to remove it permanently; it would have to be put on standby for the duration of the treatment, which this study does not do.

Technical Explanation

1. The real bottleneck is not the enzyme, it is the journey. A genome editor delivered without a virus must clear a succession of steps before acting: adhesion to the membrane, internalisation by endocytosis, escape from the endosome before it acidifies and fuses with a lysosome, transit through the cytosol, then nuclear import. Each step is a multiplicative filter. The authors describe this sequence as the "delivery-to-editing" arc, from internalisation to intracellular trafficking, then to nuclear import and DNA repair — and draw a methodological consequence from it: if the limiting factor is the journey, then the genes that govern it are targets just as much as the enzyme.

2. The screen, and what exactly it measures. The Brunello library is a collection of 77,441 guide RNAs covering 19,114 genes at four guides per gene, plus 1,000 non-targeting guides that serve as a statistical baseline. Introduced by lentivirus into a population of about 200 million HEK293 cells, it produces a population in which each cell carries a different knockout. Nonviral editing is then applied to the whole population, followed by sequencing — and this is where the design is ingenious: an amplicon of about 150 base pairs captures on the same read the identity of the guide and the outcome of the edit (insertion-deletion or intact sequence). Each guide is therefore assigned its own editing rate, without any need to sort the cells beforehand. The logic of the evidence remains statistical rather than causal — hence the next step, which is indispensable.

3. From 26 candidates to 6: validation in an arrayed panel. A pooled screen produces false positives, notably through differential growth effects. The authors therefore retained 26 candidate genes among the 50 top scores of the screen — on criteria of guide enrichment and reproducibility across iterations — then knocked them out again one by one, in an arrayed panel of polyclonal lines, and retested them across different combinations of payload, locus and cell type. Six survive this filter: BET1L, MS4A13, RBM44, SLCO1C1, GJB2 and ZNF584. According to the authors' abstract, their depletion increases editing efficiency "up to sixfold" depending on the context. The fact that the gain holds up while changing the vector, the target and the cell all at once is what distinguishes a real effect from a cell-line artefact.

4. The experiment that locates the mechanism — and it is the most important one. Knowing that six genes act as brakes does not say where they brake. For this the authors use a bypass: electroporation, which pushes the cargo directly into the cytosol and the nucleoplasm, short-circuiting endocytosis. The result is clear-cut — "no significant difference in editing efficiency between nearly all lines and the wild-type control". The reading is strong without being binding: if removing the brake no longer brings anything once you enter by force, then the brake was likely acting on the entry route, upstream of DNA cleavage and repair. The authors phrase their conclusion in terms of comparative likelihood, and they should be followed to the letter: these results are "more consistent with an upstream, delivery or trafficking role than with a dominant effect on the later steps of editing", while specifying that "contributions to Cas9 activity and DNA repair cannot be excluded". The reasoning is that of an epistasis applied to intracellular pharmacokinetics: it is worth more than a correlation, without amounting to exclusion.

5. Confirmation by imaging. It remains to see the cargo arrive. In multispectral flow cytometry — which, unlike conventional cytometry, produces an image of each cell and therefore allows measurement of a colocalisation and not only an intensity — knockouts of BET1L and MS4A13 give "an approximately 1.5-fold increase in colocalisation of Cas9-GFP with the nuclear marker". The technique measures a spatial overlap; it indicates that more editor actually reaches the nucleus, which the electroporation test only inferred. A 1.5-fold increase is a converging clue, not a demonstration — and note that GJB2, the most effective gene otherwise, does not appear among the results of this imaging experiment.

6. The two reporter systems, and why the direction of reading flips. In the EGFP reporter, successful editing destroys the fluorescence gene: success therefore reads as a drop in the number of green cells, by up to a factor of ten in the knockout lines compared with controls. In KCNJ13, conversely, editing repairs: an adenine base editor converts an A into a G to remove the premature stop codon, and corrected alleles are counted. GJB2 knockout brings this conversion to 19.3% ± 0.67, that is 6.7 times the control value; BET1L knockout to 14.5% ± 3.01, a factor of five. A third system, the introduction of a cytosine mutation into the GABAA receptor gene, serves to verify that the effect does not depend on the type of chemical conversion used.

