In vivo precision base editing to rescue mouse models of disease

Abstract

CRISPR base editing enables precise, irreversible base conversions without inducing double-stranded breaks (DSBs) and has gained significant attention in recent years. By converting cytosine to thymine (C→T) or adenine to guanine (A→G), base editors (BEs) efficiently correct pathogenic single-nucleotide variants (SNVs). This review examines in vivo mouse disease models—assessing editing efficiency, phenotypic rescue, and therapeutic potential across 66 studies. A key challenge in base editing is optimizing delivery. Most studies rely on split-intein dual adeno-associated virus (AAV) vectors due to BEs exceeding AAV packaging limits, though lipid nanoparticle (LNP) delivery is emerging. Editing efficiencies vary widely, influenced by enzyme design, delivery method, and sequence context. Many studies show significant functional gains, including extended survival in severe models such as FAH-deficient tyrosinemia type I and Hutchinson-Gilford progeria, restored dystrophin in Duchenne muscular dystrophy, and cognitive improvement in neurodegenerative models. Despite advantages such as reduced indels and increased precision, base editing is restricted to SNV correction and targets only a limited editing window relative to a protospacer adjacent motif (PAM) site. Advances in enzyme engineering, delivery strategies, and hybrid approaches incorporating prime editing could broaden its applications. As base editing evolves, its success in preclinical models positions it as a key player in next-generation gene therapies.

Graphical abstract

This review evaluates 66 in vivo studies applying CRISPR base editors to correct disease-causing SNVs in mouse models. It highlights therapeutic rescue outcomes, delivery challenges, and emerging tools such as LNPs and compact editors, which position base editing as a promising approach for precision medicine.

Introduction

CRISPR base editors (BEs) are genome editing tools that directly and irreversibly convert nucleotide bases—most commonly, cytosine to thymine (C→T) or adenine to guanine (A→G)—using a mechanism that does not feature double-stranded breaks (DSBs).¹ Before the development of BEs, precise editing of the genome primarily relied on DSB generation followed by homology-directed repair (HDR). Although HDR offers high fidelity, it is inefficient, limited to the S and G2 phases of the cell cycle, and notoriously difficult to induce in post-mitotic cells, which are often the cells implicated in human disease.² Moreover, DSB induction can lead to unwanted insertions and deletions (indels), causing missense, nonsense, and frameshift mutations that may result in dysfunctional proteins. Given that single-nucleotide variants (SNVs) are causative in many genetic diseases, BEs hold significant therapeutic potential.

Structurally, base editing enzymes consist of three main components: a single-guide RNA (sgRNA)-programmable Cas protein fused to a DNA deaminase enzyme.³ Mechanistically, the sgRNA directs the Cas protein to the target genomic locus, where RNA-DNA base pairing displaces a strand of DNA, forming an R-loop and exposing single-stranded DNA (ssDNA). This ssDNA becomes accessible for enzymatic deamination. DNA BEs are primarily categorized into two types: cytosine base editors (CBEs) and adenine base editors (ABEs). CBEs mediate cytosine-to-thymine (C→T) conversions within a ∼5-nucleotide (nt) editing window,⁴ while ABEs perform adenine-to-guanine (A→G) transitions within a ∼4–9 nt window.⁵ As with all CRISPR-Cas systems, efficient targeting requires the presence of a protospacer adjacent motif (PAM) near the target site, which constrains editable loci. The ability to edit the reverse strand (i.e., G→A and T→C) further enhances the versatility of base editing.

The first generation of CBEs (CBE1), developed by Komor and colleagues in 2016, was comprised a catalytically inactive Cas9 (dCas9) fused to the cytidine deaminase APOBEC-1.^4,6 Although CBE1 facilitated cytosine-to-uracil (C→U) conversions at efficiencies of 0.8%–7.7% in vitro, it exhibited poor final editing outcomes (C:G→T:A conversions) in cells due to uracil excision by the endogenous DNA repair enzyme uracil DNA glycosylase (UDG).⁴ To address this, a second-generation construct (BE2) incorporated a uracil glycosylase inhibitor (UGI), increasing editing efficiency to approximately 20%.⁴ The third-generation editor (BE3) replaced dCas9 with a nickase version of Cas9 (Cas9n), which cut the non-edited strand. This approach leveraged endogenous mismatch repair pathways to preferentially retain the edited base, further increasing editing efficiency by approximately 6-fold compared to CBE2.^1,4

Subsequent iterations aimed to enhance efficiency and minimize indels further. The fourth-generation cytosine editor (CBE4) added extended peptide linkers, a second copy of UGI, and Gam—a bacteriophage Mu protein that binds DNA breaks—to reduce unintended indel formation.⁶ CBE4 showed substantial improvements in editing efficiency, achieving up to 56.25% correction in mouse embryos in vivo.⁷ BE4max (a CBE) was later designed by reconstructing ancestral versions of the deaminase enzyme and achieved up to 3-fold higher editing efficiencies in HEK293T cells.⁸

Recognizing that many pathogenic mutations involve adenine residues, the development of ABEs occurred in parallel. A major challenge was the absence of naturally occurring adenine deaminases capable of acting directly on DNA. Gaudelli et al. addressed this limitation through directed evolution of the E. coli tRNA adenine deaminase TadA, an enzyme homologous to APOBEC-1.⁵ Initial constructs (ABE1.2) exhibited modest editing efficiencies (∼3.2%) across multiple human genomic sites. However, further iterative rounds of evolution yielded the significantly improved ABE7.10, which achieved editing efficiencies averaging 53% across 17 genomic loci with minimal off-target activity.⁵ The latest generation ABEs include ABEmax⁸ and ABE8e,⁹ which were both developed by David Liu’s laboratory (Broad Institute). ABEmax is a variation of ABE7.10 and includes modifications in the nuclear localization signal sequences and codon optimization. These changes yielded a 46% editing efficiency for ABEmax compared to 6.5% for ABE7.10 in cell-based assays.⁸ The development of ABE8e utilized phage-assisted continuous evolution of the deoxyadenosine deaminase, testing against different Cas enzymes showed that ABE8e has an average 6-fold higher editing efficiency compared to ABE7.10.⁹ Collectively, ABEs have broadened the precision genome-editing toolkit and hold therapeutic promise for correcting disease-causing G→A point mutations.

The structure, development, and scope of the classical CRISPR DNA BEs is summarized in Figure 1.

Figure 1.

Evolution of base editors: successive modifications to cytosine base editors led to the development of adenine base editors and further optimized variants, including ABEmax and BEmax

Shortly after its development, base editing technology was harnessed to introduce single-base substitutions in mouse embryos.¹⁰ Although base editing is typically less efficient than CRISPR-mediated gene disruption and less versatile than HDR, it remains a powerful method for generating precise disease-associated mutations. Subsequent studies have provided detailed methodologies for employing BEs to create mouse models with specific point mutations.^11,12

BEs have proven valuable for generating genetically modified mice, but their size and the challenges of in vivo delivery remain significant hurdles for therapeutic applications. The combined size of the deaminase functional domains and SpCas9 exceeds the 4.5 kb packaging limit of adeno-associated virus (AAV) vectors, the most widely used delivery system for gene therapies. To overcome this limitation, several dual-vector strategies have been developed, including fragmented, overlapping, trans-splicing, and hybrid approaches.¹³ Currently, the most widely adopted method for CRISPR nucleases relies on split-inteins.¹⁴ A split-intein system is a self-splicing protein mechanism in which two separate fragments assemble, fold into an active enzyme, and excise themselves, linking the surrounding protein segments via a native peptide bond. The first application of this system in base editing was for rescuing a murine phenylketonuria model.¹⁵ This approach used the split-intein moiety from Nostoc punctiforme (Npu) to deliver a SaKKH-BE3 transgene.

While split-intein systems enable the delivery of large SpCas9-based DNA BEs, an alternative strategy is to bypass DNA editing constraints altogether by directly modifying RNA. RNA-guided Cas13 enzymes facilitate programmable adenine-to-inosine (A→I) editing with RNA editing for programmable A-to-I replacement (REPAIR) and cytosine-to-uracil (C→U) editing with RNA editing for specific C-to-U exchange (RESCUE).^1,16 Both constructs have undergone multiple rounds of optimization, enhancing RNA-binding affinity, improving on-target editing, and minimizing off-target effects. Therapeutically, compact RNA BEs offer the advantage of smaller construct sizes. For example, Cas13bt enzymes (∼2.3 kb) are significantly smaller than SpCas9-based BEs (∼5.2–5.8 kb, including deaminase domains), allowing packaging within a single AAV vector.^17,18 However, because RNA editing is inherently transient, continuous enzyme expression is required, which may limit its suitability for applications requiring sustained therapeutic effects.

The core base editing enzymes—CBEs, ABEs, and Cas13-derived systems—have been widely applied to achieve functional rescue in preclinical mouse models of disease. However, a broader suite of alternative CRISPR-based editors has also been developed over the past several years, as illustrated in Figure 2. Cytosine-to-guanine BEs (CGBEs) are modified CBEs in which the UGI is replaced with base excision repair proteins or engineered UDGs to promote C→G transversions.^19,20,21 Editing efficiency is highly sequence-context dependent, and the ratio of desired C→G to byproduct C→T conversions is often suboptimal, limiting in vivo applicability.²² Similarly, adenine transversion base editors (AYBEs) aim to introduce A→C and A→T edits. These systems tether a methylpurine DNA glycosylase to an ABE scaffold to remove hypoxanthine and promote alternative repair outcomes beyond the canonical A→G transition.²³ However, AYBEs currently offer less control over transversion outcomes compared to conventional ABEs. With further engineering, both CGBE and AYBE platforms could extend the therapeutic reach of base editing to include transversion mutations.

Figure 2.

Alternative CRISPR editors: CGBEs, AYBEs, and RNA editors provide additional capabilities and editing outcomes to classical ABE and CBEs systems

Scope of the review

This review focuses on in vivo preclinical studies that utilize CRISPR base editing to rescue disease phenotypes in mouse models. The inclusion and exclusion criteria defining the scope of these therapeutic studies are shown in Table 1.

Table 1.

Criteria for in vivo base editing therapeutic rescue in mouse models of disease

Disease models	Studies must involve mouse models of disease. Models generated solely to induce disease or modifications in healthy mice (e.g., Pcsk6 studies in wild-type mice) are excluded
Editing modalities	All base editing modalities are included, encompassing both cytosine base editors (CBEs) and adenine base editors (ABEs), whether they are conventional systems or novel/compact variants
Phenotypic outcomes	Although most studies report on phenotypic rescue, the review will also include studies that demonstrate high levels of base editing without addressing phenotypic outcomes
Ex vivo vs. in vivo	Studies involving grafting of humanized cells are included only if the base editing is performed in vivo after transplantation; studies focused on ex vivo base editing followed by cell transplantation (e.g., many hematopoietic studies) are excluded
Complementary technologies	Studies may incorporate other editing technologies (e.g., prime editing) but must include a base editing component; importantly, in vivo rescue models that utilize gene disruption or other repair modalities without base editing are excluded
Delivery systems	The review is delivery system agnostic; any delivery method is acceptable if the editing occurs in vivo
Model systems	The focus is exclusively on mouse models, excluding studies in other preclinical systems (e.g., zebrafish), non-human primates, or clinical trials
Publication type	Only peer-reviewed articles are included; conference abstracts and non-peer-reviewed sources are excluded
Temporal coverage	The review covers studies published prior to February 2025

Overall findings

An initial PubMed database search identified over 250 papers; however, a preliminary review of abstracts for papers that indicated both precision base editing and the use of mouse models of disease narrowed the pool to 94 papers. Following a full-text analysis, 66 papers were deemed suitable for inclusion (see Table 2). Most of the later exclusions were due to studies focusing on ex vivo repair and cell transplantation or involving base editing outside the context of a disease model.

