CRISPR Base Editing Correction of TGFBI Mutations in Autosomal Dominant Corneal Dystrophies

Abstract

Purpose

Lattice and granular corneal dystrophy comprise two common TGFBI-associated autosomal dominant corneal disorders. Existing therapies are only temporizing and carry significant morbidity. Here, we develop a novel therapeutic approach using an adenine base editor (ABE) to correct common TGFBI mutations.

Method

We generated two human corneal epithelial (HCE) cell models harboring a copy of the most common disease-causing TGBFI mutations, R124C or R555W. These lines were electroporated with an ABE8e-NG–encoding mRNA and guide RNAs targeting the mutations. The resulting A•T-to-G•C editing efficiencies and off-target (OT) effects were assessed by amplicon sequencing. GFP-expressing adeno-associated viruses (AAVs) with different capsid types were transduced into HCE cells and healthy human corneal donor tissues, and GFP fluorescence was evaluated.

Results

Using all-RNA delivery for ABE8e-NG, we achieved 91% and 62% correction of the pathogenic adenines in HCE TGFBI^R124C/WT and TGFBI^R555W/WT cells, without editing the wild-type allele. Indel formation was negligible (<0.2%), bystander adenine editing was minimal (<0.7%), and editing at top computationally predicted OT sites was modest (<1.2% at all but 1 of the 20 OT sites analyzed), suggesting minimal safety concerns. Correction of TGFBI^R124C/WT in HCEs rescued the aberrant lysosomal localization of TGFBI. We further identified AAV1 as the most effective serotype for gene delivery into both human corneal donor tissue and HCE cells.

Conclusions

Our study demonstrates the feasibility and safety of CRISPR adenine base editing as a new therapeutic strategy for correcting common TGFBI mutations in corneal dystrophies, paving the way for further preclinical testing.

Transforming growth factor beta–induced (TGFBI) corneal dystrophies are a group of highly penetrant, autosomal dominantly inherited corneal diseases that generate opacities in the cornea and have variable expressivity.^1,2 TGFBI corneal dystrophies are caused by abnormal TGFBI protein variants in the corneal epithelium that are resistant to proteolysis, leading to the accumulation of protein aggregates in the cornea and impairing visual acuity progressively.³ Genetic studies have associated specific TGFBI gain-of-function variants with specific disease phenotypes; for instance, TGFBI p.R124C is the most common cause of lattice corneal dystrophy type 1 (LCD1), and TGFBI p.R555W is the most common cause of granular corneal dystrophy type 1 (GCD1).^1,4 Current treatments include symptomatic control of corneal erosions along with phototherapeutic keratectomy or corneal transplant.⁵ However, even with surgical intervention, recurrence is common, leading to vision loss and morbidity.¹ Genetic therapies—including gene supplementation, RNA interference, and CRISPR-Cas9 gene editing—have emerged as alternatives to surgical and medical intervention for genetic disorders, by tackling the underlying root causes.⁶ However, autosomal dominant mutations are not amenable to conventional gene supplement therapy, as toxic protein production from the mutant allele persists. Thus, there remains a critical need for alternative therapies for autosomal dominant disorders.^7–9 CRISPR-Cas9 enables targeted gene editing that promises one-time cures of genetic diseases through permanent changes in the patients’ genome.¹⁰ Unlike small interfering RNA (siRNA) or antisense oligonucleotide approaches, which require repeated dosing, gene editing offers the advantage of a durable, potentially lifelong therapeutic benefit from a single intervention.

In traditional CRISPR gene editing, Cas9 nuclease effectively creates genomic double-stranded breaks (DSBs), leading to gene knockouts via error-prone repair that generates a mixture of indels.^11,12 However, precise gene correction relies on inherently inefficient homology-directed repair (HDR), limiting their utilities for gene therapy.¹¹ Base editors are a new class of tools that make single-nucleotide changes in the genome of living cells with high efficiency, without causing toxic DSBs, generating unwanted indels, or requiring a donor repair template.¹³ They use a cytidine or adenosine deaminase enzyme fused to a catalytically impaired Cas9 to install targeted C•G-to-T•A or A•T-to-G•C changes within a ∼5-bp activity window in the target region specified by Cas9.^14–16 Owing to their high efficiency, precision, and safety, base editors are gaining popularity in clinical settings, with multiple clinical trials currently underway in the United States.^17,18 Adenine base editors (ABEs) are particularly useful as nearly half of all the pathogenic single-nucleotide polymorphisms can be corrected by converting an A•T pair to a G•C pair,¹⁵ such as in recent examples of progeria¹⁹ and sickle cell anemia studies.²⁰

Here, we developed the first efficient ABE-based therapeutic strategies for correcting the TGFBI p.R124C and p.R555W mutations in human corneal epithelial (HCE) cell line models for LCD1 and GCD1. By electroporating ABE8e-NG mRNA alongside synthetic guide RNAs, we achieved up to 91% and 62% corrections for the TGFBI p.R124C and p.R555W mutation, respectively. We demonstrated low indel formation, minimal off-target (OT) effects, and an effective platform for mitigating unexpected bystander editing using window-narrowing variants of Tad-A8e deaminase. Our approach ensures high-efficiency mutation correction while preserving the wild-type (WT) allele, thereby maintaining normal TGFBI function. Importantly, base editing of TGFBI^R124C/WT in HCEs reversed the aberrant lysosomal localization of TGFBI and restored wild-type localization, demonstrating functional rescue. In parallel, we evaluated adeno-associated virus (AAV) serotypes and found AAV1 to exhibit the highest infectivity across both HCE cells and human corneal tissue. Our findings pave the way for the preclinical testing and the ultimate clinical application of CRISPR base editing in treating patients with lattice and granular corneal dystrophy.