7. The move to patient cells, and the test that commits the most. The final model is a retinal pigment epithelium differentiated from induced pluripotent stem cells of a patient with Leber congenital amaurosis type 16 — a paediatric blindness caused, in this case, by the homozygous W53X nonsense mutation in KCNJ13. The editor is delivered by lipid nanoparticles. Here, silencing is not obtained by constitutive knockout as in HEK293 cells but by lentiviral transduction producing insertions-deletions at both loci — the authors speak of knockdown. The modification is nonetheless durable: nothing in this model is transient or reversible. Compared with an unperturbed control at 5.8% ± 1.68, depletion of BET1L increases correction 3.6-fold (20.9% ± 5.0) and that of GJB2 eightfold (46.6% ± 10.1); MS4A13 also reaches a factor of about eight. Above all, the authors do not stop at counting alleles: the corrected cells re-express the full-length Kir7.1 protein, correctly targeted to the apical membrane, and patch-clamp electrophysiology measures the current. This is where the figures must be read closely: the experiment covers four cells per condition, functional rescue reaches statistical significance only for GJB2, whereas for BET1L the mean change in rubidium flux does not ($p = 0{,}44$p=0,44) — the authors see in this a functional heterogeneity consistent with a mixed population, 25% of edited cells exceeding the maximum of the diseased control. Measuring the current rather than the sequence alone is what turns a percentage into an argument about function; the sample size is a reminder that this argument remains preliminary.

Why It Worked

The idea that holds everything together is a shift of target: treating the recipient cell as an engineering variable, instead of considering it a fixed terrain to be conquered with a better tool. In figures, the gap is substantial — in unperturbed patient cells, correction of the KCNJ13 locus tops out at 5.8% ± 1.68; silencing GJB2 brings it to 46.6% ± 10.1 (eightfold) and silencing BET1L to 20.9% ± 5.0 (3.6-fold). A third gene, MS4A13, also produces a gain of about eightfold in this same model. And the demonstration does not stop at sequencing: the Kir7.1 channel recovers its apical localisation and a measurable current.

It is nonetheless necessary to name the gap between what is demonstrated and what one might hear. This study shows no therapy: there is no animal, no in vivo administration, no safety track record. Functional restoration is explicitly obtained "in a subset" of the edited cells, on four cells per condition and without statistical significance for BET1L, and the authors specify that a return to the phenotype of a homogeneous hexagonal monolayer, that of healthy cells, "may require further optimisation". The heaviest limitation is biological: GJB2 encodes connexin 26 and, as the authors write, "is essential for hearing and skin development; its permanent systemic knockout would be deleterious". The therapeutic target is therefore not ablation but transient modulation — which this study does not achieve. To this are added two reservations about scope: in cell types where these genes are weakly expressed, the strategy will have to be retailored to the molecular context; and the authors acknowledge that they cannot definitively exclude that part of the BET1L effect works through altered DNA repair kinetics or intrinsic editor activity, rather than through delivery alone. Finally, a calibration remark: the article's abstract announces gains "up to sixfold" and "more than 3.5-fold" in patient cells, wordings more cautious than the 6.7 and 8 of the figures — the authors' caution probably bears on generality, not on the individual measurements.

Causal Chain

Nonviral editing avoids the immunogenicity and integration of viral vectors → but it depends entirely on a low-yield intracellular journey → this journey is governed by proteins of the recipient cell, never systematically mapped → a genome-wide CRISPR screen across 19,114 genes turns this question into a measurement → 26 candidates are retained among the 50 top scores, 6 survive individual validation → electroporation, which short-circuits entry, abolishes their effect in nearly all lines, which places the brake more likely upstream of repair than downstream → imaging indicates an approximately 1.5-fold increase in editor reaching the nucleus → applied to retinal cells from a patient carrying the KCNJ13 W53X mutation, switching off GJB2 raises correction from 5.8% to 46.6% → the re-expressed Kir7.1 channel recovers, in a fraction of the cells and on a sample of four per condition, a current measurable by patch-clamp → what remains is to convert a permanent knockout into a transient and reversible inhibition, a condition without which the approach is unusable therapeutically.