Table 2.

Summary of in vivo preclinical studies using CRISPR base editing in mouse models of disease ordered by base edited gene symbol

Reference	Journal article	Base edited gene symbol	Base edited gene name	Mutation being edited	Enzyme used for editing	Split intein	Mouse model	Vector/delivery	Dosing	Result of in vivo rescue (editing, phenotype)
Gopalappa et al.24	In vivo adenine base editing rescues adrenoleukodystrophy in a humanized mouse model (2024)	ABCD1	ATP-binding cassette subfamily D member 1	c.1534G>A, p.G512S pathogenic mutation	NG-ABE8e(V106W)	yes	humanized ALD (hALD) mouse model with a human ABCD1 pathogenic variant	tail vein injection of AAV-PHP.B	low dose: 1.0 × 1012 vg of each AAV vector (N-intein and C-intein); high dose: 6.4 × 1012 vg of each AAV vector (N-intein and C-intein)	brain: up to 5.5% (gDNA), up to 10% (cDNA); spinal cord: up to 5.1% (gDNA), up to 8.4% (cDNA); liver: ∼4.5%; adrenal gland: ∼2%; kidney/testis: <0.2%	significant reduction in very long chain fatty acids; no significant brain toxicity or histological abnormalities
Alves et al.25	In vivo treatment of a severe vascular disease via a bespoke CRISPR-Cas9 base editor (2024)	ACTA2	alpha actin isotype 2	c.536G>A, p.R179H	ABE8e-eVRQR	yes	knockin Acta2 R179H multisystem smooth muscle dysfunction syndrome (MSMDS) mouse model	neonatal injection of AAV-PR and AAV9 vectors	P3 injection: AAV-PR-ABE; P14 injection: 8 × 1010 vg of N- and C-terminal AAV9 vectors	detectable R179H correction in in vivo tissues, higher when administered at P3 compared to P14	lifespan increase from (6 weeks to median 22.6 weeks); improved vascular structure, smooth muscle contractility and cerebral vasculature; restored neurovascular anatomy and myelination; rescued body weight, locomotion, and behavioral performance; improved kidney, lung, and liver morphology
Miskalis et al.26	SPLICER: a highly efficient base editing toolbox that enables in vivo therapeutic exon skipping (2024)	APP (exon 17)	amyloid beta precursor protein	skipping of exon 17 to reduce Aβ42 peptide formation	ABEs, CBEs	yes	B6.129-Tg(APPsw)40Btla/Mmjax (transgenic mouse model)	stereotaxic injection of AAVrh10 vectors into hippocampus	AAVrh10-pAAV-CAG-N-ABE8e, pAAVrh10-CAG-C-ABE8e (5 × 109 vg) and AAVrh10-pAAV-CAG-KASH (5 × 108 vg)	6.4%–23.4% editing efficiencies in vivo; enabled 20% exon skipping in vivo	no reported phenotypic rescue in mice (reduction in Aβ42 peptide in cell model)
Li et al.27	In vivo HSPC gene therapy with base editors allows for efficient reactivation of fetal gamma-globin in beta-YAC mice (2021)	BCL11A enhancer of HBG1/2 promoter	regulator of γ-globin expression	−113A>G HPFH mutation in the HBG1/2 promoter	ABE (enzyme not specified)	no	b-YAC1/2/CD46 transgenic mice (for human HBG genes and adenovirus targeting)	intravenous (i.v.) injection of HDAd5/35++ vectors (HDAd-ABE-sgHBG-2)	not stated	∼20% base conversion in total bone marrow cells	γ-globin expression in peripheral blood increased to ∼43% of total hemoglobin by week 16 post-transduction; no hematological issues or toxicity
Lebek et al.28	CRISPR-Cas9 base editing of pathogenic CaMKIIdelta improves cardiac function in a humanized mouse model (2024)	CAMK2D	CaM kinase II beta subunit	c.A841G, c.A844G, c.A848G (oxidation-sensitive Met residues in CAMK2D regulatory domain)	ABE8e(V106W)-SpCas9 and ABE8e(V106W)-SpRY	yes	humanized CAMK2D knockin mice	intracardiac injection of MyoAAV-2A carrying base editing components	1.5 × 1011 vg/kg body weight (equal amounts of N- and C-term for IR + sgRNA1 and IR + sgRNA2)	for sgRNA1: 36.2% (gDNA), 83.2% (cDNA) at c.A841G; for sgRNA2: 37.0%, 27.8%, 36.4% (gDNA) and 82.8%, 84.8%, 83.4% (cDNA) at c.A841G, c.A844G, and c.A848G, respectively	edited mice recovered cardiac function after IR injury, showing significantly improved fractional shortening, LVIDd, and LVIVd compared to controls
Xiao et al.29	Engineered IscB-omegaRNA system with expanded target range for base editing (2024)	DMD	dystrophin	ΔmE5051 (KIhE50/Y)	IscB.m16∗-CBE-v2	no	ΔmE5051 (KIhE50/Y) (C57BL/6J background)	intramuscular injection of AAV9	30 μL of AAV9 (2.5 × 1011 vg), single dose	7% conversion of G-to-H and up to 30% level of exon 51 skipping	restoration of dystrophin protein in diseased mice
Lin et al.30	Adenine base editing-mediated exon skipping restores dystrophin in humanized Duchenne mouse model (2024)	DMD	dystrophin	exon 50 and exon 51 deletions	ABE2	yes	DMDΔmE5051 (KIhE50/Y) humanized DMD model, human exon 50 replacing mouse exon 50/51	systemic delivery (intraperitoneal [i.p.], i.v.) of dual-AAV9-BE2 expressing vectors	1 × 1014 vg/kg	A-to-G editing rates: heart: 16.41%, DI: ∼50%, TA: ∼50%; no off-target editing in potential loci	dystrophin expression restored (up to 96% in heart, ∼50%–58% in diaphragm and tibialis anterior); ALT levels normalized; blood urea nitrogen levels unchanged
Li et al.31	Engineering TadA ortholog-derived cytosine base editor without motif preference and adenosine activity limitation (2024)	DMD	dystrophin	exon 55 skipping in DMD	aTdCBE (engineered TadA variant CBE)	yes	DMDΔE54mdx humanized DMD mouse model with exon 54 deletion	intramuscular injection of dual-AAV9 vectors expressing aTdCBE	2.5 × 1011 vg of each vector (50 μL injected into TA muscle of 3-week-old mice).	over 40% DNA base editing efficiency at the exon 55 splice acceptor site (SAS)	exon 55 skipping led to dystrophin restoration, improvement in muscle histology, creatine kinase (CK) activity, and motor function
Jin et al.32	Correction of human nonsense mutation via adenine base editing for Duchenne muscular dystrophy treatment in mouse (2024)	DMD	dystrophin	c.4174C>T, p.Gln1392∗ (exon 30, DMDE30mut model); c.2977C>T, p.Gln993∗ (exon 23, DMDE23mut model)	SpG-ABE (ABEv1 and ABEv2 variants)	yes	DMDE23mut (humanized mouse with exon 23 of human DMD harboring c.2977C>T); DMDE30mut (humanized mouse with exon 30 of human DMD harboring c.4174C>T)	intramuscular injection of adult mice and neonatal (systemic) delivery with AAV9 vectors	intramuscular: 5 × 1011 vg/leg per virus (DMDE30mut and DMDE23mut); intraperitoneal (systemic): 1 × 1011 vg per virus (DMDE30mut)	local intramuscular injection achieved up to 12.7% genomic and 68.3% cDNA editing; systemic delivery showed the highest genomic (37.2%) and cDNA (89.1%) editing in the heart, with lower efficiencies in diaphragm and skeletal muscle	therapy restored dystrophin expression in skeletal and cardiac muscle and improved muscle function in rotarod tests
Wang et al.33	Develop a compact RNA base editor by fusing ADAR with engineered EcCas6e (2023)	DMD	dystrophin	c.4174C>T, p.Q1392X	ceRBE with hADAR2ddE488Q (RNA base editor)	no	DMDQ1392X (humanized exon 30 replacement)	intramuscular injection of AAV9 into the TA	5 × 1011 vg (50 μL)	68.3% A-to-I conversion in TA tissue	49.3% dystrophin restoration by immunoblot; 68.1% dystrophin restoration by histological staining
Li et al.34	Mini-dCas13X-mediated RNA editing restores dystrophin expression in a humanized mouse model of Duchenne muscular dystrophy (2023)	DMD	dystrophin	c.4174C>T, p.Gln1392∗	mini-dCas13X (used for RNA adenine base editing, mxABE)	no	DMDE30mut (humanized mouse model of DMD)	intramuscular injection into the adult TA or systemic (i.p.) injection into P3 neonates of the AAV9-mxABE vector	intramuscular: 3 × 1011 vg/mouse; systemic: 1 × 1012 vg/mouse	editing efficiency: up to 70% in tibialis anterior (TA) muscle and 77% in heart tissue after systemic AAV9-mxABE delivery.	dystrophin expression restored to 50% of wild-type levels in TA muscle, and significant improvements in muscle function, including grip strength and reduced fibrosis
Chai et al.35	Single swap editing for the correction of common Duchenne muscular dystrophy mutations (2023)	DMD	dystrophin	exon 44 deletion (base editing allows for exon skipping)	ABE8eV106W fused to nSpCas9	yes	DMD ΔEx44 mouse model (deletion of exon 44)	intramuscular injection (in P12 mice) or systemic injection (in P2 neonates) of AAV9 vectors	intramuscular injection: 5 × 1010 vg/leg of each viral half (total 1 × 1011 vg); neonatal systemic injection: 1.5 × 1014 vg/kg (low dose) and 3 × 1014 vg/kg (high dose)	in vivo intramuscular editing (TA muscle): 29.5% editing; in vivo systemic editing: 5.5% (low dose) and 8.1% (high dose) editing in TA muscle, 22.0% (low dose) and 26.2% (high dose) in the heart	the treatment restored dystrophin expression in both skeletal (19.3%–31.0%) and cardiac muscle (53.7%–59.7%), improving muscle structure and reducing markers of dystrophy such as fibrosis and inflammation
Xu et al.36	Efficient precise in vivo base editing in adult dystrophic mice (2021)	Dmd	dystrophin	premature stop codon (G-to-T transversion) at codon 23	iABE-NGA (NG-targeting adenine base editor)	yes	mdx4cv mice	tail vein (i.v.) injection or intramuscular injection of the right gastroc of dual-AAV9 vectors	1 × 1014 vg/kg (systemic) or 2 × 1011 vg in 25 μL (gastroc)	32.6% T-to-C editing in heart at 10 weeks; 84.6% T-to-C editing in heart at 10 months	at 10 months, dystrophin was nearly fully restored in the heart (95.9 ± 1.6%) and at 8.0 ± 2.6% of WT levels in the gastrocnemius muscle, with reduced fibrosis, fewer centrally nucleated fibers, and improved muscle function
Li et al.37	Therapeutic exon skipping through a CRISPR-guided cytidine deaminase rescues dystrophic cardiomyopathy in vivo (2021)	Dmd	dystrophin	4-bp deletion in exon 4 causing frameshift and premature stop codon	enhanced target-AID (eTAM) - a CBE	no	C57BL/6 DmdE4∗ mice	i.p. injection at P2-P3 (dual-AAV9 vectors but not split intein)	1.1 × 1012 vg per neonatal mouse, 1:1 vg ratio between AAV9-eTAM and AAV9-sgRNA vectors	exon 4 skipping: ∼60% in Dmd transcripts in heart	dystrophin protein restoration: ∼86% of WT levels in heart; improved cardiac function, skeletal muscle function and reduced muscle fatigue; extended lifespan (partial rescue)
Chemello et al.38	Precise correction of Duchenne muscular dystrophy exon deletion mutations by base and prime editing (2021)	Dmd	dystrophin	exon 51 deletion (ΔEx51)	ABEmax and NG-ABEmax	yes	ΔEx51 mice (mimics DMD pathology)	intramuscular injection of AAV9 dual vectors	5 × 109 viral genomes (vg) per leg for each of the two AAV9 vectors	on-target editing at A14 was 35.0%, with bystander editing at A12 (6.7%) and A18 (10.7%), and no significant off-target editing detected at the top eight predicted sites	exon skipping induced by base editing restored dystrophin expression in 96.5% of myofibers and improved muscle histology, reducing fibrosis, necrosis, and centralized nuclei, significantly alleviating the DMD phenotype