Methods

Patient Recruitment and Sequencing

Protocols were approved by the Institutional Review Board of the University of Michigan per the Common Rule of the US federal government (46CFR45). Patients provided written informed consent and were recruited based on meeting clinical criteria for lattice, granular, or Reis-Buckler corneal dystrophy. Saliva was collected in GFX-02 kits (Isohelix, Harrietsham, Kent, UK) and processed for automated DNA extraction using the FlexStar Saliva DNA kit (Autogen, Holliston, MA, USA). Exons 4 and 12 were amplified using established reported primers and Sanger sequenced at GeneWiz (Azenta Life Sciences, South Plainfield, NJ, USA).²¹

Plasmid Construction

A complete list of all plasmids and their detailed construction strategies are described in Supplementary Table S1. All oligonucleotides used were ordered from IDT (Coralville, IA, USA), with their sequences listed in Supplementary Table S2. PCR amplifications to create the vector backbone and inserts were conducted using Q5 High-Fidelity DNA Polymerase (NEB, Ipswich, MA, USA). Molecular cloning reactions were done by T4 DNA ligation, Gibson assembly, or Q5 site-directed mutagenesis. All constructs were verified by colony PCR using GoTaq Green Master Mix (Promega, Madison, WI, USA), restriction digestion, and Sanger sequencing (Eurofins Genomics, Louisville, KY, USA). Plasmids were purified using the QIAprep Spin Miniprep kit (QIAGEN, Hilden, Germany) according to manufacturer instructions.

Human Cell Culture

All cells used were cultured at 37°C and 5% CO₂ in a humidified incubator. HEK293T was cultured as described previously.²² HCE cells were kindly provided by Dr. Vinay Aakalu.²³ HCE cells were cultured in Dulbecco's Modified Eagle's Medium (DMEM)/F12 (Gibco, Waltham, MA, USA) supplemented with 10% fetal bovine serum (FBS; Corning, Corning, NY, USA). Cells were split every 5 to 6 days using TrypLE Express (Gibco). Cell lines were authenticated by their respective suppliers and tested negative for mycoplasma.

Plasmid Transfection Into HCE Cells

HEK293T cells were seeded 1 day before transfection at 1.2 × 10⁵ per well of a 24-well plate. For each transfection, 375, 125, and 50 ng of the ABE8e-NG, single guide RNA (sgRNA), and episomal target plasmids diluted in Opti-MEM (Gibco) were mixed with 1.5 µL Lipofectamine 3000 and 1 µL P3000 reagents (Invitrogen, Carlsbad, CA, USA) per the manufacturer's instructions. Transfected cells were subjected to media change 24 hours after transfection. Cells were harvested for analysis on day 4.

Production of ABE8e-NG mRNA and sgRNA by In Vitro Transcription

The 5′ capped and 3′ polyadenylated ABE8e-NG mRNAs were created by in vitro transcription (IVT) using the AmpliScribe T7-Flash Transcription Kit (Lucigen, Middleton, WI, USA) per the manufacturer's instructions with minor modifications. IVT templates were purified PCR amplifications from pYZ3143, which encodes ABE8e-NG, along with mammalian-optimized 5′ and 3′ UTR sequences. The PCR primers used incorporated a 119-base polyA tail and a modified T7 promoter to enable cotranscriptional capping. During IVT, UTP is substituted with N1-methylpseudouridine-5′-triphosphate (TriLink, San Diego, CA, USA); cotranscriptional capping is done with CleanCap Reagent AG (TriLink) added to a 7.2-mM final concentration. The resulting mRNAs were purified by LiCl precipitation and resuspended in nuclease-free H₂O as described previously.²⁴ The sgRNAs used in Cas9 ribonucleoprotein (RNP)–mediated editing experiments were generated by using the GeneArt Precision gRNA Synthesis Kit (ThermoFisher, Waltham, MA, USA) following the manufacturer's instructions.

Electroporation of ABE8e-NG mRNA, sgRNA, and Cas9-RNP Into HCE Cells

The Neon Transfection System (ThermoFisher) was used for electroporation per the manufacturer's instructions. Briefly, HCE cells were washed with 1× PBS on the plate and then dissociated with TrypLE Express and pelleted. Before resuspension in Neon buffer R, cells were washed again with 1× PBS. For RNP electroporation, 3 µg SpyCas9 protein was incubated with 0.6 µg sgRNA at room temperature (RT) for 15 minutes in 10 µL buffer R. The resulting Cas9 RNP was then mixed with 100 pmol single-stranded oligonucleotide (ssODN) and 1 × 10⁵ cells in Neon buffer R to a final volume of 11 µL. For mRNA electroporation to correct TGFBI p.R555W, 1.5 × 10⁵ cells were mixed with 500 ng ABE8e-NG mRNA and 75 pmol synthetic sgRNAs (GenScript, Piscataway, NJ, USA) in Neon buffer R in a final volume of 11 µL. To correct TGFBI p.R124C, 1.5 × 10⁵ cells were mixed with either 250 ng ABE8e-NG mRNA, along with 1 pmol R124C-G1 synthetic sgRNA, or 62.5 ng ABE8e-NG mRNA and 1 pmol R124C-G2. Mixtures were electroporated with a 10-µL Neon tip (1300 V, 20 ms, two pulses) and plated in 24-well tissue culture plates containing 500 µL DMEM/F12 with 10% FBS.

Creation of HCE-TGFBI-R124C and TGFBI-R555W Cell Line Models

The SpyCas9 nuclease-HDR method was used to create these models. The sgRNA sequences used to direct Cas9 cleavage at the TGFBI exon 4 for R124C and exon 12 for R555W are detailed in Supplementary Table S2, along with the corresponding ssODNs to install the patient mutations via HDR. The mixture of SpyCas9 RNP and ssODN was electroporated into HCE WT cells, as described in the earlier section. Half of the electroporated cells were lysed on day 4 and subjected to amplicon sequencing to assess HDR efficiency. The remaining half was sorted with a Bigfoot Cell Sorter (ThermoFisher) into 96-well plates at 1 cell per well and cultured for approximately 14 days. Surviving clones were isolated, and the targeted TGFBI locus was genotyped by amplicon sequencing. Desired clones were expanded and cryopreserved for long-term use and storage.