Anecdote

The chance of the screen placed in first position a gene that medical genetics knows well, but for an entirely different reason: GJB2 encodes connexin 26, and its mutations are the leading cause of non-syndromic hereditary deafness — up to half of the cases of genetic origin in several populations studied. A gene that has been discussed for twenty-five years because its loss causes deafness turns up here because its loss facilitates genome editing. The authors immediately draw the necessary consequence and record it among their limitations: it will never be a matter of switching it off for good. It is a useful reminder of what "therapeutic target" means — not a switch to be flipped down, but a function that would have to be suspended just long enough, and only in the part of the body being treated.

Legacy and Current Data

This team is not discovering the problem: it had already published in 2023 a nonviral correction of the same KCNJ13 mutation using silica nanocapsules carrying a base editor, which preserved vision in a model of the disease and reached 47% editing in patient fibroblasts but only 17% in iPSC-derived retinal epithelium. It is precisely this drop-off between cell types — the same tool, a yield that collapses depending on the cell targeted — that the screen sheds light on, and the 46.6% obtained here in this same refractory cell type takes on its meaning in that comparison. For the rest, it must be said plainly: the field of nonviral gene therapies has, to date, no market or adoption data to set against this, because no approach based on modulating these brakes has entered a trial. What remains to be established is clear — a transient inhibitor, an in vivo demonstration, and a safety profile.

The Researcher's View — Open Questions

Interpretation, not a result of the study. Three experiments would decide what comes next. Transient modulation first: replacing permanent knockout with reversible silencing — interfering RNA, targeted protein degradation, pharmacological inhibitor — and verifying that the editing gain survives reversibility; this is the only route compatible with therapeutic use, and the study does not test it. The double brake next, which the authors themselves call for: do GJB2 and BET1L act on the same step or on two distinct steps of the journey? Knocking them out simultaneously would say — an additive gain would indicate two independent barriers, a saturating gain a single limiting step. The portability test finally: does the same lifting of brakes work in the cells that really matter for gene therapy — hepatocytes, haematopoietic stem cells, T lymphocytes? The authors themselves note that where these genes are weakly expressed, the strategy will have to be retailored — in other words, the generality of their result is not established.

Sources

References verified during the fact-checking audit of 14 August 2026 — full text consulted: abstract, Results, Methods, sample sizes, statistical tests and limitations declared by the authors.

• S. Saxena, M. Kabra, A. A. Abdeen, D. M. Tabima, D. Sinha, P. A. Rawding, M. Zhu, R. Xie, T. Kulkarni, G. M. Hanstad, M. A. Fernandez Zepeda, D. M. Gamm, B. R. Pattnaik, S. Gong, K. Saha et al., "Genome-wide CRISPR screening identifies cellular factors controlling nonviral genome editing efficiency", Nature Communications 17, article 8086, 13 August 2026. DOI: 10.1038/s41467-026-76350-5 — peer-reviewed article, open access.

Background References

These references illuminate the context and do not come from the study discussed.

• Preliminary version of the same work deposited on bioRxiv in March 2025 (DOI 10.1101/2025.03.12.642795) — its figures differ from the published version; only the values from the peer-reviewed article are used here.
• "Nonviral base editing of KCNJ13 mutation preserves vision in a model of inherited retinal channelopathy", 2023 (PMID 37561581) — earlier work by the same team, source of the 47% yield in fibroblasts and 17% in retinal epithelium cited for perspective.
• Pattnaik et al., "A novel KCNJ13 nonsense mutation and loss of Kir7.1 channel function causes Leber congenital amaurosis (LCA16)", Human Mutation, 2015 (PMID 25921210) — description of the disease and of the role of the Kir7.1 channel in the retinal pigment epithelium.
• Literature on GJB2/connexin 26 as the leading cause of non-syndromic hereditary deafness (see in particular PMID 10704187) — for the anecdote only.