Ryu et al.39	Adenine base editing in mouse embryos and an adult mouse model of Duchenne muscular dystrophy (2018)	Dmd	dystrophin	p.Q871X (referenced but not explicitly stated)	ABE7.10	yes	Dmd knockout mice (p.Q871X)	intramuscular injection of dual AAV9 vectors	1 × 1013 viral genomes in 40 μL saline	3.3% A-to-G base substitutions in the Dmd gene 8 weeks post-injection	17% of myofibers showed dystrophin expression; sarcolemmal localization of nNOS was seen, suggesting functional recovery of muscle cells
Yoon et al.40	CRISPR-mediated ablation of TP53 and EGFR mutations enhances gefitinib sensitivity and anti-tumor efficacy in lung cancer (2024)	EGFR, TP53	epidermal growth factor receptor, p53	EGFR-T790M, TP53-R273H	ABE7.10	no	H1975 lung cancer xenograft in 6-week-old male nude mice	intratumoral injection of adenovirus (Ad) carrying ABE and SNP-specific sgRNAs	5 × 1010 viral particles (VP) in 30 μL PBS on days 0, 2, and 4; gefitinib (100 mg/kg in 10% DMSO in PBS) orally from days 1–5	15.9% base editing (Ad/ABE-TP53) and 6.1% base editing (Ad/ABE-TP53-EGFR)	increased gefitinib sensitivity; significant tumor suppression compared to gefitinib alone
Bock et al.41	In vivo base editing of a pathogenic Eif2b5 variant improves vanishing white matter phenotypes in mice (2024)	Eif2b5	eukaryotic translation initiation factor 2B subunit 5	p.Arg191His (R191H)	ABEmax, pCbh-ABE8e	yes	Eif2b5R191H knockin mouse model of vanishing white matter (VWM)	intracerebroventricular (i.c.v.) injection of AAV-PHP.eB at P1 (postnatal day 1)	4.8 × 1010 vg (pCMV-ABEmax) or 8.0 × 1010–9.6 × 1010 vg (pCbh-ABE8e) per animal via i.c.v. injection	cortex: 45.9% correction of the R191H variant; note bystander editing and inefficient delivery to deep brain regions	partial rescue in female mice: increased body weight (16% improvement) and improved grip strength (close to wild-type levels); balance and motor skills remained impaired or worsened
Yun et al.42	Engineered CRISPR-base editors as a permanent treatment for familial dysautonomia (2024)	ELP1	elongator acetyltransferase complex subunit 1	T-to-C mutation in intron 20 of ELP1	TadCBEd-SpG	yes	TgFD9 mice carrying the human ELP1 transgene with the intron 20 T>C mutation	intravitreal or temporal facial vein injection of AAV2 and AAV9 vectors	AAV2 vector cocktail (45% N-term, 45% C-term, 10% eGFP) at 1.5 × 109 or 1 × 1010 vg per eye to adult mice. AAV9 vectors systemically delivered to neonatal mice at 3 × 1010 or 1.6 × 1011 vg/g	retinal cells: 9% on-target editing, liver: 40%; brain: 5%	no reported phenotypic rescue in mice (hiPSC-derived neurons only)
Hiramoto et al.43	PAM-flexible Cas9-mediated base editing of a hemophilia B mutation in induced pluripotent stem cells (2023)	F9	coagulation factor IX	c.947T>C (p.I316T) hemophilia B mutation	SpCas9-NG	yes	KI mice expressing F9 cDNA with or without the p.I216T mutation	i.p. of AAV8 dual vectors	1 × 1011 viral genomes, administered in 10 μL to neonatal mice	1.74% correction of the mutation in the liver	significantly increase in plasma coagulation factor IX (FIX) levels and restored FIX activity in treated KI mice, compared to untreated controls
Yang et al.44	Engineering APOBEC3A deaminase for highly accurate and efficient base editing (2024)	Fah	fumarylacetoacetate hydrolase	C-to-T conversion at C1 position in FahNS/NS allele	haA3A-CBE-G, haA3A-CBE-VA	yes	FahNS/NS HT1 mouse (a tyrosinemia model)	tail vein injection (i.v.) of AAV8 (split intein vectors) or LNP-mRNA	2 × 1012 vg total (1 × 1012 vg per half) per mouse; 3 mg/kg of LNP-mRNA (single injection)	AAV: 40.7% editing at 2 weeks; 54.5% at 4 weeks; FAH protein restored; survival and weight gain after drug withdrawal; biomarkers indicate improved liver function; increased bystander edits by 9 weeks	LNP: haA3A-CBE-G: 19.6% editing at 1 week; haA3A-CBE-VA: 33.0% editing at 1 week; lower bystander edits than AAV; stable editing at 6 weeks post-injection
Guo et al.45	Engineered IscB-omegaRNA system with improved base editing efficiency for disease correction via single AAV delivery in mice (2024)	Fah	fumarylacetoacetate hydrolase	G-to-A mutation in Fah exon 8	EnIscB-derived adenine base editor (miABE)	no	Fah−/− mouse (tyrosinemia model)	tail vein injection (i.v.) of AAV8	5 × 1011 vg/mouse	15% correction of Fah mutation by miABE; 50% indel formation in Hpd by EnIscB; 35% STOP mutations in Hpd by miCBE (C-to-T base editing)	weight recovery after NTBC withdrawal; restored FAH protein expression in miABE-treated mice; reduced HPD protein expression in miCBE-treated mice; improved liver histology
Zhang et al.46	Adenine base editing in vivo with a single adeno-associated virus vector (2022)	Fah	fumarylacetoacetate hydrolase	A>G splice-site mutation in Fah exon 8	Nme2-ABE8e (compact ABE using the Nme2Cas9 variant)	no	FahPM/PM mice, which have a mutation in Fah leading to tyrosinemia (HT1, a liver disease).	hydrodynamic tail vein injection and intrastriatal injection for AAV9	hydrodynamic tail vein injection: 4 × 1011 viral genomes (vg) per mouse; intrastriatal injection: 1 × 1010 vg per mouse	before NTBC withdrawal, the Nme2-ABE8e-U6 group had 4.58% FAH+ hepatocytes, while the Nme2-ABE8e-miniU6 group had 1.71% FAH+ hepatocytes; final editing efficiency (at 1 month) in the AAV9-Nme2-ABE8e-U6 group reached 6.49% FAH+ hepatocytes	mice injected with the Nme2-ABE8e-U6 plasmid exhibited weight gain after NTBC withdrawal
Yang et al.47	Amelioration of an inherited metabolic liver disease through creation of a de novo start codon by cytidine base editing (2020)	Fah	fumarylacetoacetate hydrolase	mutation affecting the start codon (ATG) in Fah	BE4max	yes	FahNS/NS (a tyrosinemia model)	tail vein (i.v.) injection of rAAV8 vectors	1 × 1012 vg per rAAV per mouse (high dose) or 1 × 1010 vg per rAAV per mouse (low dose), 1:1 ratio of two vectors	day 7 (pre-selection): 2.7% editing in high-dose group, with only 0.7% restoring the ATG start codon; after 2 months (post-selection): 16.7% correctly edited alleles in high-dose group; 15.6% in low-dose group.	FAH-positive hepatocytes expanded clonally and repopulated most of the liver after NTBC withdrawal; mice showed gradual weight gain, indicating improved liver function
Song et al.48	Adenine base editing in an adult mouse model of tyrosinemia (2019)	Fah	fumarylacetoacetate hydrolase	A>G splice-site mutation in Fah exon 8	ABE6.3 for hydrodynamic delivery; RA6.3 (optimized ABE6.3) for LNP delivery	no	Fahmut/mut mice (HTI model)	hydrodynamic tail vein injection of plasmid or tail vein injection of LNPs	plasmid: 30 μg ABE plasmid +30–60 μg sgFah plasmid; LNP: 1 mg/kg RA6.3 mRNA +0.5 mg/kg sgRNA, injected every 3 days for 3–4 doses	plasmid: 9.5% (A>G at position 9); LNP: 0.44% corrected hepatocytes (under non-selective conditions)	plasmid: restored Fah expression, generated patches of Fah+ hepatocytes, partially restored correct splicing, and prevented weight loss; mice survived 106 days without NTBC
Rossidis et al.49	In utero CRISPR-mediated therapeutic editing of metabolic genes (2018)	Fah	fumarylacetoacetate hydrolase	Q352X (nonsense mutation)	BE3 (base editor 3), a cytosine base editor	no	Fah−/− knockout mice (a model for hereditary tyrosinemia type 1); (note: base editing to inactivate Pcks9 in wild-type mice was also performed)	in utero injection via the vitelline vein of E16 fetuses of an adenoviral (Ad) vector	1 × 108–1 × 109 viral particles per fetus	∼15% base editing in liver at 2 weeks of age, with 37%–40% at 1 month and 3 months	rescue of the lethal phenotype in Fah−/− mice with 89% survival at 3 months, normal weight gain, and improved liver function compared to control (Ad-BE3-Null) recipients
Arnaoutova et al.50	Base-editing corrects metabolic abnormalities in a humanized mouse model for glycogen storage disease type-Ia (2024)	G6PC1	glucose-6-phosphatase-α	c.247C>T (p.R83C)	ABE (enzyme not specified)	no	huR83C mice (3-week-old)	temporal vein injection to newborns, retro-orbital sinus injection of LNPs	low dose (301 L, 0.75 mg/kg); high dose (301H, 1.5 mg/kg)	up to ∼60% correction of G6PC1-R83C in liver cells	restored blood glucose control; corrected metabolic abnormalities; enabled survival beyond 1 year (untreated controls had 61% mortality at 3 weeks)
Lian et al.51	Bone-marrow-homing lipid nanoparticles for genome editing in diseased and malignant hematopoietic stem cells (2024)	HBB	β-globin	p.Glu6Val sickle cell mutation	ABE8e_NRCH	no	Townes (HBBS/S) mice	i.v. injection of bone-marrow-homing LNPs	3 mg/kg total dose, with a 21 ratio of mRNA/sgRNA given twice weekly for 2 weeks	2.43% conversion to Makassar allele with base editing, 5.2% indels (insertions and/or deletions) after CRISPR-Cas9 delivery	reduced sickling in erythrocytes to 25% of untreated group after base editing in HSPCs
Li et al.52	Introducing a hemoglobin G-Makassar variant in HSCs by in vivo base editing treats sickle cell disease in mice (2024)	HBB	β-globin	GTG (sickle mutation) to GCG (Makassar variant)	ABE8e-NRCH	no	CD46/Townes SCD mouse model	i.v. injection of HDAd (helper-dependent adenovirus)-Maka-v3	4 × 1010 viral particles per mouse.	in vivo: 88% editing efficiency in HSCs from SCD mice; 24.5% Makassar variant in long-term repopulating HSCs	normalized RBC count and hemoglobin level, decreased sickling of RBCs, reduced reticulocyte count, and splenomegaly.
Li et al.53	In vivo base editing by a single i.v. vector injection for treatment of hemoglobinopathies (2022)	HBG1, HBG2	γ-globin subunit encoding genes	−113 A>G (HPFH mutation in γ-globin promoter)	ABE8e	no	CD46/β-YAC mice (human β-globin locus, carrying the human CD46 genomic locus)	retro-orbital (i.v.) injection of Ad vector (HDAd5/35++) after AMD3100 mobilization regime	4 × 1010 viral particles per mouse	>60% conversion at the −113 A site in HBG1/2 promoters, with 30% γ-globin expression of β-globin in 70% of erythrocytes.	30% γ-globin expression in RBCs at week 4, increasing to 70% at week 16.
Smekalova et al.54	Cytosine base editing inhibits hepatitis B virus replication and reduces HBsAg expression in vitro and in vivo (2024)	HBV genes (HBs and Precore)	hepatitis B virus genes	stop codons introduced in HBV genes (HBs and Precore)	CBE4	no	C3H male mice (HBV minicircle model)	hydrodynamic delivery of HBV minicircle DNA, then tail vein injection of LNP with CBE mRNA and gRNA	not stated	BE4/g37 introduced stop codon W156∗ in HBs with ∼30% efficacy; BE4/g40 introduced stop codon W28∗ in Precore with 42% efficacy	>3log10 reduction in serum HBV DNA, >2log10 reduction in HBsAg, 4/5 mice showed HBsAg loss
Kabra et al.55	Nonviral base editing of KCNJ13 mutation preserves vision in a model of inherited retinal channelopathy (2023)	KCNJ13	potassium inwardly rectifying channel subfamily J member 13	p.W53X	ABE8e	no	Kcnj13W53X/+ΔR mouse model (heterozygous knock in with deletion component)	subretinal injection of silica nanocapsules (SNC-PEG-ATRA)	2–3 μg ABE8e mRNA (nanocapsule amount not stated)	16.8% retinal pigment epithelium (RPE) editing at 3 μg dose; 9.5% RPE editing at 2 μg dose; no significant indel formation or off-target A>G substitutions.	optical coherence tomography and multifocal electroretinography showed preserved retinal structure and regionalized recovery; no significant immune response
