Repairing More Tissue Than You Edit: What In Vivo Selection Could Mean for Gene Therapy

Key Takeaways

• The Repair Drive in vivo selection strategy increased reporter-positive, gene-targeted hepatocytes to approximately 25% in mice and raised therapeutic factor IX expression about fivefold.

• In vivo selection could help overcome limited gene-editing efficiency by allowing successfully corrected cells to expand after the initial editing event.

• The strategy relies on controlled tissue regeneration and a temporary survival advantage, making the strength and duration of selection central development variables.

• Selective expansion may also amplify cells carrying unintended edits or preexisting fitness advantages, increasing the importance of clonal tracking and tumorigenic-risk assessment.

• Translation beyond the liver will depend on tissue-specific regenerative capacity, controllable conditioning, reliable safety switches, and practical methods for long-term clonal monitoring.

The effectiveness of in vivo genome editing is often limited by the number of target cells that receive the editing machinery and complete the intended molecular repair. This constraint becomes particularly important for homology-directed repair (HDR), which can install precise insertions and replacements but functions inefficiently in many adult, terminally differentiated tissues. Even when delivery reaches the correct organ, the proportion of cells that acquire the desired edit may remain too small to produce a therapeutic effect.^1,2

Most efforts to overcome this limitation focus on improving delivery, editor activity, donor-template design, or DNA-repair efficiency. A different strategy is to accept a modest initial editing rate and then increase the representation of successfully corrected cells after the editing event. If those cells can be given a controlled survival or proliferative advantage, their descendants may eventually occupy a much larger fraction of the tissue than the cells initially corrected.

Repair Drive, an experimental in vivo selection strategy reported in mice by a multi-institutional research team in 2025, provides an early proof of concept for this approach in the liver. Rather than attempting to edit every therapeutically relevant hepatocyte, the system linked successful targeted integration to resistance against a temporary metabolic selection pressure. Correctly targeted hepatocytes survived and divided, while unprotected cells were gradually depleted. The resulting tissue contained a substantially larger proportion of gene-targeted cells than comparable tissue that had not undergone selection.¹

This mechanism does not literally repair cells that remained unedited. It expands the descendants of cells that already acquired an editing outcome capable of expressing the selectable marker. However, the therapeutic significance could be similar to correcting a much larger starting population. In a regenerative tissue, editing efficiency may define the starting point, while controlled selection determines the eventual proportion of corrected cells.

That possibility changes the development question. Instead of asking only how many cells can be edited during treatment, developers may also need to ask how a corrected population can be expanded, how that expansion can be controlled, and whether the final tissue composition remains safe over time.

Repair Drive used a temporary metabolic disadvantage to distinguish correctly targeted hepatocytes from surrounding cells. The system suppressed fumarylacetoacetate hydrolase (Fah), an essential enzyme in the final step of tyrosine catabolism, using a hepatocyte-targeted small interfering RNA (siRNA). Loss of Fah activity leads to the accumulation of toxic metabolites and hepatocyte injury.¹

The donor cassette contained a human FAH sequence that was resistant to the siRNA because of sequence differences between the human and mouse genes. This selectable marker was linked in cis to a reporter or therapeutic transgene, so cells that received a suitable targeted integration could express both the protective FAH protein and the accompanying payload.

The editing system used two adeno-associated virus (AAV) vectors. One encoded the guide RNA and Staphylococcus aureus Cas9, while the other carried the donor cassette. The cassette was designed for integration into the 3′ untranslated region of the highly expressed liver gene Apoa1. Successful integration allowed the targeted locus to drive expression of apoA1, the human FAH selectable marker, and the linked reporter or therapeutic protein.

After editing, the mice received repeated doses of the Fah-directed siRNA. Unprotected hepatocytes experienced metabolic stress and cell death, which stimulated compensatory liver proliferation. Correctly targeted hepatocytes retained FAH activity, survived the conditioning period, and expanded as the tissue regenerated. The study therefore linked three events within the same cell: targeted integration, selectable-marker expression, and therapeutic-transgene expression.

This architecture is central to the concept. Selection does not identify a corrected DNA sequence directly. It favors cells that express enough of the protective marker to withstand conditioning. The intended therapeutic outcome depends on the assumption that protection and therapeutic expression remain coupled as selected cells expand.

