Arrayed dual-gRNA CRISPR screening platform for C9orf72 repeat expansion excision in patient iPSCs

Abstract

An intronic hexanucleotide repeat expansion in C9orf72 is the leading genetic cause of both frontotemporal dementia and amyotrophic lateral sclerosis (C9-FTD/ALS). We have previously demonstrated that CRISPR-Cas9 excision of the repeat expansion in patient iPSCs reverts pathological hallmarks of C9-FTD/ALS. Here, we aim to identify efficient and safe gRNAs for CRISPR-spCas9 dual-gRNA excision of the C9-repeat expansion. Utilizing novel ddPCR and single-molecule sequencing assays, we screened 120 gRNA pairs, comparing 64 bi-allelic, intronic excisions of the repeat region to 56 allele-specific excisions of the mutant allele in patient iPSCs, ranking them by efficiency. Bi-allelic excisions of the intronic repeat region were more efficient than excisions of the mutant allele. Single gRNA indel rates can nominate likely efficient gRNA pairs, but these pairs must be tested empirically. The length of the repeat expansion did not impact excision efficiency; rather, the activity of individual gRNAs drove excision efficiencies. Using whole genome sequencing and INDUCE-seq, we found only one detectable off-target of those nominated by Cas-OFFinder and CHANGE-seq across 4 of the most efficient gRNAs. This study advances the development of targeted therapies for C9-FTD/ALS and establishes a framework for dual-gRNA screening in patient iPSCs applicable to other repeat expansions.

Graphical abstract

Clelland and colleagues advance CRISPR gene therapy for C9orf72-FTD/ALS by screening gRNA pairs for excision of the C9orf72’s intronic repeat region and selective excision of the mutant allele in patient iPSCs. They optimize off-target analysis in patient iPSCs using three unbiased nomination methods and a novel whole genome sequencing validation pipeline.

Introduction

Repeat expansion of GGGGCC in the first intron of C9orf72 is the leading known genetic cause of both frontotemporal dementia (FTD) and amyotrophic lateral sclerosis (ALS) (C9-FTD/ALS).^1,2 Currently, there is no cure for this fatal neurodegenerative disease. Phase 1 and 1b/2a clinical trials of anti-sense oligonucleotide therapy targeting the sense C9-transcripts for the treatment of C9-FTD/ALS were stopped due to futility highlighting the urgent need for additional therapeutic approaches.³ We and others have shown that removing the C9orf72 repeat expansion from the genome can reverse pathology in human induced motor neurons^4,5,6 and in mice.^4,5 We have previously shown that disrupting transcribed exons 1A or 1B flanking the repeat expansion disrupts normal gene expression in induced human motor neurons.⁷ In contrast to other approaches that excise flanking transcribed exons,^4,5 our approaches⁶ aim to preserve normal gene function and protein levels by excising only the intronic region containing the repeat expansion on both alleles or selectively excising the mutant allele, leaving the non-diseased allele intact. We previously showed that removing the repeat expansion from the genome through either approach was critical to reverting the hallmarks of C9-FTD/ALS pathology (RNA abnormalities, dipeptide repeat (DPR) expression, and TDP-43 mislocalization) in induced pluripotent stem cell (iPSC)-derived motor neurons.⁶

Dual-gRNA CRISPR-Cas9 excision has emerged as a powerful tool for addressing mutations that are difficult to target, including the C9orf72 repeat expansion.^4,5,6 Single-gRNA (sgRNA) methods for gene editing are proving safe and effective in clinical settings,^6,8,9,10,11 but sgRNA strategies can only silence or de-repress genes. We have previously shown that attempting to silence the repeat expansion through removal of exon 1A did not halt antisense pathology.⁶ Dual gRNA excision is particularly useful for knocking out genes with alternative start sites, such as C9orf72, for which indels produced by sgRNAs are incapable of inactivating all gene products. In addition, by targeting single nucleotide polymorphisms (SNPs) in cis with a mutation, dual-gRNAs can remove an entire allele harboring a mutation, even when the mutation itself cannot be targeted due to off-targets (such as for C9orf72) or inaccessibility. Screening dual-gRNAs by comparing efficiency of gRNA pairs has proven valuable in preclinical trials for CRISPR-Cas9 gene therapies, but typically involves low throughput comparisons, relying heavily on bioinformatic predictions. For example, recent repeat expansion cutting screens have screened limited gRNA pairs: C9-FTD/ALS (2–4 pairs),^4,5,6 Huntington’s disease (2 pairs),¹² Duchenne’s muscular dystrophy (8 pairs),¹³ and Hepatitis B (11 pairs).¹⁴ When dual-gRNA approaches offer dozens of unique gRNA combinations, accurate on-target editing can be quantified using arrayed dual-gRNA droplet digital PCR (ddPCR) screens, as we have shown here.

We have previously shown that removing the repeat expansion through dual-gRNA excision reverts pathology in patient-derived induced motor neurons,⁶ but what are the most efficient gRNAs that can safely achieve such an edit? This study aims to identify the most efficient dual-gRNAs for excising the C9orf72 repeat expansion and to evaluate their off-target profiles, utilizing several tools we have developed to enable screening at a difficult to sequence repeat region. First, we engineered a patient iPSC line to facilitate our screen of repeat expansion excision and mitigate reading out editing of the wild-type (WT) allele. We next screened 64 bi-allelic and 54 allele-specific dual-gRNAs using ddPCR assays to quantify excisions to enable head to head comparisons of each gRNA pair. We then validated top gRNA pairs across 4 patient iPSCs lines with both alleles intact and varying repeat lengths, and determined unwanted editing outcomes, such as indels or inversions with ddPCR and single-molecule sequencing. For CRISPR gene therapy to be safe, it is crucial to limit off-target editing, which can disrupt vital genes and cause chromosomal abnormalities, risking the development of tumors and other adverse effects.¹⁵ We tested gRNA pairs with the lowest predicted potential for off-target editing and empirically evaluated off-target profiles of the most efficient gRNAs in human iPSCs using a combination of cell-free CHANGE sequencing (CHANGE-seq), in silico bioinformatic predictions, unbiased whole-genome INDUCE sequencing (INDUCE-seq), and whole genome sequencing (WGS). Our results nominate the most efficient gRNA pairs for a human-specific CRISPR therapeutic for C9-FTD/ALS. The tools we have developed overcome the challenges of quantifying on-target editing at a difficult to sequence repeat expansion and provide a roadmap for large-scale arrayed dual-gRNA screens at any locus.