Yang et al.56	Adenine base editor-based correction of the cardiac pathogenic Lmna c.1621C > T mutation in murine hearts (2024)	Lmna	lamin A/C	c.1621C>T (cardiomyopathy-causing mutation)	ABE7.10, ABE8e, TadA8e-SauriCas9	both	LmnaRC/RC (cardiomyopathy model)	neonatal injection (P1) of AAV9 vectors	2 × 1011 vg/g or dual or single vectors	2.5%–8% correction with variable amounts of bystander effect	no information on phenotypic rescue
Koblan et al.57	In vivo base editing rescues Hutchinson-Gilford progeria syndrome in mice (2021)	LMNA	lamin A/C	c.1824C>T mutation (p.G608G)	ABE7.10max-VRQR	yes	C57BL/6-tg(LMNA∗G608G)HClns/J “progeria mice”	neonatal injection at P3 and P14 of AAV9 dual vectors (ABE and LMNA-targeting sgRNA)	5 × 1010 viral genomes (vg) for P3 injections and 5 × 1011 vg for P14 injections (total of 1 × 1011 vg and 1 × 1012 vg per mouse, respectively).	10%–60% base editing efficiency across various tissues (notably aorta and heart), with higher efficiencies in P14-treated mice.	corrected LMNA RNA splicing, reduced progerin protein, rescued vascular smooth muscle cell counts, prevented adventitial fibrosis, and extended the median lifespan of treated progeria mice from 215 to 510 days
Gee et al.58	CRISPR base editing-mediated correction of a tau mutation rescues cognitive decline in a mouse model of tauopathy (2024)	MAPT	gene encoding Tau	P301S (A-to-G correction)	NG-ABE8e	trans-splicing	PS19 transgenic mice expressing human MAPT P301S	intracranial injection of trans-splicing AAV9 (tsAAV-NG-ABE8e)	not stated	A-to-G editing at the MAPT-P301S site was 5.7% in hippocampal DNA; bystander editing at A13 occurred at 0.35%.	treatment reduced insoluble tau aggregation and phosphorylated tau levels without altering total MAPT expression; cognitive function improved in the Morris water maze and passive avoidance test
Reichart et al.59	Efficient in vivo genome editing prevents hypertrophic cardiomyopathy in mice (2023)	Myh6	α-cardiac myosin heavy chain	p.R403Q (dominant missense mutation)	ABE8e	yes	R403Q-129SvEv (slow-onset hypertrophic cardiomyopathy) and R403Q-129SvEv/S4 (rapid-onset hypertrophic cardiomyopathy)	intra-thoracic injection of AAV9 into neonatal and adult mice	1.25 × 1013 vg/kg (of each AAV9 vector) in a single dose	genomic DNA (LV tissue): ∼16.0% of cardiac cells edited; transcript-level (Myh6 RNA): LV ∼81%, RV ∼61%, LA ∼26%, RA ∼34%	treated mice maintained normal heart structure and function, preventing hypertrophy and fibrosis, with effects lasting at least 32 weeks
Chai et al.60	Base editing correction of hypertrophic cardiomyopathy in human cardiomyocytes and humanized mice (2023)	MYH7	β-myosin heavy chain	c.1208G>A (p.R403Q), a pathogenic variant associated with hypertrophic cardiomyopathy	ABEmax	yes	humanized Myh6h403/+ mouse model (with MYH7 c.1208G>A variant integrated into the Myh6 gene)	intrathoracic injection of AAV9 vectors at P0	low dose: 4 × 1013 vg/kg (each AAV9); high dose: 1.5 × 1014 vg/kg (each AAV9)	32.3% editing of the target pathogenic adenine in heart tissue	high-dose ABE treatment improved survival in homozygous mutant mice (from 7 to 15 days), reduced cardiac hypertrophy and fibrosis, and normalized echocardiographic and histological measures
Xiao et al.61	Rescue of autosomal dominant hearing loss by in vivo delivery of mini dCas13X-derived RNA base editor (2022)	Myo6	myosin VI	p.C442Y	RNA base editor mxABE (mini dCas13X.1-based adenosine base editor)	no	Myo6C442Y/+ and Myo6C442Y/C442Y mouse models of dominant inherited deafness	cochlear injection of AAV-PHP.eB	not stated	4.22% in Myo6C442Y/C442Y mice (homozygous mutation), and a 2.01-fold increase in the wild-type (Myo6+) allele in Myo6C442Y/+ mice (heterozygous mutation)	improved auditory brainstem response and otoacoustic emission up to 3 months after treatment; hair cell survival was increased, and hair bundle morphology was preserved compared to untreated control ears
Cui et al.62	A base editor for the long-term restoration of auditory function in mice with recessive profound deafness (2024)	OTOF	otoferlin	c.2482C>T (p.Q828X)	ABE7.10max (SpCas9-NG)	yes	OtofQ828X/Q828X (homozygous)	round-window membrane injection of AAV-PHP.eB	1.5 × 1010 vg (N-terminal) + 1.5 × 1010 vg (C-terminal) per cochlea	2.1% gDNA, 88% cDNA correction	improved synaptic exocytosis in inner hair cells; auditory function rescued to near-wild-type levels for 1.5+ years
Yin et al.12	Amelioration of metabolic and behavioral defects through base editing in the Pah(R408W) phenylketonuria mouse model (2025)	Pah	phenylalanine hydroxylase	R408W (A5>G)	ABE8e-SpRY	yes	PahR408W PKU mouse model	i.v. AAV8 for adult mice, i.p. AAV8 for neonatal mice	low-dose: 1 × 1012 vg/mouse (adult); high-dose: 4 × 1012 vg per mouse (adult); 4 × 1011 vg per mouse (neonatal)	A5-to-G correction efficiency: 21.2% (low-dose), 34.6% (high-dose) after 52 weeks	normalized blood phenylalanine (Phe) levels, restored weight and fur color, improved brain function (neonatal)
Brooks et al.63	A base editing strategy using mRNA-LNPs for in vivo correction of the most frequent phenylketonuria variant (2024)	PAH	phenylalanine hydroxylase	c.1222C>T (p.Arg408Trp)	SpRY-ABE8.8	no	humanized PKU mouse model with the c.1222C>T PAH allele	retro-orbital injection of LNPs containing SpRY-ABE8.8 mRNA and PAH4 guide RNA	2.5 mg/kg and 5 mg/kg doses	26%–29% desired on-target editing and 3%–4% undesired bystander editing	blood Phe normalization; blood Phe levels <125 mmol/L by 1 week, no significant liver pathology
Brooks et al.64	Rapid and definitive treatment of phenylketonuria in variant-humanized mice with corrective editing (2023)	PAH	phenylalanine hydroxylase	c.842C>T (p.P281L)	ABE8.8	no	humanized PKU mice with the P281L variant in the PAH gene	retro-orbital injection of LNPs containing ABE8.8 mRNA and either PAH1 or PAH2 guide RNAs	2.5 mg/kg	corrective editing in the liver ranged from 28% to 46% for ABE8.8/PAH1 LNPs and 47%–58% for ABE8.8/PAH2 LNPs	blood Phe levels normalized within 48 h; treated PKU mice had normal Phe levels 1 week after treatment; hypopigmentation resolved by 8 weeks
Villiger et al.15	Treatment of a metabolic liver disease by in vivo genome base editing in adult mice (2018)	Pah	phenylalanine hydroxylase	c.835T>C (p.Phe263Ser) loss of function leading to PKU	nSaKKH-BE3	yes	Pahenu2 mice, a model for PKU	tail vein (i.v.) injection of AAV8	low dose: 5 × 1010 vector genomes (vg) per AAV per mouse; high dose: 5 × 1011 vg per AAV per mouse	Pah gene correction rates: 9.7% at 4 weeks, 18.6% at 8 weeks, 22.1% at 14 weeks, and 25.1% at 26 weeks post-injection; mRNA correction rates of up to 63% with the high dose	blood Phe levels were reduced to physiological levels (<120 μmol/L), and the light fur phenotype was reversed
Peters et al.65	Rescue of hearing by adenine base editing in a humanized mouse model of Usher syndrome type 1F (2023)	PCDH15	protocadherin-15	R245X (CGA → TGA, a stop codon)	ABEmax and ABE8e	yes	Pcdh15R245X/R245X humanized knockin (constitutive knockout and conditional knockout variants)	cochlear injection into neonatal mice (P0-P1) of AAV9-PHP.B, dual-AAV vectors	3.4 × 1010 GC each (N-intein and C-intein) per ear	cochlea gDNA: ∼1.40% (ABEmax) and ∼1.52% (ABE8e); cochlear cDNA: 1.60% (ABEmax) and 3.3% (ABE8e)	no hearing restoration in constitutive Pcdh15R245X/R245X KO mice; significant hearing rescue in Pcdh15R245X/fl, Myo15-Cre+ cKO mice; more effective rescue with ABEmax than ABE8e in some frequencies; rescue sustained at P60+ in about half of responding mice.
Liu et al.66	Efficient rescue of retinal degeneration in Pde6a mice by engineered base editing and prime editing (2024)	Pde6a	phosphodiesterase 6A	c.2009A>G (p.D670G)	AID-N51A-CBE (and prime editors)	both	Pde6anmf363/nmf363 retinal degeneration mouse model	in vivo electroporation of mouse retinas at P0; (and subretinal injection of split-intein dual-AAV2.NN vectors for PE constructs)	subretinal injection of 300–500 ng AID-N51A-CBE plasmids into P0 pups followed by electroporation (5 pulses, 80 V, 50 ms, 950 ms intervals, 10 mm electrode)	23.8% editing efficiency at target mutation, bystander editing (13.6%) but no off-target effects	restoration of PDE6A protein expression; preserved photoreceptors; retinal function rescued
Wu et al.67	AAV-mediated base-editing therapy ameliorates the disease phenotypes in a mouse model of retinitis pigmentosa (2023)	Pde6b	phosphodiesterase 6B	c.1678C>T (p.R560C)	SpRY-ABE8e	yes	rd10 (retinitis pigmentosa model)	subretinal injection of dual-AAV5 system	1 μL of 1012 vg/mL	gDNA: 13.06% (avg), up to 17.49%; cDNA: 34.07% (avg), up to 49.11%	preservation of photoreceptors, restored PDE6B expression, improved electroretinogram signals, improved vision-guided behavior
Su et al.68	In vivo base editing rescues photoreceptors in a mouse model of retinitis pigmentosa (2023)	Pde6b	phosphodiesterase 6B	c.1678C>T (p.R560C)	NG-ABE8e	yes	rd10 (retinitis pigmentosa model)	subretinal injection of AAV8	3 × 109 genome copies (GC) per eye for each AAV8 vector	20.79% efficiency at the DNA level and 54.97% efficiency at the cDNA level without bystanders	restored PDE6B protein expression; preserved photoreceptors; rescued visual function; RNA-seq showed preservation of phototransduction and photoreceptor survival genes
An et al.69	In vivo base editing extends lifespan of a humanized mouse model of prion disease (2025)	PRNP	prion protein	R37X (C>T) and Q91X (C>T)	BE3.9max (CBE), enCjCas9-TadCBEd, SauriCas9-TadCBEd, SauriCas9-ABE8e, SpCas9-ABE8e(V106W)	both	Tg66 mice (for injection route optimization), Tg25109 mice (human PRNP, endogenous Prnp knockout)	systemic retro-orbital injection, i.c.v. injection of AAV-PHP.eB	initial: 4 × 1012 vg/kg for retro-orbital injection, 5.5 × 1011 vg/kg for i.c.v. injection; optimized: 1.5 × 1013 vg/kg for single-AAV	55%–63% PrP reduction, minimal off-target effects with optimized dose	lifespan extension by 52%
Bock et al.70	Base editing of Ptbp1 in neurons alleviates symptoms in a mouse model of Parkinson’s disease (2024)	PTBP1	polypyrimidine tract binding protein 1	loss-of-function splicing mutation (precise mutation not stated)	ABE8e-SpG	yes	C57BL/6J mice with a Parkinson’s disease model induced by 6-hydroxydopamine (6-OHDA) brain injection	stereotaxic injection of AAV-PHP.3B vectors	3.6 μg of 6-OHDA were injected to induce the PD model; AAV vectors dosed at 2 × 108 vg/mouse	AAV-ctrl: gDNA: 0.04%; cDNA: not reported; AAV-GFAP: gDNA: 14.7%; cDNA: 23.7%; protein: 10.8%; AAV-hsyn: gDNA: 15.5%; cDNA: 24.8%; protein: 13.8%.	downregulating PTBP1 in striatal neurons caused them to express tyrosine hydroxylase (TH); this was associated with improved motor symptoms, including reduced forelimb akinesia and spontaneous rotations