Next-Generation Sequencing Library Construction and High-Throughput Sequencing

Cells were lysed 4 days after plasmid transfection or electroporation in 250 µL of lysis buffer (10 mM Tris-HCl, pH 8.0; 0.05% SDS; 40 µg/mL proteinase K) per well. The cell lysis mixture was incubated at 37°C for 60 minutes and 80°C for 30 minutes and used directly for downstream analysis. Next-generation sequencing (NGS) library construction was performed as previously described¹⁶ with minor modifications. Briefly, a 200- to 300-bp region surrounding the target site was PCR amplified and then barcoded. Final PCR amplicons were pooled and purified by 1% agarose gel extraction using the Zymoclean Gel DNA Recovery Kit (Zymo Research, Irvine, CA, USA). DNA concentrations were quantified using a Qubit dsDNA High Sensitivity Assay Kit (ThermoFisher) and sequenced on an Illumina (San Diego, CA, USA) MiSeq instrument (300-bp single-end read) according to manufacturer's protocol. Primers used for PCR amplifying the targeted amplicons are listed in Supplementary Table S2.

Bioinformatic Analysis of NGS Datasets

Sequencing reads were demultiplexed using BaseSpace (Illumina), and the FASTQ files were analyzed using CRISPResso2²⁵ and in Excel (Microsoft, Redmond, WA, USA) as described previously.²⁶ Base editing efficiencies were reported as the percentage of reads with desired base conversion at a specific location over the total aligned read. For editing experiments conducted in the HCE cell line models, base editing efficiencies were calculated using the following equation: [(guanine % of a transfected sample – guanine % of the untransfected control)/adenine % in untransfected control]. Indel frequencies were calculated as the ratio between the number of reads with indels and the total aligned reads.

Profiling Guide-Dependent OT Effects

Genome-wide OT sites were predicted using Cas-OFFinder²⁷ with protospacer adjacent motif (PAM) set to “NGN.” The OT list generated was further ordered based on the number and location of mismatches identified. Sites with a lower number of mismatches and with mismatches outside of the PAM-proximal half of the guide sequence (i.e., seed) were ranked higher. For each guide, the top five nominated OTs were profiled by amplicon sequencing to define OT editing efficiencies.

AAV Transduction of Human Cells

HCE cells were washed with 1× PBS, dissociated with TrypLE Express, and neutralized with growth media. These cells were then pelleted and resuspended in serum-free DMEM/F12 to a final concentration of 2.5 × 10⁴ cells per 100 µL. GFP-expressing AAV viruses (prepared by University of Michigan Vector Core using standard methods) were then added to 100 µL cells at an multiplicity of infection (MOI) of 1 × 10⁵. Transduced cells were analyzed by flow cytometry with a ZE5 Cell Analyzer (BioRad, Hercules, CA, USA) 6 days after transduction. Transduction efficiency was determined by quantifying the percentage of GFP-positive cells over the total cell population.

AAV Transduction and Imaging

The human tissue experiments complied with the guidelines of the ARVO Best Practices for Using Human Eye Tissue in Research (November 2021). Donor corneas with an endothelial cell count >2000 per mm² and an absence of epithelial or stromal scarring or infiltrate were obtained from an accredited eye bank (Eversight, Ann Arbor, MI, USA) and maintained in Optosol-GS at 4°C until use, which occurred at <24 hours after procurement. Corneal donor tissue information for the two donors is as follows. Donor 1 was a white/Caucasian man aged 67 years who had chronic obstructive pulmonary disease as the primary cause of death, with a time from death to preservation of 6 hours, 27 minutes. Donor 2 was a black woman aged 74 years who had breast cancer as the primary cause of death, with a time from death to preservation of 16 hours 12 minutes. Four corneas from the two separate donors were cut into four wedges per cornea, transferred into DMEM 10% FBS tissue culture media, and cultured as a suspension. Each AAV (serotypes 1, 2, 5, 8, 9, B10, PhPeB, and DJ) concentration was normalized to a targeted concentration of 5 × 10¹² vg/mL, and 2 µL of each respective AAV was systematically injected into the anterior stroma of corneal wedges approximately 1 mm from the limbus utilizing a 34-gauge nanofil needle (World Precision Instruments, Sarasota, FL, USA) in duplicate. The wedges were maintained at 37°C for 14 days in DMEM 10% FBS media with antibiotics, with media changes every 2 days. On days 7 and 14, the corneal wedges were imaged using the EVOS FL Fluorescence Microscope (Invitrogen) at 10× and 4× magnification at 40% intensity using the GFP fluorescent channel. On day 14, transduced wedges were fixed in 4% paraformaldehyde overnight and washed in PBS. Corneal wedges were permeabilized with 1% Triton-X100 in PBS for 30 minutes at RT and washed with 1× PBS for 5 minutes; transferred to a 24-well plate; blocked in 2% BSA, 5% normal donkey serum, 5% normal goat serum, and 1% Triton-X100 in 1× PBS for 2.5 hours at RT; and incubated in primary antibody solution (1: 500 chicken anti-GFP in blocking solution) at 4°C for 2 days with gentle rocking. Corneas were washed three times over 3 hours with 1× PBS, followed by incubation overnight at 4°C with secondary antibody (1:400 goat anti-mouse Alexa 647, 1:500 donkey anti-chicken Alexa 488, 1:1000 DAPI in blocking solution). Corneas were washed four times over 3 hours with 1× PBS, excess sclera was removed from the wedges, three relief cuts were made along the outer edge of the wedges to allow flatmounting, and the corneas were mounted in Prolong Gold Antifade Mountant (ThermoFisher). Corneal wedges were tile-imaged on a Leica DM6000, and high-power imaging was done on the Leica SELLARIS 8 Falcon confocal microscope. Images were processed using the LAS X software (Leica, Boston MA, USA), followed by annotation in Photoshop (Adobe, San Jose, CA, USA). Epithelium Z-slices in the 20× images were identified using cell/nucleus morphology. The remaining Z-slices were split in half to represent the anterior and deep stroma.