Results

Bi-allelic gRNA screen design and engineered cell line to facilitate gRNA screening

We designed a bi-allelic gRNA screen to assess gRNA pairs that could excise the pathogenic repeat expansion with high efficiency while preserving surrounding coding regions. The repeat region lies in the intronic region between alternative start sites exons 1A and 1B (Figure 1A). Due to the absence of SNPs in this region, the gRNAs targeted both the wild-type (WT) and mutant alleles. We identified all intronic spCas9 gRNA pairs flanking the repeat expansion, selecting the 20 gRNA pairs with the lowest predicted off-targets as determined by CRISPOR.¹⁶ Sixteen gRNAs were proximal to exon 1A, and four gRNAs were proximal to exon 1B (Table S1). Each gRNA was tested in combination with every other gRNA, resulting in a total of 64 combinations (Figure 1A). Screening in C9-patient iPSCs is important, rather than WT or artificial systems, because the expanded repeat and the surrounding genomic context may impact editing efficiency. To prevent excision of the WT allele, which only has two repeats, from interfering with our ability to detect edits on the mutant allele, we utilized a C9^(–/282) engineered patient iPSC line with a 21 kb excision on the WT allele, leaving only the mutant allele intact⁶ (Table S2; Figure S1A). This line has a repeat expansion of approximately 280 repeats.

Figure 1.

Bi-allelic CRISPR-spCas9 gRNA screen for the excision of C9orf72 repeat expansion region in patient iPSCs

(A) The repeat expansion (red triangle) lies in the intronic region between exons 1A and 1B (red box). We tested 16 target spCas9 gRNA sites 5′ and 4 gRNA sites 3′ of the repeat expansion in the intronic region totaling 64 bi-allelic gRNA pairs. ddPCR primers and probes (turquoise arrows and line) detect excisions on either allele as a gain-of-signal. (B) Average on-target excision efficiency for each of 64 gRNA pairs tested in patient iPSCs across biological duplicates. (C) Heatmap displaying the on-target excision efficiency with each square representing the average editing efficiency across two replicates. (D) Excision efficiency of 4 top gRNA pairs, C4, C11, C12, and C13 were not significantly different across 4 unrelated and unedited patient iPSC lines (one-way ANOVA, F(3,12) = 1.08, p = 0.4. Triangles represent individual patient lines and an average of biological duplicates). (E and F) PacBio single molecule sequencing analysis of editing outcomes at the C9-locus, averaged across the 4 patient lines used in (D) comprising of excisions of the repeat expansion in the intronic region (light green), excisions of the repeat expansion extending into either exon 1A or 1B (dark green), inversion (gray) or no editing (yellow) (one-way ANOVA, mutant allele: F(3,8) = 0.64, p = 0.61; wild-type allele: F(3,8) = 10.24, p = 0.004). (G) MSD-immunoassay of poly-GA in edited pools 5 days post nucleofection in C9^(2/282), which has been previously validated to express poly-GA in iPSCs (one-way ANOVA, F(4,15) = 13.52, ∗∗∗p < 0.0001 compared to control treated with non-targeting gRNA (NT) via Dunnett’s multiple comparisons test). Error bars = SEM.

Quantification of intronic repeat expansion region excision efficacy

To quantify excisions of the intronic repeat expansion region, we designed a novel gain-of-signal ddPCR assay specific to the C9 repeat region (Figure 1A; Table S3). This assay reports a signal when the repeat expansion region is excised (Figures 1B and 1C). We validated this ddPCR assay by spiking-in ratios of unedited and edited genomic DNA (gDNA) from clonal iPSC lines: a WT line with all repeats removed (WT^(0/0)) (where excisions are expected to be measured at 100%) and a C9-knockout line with 21 kb excisions on both alleles (C9^(−/−)) (where excisions are expected to be measured at 0%) (Figure S1B). Our assay was linear from 0 repeats to 100% mixture containing expanded repeats (Figure S1B), demonstrating that our assay can detect excisions across the entire possible range of editing. We further validated this ddPCR assay by quantifying excisions in clonal lines with varying repeat lengths, including C9^(2/282), WT^(2/10), C9^(2/0), and C9^(−/−) and found that the assay could readout excision of the repeat expansion with high precision (Figure S1C).

Bi-allelic gRNA screen outcome

We tested each gRNA pair via electroporation of spCas9 ribonucleoproteins (RNPs) into the screening patient iPSC line and collected DNA 48 h later. We chose this time point to allow for editing and repair of the double-stranded break to occur and cellular recovery from electroporation but with the shortest post-electroporation interval to minimize the impact of cell division on editing outcomes. Excision efficiency of the 64 gRNA pairs ranged from 11% to 85% (Figures 1B and 1C). The most efficient gRNA pairs included the 3ꞌ guide C (Figures 1B and 1C). The top 5ꞌ gRNAs 3, 4, 11, 12, and 13 when paired with 3ꞌ gRNAs other than C (A, B, and D) had reduced excision rates by 15%–56%, indicating that excision efficiency of a gRNA pair can be promoted or limited by individual gRNAs in the pair.

Because our screening line contained only the mutant allele, we next validated the top gRNA candidates for excision in multiple patient cell lines with both alleles intact. This validation is particularly important in repeat expansion disorders because it is possible that variability in the repeat length could alter the efficiency of excision. We therefore included patient cell lines with a wide range of repeats: 282, 781, 1,063, and 1,243 repeats (Figures 1D; Table S2). We chose gRNA pairs C4, C11, C12, and C13 for validation across patient iPSC lines because these gRNA pairs had the highest editing efficiency of ∼80% in our screen. Notably, C3 also had high excision efficiency but we chose not to further test this gRNA pair because of the higher number of off-targets predicted by CRISPOR. Editing was not significantly different between the gRNA pairs across patient lines (Figure 1D). Editing efficiency was reduced in the validation iPSC lines with both alleles intact (Figure 1D) compared to the screening iPSC line with only one allele intact (Figures 1B, 1C, and S2A).

Since our ddPCR assay measuring excisions in the repeat expansion region is agnostic to the allele-specificity of editing, we next used single-molecule sequencing to determine editing rates on each allele (Figures 1E and 1F). This method allowed us to quantify excisions within the first intron housing the repeat region, as well as excisions extending into exon 1A or 1B. We also measured inversion and indel rates. For excisions within the first intron, gRNA pair C4 had rates of 65% on the mutant allele (Figure 1E, light green) and 69% on the WT allele (Figure 1F, light green). gRNA pairs C11, C12, and C13 showed higher excision rates on the mutant allele (54%–65%) (Figure 1E) compared to the WT allele (18%–24%) (Figure 1F). Exonic excisions were rare, occurring only with gRNA pair C13 on the mutant allele (0%–5%) (Figure 1E, dark green). Inversions were observed for all 4 gRNA pairs on the mutant allele (4%–14%) (Figure 1E, gray), which is in-line with previous reports.^17,18 We manually reviewed each inversion and noted that each inversion started and ended at the gRNA cut sites and did not disrupt the surrounding intron regions or exons (i.e., each inversion was a “perfect” inversion).

Next, we sought to determine whether the top bi-allelic gRNA pairs would lead to a reduction in C9-specific pathology. Poly-GA is eliminated in C9-patient cells after removal of the repeat expansion⁶ and is a biomarker used in clinical trials.^19,20 We previously validated an antibody for the detection of C9-specific poly-GA levels in patient iPSCs.⁷ We tested the top gRNA pairs in the C9^(2/282) line which produces high levels of DPRs in neurons.^6,7 All four of the bi-allelic gRNA pairs tested led to a reduction in poly-GA, with no significant difference in reduction among them (Figure 1G).