Under the initial conditioning protocol, Repair Drive increased the proportion of hepatocytes expressing tdTomato, a red fluorescent reporter included in the donor cassette to identify cells carrying the targeted integration, from 0.3 ± 0.3% in unselected mice to 9 ± 3%. The positive cells appeared in colonies, which was consistent with clonal expansion after selection. HDR-modified Apoa1 alleles reached 3 ± 1.4% in the same experiment.¹

The strategy also produced expansion in female mice. TdTomato-positive hepatocytes reached 13.2 ± 3.1%, while HDR-modified Apoa1 alleles reached 6.4 ± 1.9%. These findings indicated that the effect was not limited to the initial male-mouse experiment.¹

The investigators then increased selection pressure by intermittently feeding mice a high-protein diet, which increased tyrosine catabolism and the accumulation of toxic metabolites during Fah suppression. Under this intensified protocol, tdTomato-positive hepatocytes reached 24.6 ± 6.4%, compared with 2.8 ± 1.6% in unselected animals. HDR-corrected Apoa1 alleles increased approximately sixfold, from 1.2 ± 0.6% to 7 ± 4%.

The distinction between reporter-positive cells and HDR-allele measurements matters. The study did not show that one-quarter of all liver alleles carried a perfect intended integration. Bulk liver DNA included nonparenchymal cells that were not targeted efficiently by the AAV system, and a hepatocyte could express the reporter from a single corrected allele. Long-read sequencing also showed that the selected population contained heterogeneous integration structures rather than a uniform molecular product.

The therapeutic experiment replaced the reporter with human factor IX (FIX). Repair Drive increased circulating FIX approximately fivefold, reaching 342.0 ± 162.2 ng/mL at the endpoint compared with 74.1 ± 26.6 ng/mL in unselected mice. Expression remained elevated through 52 weeks. The study therefore demonstrated that post-editing selection could increase functional transgene output, not merely the abundance of fluorescently labeled cells.

The result is best understood as biological amplification of an initially limited editing event. It does not establish efficient, uniform correction throughout one-quarter of the liver, but it shows that a selected population can become large enough to alter therapeutic-protein production substantially.

The liver offers an unusually favorable setting for this strategy. Hepatocytes can proliferate in response to tissue injury and cell loss, providing a regenerative mechanism through which protected cells can expand. The organ is also accessible to both AAV-based gene delivery and N-acetylgalactosamine–conjugated oligonucleotides, allowing the editing and conditioning components to be directed toward the same cell population.¹

The broader principle has a long biological precedent in hereditary tyrosinemia type 1. In Fah-deficient mouse models, hepatocytes that regain FAH activity acquire a strong survival advantage over neighboring diseased cells. Even a small population of corrected hepatocytes can expand and repopulate substantial portions of the liver because the surrounding cells remain metabolically impaired.³

That natural selection process demonstrated that gene correction and tissue repopulation could reinforce each other. Its usefulness, however, depended on a disease environment that inherently favored corrected cells. Most genetic liver disorders do not create such an advantage.

The Repair Drive approach was an attempt to manufacture the selective environment in otherwise healthy wild-type liver. By temporarily suppressing Fah in all hepatocytes while supplying an siRNA-resistant FAH sequence only to correctly targeted cells, the system reproduced some of the competitive conditions present in Fah deficiency. The survival advantage became part of the therapeutic design rather than a feature of the underlying disease.^1,3

Earlier work had also shown that protected gene-modified hepatocytes could expand during transient acetaminophen conditioning. In that approach, hepatocytes carrying a protective construct resisted acetaminophen toxicity and repopulated the liver, producing therapeutic effects in mouse models of hemophilia B and phenylketonuria. The method achieved substantial expansion but relied on a different metabolic-protection strategy and a hepatotoxic drug regimen.⁴

Repair Drive’s principal conceptual advance was the use of transient inhibition rather than permanent disruption of the metabolic gene that created the selective advantage. Endogenous Fah expression recovered after the selection period, and the targeted cells did not require permanent deletion of Fah, Hpd, or Cypor to remain viable.¹

A selection-based gene therapy would operate at two biological levels. At the molecular level, the treatment must deliver the editing machinery, produce the intended integration, preserve the selectable marker, and maintain expression of the therapeutic sequence. At the tissue level, it must alter competition among cells so that the desired population expands at a predictable rate.

The final therapeutic effect could therefore depend on more than the initial percentage of edited cells. Selection intensity, duration of conditioning, regenerative capacity, cassette integrity, clonal fitness, and functional output per corrected cell could all influence the endpoint. A low initial editing rate might still produce a meaningful result if selected cells expand efficiently, while a higher initial rate might provide limited benefit if the edited cells do not survive or proliferate.