Corneal Fluorescence Quantification

Whole cornea immunofluorescence images were analyzed using ImageJ2. The GFP channel for all images was imported into ImageJ (National Institutes of Health, Bethesda, MD, USA), and a selection was made containing the entire cornea minus the outer edges to avoid edge effects. Integrated density was measured for each slice in the Z-stack and then averaged to determine the integrated density of each corneal wedge (https://theolb.readthedocs.io/en/latest/imaging/measuring-cell-fluorescence-using-imagej.html). This value was then divided by the mean integrated density of the AAV2-transduced wedge, matching the replicate number (1 or 2) to normalize the integrated density. Mean integrated density and standard deviation for each AAV serotype were calculated and plotted using Excel.

Immunofluorescence

HCE cells were seeded in 8-well chamber slides (Falcon, 354108, Corning, NY, USA) at 1 × 10⁴ cells/well. Cells were fixed for 15 minutes in 4% paraformaldehyde (PFA)/PBS at RT and then rinsed in PBS. The cells were permeabilized for 15 minutes in 0.3% Triton-X100/PBS at RT and then incubated in blocking solution for 1 hour at RT (10% normal donkey serum and 1% BSA in PBS). Goat anti–human beta IG-H3 (1:100 dilution, cat. AF-2935; R&D Systems, Minneapolis, MN, USA) and mouse anti-human lysosomal-associated membrane protein 2 (LAMP2; 1:70 dilution, clone H4B4; Developmental Studies Hybridoma Bank, Iowa City, IA, USA) primary antibodies were incubated overnight at 4°C. After washing with PBS three to four times, the cells were incubated with secondary antibodies, donkey anti-goat AF488 and donkey anti-mouse AF647, for 1 hour at RT. Nuclei were counterstained with Hoechst (Vector Laboratories, Burlingame, CA, USA). The slides were mounted with Prolong Gold antifade reagent (Invitrogen). The images were captured using the Leica SP8 confocal microscope. For quantification, four to six 40× fields were imaged per well, and cells were individually scored for colocalization of LAMP2 and TGFBI across genotypes and conditions.

Statistics and Reproducibility

Statistical evaluations were conducted with GraphPad Prism 10.2.2 (GraphPad Software, La Jolla, CA, USA). In gene editing experiments for bystander mitigation using TadA-8e variants, P values were calculated using unpaired two-tailed t-tests, with data derived from n = 3 to 4 biologically independent experiments (NS, P > 0.05; *P ≤ 0.05; **P ≤ 0.01). For TGFBI immunofluorescence experiments, statistical significance was assessed using one-way ANOVA, followed by Tukey's test, with n = 4 to 8 samples per condition (*P ≤ 0.05, **P ≤ 0.01, ****P ≤ 0.0001).

Results

Genotypic and Phenotypic Spectrum of Our Patient Population

To define a target population and the specific patient demographics in our local University of Michigan patient population, 26 individuals from 24 families with lattice (13), granular (8), Reis-Buckler (1), or Avellino (1) corneal dystrophy were recruited and sequenced for exon 4 and exon 12 for the most common variants. We identified the R124C variants in 10 families with lattice corneal dystrophy and the R555W variants in 4 families with granular corneal dystrophy, with an overall diagnostic rate of 92% (22/24) for known TGFBI variants (Figs. 1A, 1B). Specific details of genotype–phenotype correlation and clinical phenotyping, including recurrence risk, will be reported elsewhere.

Figure 1.

The genotypic spectrum of the local patient population that would benefit from a base editing correction strategy for TGFBI-linked corneal dystrophies. (A) Genetic testing outcomes showing the fraction of cases with p.R555W and p.R124C variants among 24 corneal dystrophy probands. Together, these two common mutations account for 54% (13/24) of familial cases. (B) Breakdown of genotypes by dystrophy subtype. R555W was seen only in the granular dystrophy subtype, while R124C explained a large fraction of the lattice dystrophy cases in our cohort.

TGFBI p.R124C and p.R555W Are Amenable to ABE Correction in Human Cells

The p.R124C (c.370C>T) and p.R555W (c.1663C>T) mutations are located in exon 4 and exon 12 of the TGFBI gene, respectively. We designed multiple sgRNA-PAM strategies to position ABE's editing window over the pathogenic adenines to enable conversion back to the WT sequences (Figs. 2A, 2B). We used ABE8e-NG, which recognizes NGN PAM via a Cas9-NG variant²⁸ and uses a laboratory-evolved transfer RNA adenosine deaminase TadA-8e known for its high A•T-to-G•C base-editing efficiency.¹⁶ For the TGFBI p.R124C mutation, two guide RNAs place the pathogenic adenine at the fifth and seventh nucleotide (nt) positions of the target sequence, respectively, counting from the PAM-distal end (Fig. 2A). For the TGFBI p.R555W mutation, three guide RNAs place the pathogenic adenine at the fifth, sixth, and seventh nt position, respectively (Fig. 2B). Importantly, the absence of additional adenines within the editing window prevents undesired modification on the WT chromosome and would enable allele-specific correction, an essential feature for treating autosomal dominant disorders.

Figure 2.