Grosch et al.71	Striated muscle-specific base editing enables correction of mutations causing dilated cardiomyopathy (2023)	Rbm20	RNA-binding motif protein 20	p.P635L and p.R636Q mutations	SpRY, NRTH, NRCH, 8e-NRCH	yes	Rbm20P635L/P635L and Rbm20R636Q/R636Q mice	tail vein (i.v.) injection of AAVMYO dual-vectors	5 × 1011 vg (each AAV vector)	highest editing efficiency was observed with 8e-NRCH (21.4%); reported up to 71% of Rbm20 mRNA edited after 12 weeks	partial phenotypic rescue, including improved cardiac function and reduced fibrosis
Nishiyama et al.72	Precise genomic editing of pathogenic mutations in RBM20 rescues dilated cardiomyopathy (2022)	RBM20	RNA binding motif protein 20	p.R636Q (corresponding to the mouse Rbm20 p.R636Q mutation)	ABEmax-VRQR-SpCas9	yes	Rbm20R636Q/R636Q mice, which develop severe cardiac dysfunction	i.p. of AAV9 vectors	2.5 × 1014 vg/kg total AAV dose	19% DNA editing efficiency in the hearts of R636Q/R636Q mice, with 66% correction in RBM20 cDNA transcripts (from treated cardiomyocytes)	improved cardiac function, reduced left ventricular dilation, extended lifespan, restored RBM20 nuclear localization, eliminated RNP granules, and partially corrected splicing of the Ttn gene
Du et al.73	In vivo photoreceptor base editing ameliorates rhodopsin-E150K autosomal-recessive retinitis pigmentosa in mice (2024)	RHO	gene encoding Rhodopsin	E150K (causing autosomal recessive retinitis pigmentosa)	ABEmax	yes	Rho-E150K homozygous knockin mouse model	subretinal delivery of dual-adeno-associated virus (AAV2/8) vectors	7.65 × 109 total vg/mouse each for the N- and C-terminal AAVs (1 μL of AAV solution in each eye)	genomic DNA: 10.5% total editing at the on-target base, with 4.6% precise single-base correction, and a maximal correction of 11.4%	partial restoration of rhodopsin expression, but no phenotypic rescue of the electroretinography (ERG) a-wave; partial preservation of retinal structure
Lee et al.74	Bystander base editing interferes with visual function restoration in Leber congenital amaurosis (2024)	RPE65	retinal pigment epithelium-specific 65 kDa protein	premature stop codon (rd12 mutation) causing Leber congenital amaurosis	ABE8e, ABEmax, and ABE8eWQ	yes	rd12 LCA mouse model	subretinal injection of AAV vectors carrying ABE	4.3 × 1010 viral genomes (AAV2/2 and AAV2/9 each, in 3 μL PBS)	7.2% on target base correction (A6) and up to 4.6% bystander editing (A3) 6.5% (A8) and 0.3% (A11).	RPE65 protein expression restored in edited mice (50%–54% positive cells); bystander mutations (L43P) from ABE8e disrupted RPE65 function, preventing visual restoration
Jo et al.75	Visual function restoration in a mouse model of Leber congenital amaurosis via therapeutic base editing (2023)	Rpe65	retinal pigment epithelium-specific 65 kDa protein	p.R44X (C-to-T nonsense mutation)	NG-ABEmax	yes	rd12 mouse model	subretinal injection of dual-AAV2/9 vectors	5.4 × 1010 vg of AAV2/2 per eye (early experiment, low editing efficiency); 7.3 × 1010 vg of AAV2/9 per eye (used in follow-up experiments for improved delivery)	4-month-old rd12 mice (AAV2/9): 13.5% on-target editing; 3-week-old rd12 mice (AAV2/9): 11.8% (range 3.2%–25.3%) at 6 weeks post-injection; editing maintained at 6.1% after 3 months	restored wild-type Rpe65 mRNA and RPE65 protein expression; restored light-induced retinal electrical responses; bystander edits noted (L43P, C45R) but no in vitro toxicity
Choi et al.76	In vivo base editing rescues cone photoreceptors in a mouse model of early-onset inherited retinal degeneration (2022)	RPE65	retinal pigment epithelium-specific 65 kDa protein	premature stop codon (rd12 mutation) causing Leber congenital amaurosis	NG-ABE variant (evolved for wider PAM compatibility)	both	rd12, rd12Gnat1−/−mice	subretinal injection: LV single vector system or dual-AAV (split base editor) system	Not stated	up to 57% A-to-G conversion in RPE cells (lentiviral delivery); up to 2.7% editing efficiency with AAV (much lower)	restored visual pathway function (VEPs) and improved single-neuron activity in the visual cortex; AAV showed slower effects compared to lentivirus
Suh et al.77	Restoration of visual function in adult mice with an inherited retinal disease via adenine base editing (2020)	RPE65	retinal pigment epithelium-specific 65 kDa protein	c.130C>T; p.R44X nonsense mutation that impairs visual function	ABEmax	no	rd12 mice	subretinal injection of lentivirus expressing ABEmax and sgRNA targeting the Rpe65 mutation (co-injection with AAV1-CMV-eGFP for tracking)	LV dose: 1 × 106 transducing units per eye; AAV1-CMV-eGFP dose: 5 × 107 genome copies per eye	A5-treated mice: up to 29% correction efficiency of the mutation (15.95% average); A6-treated mice: up to 11% correction efficiency (5.22% average)	restoration of RPE65 protein in treated mice and near-normal retinal and visual functions in A5-treated mice
Qi et al.78	In vivo base editing of Scn5a rescues type 3 long QT syndrome in mice (2023)	Scn5a	sodium voltage-gated channel alpha subunit 5	T1307M pathogenic variant (T→M)	ABEmax	yes	Scn5a-T1307M mouse model	i.p. of AAV9-ABEmax	3 × 1014 GC/kg of AAV9-ABE, administered at postnatal day 14	Scn5a mRNA correction rates of up to 99.20%	reduced QT prolongation, normalized cardiac rhythm, eliminated carbachol-induced Torsades de Pointes and ventricular tachycardia
Alves et al.79	Optimization of base editors for the functional correction of SMN2 as a treatment for spinal muscular atrophy (2023)	SMN2	survival of motor neuron 2	C6>T (base edit A>G in exon 7)	ABE8e-SpRY	yes	SMNΔ7 mice (FVB.Cg-Grm7Tg(SMN2)89Ahmb Smn1tm1Msd Tg (SMN2∗delta7) 4299Ahmb/J)	i.c.v. injection at P1 (± i.v. co-injection in some cohorts) of AAV9	3 × 1010 vector genomes (vg) of each N- and C-terminal AAV per route	2%–10% A-to-G editing in brain and spinal cord at P13; up to 20% in brain and 30% in liver at 12 weeks; 2–4× increase in SMN transcript levels	increased body mass; improved motor function; 30% increase in survival (p = 0.01)
Arbab et al.80	Base editing rescue of spinal muscular atrophy in cells and in mice (2023)	SMN2	survival of motor neuron 2	C6>T (exon 7)	ABE8e-SpyMac	yes	Δ7SMA mice (SMA model with exon 7 deleted in SMN1)	i.c.v. injection of AAV9 dual-vectors	2.7 × 1013 vg/kg AAV9-ABE vectors (for base editing); 2.7 × 1012 vg/kg AAV9-Cbh-eGFP-KASH (for viral transduction control)	87% T6>C conversion in GFP+ (transduced) neurons	improved motor function, extended survival (111 days), and normal behavior
Alves et al.81	Base editing as a genetic treatment for spinal muscular atrophy (2023)	SMN2	survival of motor neuron 2	C6>T (exon 7)	ABE8e-SpRY	yes	SMN Δ 7 mice (SMA model with human SMN2 gene and severe SMA phenotype)	i.c.v. injection of dual-AAV9 vectors	3 × 1010 vector genomes (vg) for each of the N- and C-terminal AAV vectors.	on-target editing in the spinal cord and brain, with editing levels ranging from ∼2% to 10% (average 4% in spinal cord, 6% in brain)	no obvious improvement in SMA phenotype within the short time frame of the study (mice sacrificed at P13, typical lifespan <15 days)
Lim et al.82	Treatment of a mouse model of ALS by in vivo base editing (2020)	SOD1	superoxide dismutase 1	c.93G>A nonsense mutation	CBE (with APOBEC1, UGI and SpCas9 D10A nickase)	yes	G93A-SOD1 mouse (B6SJL-Tg(SOD1∗G93A)1Gur/J) expressing a mutant human SOD1 transgene	intrathecal injection of AAV9 vectors into the lumbar subarachnoid space (L5-L6)	8 × 1010 vector genomes each (total 1.6 × 1011 vg per mouse)	editing estimated at 1.2% across all SOD1 reads, corresponding to an “effective editing rate” of approximately 20% in dual-transduced cells	base editing in G93A-SOD1 mice extended survival, slowed disease progression, improved motor function, and reduced SOD1-positive inclusions by ∼40% at end-stage
Li et al.83	Programmable base editing of mutated TERT promoter inhibits brain tumor growth (2020)	TERT	telomerase reverse transcriptase	−124C>T TERT promoter mutation	CjABE (Campylobacter jejuni Cas9-fused ABE)	no	Athymic nude mice (orthotopic brain tumor model)	intracranial injection of AAVs (serotype/variant not specified)	1 × 1010 viral particles (in 10 μL of PBS), 3 injections	85.2%–90.4% conversion of the −124C>T mutation to −124C in tumors from both U87 and PDX GBM cells	in vivo treatment inhibited glioma growth, as evidenced by reduced tumor bioluminescence and improved survival
Yeh et al.84	In vivo base editing restores sensory transduction and transiently improves auditory function in a mouse model of recessive deafness (2020)	Tmc1	transmembrane channel-like 1	c.A545G (p.Y182C)	AID-BE3.9max	yes	Tmc1Y182C/Y182C (Baringo mice)	inner ear injection of dual-AAVs (Anc80L65 capsid)	9.7 × 108 viral genomes (vg) per ear, corresponding to ∼1.8 × 109 vg/kg	2.3% editing at the DNA level in whole cochlear tissue; 10%–51% editing at the mRNA level in hair cells	restored hair cell sensory transduction; preserved inner hair cell stereocilia morphology; partially and transiently rescued low-frequency auditory function at 4 weeks
Tachida et al.85	Systematic empirical evaluation of individual base editing targets: validating therapeutic targets in USH2A and comparison of methods (2024)	USH2A	usherin	c.11864G>A p.Trp3955∗(nonsense mutation)	ABE8e	yes	USH2A humanized knock-in mouse model	subretinal injection of split-intein AAV9 vectors	high-dose: 1 × 1010 gc/μL, low-dose: 5 × 109 gc/μL, 1 μL per eye	correction rate: 65% at mutant site; 52% excluding bystander edits; high editing efficiency compared to clinical retinal gene editing trial	restoration of usherin protein
Fry et al.86	Comparison of CRISPR-Cas13b RNA base editing approaches for USH2A-associated inherited retinal degeneration (2025)	USH2A	usherin	c.11840G>A (p.W3947X)	RNA base editors PspCas13b or Cas13bt3 with ADAR2	no	Ush2aW3947X/W3947X knockin mouse (C57BL/6J background with repaired Cdh23 allele)	intravitreal injection of AAV8-Y733F	1 × 109 gc/eye	mean RNA editing rates ranged from 0.32%–2.04% with greater efficiency using PspCas13b compared to Cas13bt3	restored usherin protein localization to the connecting cilium in photoreceptors; retinal thinning observed (AAV or editor toxicity?)
Liu et al.87	Screening an effective dual-adeno-associated virus split-cytosine base editor system for C-to-T conversion in vivo (2023)	Vegfa	vascular endothelial growth factor A	C-to-T conversion at splicing sites or early stop codons (specific mutations are not detailed)	CBE3	yes	CD-1 mice (5-week-old males)	subretinal injection of AAV8 vectors	1 × 1010 vg of N-terminal half of split-BE3; 1 × 1010 vg of C-terminal half of split-BE3; 1 × 109 vg of GFP	editing efficiency in RPE cells: 23.29%–10.98% for split-BE3-Rma674 with Vegfa-gRNA1, 21.78%–6.09% for split-BE3-Rma713 with Vegfa-gRNA1, and 6.01%–2.45% for split-BE3-Rma674 with Vegfa-gRNA2	no reported phenotypic rescue in mice