This framework also makes tissue injury part of the dose-response relationship. Repair Drive worked because unprotected hepatocytes were placed at a disadvantage. Fah knockdown caused hepatocyte death, transient elevations in alanine transaminase and aspartate transferase, modest inflammatory responses, temporary collagen-gene induction, and compensatory proliferation. These effects largely resolved by later time points, but they were integral to the selection mechanism.¹

The intensified protocol produced greater expansion by adding periodic high-protein feeding. It also caused transient body-weight loss and elevations in liver enzymes. Stronger selection therefore generated a larger corrected-cell population at the cost of greater physiological pressure.

Clinical translation would require a therapeutic window in which enough unprotected cells are disadvantaged to permit useful expansion without causing unacceptable organ injury. That window may differ sharply among patients. A relatively healthy liver may tolerate temporary metabolic conditioning better than a liver affected by fibrosis, inflammation, impaired reserve, or ongoing injury.

The amount of repopulation required would also depend on the therapeutic mechanism. A disease treated through secretion of a circulating protein might require correction of only a minority of hepatocytes. A disorder caused by a cell-autonomous metabolic defect throughout the liver could require much broader tissue replacement. The acceptable balance between injury and expansion would therefore need to be defined for each indication rather than for the platform as a whole.

Repair Drive introduces several different forms of control that should not be grouped under a single claim of reversibility.

The first is control of the conditioning pressure. Fah suppression depended on repeated siRNA administration, so the imposed disadvantage was temporary. Once dosing stopped and the siRNA effect waned, endogenous Fah expression recovered.¹

The second is control over selection intensity. The study increased pressure through a high-protein diet, and the investigators proposed that siRNA dose could also be adjusted. These results suggest that the rate and extent of expansion may be tunable rather than fixed.

The third issue is the persistence of the tissue composition created during selection. Stopping the siRNA did not reverse the divisions that had already occurred or return expanded clones to their original abundance. The selective pressure was transient, but the altered population structure was durable.

This distinction matters for safety. A treatment could include a reliable way to stop further selection without providing a way to remove a clone that had already expanded. The study proposed that nitisinone, also known as NTBC, might reduce selection pressure by blocking tyrosine catabolism upstream of Fah. It also proposed development of an anti-siRNA intervention that could remove or neutralize the conditioning agent. Neither approach was tested.

A clinically useful system would need a rapid and predictable method for halting selection during excessive liver injury or another adverse response. It would also need evidence that expansion stops when the conditioning pressure is removed.

The more difficult question is whether developers would need a second safeguard capable of eliminating an undesirable selected clone. Such a mechanism would add complexity and might introduce new risks, but the durability of clonal expansion makes the issue difficult to ignore. The ability to stop the pressure that favors a clone is not equivalent to the ability to reverse the clone’s contribution to the tissue.

The sequencing results show why cell selection cannot be treated as a simple multiplier of perfect editing.

Long-read analysis identified several outcomes at the intended Apoa1 locus. These included complete HDR, HDR with partial truncation of the reporter sequence, and hybrid structures combining HDR at one boundary with nonhomologous end-joining (NHEJ) at the other. The study also detected indels and NHEJ-mediated capture of AAV sequences at the cut site.¹

Genome-wide analysis found donor integration outside the intended locus, including integrations in intergenic and intronic regions, known AAV-integration hotspots, and highly expressed liver genes. Repair Drive enriched intended HDR outcomes relative to several unintended outcomes, but the unwanted events did not disappear.

Some heterogeneous integration structures increased during selection because they retained functional expression of the FAH selectable marker. This finding highlights a fundamental property of the method: selection responds to cellular phenotype, not to whether the underlying DNA structure matches the intended design. A cell that expresses enough protective FAH can survive and expand even if the associated integration event is incomplete or structurally atypical.

The risk must also be considered within the existing clonal biology of the liver. Human livers accumulate somatic mutations and contain expanding hepatocyte clones during aging. Cirrhotic livers show altered clonal architecture, increased structural variation, and millimeter-scale expansions shaped partly by fibrotic tissue. Some mutations can increase hepatocyte fitness in chronic liver disease.^5,6

These observations do not demonstrate that Repair Drive will preferentially expand oncogenic cells. They indicate that a future therapy would act on a genetically and functionally heterogeneous population rather than a uniform field of hepatocytes. A preexisting high-fitness clone could receive the selectable cassette, an unintended editing event could alter cellular behavior, or chronic disease could change which cells respond most strongly to conditioning.