ABE8e-NG can correct TGFBI R124C and R555W in human cells. (A, B) Sequence of TGFBI regions at the p.R124C or p.R555W mutation. Pathogenic adenines are in red, and intended amino acid corrections are boxed. PAMs and sgRNA-paired regions are denoted by pink and blue lines, respectively. Numbering: nucleotide positions 1 to 20 of the guide sequence are counted from the PAM-distal end. The bystander adenine edited at the R124C site in F and G is shown in orange. (C) Experimental workflow for feasibility assays with episomal targets. Episomal target plasmid bearing a TGFBI segment containing the p.R124C or p.R555W mutation was cotransfected into HEK293T cells with plasmids of ABE8e-NG and a targeting guide. Base editing efficiency was quantified by deep sequencing amplicons derived from the episomal target. (D) Heatmap of mean A•T-to-G•C editing efficiencies on the episomal TGFBI p.R555W targeted by guides G1, G2, and G3. Pathogenic adenine is highlighted in red. (E) Bar graph quantification of A•T-to-G•C editing efficiencies and indel rates from the same experiment as in D. Data shown are mean ± SEM, n = 3. (F) Heatmap of mean A•T-to-G•C editing efficiencies on the episomal TGFBI p.R124C targeted by guides G1 and G2. The pathogenic adenine is in red. (G) Bar graph quantification of A•T-to-G•C editing efficiencies and indel rates from the same experiment as in F. Data shown are mean ± SEM, n = 3.

To rapidly assess the base conversion efficiencies of these guides, we transfected HEK293T cells with plasmids encoding ABE8e-NG, a sgRNA targeting either p.R124C or p.R555W, and an episomal reporter plasmid containing a ∼200 nt TGFBI sequence bearing either mutation; 4 days posttransfection, A•T-to-G•C editing efficiencies were quantified by amplicon sequencing (Fig. 2C). On the p.R555W-containing episomal target, guides R555W-G1, G2, and G3 achieved desired A•T-to-G•C editing of 40.0% ± 2.2%, 18.8% ± 2.6%, and 29.9% ± 0.6%, respectively, and these robust outcomes were accompanied by modest indel rates of 3.9% ± 0.2%, 5.4% ± 0.7%, and 0.6% ± 0.1% (Figs. 2D, 2E). Based on their superior performance, R555W-G1 and G3 were selected for subsequent experiments. For the TGFBI p.R124C-containing episomal target, guides G1 and G2 yielded 31.8% ± 1.2% and 30.3% ± 0.5% A•T-to-G•C conversion efficiencies at the pathogenic adenine, with indel rates of 2.3% ± 0.3% and 3.4% ± 0.6%, respectively (Figs. 2F, 2G). Altogether, these pilot tests demonstrate that the two common TGFBI mutations are suited for ABE-mediated gene correction.

We observed bystander editing at an unexpected adenine position: A16 for R124C-G1 at an efficiency of 1.7% ± 0.01% and A14 for R124C-G2 at an efficiency of 7.4% ± 0.5% (Figs. 2F, 2G). These positions lie well outside ABE8e's typical editing window (protospacer positions 4–8)¹⁶ and are not readily explained. To mitigate potential phenotypic consequences from the resulting tyrosine-to-histidine change, we later explored using TadA-8e variants previously shown to narrow the editing window and reduce bystander activity^29,30 in HCE cell models.

Creating HCE Cell Models for TGFBI p.R124C and p.R555W

Encouraged by the promising episomal targeting results, we next evaluated our ABE correction strategies in physiologically relevant HCE cells. To generate cell models heterozygous for the TGFBI p.R124C or p.R555W mutation, we employed a Cas9-HDR approach. WT HCE cells were electroporated with Cas9 nuclease RNPs targeting the p.R124C or p.R555W locus, along with ssODN donor templates carrying the desired mutations (Figs. 3A–C). HDR leading to the p.R124C or p.R555W mutation occurred at efficiencies of 7.0% and 3.1%, respectively, based on amplicon sequencing of the bulk-edited cell populations. These were accompanied by Cas9-induced indel rates of 44.7% and 20.9% (Fig. 3D).

Figure 3.

Creating human corneal epithelial cell models for TGFBI^R124C/WT and TGFBI^R555W/WT via the Cas9 nuclease-HDR method. (A) Schematics of the HCE cell model creation workflow. Cas9 RNPs were co-electroporated into WT HCE cells with a corresponding ssODN donor. The edited cell pool was sorted into single cells, expanded, and genotyped by amplicon sequencing to obtain the desired heterozygous clones. (B, C) Schematics of Cas9-HDR targeting and repair strategies to generate the TGFBI p.R124C and p.R555W mutations. The 5′-NGG PAM for Cas9 and sgRNA-paired regions are denoted by the orange and blue lines, respectively. Cas9 nuclease cleavage sites, blue arrowheads. TGFBI point mutations encoded on the ssODN donors are highlighted in red. (D) NGS quantifications of desired HDR efficiencies and accompanying indels in bulk-edited cells. Data represent a single experiment for each model that proceeded to single-cell sorting. n = 1. (E) Genotyping results categorizing isolated single-cell clones into four groups: unedited, het (HDR + WT), indels + HDR, and indels or indels + WT. TGFBI^R124C/WT and TGFBI^R555W/WT single clones were expanded and used in later studies. n = 1.

We then isolated single-cell clones and genotyped them by amplicon sequencing (Fig. 3A). Among 80 clones isolated from the TGFBI p.R124C modeling experiment, 2 were heterozygous for p.R124C (TGFBI^R124C/WT); the remaining clones were unedited (16 clones), were HDR with indels (14 clones), or had indels or indels with WT only (48 clones) (Fig. 3E). For the TGFBI p.R555W modeling experiment, we analyzed 30 clones and obtained 4 heterozygous for p.R555W (TGFBI^R555W/WT); the rest were unedited (12 clones), were HDR with indels (3 clones), or had indels or indels with WT only (12 clones) (Fig. 3E). Together, the findings confirm our successful establishment of two HCE models for corneal dystrophies. These results also highlight the challenges of achieving precise editing via Cas9-HDR, particularly due to limited HDR efficiency and the accompanying high indel rates.

Highly Efficient and Precise Mutation Correction in the HCE-TGFBI-R555W Model

To evaluate the efficacy of our therapeutic strategies outlined in Figures 2A and 2B, we electroporated ABE8e-NG mRNA and chemically modified synthetic guide RNAs into the HCE models (Fig. 4A). In HCE-TGFBI^R555W/WT cells, R555W-G1 and G3 yielded 61.9% ± 3.6% and 44.6% ± 7.5% A•T-to-G•C correction of the pathogenic adenine, respectively, both with minimal indel formation (0.1% ± 0.03%), and no nearby adenine exhibited bystander editing (Figs. 4B, 4C).