Allele-specific gRNA screen design

Our orthogonal approach to reverting pathology from C9orf72 repeat expansion is to excise the mutant allele by targeting SNPs in cis with the mutation. By targeting the SNP with a gRNA, we can preferentially cut the mutant allele, leaving the WT allele intact. This approach reverts pathological hallmarks of C9-FTD/ALS and preserves a normal amount of WT protein from the WT allele in patient iPSC-derived motor neurons.⁶ By targeting SNPs that are frequent in the patient population, we can target the greatest number of patients with our proposed therapy. Targeting SNPs in cis with the mutation also allows us to expand the search space for gRNAs, since we are no longer restricted to the 227 bp intronic region. The 5ꞌ allele-specific gRNA used to make the screening cell line previously described was used as the uppermost bound for our guide screen, which is 12.3 kb from the exon 1A start site (Figure S3A). The lower bound was the intergenic region before the next proximal genes, XM_047424279.1 and MOB3B.

We included all allele-specific spCas9 gRNAs within this 55.5 kb region with the furthest ranging gRNA pair spanning 41.2 kb (not including the repeat expansion). Criteria for gRNA selection were: (1) the gRNA pair fell within genomic coordinates Ch9:27530701–Ch9:27586178, (²) SNPs for allele-specific CRISPR-Cas9 gRNA were in phase with the repeat expansion, (³) SNPs were located within 7 bp of the NGG protospacer adjacent motif (PAM) to promote allele-specific cutting,²¹ and (⁴) SNPs were present in >20% of the C9 patient population or additionally created a PAM exclusively on the mutant allele. For SNPs that produced multiple allele-specific gRNAs, we used CRISPOR to choose the gRNA with the lowest predicted number of off-targets. SNPs were phased to the repeat expansion by Sanger sequencing and PacBio single-molecule sequencing. We also included the gRNAs (gRNA 1′ and B′) that we have previously confirmed to produce allele-specific excisions.⁶ Using these criteria, we identified 7 allele-specific gRNAs upstream of the repeat expansion which we henceforth refer to as 1′–7′ and 8 allele-specific gRNAs downstream of the repeat expansion, which we refer to as A′–H’ (Figure 2A; Table S1).

Figure 2.

Allele-specific CRISPR-spCas9 gRNA screen for the excision of the mutant C9orf72 allele in patient iPSCs

(A) Schematic of a 40.3 kb region 5′ to the C9orf72 start site at exon 1A through the end of the gene targeted for screening allele-specific gRNAs. The repeat region is depicted by the red triangle. Small red arrows depict the location of 7 SNP-targeting gRNAs 5′ to the exon 1A start site and 8 SNP-targeting gRNAs 3′ of the repeat expansion totaling 56 allele-specific gRNA pairs. ddPCR primers and probes (turquoise arrows and line) detect excisions of this locus as a loss-of-signal assay. (B) Average on-target excision efficiency for each of 56 gRNA pairs tested in patient iPSCs across biological duplicates. (C) Heatmap displaying the on-target excision efficiency of the 56 gRNA pairs with each square representing the average editing efficiency across replicates. (D) Excision efficiency of 4 top gRNA pairs, C′1′, C′7′, E′1′, and E′7′ was not significantly different across 4 unrelated patient iPSC lines with both C9-alleles intact (two-way ANOVA F(3,12) = 0.24, p = 0.9; each triangle and circle represent an average of biological triplicates). Excision efficiency is quantified for the mutant allele (triangles, left bars) and wild-type allele (circles, right bars) using a combination of loss-of-signal ddPCR and allele-discrimination ddPCR assays. There was no detectable editing on the WT allele. (E) Comparison of excision, inversion, and indel rates at the C9-locus for the four top gRNA pairs using ddPCR and ICE analysis. C′7′ had the highest excision efficiency (one-way ANOVA F(3,12) = 8.67, p = 0.003; Holm-Šídák’s multiple comparisons test p < 0.05). (F) MSD-immunoassay of poly-GA in edited pools of C9^(2/282) patient iPSCs 5 days post nucleofection (one-way ANOVA, F(4,15) = 4.96, p < 0.0095; ∗p < 0.05, ∗∗p < 0.01 compared to control treated with non-targeting gRNA (NT) by Dunnett’s multiple comparisons test). (G) Heterozygous and homozygous frequencies of SNP targeting gRNAs included in the allele-specific gRNA screen from C9-FTD/ALS patients (n = 83) and general population sample, expanded cohort of the 1,000 Genomes Project (n = 3202). Error bars = SEM.

Quantification of allele-specific excision efficiency of the mutant C9-allele

To quantify excision efficiency of large, allele-specific excisions, it is impractical to use a gain-of-signal ddPCR assay as we used in the bi-allelic screen because each gRNA pair would require a unique ddPCR assay. Instead, we designed a loss-of-signal ddPCR assay positioned centrally so that it can be used to quantify excision rates from all the gRNA pairs through a single assay. Additionally, we designed this assay to target an SNP in the probe binding region to preferentially bind to the mutant allele (Figures 2A and S3B–S3E). This allowed us to use this ddPCR assay for screening in our cell line with only the mutant allele present and quantify editing in cell lines with both alleles present.

Allele-specific gRNA screen outcome

We tested all 56 unique combinations of allele-specific gRNA pairs (Figure 2A). Excision efficiency on the mutant allele was between 0% and 35% and notably highest pairs including the 5′ gRNA 1′ or 7′ and the 3′ gRNA E′ or C’ (Figures 2B and 2C). We next validated the excision efficiency of the top allele-specific gRNA pairs in the same 4 patient iPSC lines (Figure 2D) used for validation of the bi-allelic screen, with repeat lengths ranging from 282 to 1,243. We chose gRNA pairs C′1′, C′7′, E′1′, and E′7′ for validation because these gRNAs yielded the highest rates of excisions in our screen (Figures 2B and 2C). We noted that the excision efficiency across patient cells with both alleles intact (Figures 2B and 2C) was reduced when compared to excision efficiency across the engineered screening line (20% vs. 30%; Figures 2D and S2B). This lower editing was not attributable to repeat length as cell lines with longer repeats had higher efficiency. Lower excision rate in C9^(2/781) was restricted to allele-specific editing (Figure 2D), as bi-allelic editing was comparable to other patient cell lines (Figure 1D). Subsequent allele-discrimination ddPCR assay revealed all excisions exclusively occurred on the mutant allele (Figure 2D).

Since dual-gRNA cutting could increase the risk of inversions, we measured inversion rates using a ddPCR assay with primers oriented so that amplification only occurs in the event of an inversion (Table S3).¹⁸ The rate of inversions ranged from 2% to 6% and had a similar distribution across all four patient lines (Figure 2E, gray), in line with prior reports.^17,18 Interestingly, C9^(2/781) did not have markedly lower rates of inversions despite the lower excision efficiency.