The one-year safety findings were encouraging but require careful interpretation. At the endpoint, Repair Drive mice did not show treatment-associated changes in body or organ weight, liver enzymes, hepatocyte death, hepatocyte proliferation, fibrosis, inflammatory-marker expression, or immune responses against FIX.

Four localized proliferative liver lesions were identified: two among 14 unselected mice and two among 14 Repair Drive mice. The two Repair Drive lesions showed partial loss of Fah expression, suggesting dedifferentiation and possible dysplasia, but none of the lesions displayed the reported histologic features of malignancy. The study did not show an increased incidence of proliferative lesions in the treated group, but its size and duration cannot establish lifetime or human tumor safety.

For a selection-based product, the safety question extends beyond whether an unintended alteration occurs. It includes whether the treatment could increase the representation of the cell carrying that alteration.

Selective expansion makes tissue mosaicism dynamic. The proportion and distribution of edited cells shortly after dosing may differ substantially from the composition months or years later. A clone that is rare at the end of the editing phase may become prominent after repeated conditioning, while other corrected clones may contribute little to long-term tissue function.

Bulk measurements can show the average frequency of an editing outcome, but they cannot establish whether the corrected population remains broadly polyclonal or becomes dominated by a small number of expanding clones. Two tissues with the same overall percentage of edited cells could have very different risk profiles if one contains thousands of small clones and the other depends on a few large ones.

The Repair Drive study observed colonies of reporter-positive hepatocytes and identified repeated integration junctions that could reflect clonal expansion.¹ These findings support the need to measure not only how much editing occurred but which edited cells supplied the final population.

Hematopoietic gene therapy studies provide the clearest precedent for longitudinal clonal tracking. Stable integration sites have been used as molecular barcodes to follow tens of thousands of marked clones across blood lineages and over several years. These analyses can identify changes in population diversity, persistent dominant clones, and integrations near genes associated with cancer.^7–9

Regulatory guidance also recognizes the significance of predominant clones, oligoclonality, monoclonality, persistent clonal expansion, and integration near oncogenic loci in products for which those risks are relevant. Such findings do not automatically indicate malignancy, but they can justify more intensive investigation and follow-up.¹⁰

The liver presents a harder monitoring problem than blood. Peripheral blood can be sampled repeatedly, while liver biopsy captures only a small and spatially limited portion of the organ. A dominant clone located elsewhere could be missed, particularly if expansion is patchy or constrained by fibrosis.

A future monitoring strategy might need to combine occasional tissue sampling with circulating biomarkers, cell-free DNA, imaging, liver-function measurements, and long-term clinical surveillance. None of these approaches currently provides a validated substitute for comprehensive clonal mapping throughout the human liver.

This limitation is especially important because the consequences of treatment could continue evolving after the editor and conditioning agent were no longer active. Monitoring would need to follow the evolving cellular population, not merely the persistence of the administered components.

The underlying idea of selecting genetically modified cells after treatment is not unique to hepatocytes. Experiments using the MGMT P140K resistance gene showed that gene-modified hematopoietic stem and progenitor cells could be enriched through drug selection in mice, dogs, and nonhuman primates. In nonhuman primates, selected cells contributed to long-term hematopoiesis, demonstrating that post-transfer selection can durably increase a modified population in another renewable tissue.^11,12

These studies also illustrate the clonal consequences of selection. Extended selection can reduce the repertoire of clones contributing to hematopoiesis and increase the proliferative burden carried by surviving clones, even when stable long-term function is maintained.¹³

The comparison does not show that the Repair Drive architecture can be transferred directly to blood or another organ. It supports the broader principle that a genetically defined survival advantage can amplify a modified cell population after gene transfer.

Application elsewhere would require several biological conditions. The target tissue would need a renewable or expandable cell population, a selectable function that could be linked tightly to the intended edit, and a conditioning mechanism that did not cause unacceptable organ damage. Selected cells would need sufficient time and spatial opportunity to repopulate the tissue, and developers would need credible methods for stopping selection and monitoring clonal behavior.

The hematopoietic system satisfies several of these requirements and has direct experimental precedents. Other tissues may contain renewable stem or progenitor compartments, but the current evidence does not establish that selective expansion can be safely or effectively applied to intestine, skin, lung, muscle, or the central nervous system.

The concept may therefore extend beyond the liver, but its applicability will depend less on the availability of an editor than on the regenerative and competitive biology of each tissue.

A Repair Drive–like treatment would derive its activity from the coordinated performance of several components: an editing system, a donor template, a selectable marker, a therapeutic sequence, and a conditioning agent. Its full biological effect would emerge over time as selected cells expanded.