Figure 4.

Efficient and precise mutation correction of TGFBI p.R555W in an HCE model. (A) Experimental workflow of all-RNA delivery of ABE8e-NG and its guide RNA into the HCE model for corneal dystrophies used in Figures 4 and 5. (B) Heatmap of mean A•T-to-G•C editing efficiencies at the genomic TGFBI p.R555W site by guides G1 and G3. Pathogenic adenine is highlighted in red. (C) Bar graph of A•T-to-G•C editing efficiencies and indel rates from the same experiment as in B. Data shown are mean ± SEM, n = 3. (D, E) Top five Cas/guide-dependent OTs predicted computationally for each guide. Mismatch nucleotides are highlighted in red. (F, G) Peak A•T-to-G•C editing efficiencies detected at on-target and OT sites; data shown are mean ± SEM, n = 3.

To assess potential safety risks from OT editing at unintended genomic sites, we used Cas-OFFinder²⁷ to computationally predict guide-dependent OT sites. All predicted sites, except for one, contained two or more mismatches relative to the intended on-target sequence (Figs. 4D, 4E). We then performed targeted amplicon sequencing at the top five OT sites for each guide to profile the total editing outcomes, including base changes and indels. For R555W-G1, all OT sites showed peak editing below 0.12%, except OT2, which exhibited 1.2% ± 0.03% A•T-to-G•C conversion that would result in a silent mutation (Fig. 4F). For R555W-G3, OT1 displayed 9.4% ± 1.6% A•T-to-G•C modification, likely due to high sequence similarity to the on-target site, differing by only 1 nucleotide. No discernible editing was detected at the other four OT sites (Fig. 4G). Since OT1 is in an intergenic region, we consider its safety risks minimal. Overall, these results support the feasibility and safety of using ABE8e-NG for efficient correction of TGFBI p.R555W correction in an HCE granular corneal dystrophy model.

Efficient Correction of TGFBI p.R124C and Mitigation of Bystanding Editing

To assess our strategy for correcting the TGFBI p.R124C mutation, we electroporated ABE8e-NG mRNA along with the corresponding synthetic guides into the HCE-TGFBI^R124C/WT model and measured the desired A•T-to-G•C editing, indel formation, bystander events, and OT effects. R124W-G1 and G2 guides yielded 60.1% ± 0.76% and 97.6% ± 0.49% A•T-to-G•C correction of the pathogenic adenine, with minimal indel frequencies (0.2% for both; Fig. 5A).

Figure 5.

Efficient correction of TGFBI p.R124C and bystanding editing mitigation in an HCE model. (A) Quantification of A•T-to-G•C editing efficiencies and indel rates for p.R124C-targeting guides G1 and G2, following their co-electroporation with ABE8e-NG into HCE-TGFBI^R124C/WT model. Data shown are mean ± SEM, n = 3. (B) Mitigation of unwanted bystander editing at A14 for R124C-G2, via the use of deaminase variants TadA-8e-V28F and TadA-8e-M151D. Data shown are mean ± SEM, n = 3. P values were calculated using two-tailed t-tests, with data derived from n = 3 biologically independent experiments. NS, P > 0.05; **P ≤ 0.01. (C, D) Top five Cas/guide-dependent OT sites predicted computationally nominated for each guide. Mismatch nucleotides are highlighted in red. (E, F) Peak A•T-to-G•C editing efficiencies detected at on-target and OT sites. Data shown are mean ± SEM, n = 3. For R124C G2, samples from the experiment using TadA-8e-V28F from B were used.

Bystander editing at the out-of-the-window adenine, as previously noted in episomal assays in Figure 2D, was minimal for R124C-G1 (0.67% ± 0.08%) but substantial for R124C-G2 (9.56% ± 1.3%) in the genomic context (Fig. 5A). To address this, we evaluated two TadA-8e variants, V28F and M151D, known to sharpen the editing window.^29,30 For R124C-G2, both variants substantially reduced bystander A14 editing from 9.6% ± 1.3% to 0.51% ± 0.11% and 1.0% ± 0.07%, while maintaining the high desired editing at 91.4% ± 3.7% and 69.6% ± 5.9%, respectively (Fig. 5B). Based on its near-complete mitigation of bystander activity, V28F was deemed the most suitable for therapeutic applications using R124C-G2 and was subsequently evaluated for OT effects.

We identified the top OT sites for R124C-G1 and G2 using Cas-OFFinder, all of which contained 2-nt mismatches relative to their respective on-target sequences (Figs. 5C, 5D). Amplicon sequencing from two editing experiments, one with TadA-8e-NG and R124C-G1, the other with TadA-8e-NG V28F and R124C-G2, revealed negligible editing at all top five OT sites, indicating a favorable safety profile (Figs. 5E, 5F). Collectively, our findings establish a framework for evaluating and mitigating unwanted editing events in base editing therapies in HCE corneal dystrophy models.

Correction of TGFBI p.R124C Rescues TGFBI Localization

We next assessed the functional impact of pathogenic TGFBI mutations and their genetic correction by examining TGFBI subcellular localization using immunofluorescence microscopy. In HCE cells, the p.R124C mutant protein showed increased colocalization with LAMP2, indicating aberrant lysosomal sequestration of TGFBI, whereas p.R555W showed no such phenotype (Supplementary Figs. 3A, 3B). Consistent with our result, mislocalization to the lysosome has also been reported for the R124H TGFBI variant.³¹ Following ABE-mediated correction of p.R124C, normal TGFBI localization was restored in edited p.R124C HCE cells, suggesting rescue of TGFBI intracellular trafficking (Figs. 6A, 6B). Together, these data demonstrate functional restoration of TGFBI trafficking following our base editing strategy, supporting its therapeutic potential.