Another possible editing outcome is the formation of an insertion or deletion (indel) at either gRNA’s cut site, which may prevent the excision from occurring. We measured indel formation at each cut site using Inference of CRISPR Edits (ICE) analysis for each dual-gRNA edited sample (Figure 2E, yellow). The indel rate was lowest for gRNAs 7′ (0%–2%) and C′ (1%–5%) and highest for gRNA 1′ (7%–16%) and E’ (11%–35%). As a result, gRNA pair C′7′ had the highest ratio of excisions (desired on-target editing outcome) to inversions and indels (unwanted on-target changes).

Finally, we examined the reduction in the C9-specific pathology, poly-GA levels, in pools of patient iPSCs treated with the top allele-specific gRNA pairs. Of the four allele-specific gRNA pairs, we found a significant reduction in poly-GA for C′7′ and E′7′ (Figure 2F).

Population genetics identifies common variants to target for allele-specific editing

Because C9orf72 repeat expansion results from a single or limited number of founder events,²² we hypothesize there could be allele-specific gRNAs that could treat most or all C9-ALS/FTD-patients. We assessed the heterozygous and homozygous frequencies of the SNPs we targeted for allele-specific editing in a cohort of 83 C9orf72 repeat expansion carriers. This cohort combined sequencing data from Target ALS,²³ Answer ALS,²⁴ and sequencing from our patient iPSC lines. We removed all individuals who were duplicated in the datasets. Heterozygosity ranged between 48% and 59% for the 5ꞌ SNPs. For the 3ꞌ SNPs, heterozygosity was highest at 81% for A′, C′, D′, F′, and G′ (Figure 2G). Comparison of SNPs between the 83 C9orf72 repeat expansion carriers and 3,202 individuals from the general population²⁵ confirmed that SNPs most frequently heterozygous were distinctly so in the C9-FTD/ALS population (Figure 2G). In summary, we found five Cas9 gRNAs (A′, C′, D′, F′, and G′) containing SNPs with high heterozygous cohort prevalence (81%) and relatively low population prevalence (26%–30%).

Enhancing excision efficiency

We conducted a head to head comparison of dual-gRNAs in patient-derived iPSCs to identify the most efficient candidates for further preclinical and potential clinical testing. While the excision efficiencies observed in iPSCs may not directly reflect those we hope to ultimately achieve in patients, this screen was designed to rank gRNAs under controlled conditions. Clinical excision efficiency depends on several factors, including delivery method, Cas9 protein and gRNA expression levels in the nucleus, edited cell survival, and synchronization of DNA repair following excision. Once top-performing gRNAs are identified, excision efficiencies can be further optimized by addressing these factors. To assess key factors influencing excision efficiency with nucleofection in iPSCs, we tested whether increasing the amount of RNP or co-nucleofection of a B-cell lymphoma-extra large the name of the protein/gene and plasmid (BCL-XL), known to inhibit nucleofection-induced apoptosis,²⁶ and the DNA-PK inhibitor KU57788, which synchronizes DNA repair via inhibition of non-homologous end joining (NHEJ),^26,27 could enhance the excision efficiency for the allele-specific gRNA pair C′1′. There was a dose-responsive increase in editing efficiency from 20 to 200 pmol Cas9 RNP (Figure 3A). BCL-XL increased excision by 35%, KU57788 increased excision by 51%, and combining BCL-XL with KU57788 led to a 78% increase (Figure 3B). Neither BCL-XL nor KU57788 alone nor together affected inversion rates (Figure 3C) or indels (Figure 3D). Overall, we observed the greatest improvement in excision efficiency in iPSCs by inhibiting apoptosis with BCL-XL and synchronizing DNA repair through NHEJ inhibition, which demonstrates that lead gRNA pairs can be further optimized to enhance excision rates (Figures 3B and 3E). When we co-nucleofected with both BCL-XL and KU57788 with C′1′ gRNAs, we observed a significant reduction in poly-GA levels compared to the treated pool without BCL-XL and KU57788 (Figure 3F).

Figure 3.

Enhancing CRISPR excision efficiency in patient iPSCs with RNP dosage and pharmacologic treatment

(A) C′1′ RNP dose-response curve in C9^(2/282) patient iPSCs. Cas9 RNPs were assembled with a 2.5-fold molar excess of sgRNA. (one-way ANOVA, F(3,4) = 88.26, p < .0001; ∗∗p < 0.01, ∗∗∗p < 0.001 compared to sample treated with 20 pmol RNP amount by Dunnett’s multiple comparisons test) (B) C9^(2/282) patient iPSCs were treated with anti-apoptosis plasmid, BCL-XL or DNA-PK inhibitor, KU57788 or both peri-electroporation of the C′1′ allele-specific gRNA pair. Excision efficiency was enhanced by both treatments (one-way ANOVA F(3,4) = 25.25, p = 0.005; Holm-Šídák’s multiple comparisons test, ∗p < 0.05, ∗∗p < 0.01). (C) Inversion rate (one-way ANOVA F(3,4) = 0.76, p = 0.57) and (D) indels (one-way ANOVA F(7,8) = 1.72, p = 0.23) were not significantly altered by treatment. (E) Comparison of total editing events from (A)–(C) with or without BCL-XL and/or KU57788. Each experiment (A)–(D) was performed using biological duplicates (independent electroporations). (F) MSD-immunoassay of poly-GA in C9^(2/282) edited pools 5 days post nucleofection with or without BCL-XL and KU57788 treatment (two-tailed unpaired t test, t⁴ = 14.59, ∗∗∗p < 0.001). Experiment performed using biological triplicates. Error bars = SEM.

Using sgRNA indel rates to predict dual-gRNA excision rates

Screening 120 gRNA pairs in an arrayed format, as conducted in this study, is a labor-intensive and resource-demanding process. We wondered whether insights from our data could be used to narrow the scope of future dual-gRNA screens. First, we asked whether the size of the excision predicted the dual-gRNA efficiencies. Limiting the size of the excision could dictate the lower or upper limits of the search space when designing gRNAs. However, we found no correlation between excision size and efficiency of excision within each strategy (Figures 4A and 4B). This included small excisions of 20 nucleotides in the bi-allelic screen (Figure 4A) and large excisions over 40 kb in the allele-specific screen (Figure 4B).

Figure 4.

Comparison of excision efficiency by size and indel rates of individual gRNAs to excision efficiencies

(A and B) Excision efficiency did not correlate to length of excision across gRNA pair for small bi-allelic excisions (A) and larger allele-specific excisions (B). (C) Indel rate of the 20 bi-allelic gRNAs tested individually and measured by amplicon sequencing in a C9^(2/−) patient iPSC line. Only the C9orf72 wild-type allele is present in this engineered cell line to allow for amplicon sequencing and indel detection across the repeat region. Each dot represents an independent electroporation biological replicate. Error bars = SEM. (D) Indel rates averaged for each gRNA in a pair moderately correlated with dual-gRNA excision rates. Red dots indicate top gRNA pairs tested in Figures 1D–1G. Simple linear regression with regression lines and 95% confidence interval.