In terms of disease systems, the final selection included 12 papers (18.2%) on liver diseases, 12 papers (18.2%) on eye disorders, 11 papers (16.7%) on muscular dystrophy, 9 papers (13.6%) on heart disease, 8 papers (12.1%) on neural conditions, 5 papers (7.6%) on hematologic conditions, 4 papers (6.1%) on deafness, 3 papers (4.5%) on rare disorders, and 2 papers (3.0%) on cancer. This breakdown underscores a predominant focus on liver, eye, and muscle models, with other organs/tissues still substantially underrepresented in the literature.

At a genetic level, 11 studies targeted the DMD gene (Duchenne muscular dystrophy [DMD] models), 6 targeted the FAH gene (tyrosinemia models), 4 targeted the PAH gene (phenylketonuria models), and 4 targeted the RPE65 gene (Leber congenital amaurosis models). While these were the most modeled repair scenarios, a total of 38 different genes were base edited across the 66 studies, highlighting the broad potential of this technology. Notably, many studies employed mouse models featuring humanized versions of the gene of interest (e.g., DMD), whereas others used the native murine gene (e.g., Dmd).

In terms of vector delivery systems, 43 (65.2%) featured split-intein systems delivered via dual AAV vectors. The remaining studies employed alternative strategies: (1) compact base-editing enzymes delivered via single AAV vectors; (2) vectors with larger packaging capacities (e.g., lipid nanoparticles [LNPs] or adenoviral vectors); or (3) plasmid DNA delivered via hydrodynamic injection or in vivo electroporation.

The studies were parsed (full-text) to extract the gene and mutation being edited, the enzyme utilized for base editing, whether a split-intein system was used for delivery, the mouse model used for rescue studies, the vector system and how it was delivered into the animals, the vector dose (typically in vg/mouse or vg/kg for AAV vectors), and the results of in vivo rescue including the editing efficiency and the phenotypic effects. These data are presented comprehensively in Table 2.

“Standardizing” dosing with split-intein dual AAV vectors

The limited packaging capacity of AAV vectors precludes the use of canonical SpCas9-based ABEs and CBEs in single-vector systems.¹³ Intein systems offer a transactivating approach, allowing protein-level reconstitution rather than genetic-level assembly. A dual-vector approach also facilitates the inclusion of the targeting sgRNA within one of the vectors. However, because both vectors are required for efficacy, low transduction rates in the target tissue can significantly diminish therapeutic outcomes. For instance, in a mouse model of amyotrophic lateral sclerosis (ALS), dual-vector delivery in the spinal cord resulted in less than 7% of cells being transduced with both vectors, with the authors reporting an “effective editing rate of ∼20%.” However, given the low transduction efficiency in the model, the overall editing rate across the entire tissue was much lower—approximately 1.2%.⁸²

In most studies using this dual AAV approach, both vectors are dosed at an even ratio, though it is possible to also include a fluorescent transduction control vector, such as eGFP.⁴² The overall dosages can sometimes be challenging to directly compare, as doses are occasionally given in vg/mouse versus vg/kg; additionally, some studies involve neonates versus adult mice, and others involve systemic versus local dosing. In most cases, the papers specified the dosage as total or per vector, but in some cases, the dose was either not stated or ambiguous.