Figure 6.

ABE8e-NG correction of p.R124C rescued TGFBI protein localization in edited HCE cells. (A) WT, unedited HCE-R124C cells, and R124C cells edited with R124C-G2 and ABE8e-NG carrying either the standard or the V28F window-narrowing variant of TadA-8e were assessed for TGFBI-lysosomal colocalization by immunostaining of TGFBI (green) and LAMP2 (red) 2 days after seeding in 8-well chambers. Images were taken under a confocal microscope at 40×. Scale bar: 50 µm. (B) Quantification of TGFBI and LAMP2 colocalization (white arrow) from A demonstrates restoration of normal localization of TGFBI by our gene editing strategies. White arrowhead indicates non-colocalization of TGFBI and LAMP2. n = 4 to 8 samples per condition. NS, P > 0.05; **P ≤ 0.01; ****P ≤ 0.0001. P values were assessed using one-way ANOVA, followed by Tukey's multiple comparison.

Defining AAV Serotypes for Effective Transduction Into HCE and Eyebank Corneal Tissue

In gene therapy, effective delivery of gene editing agents is crucial for reaching the intended tissue and cell types while ensuring therapeutic safety. AAV is a leading vehicle for ocular gene therapy due to its effective ocular transduction, low genomic integration risk, long-term expression, and proven clinical safety.³² To start optimizing conditions for the ultimate delivery of our base editing agents into human cornea, we packaged an eGFP reporter gene into eight different AAV serotypes and screened for their transduction into HCE cells and eyebank human corneal tissue. We first transduced WT HCE cells with each of eight AAV serotypes carrying a GFP cassette and assessed infectivity by flow cytometry. High infectivity was observed with AAV1, AAV2, AAV5, and AAV-DJ in HCE cells, with 90.8% ± 1.1%, 99.7% ± 0.1%, 83.9% ± 4.6%, and 99.7% ± 0.1% of the treated cells exhibiting GFP fluorescence, respectively (Fig. 7A). In contrast, AAV8, AAV9, AAV-B10, and AAV-PHPeB showed poor infectivity, with only 5.6% ± 0.4%, 2.8% ± 0.3%, 0.7% ± 0.1%, and 1.1% ± 0.1% efficiency, respectively (Fig. 7A).

Figure 7.

Defining AAV serotypes for effective transduction in HCE cells and in human donor corneal tissue. (A) Flow cytometry of WT HCE cells infected with various AAV-GFP serotypes. The percentages of GFP-positive cells are plotted as mean ± SEM, n = 3. (B) Schematic of the transduction strategy for human corneal wedges. Corneal wedges injected with AAV encoding eGFP under the control of a CAG promoter. Wedges were cultured for 2 weeks prior to fixation and immunostaining. (C) Tiled immunofluorescence images of transduced corneal wedges stained with an anti-GFP antibody. Inset images display a region of high fluorescence intensity. 10× magnification. Scale bars: overview, 1 mm; inset, 100 µm. n = 2 corneal wedges. Shown are results from a representative experiment. (D) GFP intensity analysis on immunolabeled transduced corneal wedges from C. The integrated density relative to AAV2 for each pair of transduced corneal wedges is plotted as mean ± SD.

To further refine AAV serotype selection in human tissue, we systematically transduced human corneal tissue using AAV-eGFP via an intrastromal injection method (Fig. 7B), which has been safe and effective for other corneal diseases.^33,34 Cultured corneas showed active GFP expression by 1 week (Supplementary Fig. S1), with increasing expression after 2 weeks (Supplementary Fig. S2). After harvest, fixation, and immunostaining, we quantified transduction efficiency through the corneal stroma and epithelium on tiled images throughout the entire corneal wedges (Fig. 7C). The quantitative analysis revealed good expression across most serotypes, with AAV1 showing the strongest overall expression (Fig. 7D). To further define the penetration depth and cell-type specificity of AAVs, we conducted confocal microscopy imaging on corneas transduced with AAV1 and AAV-DJ, two serotypes that showed high infectivity in both HCE cells and corneal donor tissues. Imaging through the epithelium revealed stronger expression with AAV1, while both serotypes exhibited robust expression throughout the anterior and deep stroma (Figs. 7D, 8A–F). Together, our results support AAV1 as the leading candidate for in vitro and in vivo corneal transduction for further clinical testing.

Figure 8.

Determining the depth of corneal infectivity for highly effective AAV serotypes. (A–C) Confocal images for AAV1-CAG-GFP transduced cells at various depths of imaging: A, epithelium; B, anterior stroma; and C, deep stroma. (D–F) Confocal images for AAV DJ-CAG-GFP transduced cells at various depths of imaging: D, epithelium; E, anterior stroma; and F, deep stroma. Corneas were stained with DAPI (4′,6-diamidino-2-phenylindole) to label nuclei and an anti-GFP antibody. Merge is an overlay of the DAPI and anti-GFP images. Scale bar: 100 µm.

Discussion

Unmet Medical Need in Corneal Therapeutics

Corneal dystrophies affect approximately 8 in 10,000 people in the United States,³⁵ and current treatments carry significant morbidity while failing to address the underlying disease pathology.⁵ Here, we present compelling data on efficient base editing strategies to tackle TGFBI-linked corneal dystrophies caused by the p.R555W and p.R124C mutations. In our patient population, consistent with previous reports in other demographic groups,⁴ these two variants together account for over 50% of cases, with p.R124C accounting for most lattice dystrophy cases and p.R555W accounting for most granular dystrophy cases. As such, these two mutations represent the most promising initial targets for therapeutic development, particularly given their amenability to the adenine base editing approach.