We noted that a sgRNA could enhance or impair the performance of a pair of gRNAs (Figures 1C and 2C). We therefore hypothesized that pre-screening individual gRNAs for editing efficiency could reduce the number of gRNA pairs requiring empirical validation. Specifically, we aimed to determine whether bioinformatic predictions or indel rates from sgRNAs could predict dual-gRNA excision efficiency. To facilitate this analysis, we quantified the indel rates of each gRNA used in the bi-allelic screen in a C9^(2/-) iPSC line. The C9^(2/-) iPSC line permits amplification in the first intron of C9orf72 because of a lack of a repeat expansion. The close proximity of bi-allelic gRNAs enabled precise indel quantification through a single next-generation sequencing (NGS) assay, minimizing the influence of primer efficiency or variable PCR amplification rates due to the repeat expansion on editing quantification. We excluded allele-specific gRNAs from the analysis due to their limited number within the repeat region, which precluded meaningful statistical comparison. Additionally, separate PCRs for each allele-specific gRNA, as required for the allele-specific gRNAs, would also have introduced potential bias, undermining the robustness of the results. We quantified the editing efficiency of individual gRNAs as indel rates using NGS for the 20 gRNAs included in the bi-allelic screen (Figure 4C).

We then turned to bioinformatic predictions of editing efficiency and asked whether a deep learning-based prediction algorithms, deepHF,²⁸ could predict the cutting rates of individual gRNAs. These algorithms incorporate features such as sequence, structure, and nucleotide context to predict indel rates. While deepHF is reported to have high prediction accuracy and broad generalizability across datasets,^28,29 it did not correlate with our experimentally derived sgRNA indel rates (r² = 0.01, p = 0.736; Figure S4).

We turned to our experimental data from individual gRNA indel rates to determine if they could be used to predict dual-gRNA efficiencies. We calculated the average indel rate for each pair of 5′ and 3′ bi-allelic sgRNAs and correlated these averages with experimentally determined excision efficiencies from Figure 1B (Figure 4D). Our analysis revealed a moderate but significant positive correlation between the average of indel rates between the two gRNAs and experimentally determined excision efficiency of the dual-gRNAs (Figure 4D; r² = 0.34, p < 0.0001). Importantly, sgRNAs with high average indel rates (>80%) were able to predict true-positive dual-gRNA excision rates (Figure 4D, red dots), though they also included a number of false positives.

Off-target editing rates of lead bi-allelic and allele-specific gRNA candidates

We chose the most efficient gRNA pairs—bi-allelic gRNA pair C4 and allele-specific gRNA pair C′7′ for off-target analyses. These gRNAs had high excision efficiencies in our screen, had the highest excision:indel ratio, and reduced poly-GA levels. All gRNAs were predicted to have low predicted off-target homology by CRISPOR.¹⁶ We nominated potential off-target sites bioinformatically using Cas-OFFinder³⁰ and empirically using CHANGE-seq³¹ (Tables S4 and S5). CHANGE-seq is a cell-free assay, which uses cutting of circularized naked DNA to nominate potential Cas9 cutting sites. We performed CHANGE-seq on Coriell Genome-in-a-Bottle gDNA3³² across two replicates for each of the 4 gRNAs and an unedited control. Off-target sites nominated by CHANGE-seq were included if the cleavage site had an edit distance of less than 6 from the gRNA sequence, allowing up to one gap. Off-target sites were scored using read counts normalized to the median read count per sample and site. We did not achieve high on-target editing rates in our CHANGE-seq experiment because even a normal number of C9-repeats (2 and 3 repeats on each respective allele of Genome-in-a-Bottle) are difficult to amplify (yielding poor detection of indels near the C9 repeat region from bi-allelic gRNAs C and 4) and Genome-in-a-Bottle did not contain the SNPs targeted by our allele-specific gRNAs C′ and 7′ (Table S5). Because it uses naked DNA, CHANGE-seq greatly overestimates the number of off-targets detected by cell-based assays and in vivo.³¹ Furthermore, the lack of on-target editing increases the likelihood of off-target editing as the Cas is not occupied by its target loci. We were left with thousands of low-confidence potential off-targets. Our inclusion criteria for further analysis were permissive: we included potential off-target sites which had an indel in at least 6 sequencing reads in at least one biological replicate for each gRNA. This left 181 and 2,579 nominated off-targets for bi-allelic gRNAs 4 and C, respectively, and 2,043 and 332 nominated off-targets for allele-specific gRNAs 7′ and C′, respectively. We additionally included 93 (gRNA 4), 347 (gRNA C), 2400 (gRNA 7′), 191 (gRNA C′) bioinformatically predicted off-targets, allowing for 4 mismatches and 0 bulges as predicted by Cas-OFFinder (Figure 5A).

Figure 5.

Off-target nomination and validation

(A) Number of off-target sites nominated by CHANGE-seq (yellow) and Cas-OFFinder (blue) for the top bi-allelic and allele-specific gRNAs. Circles are scaled proportional to the number of sites. (B–E) Verification of nominated on and off-target sites using whole genome sequencing (WGS). The x axis indicates chromosomal location of each nominated site. The y axis shows -log₁₀(p) values for differences in indel rates relative to unedited control WGS (Fisher’s exact test, Benjamini-Hochberg correction p < 0.05). (F) INDUCE-seq recurrency plot for gRNA 7′ showing the number of sites (y axis) with a given number of double-stranded breaks (x axis). The threshold for CRISPR editing induced breaks was determined by comparing break scores in control (yellow) and edited (blue) samples. Two sites significantly differ in the edited sample compared to the control: the on-target site and an off-target site at chr13:45277321-45277328 (Wald’s test, p < 0.05; Fisher’s exact test, p < 0.001; Benjamini-Hochberg correction). (G) Mismatch plot for allele-specific gRNA 7′ showing the sequence and number of breaks at the off-target site, the on-target mutant allele, and the wild-type allele.

It would be impractical to perform amplicon sequencing on the 5,135 total CHANGE-seq nominated off-target sites. We therefore turned to WGS to a depth of 300× to verify off-target sites nominated by Cas-OFFinder and CHANGE-seq (Figure 5A; Table S6). We created a custom reference genome for the pre-edited, electroporated C9^(2/282) iPSC pool to avoid natural variation in the patient cells appearing as edits against a standard reference genome such as hg38. We compared indel rates at each of the nominated off-target sites to this custom reference (Figures 5B–5E) using custom software we developed for this purpose.³³ We performed WGS 4 days after electroporation to allow cells to recover and double-stranded breaks to repair. Among all the candidate off-target sites, we found only one statistically significant off-target site across all four gRNAs (Figure 5E). This site is located within an intronic region of the general transcription factor IIF subunit 2 (GTF2F2) gene and was predicted by both Cas-OFFinder and CHANGE-seq. GTF2F2 is an essential component of transcription factor IIF, which facilitates RNA polymerase II-mediated transcription and elongation.^34,35 We found all other nominated off-target sites to not be statistically significant.