This creates an unusual product-definition problem. Potency might need to reflect more than the frequency of targeted integration. It could also depend on selectable-marker function, linkage between the selectable and therapeutic sequences, resistance to conditioning, and the capacity of correctly targeted cells to expand while maintaining therapeutic expression.

Cassette integrity would be particularly important. The mouse study showed that structurally heterogeneous edits could survive selection if they retained functional FAH expression. Analytical strategies would therefore need to identify which intended and unintended integration structures remain selectable and how strongly each contributes to the final population.

Manufacturing variability could also have amplified consequences. A small difference in initial editing efficiency, donor integrity, vector ratio, or oligonucleotide activity might become a much larger difference in final tissue composition after repeated selection. Comparability assessments may ultimately need to consider not only immediate molecular performance but also the downstream expansion trajectory.

Nonclinical development would require models that reproduce the intended human disease environment. Studies in healthy young mice cannot fully predict selection in fibrotic, inflamed, metabolically impaired, or older human liver. Larger and longer-lived animals would be needed to evaluate durability, clonal diversity, tissue injury, and tumorigenic risk over more clinically relevant timeframes.

The most important next evidence would include deeper mapping of expanded clones, evaluation in diseased liver models, testing of rapid methods for reducing selection pressure, comparison of repeated mild conditioning with stronger short-term conditioning, and development of minimally invasive long-term monitoring.

This approach does not remove the need for better delivery, higher editing precision, or more efficient repair. It introduces a complementary possibility. In regenerative tissues, editing could establish the starting population, while controlled selection determines the eventual therapeutic population.

That possibility could make modest editing efficiencies useful in settings where they would otherwise be inadequate. It would also require gene therapy to manage not only the installation of a genetic change but the subsequent evolution of the treated tissue.

1. De Giorgi, Marco, et al. “In Vivo Expansion of Gene-Targeted Hepatocytes through Transient Inhibition of an Essential Gene.” Science Translational Medicine. 17: eadk3920 (2025).

2. Human Gene Therapy Products Incorporating Human Genome Editing: Guidance for Industry. U.S. Food and Drug Administration. Jan. 2024.

3. Paulk, Nicole K, et al. “Adeno-Associated Virus Gene Repair Corrects a Mouse Model of Hereditary Tyrosinemia In Vivo.” Hepatology. 51: 1200–1208 (2010).

4. Vonada, Samantha, et al. “Therapeutic Liver Repopulation by Transient Acetaminophen Selection of Gene-Modified Hepatocytes.” Science Translational Medicine. 13: eabg3047 (2021).

5. Brunner, Simon F, et al. “Somatic Mutations and Clonal Dynamics in Healthy and Cirrhotic Human Liver.” Nature. 574: 538–542 (2019).

6. Zhu, Meng, et al. “Somatic Mutations Increase Hepatic Clonal Fitness and Regeneration in Chronic Liver Disease.” Cell. 177: 608–621.e12 (2019).

7. Biasco, Luca. “In Vivo Tracking of Human Hematopoiesis Reveals Patterns of Clonal Dynamics during Early and Steady-State Reconstitution Phases.” Cell Stem Cell. 19: 107–119 (2016).

8. Biasco, Luca. “Integration Site Analysis in Gene Therapy Patients: Expectations and Reality.” Human Gene Therapy. 28: 1122–1129 (2017).

9. Berry, Charles C, et al. “INSPIIRED: Quantification and Visualization Tools for Analyzing Integration Site Distributions.” Molecular Therapy—Methods & Clinical Development. 4: 17–26 (2017).

10. Long Term Follow-Up After Administration of Human Gene Therapy Products: Guidance for Industry. U.S. Food and Drug Administration. Jan. 2020.

11. Beard, Brian C, et al. “Efficient and Stable MGMT-Mediated Selection of Long-Term Repopulating Stem Cells in Nonhuman Primates.” Journal of Clinical Investigation. 120: 2345–2354 (2010).

12. Gori, Jennifer L, et al. “In Vivo Selection of Autologous MGMT Gene-Modified Cells Following Reduced-Intensity Conditioning with BCNU and Temozolomide in the Dog Model.” Cancer Gene Therapy. 19: 523–529 (2012).

13. Ball, Claudia R, et al. “Stable Differentiation and Clonality of Murine Long-Term Hematopoiesis after Extended Reduced-Intensity Selection for MGMT P140K Transgene Expression.” Blood. 110: 1779–1787 (2007).