Base Editing as a Superior Therapeutic Approach for Corneal Dystrophies

In the development of gene correction therapies, base editing offers distinct advantages over traditional Cas9 nuclease-mediated HDR, which has been explored for correcting gain-of-function TGFBI variants in corneal dystrophies. For example, Taketani and colleagues³⁶ reported Cas9-HDR correction in patient-derived primary corneal keratocytes heterozygous for TGFBI p.R124H. However, treating autosomal dominant diseases using Cas9-HDR faces fundamental challenges. First, achieving allele discrimination, targeting the mutant allele while sparing the WT allele from Cas9 cleavage, is inherently difficult. While placing the single-nucleotide polymorphism within the PAM of Cas9 seems an attractive way to bias cleavage toward the mutant allele, this strategy depends on the mutation converting a non-PAM into a good PAM. In practice, this is often constrained by the requirement for a mutation-created PAM, and protection of the WT TGFBI allele is often incomplete.³⁷ Second, Cas9-induced DSBs raise safety concerns, including indel by-products, chromosomal rearrangements, and activation of p53 stress responses.^38–40 Finally, the intrinsic low efficiency and cell cycle dependence of HDR significantly constrain utility for in vivo therapeutic gene correction.

An alternative to Cas9-HDR correction is to induce Cas9-mediated indels to selectively ablate the pathogenic allele in haplosufficient autosomal dominant diseases.⁴¹ But this approach also faces challenges, including DSB-associated toxicity and insufficient allele discrimination at certain mutations, such as TGFBI-124C and R555W.³⁷ Moreover, WT TGFBI plays crucial roles in cell adhesion, migration, and corneal wound healing,^42–44 and it often functions as a tumor suppressor in various contexts.^45,46 Thus, stringent protection of the WT allele to preserve normal TGFBI protein function is critical for patient safety.

Our base editing strategies were comprehensively evaluated for efficiency, precision, and safety, including assessments of on-target efficacy, bystander editing, and OT effects. Correction of the pathogenic adenines for p.R124C and p.R555W reached 91% and 62%, with undesired bystander editing kept below 0.5% on both alleles (Figs. 4, 5). We observed phenotypic rescue of TGFBI localization in edited p.R124C HCE cells, further supporting base editing as a viable therapeutic strategy (Fig. 6). OT profiling revealed negligible editing (<0.1%) at 18 of 20 computationally predicted top OT sites. The remaining two sites showed 1.2% and 9.4% editing, respectively, but resulted in silent mutations or occurred within intergenic regions, thereby dampening safety concerns. Nevertheless, low levels of OT and bystander editing could still post safety concerns for clinical translation, underscoring the need for further mitigation using high-fidelity Cas9 variants⁴⁷ or alternative window-narrowing TadA-8e variants,⁴⁸ along with comprehensive OT evaluation in preclinical studies.

To date, more than 70 pathogenic TGFBI variants have been reported across multiple corneal dystrophy subtypes.⁴⁹ In addition to p.R124C and p.R555W, other recurrent pathogenic variants include p.R124L, p.R124H, and p.R555Q.⁴⁹ Although these variants are not well suited to our therapeutic approach owing to mutation-specific or bystander adenine constraints, they may be amenable to alternative gene editing strategies, such as prime editing or adenine-to-cytosine base editing, as outlined in Supplementary Figure S4.

Future Therapeutic Deployment

Our findings support AAV delivery as a viable strategy for therapeutic targeting of the cornea. We observed high infectivity across multiple AAV subtypes following intrastromal delivery (Figs. 7, 8). While previous studies in mice have demonstrated the potential of AAVs for ocular gene therapy in the cornea,^34,50,51 systemic analysis across AAV serotypes in human corneal cells and tissues has been limited. We show that AAV1 enables robust transduction in both corneal epithelial cell lines and human donor tissue, supporting its suitability for epithelial and stromal delivery into the cornea. We further showed that a single anterior stromal injection of AAV1 leads to viral transduction throughout the epithelium and corneal stroma, supporting AAVs as a viable delivery method for epithelial and anterior stromal dystrophies. Importantly, we did not see penetration through the full cornea, suggesting a minimal risk of OT infectivity with intrastromal injection. Notably, the scope of the current work is limited to in vitro cell models and ex vivo human donor tissues. These results lay the groundwork for future preclinical studies evaluating in vivo AAV delivery, therapeutic efficacy, OTs, immune reactions, and potential cornea nerve responses associated with our base correction strategy. Such future studies will benefit from the development of humanized animal models carrying the specific pathogenic TGFBI variants investigated in this work.

More broadly, our work outlined a generalizable base editing strategy to treat autosomal dominant ocular disorders. It is applicable when only the pathogenic adenine or cytosine, with no additional editable bases, falls within the 5-nt editing window of a Cas9-derived base editor. This approach safeguards both the WT and corrected alleles from undesired cleavage or sequence modifications, an essential measure to ensure safety and prevent unforeseen oncogenic effects. Moreover, our approach is suitable for disorders with limited genetic heterogeneity, such as TGFBI-linked corneal dystrophies, where a few hotspot mutations, targetable with validated guides, cover most patient cases. In contrast, ocular conditions with extensive genetic heterogeneity, such as RHO-linked autosomal dominant retinitis pigmentosa, which involves over 150 distinct gain-of-function RHO mutations, each affecting only a small percentage of patients, demand a different philosophy.⁵² Developing a separate therapy for each mutation is not economically feasible, underscoring the need for mutation-independent, allele-specific ablation strategies.^52,53

Genome editing therapies may carry long-term risks, including unintended sequence alterations, structure rearrangements, and genome instability, which could lead to oncogenic effects or disruption of normal cellular functions.^54,55 Accordingly, translational development for TGFBI-linked corneal dystrophies will require long-term editing and safety assessments in preclinical animal or corneal organoid models to assess oncogenic transformation, genome instability, and broader cellular consequences of editing. Immunogenicity is another key safety consideration for gene therapy, as immune responses may be triggered by AAV vectors and the gene editing proteins.^33,56 Although the cornea benefits from relative immune privilege and localized vector delivery,³³ in vivo translation of TGFBI editing will require careful evaluation of potential inflammation, vector biodistribution, and long-term corneal integrity.