We also used a second unbiased, whole-genome method, INDUCE-seq,³⁶ to empirically test for potential off-target sites. INDUCE-seq is a cell-based assay that captures sites of double-stranded break in situ. Because it uses PCR-free NGS to identify and quantify double-stranded breaks, it is suited to detecting off-targets at regions that are hard to amplify by PCR. We performed INDUCE-seq 2 h after nucleofection of each gRNA, across 3 independent biological replicates. We compared INDUCE-seq identified break sites to a C9^(2/282) electroporated control to differentiate edited sites from breaks due to cell division or electroporation (Table S7). We compared the total break number at each site in treated samples and the electroporation-only control. To identify break sites significantly enriched in the edited lines, we performed differential analysis of break counts across replicates, considering only sites with at least 3 reported breaks in each replicate. The data were modeled using a negative binomial distribution to account for overdispersion, implemented with the DESeq2 package.³⁷ Statistical significance of differences in break site representation was assessed using the Wald test, and p values were adjusted for multiple testing using the Benjamini-Hochberg procedure to control the false discovery rate. Concurring with our WGS data, we found only 1 statistically significant off-target site for gRNA 7′, and no off-targets for the other 3 gRNAs (Figures 5F, 5G, and S5A–S5C). The 7′ off-target site is chr13:45279425-45279448, within the same intronic region of GTF2F2 predicted by Cas-OFFinder and CHANGE-seq and confirmed by WGS. INDUCE-seq also validated the specificity of our allele-specific gRNAs as, for example, we found only 1 break on the WT allele while there were 108 breaks on the targeted mutant allele (Figure 5G).

Discussion

In this study we ranked 64 bi-allelic gRNA pairs and 56 allele-specific spCas9 gRNA pairs by their ability to efficiently excise the mutant C9orf72 repeat expansion in patient iPSCs and validated them across patient iPSCs with varying repeat expansion lengths. To facilitate this screen, we developed several tools, including loss- and gain-of-signal ddPCR assays to measure editing efficiencies at the C9orf72 locus and employed engineered iPSC lines to overcome sequencing limitations in this region. The engineered screening cell line also allowed us to screen for excision efficiency exclusively on the mutant allele, which we found to be important as the bi-allelic dual-gRNA had different rates of editing between the mutant and WT alleles (Figures 1E and 1F). The novel ddPCR assays accurately quantified excision efficiency (Figures S1B, S1C, and S3B–S3E), avoiding overestimation common in alternative methods using PCR amplification and agarose gel electrophoresis.¹²

Summative effect of individual gRNAs on efficiency of dual-gRNA pairs

We found that individual gRNAs had both summative and subtractive effects on efficiency. gRNAs with the highest excision rates increased the efficiency of their paired gRNAs, while those with the lowest rates reduced the efficiency (Figures 1C and 2C). This suggests that the efficiency of a gRNA pair is determined by the more efficient gRNA and limited by the less efficient one. In addition, the bi-allelic screen showed that gRNAs in close proximity, even overlapping, could yield significant differences in editing efficiency. This suggests that excision efficiency might depend more on sequence than location. The fact that the average of the ability of individual gRNAs to form indels predicted the efficiency of dual-gRNA excision (Figure 4D) further supports the impact of the gRNA itself on editing efficiencies. We found empirically determined indel rates were a better predictive tool than excision lengths (which did not correlate with editing efficiencies) (Figures 4A and 4B), repeat expansion length (which did not impact editing efficiencies) (Figures 1D and 2D), or bioinformatic gRNA prediction tools (Figure S4).

Optimal gRNA screening methods

The positive correlation between empirically tested excision efficiency using dual-gRNA and the average of two gRNA’s empirically tested indel rates suggests that screening individual gRNA cutting efficiency (indel rate) may be the most time- and cost-effective screening approach (Figure 4D). However, while a sgRNA screening approach could be used to identify candidate gRNAs with the potential to achieve high dual-gRNA excision rates, this method would likely generate a considerable number of false positives, leading to the inclusion of many gRNA pairs with suboptimal excision efficiency, as we found (Figure 4D). We recommend a stepwise approach for future gRNA screens: screening gRNAs individually to select high indel forming gRNAs (i.e., “good cutters”) followed by empiric testing of excision efficiencies of pairs of gRNAs nominated by sgRNA screening. Surprisingly, neither size of the excision (Figures 4A and 4B) nor size of the repeat expansion (Figures 1D and 2D) impacted bi-allelic or allele-specific excision efficacy. In addition, computational predications did not predict dual-gRNA excision rates at the C9-locus (Figure S4). Developing a computational tool with greater accuracy and precision in predicting excision efficiency would be highly beneficial.

Optimal gRNA pairs

Both the average and maximum editing efficiencies were higher for bi-allelic gRNAs compared to allele-specific gRNAs (Figure 1C vs. 2C). Additionally, both bi-allelic and allele-specific gRNA pairs showed higher excision rates in our screening iPSC line in which the non-diseased allele had been removed, compared to unedited patient iPSC lines with both C9-alleles intact (Figure 1C vs. 1D and 2C vs. 2D; Figure S2). This finding likely suggests a dose-reduction at the on-target locus by binding of the Cas-gRNA to the WT allele in cells with both alleles present. An alternative explanation is that the WT allele alters editing on the mutant allele. As our goal was to rank gRNAs under controlled conditions, not optimize dosing, this finding does not alter our gRNA ranking.

It remains unclear how screening in iPSCs will translate to other models, such as mouse models of C9-FTD/ALS or human in vivo settings. Current animal models for C9-FTD/ALS are imperfect, as they only have a single partial or full copy of the human transgene with variable surrounding human C9orf72 regulatory DNA.^38,39 Given that our data suggests editing efficiencies may be impacted by the presence of two alleles, studies in mice should be interpreted cautiously. Editing efficiency could also vary across cell types due to differences in chromatin state, divergent DNA repair propensities, and stages of cell cycle.⁴⁰ Delivery reagents that impact the temporal expression of Cas9 will also impact both on- and off-target editing rates. Although overall editing efficiency may change, the comparative efficiency of gRNAs is likely to remain consistent, suggesting that top-performing gRNAs in iPSCs will likely remain the top-performing gRNAs in other settings. Finally, dosing of editing reagents will need to be assessed in the final therapeutic formulation, which includes the delivery reagent for in vivo editing. Here, we aimed to rank the possible gRNAs by their efficiency.