For most studies involving adult systemic administration (including tail vein, intraperitoneal, and intrathoracic injection), a total dose of 5 × 10¹¹–1 × 10¹² vg/mouse was the most common. In split-intein systems, this reflects the combined dose of both vectors. One of the higher doses was seen in a study rescuing pathogenic RBM20 mutations,⁷² which featured 2.5 × 10¹⁴ vg/kg (5 × 10¹² vg/20 g mouse). For neonatal studies the typical dose is less than for an adult mouse (1 × 10¹¹–2 × 10¹¹ vg/neonate); however, it is not necessarily proportional to their lower body weight (∼1.5–2 g). The most common dosing techniques were injection into the facial vein, superficial temporal vein, or peritoneum. In some cases, neonates were treated with relatively high AAV doses, such as a therapeutic exon skipping approach that used 1.1 × 10¹² vg in P2-P3 neonates,³⁷ or a treatment for preventing cardiomyopathy that used 1.5 × 10¹⁴ vg/kg (∼4 × 10¹² vg per P0 neonate).⁶⁰ From a translational perspective, it is unclear whether safety concerns would preclude the application of equivalent human doses in a clinical setting. Recent studies have highlighted potential risks, including liver toxicity and dorsal root ganglion pathology, raising questions about the tolerability of high-dose AAV delivery in patients.⁸⁸

For non-systemic injections, dosages differ largely depending on the target tissue. In the eye (intravitreal or subretinal injections), the reported doses can vary widely (1 × 10⁹–1 × 10¹¹ gc/eye—noting the more common terminology of “genome copies” in the retinal space). Cochlear injections are often around ∼5 × 10¹⁰ vg/ear, although a study restoring auditory function in deafness achieved phenotypic rescue with a 10-fold lower vector dose.⁸⁴ Intramuscular injections also showed a range (most between 1 × 10¹¹ and 2 × 10¹¹ vg/mouse); however, there are reports using extremely high local doses, such as 1 × 10¹² and 1 × 10¹³, for treatment of DMD.^31,32,39 Several studies (7/66) describe intracerebroventricular (i.c.v.) injection or stereotactic injection into the brain, which often reported smaller doses (5 × 10⁸–5 × 10¹⁰). Nevertheless, high doses appear tolerable, such as in a treatment for spinal muscular atrophy, where a dose of 2.7 × 10¹³ vg/kg was used for i.c.v. injection in mice.⁸⁰

In summary, the wide variability in dosing strategies observed across in vivo base-editing rescue studies highlights the critical need for increased consistency and standardized dosing parameters. Clearly defined, comparable dosing regimens are essential not only to facilitate accurate comparisons between independent studies but also to guide translational research and inform clinical protocols. Establishing consistent and transparent reporting of dose metrics, administration routes, and safety outcomes will be crucial for reliably predicting equivalent safe and effective dosing regimens for future human clinical trials.

Emergence of LNP delivery systems

While most studies employed viral vector systems, LNPs have been increasingly explored as an alternative approach. LNPs mediate the delivery of mRNA, and due to their higher cargo capacity, eliminate the need for split-intein BEs. In total, seven studies within the scope of this review used LNPs for BE delivery.^{44,48,50,51,54,63,64} LNPs also offer several advantages over AAVs, including a reduced risk of insertional mutagenesis, lower immunogenicity, and the potential for redosing (detailed comparison in Table 3).

Table 3.

Comparison of AAVs and LNPs as gene therapy delivery platforms

Feature	Adeno-associated virus (AAV)	Lipid nanoparticles (LNPs)
Payload capacity	∼4.7 kb	>10 kb (depends on formulation)
Cargo type	primarily DNA	versatile: DNA, RNA, protein, CRISPR RNP
Targeting	capsid tropism enables efficient delivery to a range of tissues (e.g., liver, CNS, muscle, retina)	efficient targeting primarily limited to liver (and lungs to a lesser extent); ligand-based targeting remains limited in vivo
Immune response	pre-existing immunity to natural serotypes common; long-term expression may trigger responses	generally less immunogenic; transient expression reduces risk
Duration of expression	long-term (months to years, especially in non-dividing cells)	transient (typically days to weeks)
Genomic integration risk	low (mostly episomal; rare integration events)	none (does not enter the nucleus unless specially designed to)
Repeat dosing	limited due to neutralizing antibodies	amenable to repeat dosing
Clinical experience	extensive clinical trials; multiple FDA-approved therapies	Growing rapidly; approved for mRNA vaccines; gene editing trials underway
Manufacturing	complex, costly; cell-based production; scalability remains a challenge	scalable, cell-free synthesis; simpler manufacturing pipeline
Stability and storage	stable at −80°C	generally more stable; lyophilization possible

The first study employing LNPs for base editing targeted an A>G splice-site mutation in the Fah gene.⁴⁸ However, correction rates in the liver were low (∼0.44%), performing significantly worse than hydrodynamic tail-vein injections using plasmid constructs. The first effective demonstration of phenotypic rescue using LNPs was achieved in a humanized mouse model of phenylketonuria harboring a point mutation in the PAH gene.⁶⁴ This study reported 26%–29% on-target editing, normalization of blood phenylalanine levels, and demonstrated the therapeutic potential of LNPs for targeting hepatocytes. Such an approach may be particularly suitable for disorders in which therapeutic benefit can be achieved even at relatively low levels of base editing, as exemplified by the correction of mutations in the G6PC1 gene.⁵⁰ In this study, editing rates of up to ∼60% were achieved, but 10% was sufficient for phenotypic rescue.

However, challenges remain for LNP delivery, such as optimizing tissue-specific delivery and improving stability in circulation.⁸⁹ By modifying the composition and ratios of ionizable cationic lipids, helper lipids (phospholipids), cholesterol, and polyethylene glycol (PEG) lipids, it is possible to significantly influence LNP size, stability, and cellular uptake specificity. While some organs, such as the liver, lungs, muscles, and brain, have proven feasible targets for efficient LNP delivery, others, notably the heart and bone, require substantial optimization. Furthermore, many studies to date have relied on local delivery methods (e.g., direct intramuscular injection), raising concerns about their practicality compared with systemic administration achievable with dual AAV vectors for these tissues. Therefore, further optimization of LNP formulations and delivery strategies is essential to fully realize their therapeutic potential and establish them as broadly applicable alternatives to viral vector-based methods.

Compact enzymes for base editing

Ten rescue studies utilized “compact” enzymes that could be packaged into a single AAV vector.^{29,33,34,37,45,56,61,69,83,86} These included SauriCas9-TadCBEd and SauriCas9-TadABEd (from Staphylococcus auricularis),^56,69 enhanced target-AID (eTAM) (from Staphylococcus aureus),³⁷ and CjABE (from Campylobacter jejuni)⁸³ as “compact” DNA BEs. Despite the simplified delivery approach, SauriCas9-ABE8e was less effective than the dual-AAV compatible SpCas9-ABE8e(V106W) in reducing PrP in the mouse brain (12% versus 26%), and the compact CBE was completely ineffective in vivo.⁶⁹ Similarly, when treating a mouse model of Lmna c.1621C>T cardiomyopathy, the all-in-one ABE system was inferior to the dual-vector system.⁵⁶ These results highlight a trade-off between packaging constraints and editing activity, underscoring the need for further optimization of compact BEs.

Notably, the remaining BEs featured in disease rescue studies are RNA BEs, most of which are based on Cas13. The class 2 type VI-A and IV-B RNA-guided, RNA-targeting CRISPR-Cas effectors have been engineered for RNA binding and knockdown in mammalian cells.^16,18 RNA BEs offer the advantage of transient modifications, reducing the risk of permanent off-target effects and avoiding heritable changes. This transient nature is beneficial in scenarios where temporary gene modulation is desired. However, for diseases requiring long-term correction, such as genetic disorders where permanent gene correction is necessary, RNA-based approaches may be less suitable due to the need for repeated administrations to maintain therapeutic effects.

In vivo rescue studies demonstrate a wide range of correction efficiencies

Since their inception, BEs have been reported to achieve genomic DNA (gDNA) correction efficiencies typically ranging from ∼15% to 75%, accompanied by minimal indel formation.⁴ Subsequent advancements in BE design have further increased editing performance, achieving even higher efficiencies. Notably, zebrafish models treated with contemporary ABEs and CBEs have also demonstrated impressive editing outcomes.⁹⁰ Most conventional in vivo rescue studies initially validate or optimize corrective base-editing strategies in vitro, using engineered HEK293 cells or lineage-specific cell lines. This step facilitates rapid screening and refinement of editing methods, including strategies to minimize problematic bystander edits—unintended nucleotide modifications at adjacent, non-targeted sites within the editing window. However, these optimized in vitro conditions typically represent “best-case” scenarios, where delivery efficiency approaches ideal levels, and thus may not fully reflect the challenges encountered during in vivo applications.

Editing outcomes are commonly reported at multiple levels—genomic DNA (gDNA) correction, transcript (mRNA) expression, and protein rescue—though the consistency of these measurements varies across studies. Table 2 summarizes genomic editing efficiencies across different studies, which range from low (<10%)^24,39,61 to moderate (10%–40%)^63,75 to high (>40%).^41,85 Editing efficiency is typically assessed shortly after gene therapy treatment; however, it is important to recognize that edited cell rates can change over time, often due to selective advantage. For example, T>C editing in a dystrophic mouse model increased significantly in the heart over time,³⁶ as did base editing correction in the liver in a Fah model.⁴⁹

The variation in base editing efficiency across studies arises from multiple factors. First, different BEs exhibit varying levels of activity, editing windows, and specificities, which influence their effectiveness at a given target site. Second, delivery method and efficiency play a critical role; viral vectors such as AAV may provide sustained expression but face packaging constraints, while non-viral methods such as LNPs can yield variable delivery across different tissues. Additionally, the genomic context of the target site, including chromatin accessibility and sequence composition, can significantly impact editing success.^91,92 Finally, guide RNA design is a key determinant, as poorly optimized guides may result in suboptimal editing efficiency or increased bystander edits.

Therapeutic safety: Off-target edits, bystander edits, AAV integration, and RNA editing

Off-target editing remains a significant concern for in vivo base editing, though their frequency can vary widely depending on the editor and context. For example, a genome-wide analysis in mice found that ABEs induced off-target SNVs at frequencies near the spontaneous mutation rate, whereas a CBE (BE3) generated more than a 20-fold higher off-target SNV burden.⁹³ Recent studies have also shown that BEs, while designed to avoid DSBs, can still induce them at low frequencies.⁹⁴

To identify the sites at high likelihood of off-target edits, a variety of digital tools have evolved to find near-match genomic targets.⁹⁵ Amplicon sequencing of target sites can determine whether off-target edits are a concern and is often preferred over whole-genome sequencing, which lacks the sensitivity to detect low-frequency edits. Techniques such as GUIDE-seq (genome-wide unbiased identification of DSBs enabled by sequencing) are ill-suited for base editing because they rely on detecting off-target indels by inducing DSBs, which do not occur in the context of base editing. This also holds true for alternate methods that also depend on DSB detection (e.g., CIRCLE-seq, SITE-seq, and DISCOVER-seq).⁹⁵ Emerging technologies such as GOTI (genome-wide off-target analysis by two-cell embryo injection) and Detect-Seq have been devised to detect off-target deaminase activity in the genome.^96,97 While reporting on off-target edits is standard, the sequencing techniques and model systems are inconsistent. Some studies exclusively used cell lines such as HEK293 cells, while others analyzed tissues of interest. Since chromatin structure varies across cell lineages, assessing off-target effects in the target tissue is more relevant. Moreover, Grosch et al. argued that restricting on-target editing to specific tissues also limited off-target editing.⁷¹ This was based on data illustrating striated muscle targeting reduced off-target effects in non-muscle tissues including the liver.

Bystander edits are unintended base changes at or near the on-target site caused by a BE’s activity window. Highly active editors can introduce bystander mutations that impact therapeutic outcomes. For instance, in a mouse model of Leber congenital amaurosis (LCA), ABE8e corrected the pathogenic nonsense mutation in the Rpe65 gene but simultaneously induced multiple bystander missense mutations within the editing window, ultimately preventing restoration of visual function despite correcting the pathogenic mutation.⁷⁴ Even when therapeutic rescue is achieved, bystander edits often occur at significant levels. For example, in a DMD mouse model, intramuscular injection led to bystander editing at multiple sites (A₁₂: 6.7%, A₁₈: 10.7%), while on-target editing at A₁₄ reached 35.0%, ultimately ameliorating the dystrophic phenotype.³⁸ As more precise BEs are developed, the impact of bystander edits is expected to diminish. A recent study comprehensively describes methods of refining the editing window by inserting specific peptide fragments into the substrate-binding pocket of deaminases.⁹⁸

Historically, AAV vectors were considered safe for gene therapy due to their predominantly episomal existence and low genomic integration rates. However, recent studies have raised concerns about unintended AAV integration events. Dalwadi et al. reported a 1%–3% integration frequency of AAV vectors in human hepatocytes, both ex vivo and in vivo, with integrations often accompanied by genomic rearrangements and deletions at the integration sites.⁹⁹ AAV vectors can integrate into CRISPR-Cas9-induced DSBs at frequencies up to 47% in vitro and in vivo,¹⁰⁰ noting that BEs generally avoid DSBs but can induce low levels, particularly at off-target sites.

Lastly, several 2019 studies raised concerns about unintended off-target RNA editing by ABEs and CBEs. CBEs were shown to induce widespread off-target RNA edits, primarily due to the deaminase enzyme’s activity on RNA substrates.¹⁰¹ Similarly, a study reported potential disruptions to normal RNA function with both ABEs and CBEs.¹⁰² Engineering BE variants with reduced RNA-binding affinities was proposed as a solution¹⁰³; however, because BE expression is often transient and no adverse phenotypic outcomes have been linked to off-target RNA editing, these modified variants have not been widely adopted.

Therapeutic base editing shows a remarkable potential to rescue murine disease models

Across the 66 analyzed studies, the majority demonstrated at least partial restoration of a disease phenotype in genetically modified mouse models. No standardized classification system exists for rescue efficiency, however studies generally reported outcomes in terms of DNA editing rates, restoration of messenger RNA transcripts, protein rescue, partial phenotypic rescue, or complete phenotypic rescue.

While some studies have demonstrated successful in vivo base editing, they can fall short of providing clear evidence for functional rescue in a relevant disease model. For example, the SPLICER toolbox study targeted the amyloid beta precursor protein (APP) using AAVrh10-ABE8e vectors injected into the hippocampus, which led to exon skipping in vivo. However, a reduction in Aβ42 peptide levels was only demonstrated in a cell model.²⁶ In a DMD model, intramuscular injection of AAV9 IscB.m16∗-CBE-v2 restored dystrophin protein expression, though comprehensive muscle function analysis was not reported.²⁹ Similarly, base editing of mice carrying a hereditary persistence of fetal hemoglobin (HPFH) mutation in the γ-globin promoter increased γ-globin expression, though this model did not fully recapitulate a hemoglobinopathy.⁵³

Thus, robust therapeutic base editing studies not only employ rigorous molecular techniques but also physiologically relevant disease models with detailed functional assessments. Notable examples include Li et al., who used LNPs to treat a sickle cell disease model, demonstrating a reduction in sickling to 25% of the untreated group following base editing.⁵² Studies targeting complex sensory organs, such as the eye⁷⁶ and cochlea,⁸⁴ required sophisticated methodologies to assess function, showing partial or transient restoration through base correction. In cases where only partial functional rescue was achieved, this was typically attributed to incomplete gene correction. A key advantage of mouse models is their ability to establish the threshold of gene correction necessary for meaningful functional improvement, informing future therapeutic strategies.

Several studies stand out for their remarkable effectiveness in restoring function in severe disease models. One of the earliest demonstrations of therapeutic base editing rescued a lethal Fah nonsense mutation, achieving 89% survival at three months following in utero delivery via an adenoviral vector.⁴⁹ Similarly, neonatal base editing of a severe vascular disease caused by an ACTA2 mutation dramatically improved function and extended lifespan from six weeks to approximately 23 weeks.²⁵ In a DMDΔE54 mdx humanized DMD mouse model, in vivo base editing achieved over 40% efficiency, leading to improved dystrophin expression, histological restoration, and enhanced motor function.³¹ Editing of a Hutchinson-Gilford progeria syndrome model doubled median lifespan,⁵⁷ while cognitive decline was reduced in a tauopathy (MAPT mutation) model using a trans-splicing NG-ABE8e vector system, as demonstrated by behavioral testing. Additionally, multiple studies have shown robust phenotypic rescue in mouse models of phenylketonuria.^15,64 Notably, in a mouse model of prion disease—an incurable and fatal condition—base editing extended lifespan by 52%.⁶⁹

These findings highlight the broad spectrum of outcomes observed in therapeutic base editing studies, ranging from minimal functional improvement to profound disease reversal. In cases where editing is both efficient and well-integrated into disease models, the impact has been striking, underscoring the potential for future clinical translation. While challenges remain, particularly in delivery and efficiency, the ability of base editing to rescue severe genetic disorders demonstrates its promise as a transformative tool for precision medicine.

Notable exclusions

This review applies specific inclusion and exclusion criteria, but it is also important to acknowledge recent and prominent base editing studies that fall outside its scope. A range of seminal studies explored its potential for modifying proprotein convertase subtilisin/kexin type 9 (PCSK9), a key regulator of cholesterol. Notably, the Q35X mutation can be introduced via base editing, which has been shown in preclinical models to reduce serum cholesterol in mice.^{104,105,106,107} This technology has the potential to help manage hypercholesterolemia, and clinical trials with VERVE-101—an LNP delivery system for a PCSK9-targeting ABE—showed promising results,¹⁰⁸ but was halted early due to an adverse event. In response, the next-generation candidate VERVE-102 was developed using a GalNAc-LNP delivery system, with interim results suggesting improved safety and durable cholesterol lowering. Together, these studies highlight the clinical translatability of base editing for cardiometabolic diseases, though key challenges around safety, delivery, and long-term efficacy remain.

Other notable studies have focused on innovative delivery strategies for base editing, though these were conducted in wild-type mice rather than disease models. Examples include subretinal injection of lentivirus-derived nanoparticles for modifying the Vegfa gene, and focused ultrasound-mediated microbubble destruction for modifying the Pde3b gene.^109,110

Additionally, many excluded gene-cell therapy studies involve ex vivo base editing followed by transplantation or engraftment of the modified cells. This approach is particularly relevant in the hematopoietic cell space, where transplantation of base-edited CD34⁺ stem and progenitor cells has been used to restore models of sickle cell disease, chronic granulomatous disease, and X-linked combined immunodeficiency.^{111,112,113,114,115} However, its applications extend beyond blood disorders. For example, human fibroblasts carrying a COL7A1 mutation were base-edited and engrafted into immunodeficient athymic nude mice as a model of recessive dystrophic epidermolysis bullosa.^116,117 Given that the first commercial CRISPR gene therapy product (Casgevy) involves bespoke ex vivo cell repair and transplantation,¹¹⁸ these strategies have clear translational relevance.

While this review includes two studies that incorporate both prime editing and conventional base editing,^38,66 it excludes studies focused solely on prime editing, such as Rpe65 mutation correction in the retina using PE3.¹¹⁹ Prime editing is a powerful and emerging technology with similarities to base editing and potentially greater versatility. However, its broader adoption is currently limited by factors such as the large size of the editing machinery, lower correction efficiencies, and more unpredictable targeting and construct design.

Finally, while this review centers on disease modeling in mice, base editing has been successfully used to introduce or correct mutations in a variety of mammals, including rats,¹²⁰ pig cells,¹²¹ sheep,¹²² and non-human primates.¹⁰⁵ These broader applications highlight the growing impact of base editing across biomedical research and its potential to extend beyond preclinical models toward clinical translation.

Conclusions

As base editing continues to evolve, several key challenges and opportunities shape its future trajectory. A major focus is improving editing efficiency, particularly in therapeutic contexts where even small gains in correction rates could significantly enhance clinical outcomes. Strategies such as optimizing guide RNA design, engineering more active deaminases, and refining prime editing hybrid approaches may help address current limitations. However, further expansion and refinement of BEs will be necessary to broaden the range of correctable mutations. Given their inherent design, BEs will remain largely restricted to SNVs, highlighting the need for complementary gene-editing strategies. From a translational perspective, the requirement for customized reagents tailored to each specific mutation presents additional regulatory and manufacturing challenges, particularly in the context of personalized medicine. Standardizing base editing platforms and developing broadly applicable delivery systems will be critical to facilitating clinical adoption.

Delivery remains another critical hurdle. While BEs have been successfully delivered via single and dual-AAV systems, these approaches face constraints related to packaging size and immune responses.¹²³ LNPs and other non-viral vectors offer promising alternatives, potentially improving tissue targeting and reducing immunogenicity.⁸⁹ Continued advances in delivery platforms will be essential for translating base editing into a viable therapeutic strategy.

BEs offer a versatile platform for therapeutic genome engineering, with applications extending well beyond SNV correction. ABEs and CBEs have been deployed for gene silencing, modulation of post-translational modification (PTM) sites,¹²⁴ and the functional reprogramming of regulatory elements. These strategies enable broad therapeutic targeting—such as PCSK9 repression in hypercholesterolemia¹⁰⁸—without requiring precise correction of rare pathogenic alleles. From a translational standpoint, these generalizable applications may offer greater clinical and commercial viability than SNV correction, given the limited prevalence of many individual mutations.

Compared to other CRISPR-based approaches, base editing allows for precise, irreversible nucleotide conversions without inducing DSBs, thereby reducing the risk of genomic instability. Prime editing, by contrast, offers broader sequence flexibility—enabling small insertions, deletions, and all base substitutions—but remains limited by lower efficiency and greater complexity in construct design and delivery. Meanwhile, nuclease-driven strategies (e.g., Cas9) support gene knockouts or large genomic modifications but carry higher risks of indels and off-target effects. The choice of editing modality ultimately depends on the disease’s genetic architecture, the required therapeutic mechanism (e.g., correction vs. repression), and practical considerations such as delivery, durability, and scalability. While BEs are well-suited for efficient, precise correction at accessible loci, their long-term impact will depend not only on their capacity to repair mutations, but also on their broader potential to modulate gene function across diverse clinical contexts.

Notably, the first clinical use of base editing was reported during the preparation of this review.¹²⁵ An LNP-based CRISPR BE therapy, employing a Cas9 NGC-ABE8e-V106W variant, was used to treat a 6-month-old infant with carbamoyl-phosphate synthetase 1 deficiency. The use of LNPs enabled a staged dosing regimen, with a higher dose administered 22 days after the initial infusion. Although only a 7-week follow-up was available and hepatic editing efficiency could not be directly assessed due to patient risk, the marked improvement in the child’s health strongly suggests therapeutic benefit. This single-patient trial exemplifies many of the challenges outlined in this review and offers a model for future bespoke gene editing approaches.

Acknowledgments

A.S. and S.L.G. currently oversee a base editing research program aiming to treat neurofibromatosis that is funded by the National Health and Medical Research Council, Australia (no. 2027987).

Author contributions

All authors contributed to writing of the original draft, and reviewing and editing of this manuscript.

Declaration of interests

The authors declare no competing interests.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this work the author(s) used ChatGPT 4o and 4.5 in order to copy edit text and to assist with extracting paper methods for Table 2. After using this tool/service, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the content of the publication.