We found that bi-allelic gRNA pair C4 had the greatest excision efficiency (Figures 1C and 1D) with no confirmed off-targets (Figures 5B, 5C, S5A, and S5B) and significantly reduced poly-GA pathology (Figure 1G). However, infrequent sequence variation has been reported⁴¹ in the region 3′ of the repeat expansion that could affect gRNA C binding. This variation, described as a deletion, is characterized by the loss of T nucleotides at 27,573,519 and 27,573,522 (hg38) and has been observed in published long-read sequencing data from a patient-derived cell line⁴² and brain tissue from a second patient.⁴³ We detected this variation in 3 of 85 C9 expansion carriers in short-read sequencing datasets from Target ALS²³ and Answer ALS,²⁴ but did not observe it in 20 C9 carriers in a long-read sequencing cohort from Target ALS.²³ This sequence variation has also been suggested by PCR-based assays.⁴⁴ The subset of patients carrying this variant would not be eligible for gRNA C and would thus require an alternative gRNA or editing strategy. We ruled out further preclinical advancement of allele-specific gRNA pair C′7′ because gRNA 7′ has a detectable off-target (Figures 5D–5G; S5C).

Enhancing excision efficiency

Screening gRNA efficacy head to head in patient iPSCs, as we have done here, is crucial for identifying top-performing gRNAs. The efficiencies achieved in this study are best used to rank gRNAs but do not represent the ultimate editing efficiencies we hope to achieve in patients. Optimization of factors such as delivery method, local Cas9 protein and gRNA concentrations in target cells, survival of edited cells, and synchronization of repair for excision is needed. To demonstrate this point, we showed that excisions could be enhanced by increasing the dose of RNP, inhibiting apoptosis, and synchronizing DNA repair through inhibition of NHEJ (Figures 3A, 3B, and 3E). While we show that editing efficiencies increased with pharmacologic enhancements, there was no significant effect on inversions or indels (Figures 3C–3E). Enhancing excisions also led to a larger decrease in poly-GA (Figure 3F). These strategies and others can be applied in future studies to increase excision rates, but should be tailored to the editing context. For instance, we used BCL-XL to counteract nucleofection-induced apoptosis; therefore, this inhibitor would likely not be appropriate for in vivo editing. The lack of efficient and non-toxic delivery reagents for post-mitotic cells currently limits editing across cell types. Future testing in various cell types, both diseased and non-diseased, will be important as better delivery agents become available.

Inversions

Others have noted inversions when performing dual-gRNA excisions,^17,18 as we have detected here (Figures 1E and 2E). It is not clear what the impact of an inversion of a repeat expansion would be. Only the DNA between the cut sites and not the surrounding intronic DNA were inverted (Figure 1E). Because these inversions were restricted to the intron, we do not expect them to disrupt the function of the surrounding 1A and 1B exons. The top allele-specific gRNA pairs excise exons 1A-5 (C′1′ and C′7′) or 1A-8 (E′1′ and E′7′) and regulatory elements (Figure 2A). The effect of a large inversion is not known but would lead to inactivation of that allele without removal of the repeat expansion (Figure 2E). However, an inversion could alter expression of pathology from the repeat expansion itself, such as DPRs. Reassuringly, we did not detect an increase in poly-GA (Figures 1G and 2F). Whether pathology of an inverted repeat expansion is altered is an interesting question out of the scope and tools deployed in this study, but one that should be further investigated.

Off-target detection

When designing bespoke genome therapies targeted to patient mutations or haplotypes, we have found it is important to test both and on- and off-target editing rates in patient cells. Off-targets may be overestimated in genomes or cells without the on-target site (such as Genome-in-a-Bottle and other standard cell lines) because (1) the Cas is not occupied at the on-target site and can therefore scan the genome potentially leading to increased off-targets, and (2) it is difficult to exclude noise without an on-target benchmark rate. To verify which of the over 5,135 low-confidence off-targets nominated by CHANGE-seq and 3,031 nominated by Cas-OFFinder were true off-targets, we developed a new high-throughput method using 300× WGS to detect off-targets in edited iPSC compared to electroporated but unedited control iPSCs. We had to turn to WGS because other validated cell-based methods did not work in iPSCs (e.g., GUIDE-seq⁴⁵ caused iPSC death, which prevented data collection) or were impractical across thousands of low-confidence off-targets (rhAmpSeq⁴⁶). Previous studies have employed WGS to assess off-target effects in genome-edited cells.^47,48,49 However, most approaches compare edited cells to a reference genome, requiring extensive filtering to account for natural genetic variation, mosaicism, or mutations acquired during cell culture. This complicates the identification of true editing-induced variants and reduces sensitivity. To address this, we performed deeper WGS (300×) on pooled edited and unedited cells of the same genetic background for verification of off-targets nominated by cell-free methods. Our sequencing depth is comparable to that used in clinical cancer panels for solid tumors, enabling reliable detection of low-frequency variants.^50,51 By directly comparing isogenic edited and control samples, our approach improves specificity by eliminating the need to account for germline variation or clonal mutations unrelated to editing, enhancing the detection of clinically relevant on- and off-target edits.

We used a second cell-based method, INDUCE-seq, to confirm our findings. Across all these nomination and validation measures, we found only a single confirmed off-target for gRNA 7′ in an intron of GTF2F2 (Figures 5E and 5F). INDUCE-seq uniquely detected the true off-target site without requiring prior nomination and without identifying false positives, making it a more effective nomination method than both Cas-OFFinder and CHANGE-seq. Given the essential role of GTF2F2 in transcriptional regulation, we approach any potential perturbation, even in an intronic region, with caution. We therefore will exclude gRNA 7′ as a candidate for our therapy. It is important to note that 7 Cas-OFFinder predicted off-targets for gRNA 4,¹ gRNA 7,⁴ and gRNA C′² located on the Y chromosome (Table S4) could not be evaluated by our study as both the Coriell Genome-in-a-Bottle and the patient cell line used for WGS and INDUCE-seq originate from female donors. In addition, our patient cell line had intronic deletions in the cell line, prior to gene editing, that precluded the analysis of 5 additional nominated off-targets (Table S6). Inclusion of additional cell lines is required to investigate these 9, low confidence nominated sites in cells.

In conclusion, this study provides a comprehensive analysis of spCas9-gRNA pairs with potential to excise the C9orf72 repeat expansion, identifying top candidates for both bi-allelic and allele-specific editing. The screening methods and tools we developed offer a robust platform for evaluating CRISPR-Cas9 editing efficiencies, which can be applied to other genomic loci and model systems. Future work should focus on optimizing delivery methods and testing lead gRNA candidates in vivo to ensure the therapeutic potential of these gRNAs is fully evaluated. The promising gRNA candidates identified here pave the way for further development toward potential gene therapy treatments for C9orf72-FTD/ALS.

Materials and methods

Cell line maintenance

We used four de-identified patient iPSC lines harboring C9orf72 mutation (C9^(2/282), C9^(2/781), C9^(2/1063), and C9^(2/1243)), four modified patient lines (C9^(2/–), C9^(–/282), C9^(−/−), and C9^(2/0)), one WT line (WT^(2/10)), and two modified WT lines (WT^(0/0) and WT^(−/−)) generated as described in our previous study^6,52 (Table S2). We maintained iPSCs in mTeSR Plus (STEMCELL Technologies 1000276) plated on Matrigel (Corning Life Sciences 356231), passaging at 60%–80% confluency.⁵³ All cell lines had a normal karyotype and negative quarterly mycoplasma testing.

iPSC arrayed CRISPR editing screen

We designed gRNAs (Table S1) using CRISPOR (Homo sapiens - USCS December 2013 [GRCh38/hg38]).¹⁶ sgRNAs (100 μM) were purchased from Synthego. We delivered Cas9-gRNA RNP (HiFi spCas9 proteins from MacroLab, UC Berkeley, 40 μM) into iPS cells by nucleofection (Lonza AAF-1003×, pulse code = DS138) using our published protocol.⁵⁴ We nucleofected 150 k cells in 20 μL P3 buffer for each reaction. The nucleofected cells were plated in one well of a 24-well plate and were recovered in mTeSR Plus with 10 μM ROCK1 inhibitor (Selleckchem S1049) and CloneR (STEMCELL Technologies 05888). We collected the pool of edited cells after 48 h and extracted DNA using QuickExtract DNA Extraction Solution (LGC Biosearch 76081–768) or Quick-DNA Microprep Plus Kit (ZymoGenetics D4074).

Editing efficiency of dual-gRNA quantified by ddPCR

For each ddPCR reaction, we added 11 μL 2× Supermix for Probes (No dUTP) (Bio-Rad 1863024), primers/probe (Table S3) (IDT), and nuclease-free water up to 22 μL. About 20 μL of the reaction mixture and 70 μL of oil was used to generate droplets using Q× 100 Droplet Generator (Bio-Rad 1863001). The ddPCR reactions were run in a Deep Well C1000 Thermal Cycler (Bio-Rad 1851197) with the following protocol¹: 95°C for 10 min²; 94°C for 30 s³; 62.5°C (bi-allelic), or 52°C (allele-specific) for 1 min⁴; steps 2 and 3 repeated for 39 times⁵; 98°C for 10 min⁶; hold at 4°C. We quantified probe binding, copy number and ratio of probe binding to control region (RPP30, Bio-Rad Assay ID dHsaCP2500350) with a known copy number of 2 (Bio-Rad QuantaSoft Analysis Pro Software). We ran ddPCR in technical triplicates for each of 3 biological replicates (independent nucleofected pools) with our protocol.⁵⁵ Percent excision was calculated from the target-to-reference ratio. When the screening cell line (only one intact allele) was used (Figures 1B, 1C, 2B, 2C, 4A, and 4D; S2), or when we were quantifying allele-specific excisions (Figures 2D, 2E, 3A, 3B, 3E, and 4B; S2), we normalized to the diploid copy number of the reference. For the allele-specific loss-of-signal assay (Figures 2D, 2E, 3A, 3B, 3E and 4B; S2), percent intact alleles were subtracted from 100 to get percent excision values.

ICE analysis

We PCR-amplified the sgRNA target regions for the top allele-specific guides using the primers listed in Table S3. Nested PCRs were used to enhance on-target amplification and samples were Sanger sequenced. We used Synthego ICE analysis⁵⁶ to quantify indel rates of dual-gRNA edited samples compared to unedited samples.

Editing efficiency of sgRNA quantified by NGS

We PCR-amplified the sgRNA target regions using the primers listed in Table S3. Nested PCRs were used to enhance on-target amplification in regions that are difficult to sequence or went off-targets were present. Samples were deep sequenced on an Illumina MiSeq at 300 bp paired end reads to a depth of at least 9,000 reads per sample. Cortado⁵⁷ was used to analyze editing outcomes. Briefly, reads were adapter trimmed and then merged to single reads. These joined reads were then aligned to the target reference sequence to identify editing events at the cut site. To overcome artifacts when trying to align reads across a repetitive region, percent identity was set to 85% between reads and the reference sequence, and due to this the averaged percent aligned reads across samples was ∼60% high quality alignments. NHEJ rates are calculated by counting any reads with an insertion or deletion overlapping the cut site or occurring within a 3 base pair window on either side of the cut site. SNPs occurring within the window around the cut site are not counted toward NHEJ. Total NHEJ reads are then divided by the total number of aligned reads to get percentage NHEJ.

Population variant frequencies

Whole genome sequencing data from 85 C9orf72 repeat expansion carriers, with or without clinical diagnosis of FTD or ALS, were sourced from TargetALS²³ and AnswerALS.²⁴ Duplicate data originating from iPSCs were identified by applying Hail v.0.2.122’s⁵⁸ implementation of identity by descent on chromosome 9, where samples with a PI_HAT approximately greater than or equal to 0.99 were deemed identical. We removed sequencing originating from the same individual, leaving a total of 83 samples. The prevalence of 16 variants, used for 15 allele-specific gRNAs, was quantified for the 83 C9orf72 repeat expansion carriers. In addition, we analyzed WGS of 3,202 individuals from the 1,000 Genomes dataset.²⁵ We determined heterozygous and homozygous variant frequency in these datasets using the following open-source software: Python v.3.11.4,⁵⁹ pandas v.2.1.0,⁶⁰ pysam v.0.21.0,⁶¹ and SAMtools/BCFtools/HTSlib v.1.18.^62,63 The analysis is compiled in a Jupyter notebook titled “Nucleotide_Frequency.”³³ Data are presented in Figure 2G.

BCL-XL and KU57788

In conditions using BCL-XL, 150,000 iPSCs C9^(2/282) were nucleofected with 75 ng of BCL-XL plasmid²⁶ in P3 buffer, as previously described.⁵⁴ In conditions using KU57788, 10 mM KU57788 (a gift from the Zhang Lab) was added to the media immediately after nucleofection.^26,27 Media was changed 24 h post nucleofection, and samples were collected at 48 h and 7 days post nucleofection.

Poly-GA quantification by MSD

As indicated in Figures 1G and 2G, 150,000 iPSCs C9^(2/282) were nucleofected (Lonza AAF-1003×, pulse code = DS138) with dual-gRNAs or non-targeting sgRNA (100 μM) and spCas9 (40 μM). Protein was collected 7 days post nucleofection. Electroporation with spCas9 and a non-targeting gRNA (Table S1) was used as a negative control. Meso Scale Discovery immunoassay (L45SA) was performed with poly-GA (MABN889 Millipore) as both capture and detection antibody as previously published,⁷ with the following modifications: We coated the plate with capture antibody at 1 μg/mL overnight at 4C. The plate was blocked with 3% MSD Blocker A (R93BA, MSD) in 1× DPBS + 0.2% Tween for 1 h at 600 rpm. A total of 18 μg protein was added to each well for 2 h at 4C at 600 rpm. Detection antibody at 2 μg/mL was added for 1 h. Three washes with 1× DPBS + 0.2% Tween were performed between steps. MSD Read Buffer B (R60AM-2, MSD) was added, and the plate was immediately read using the MSD Model 1300 Sector Imager 2400 plate reader. Signal was calculated by comparing luminescence intensity for each sample. Data were presented as a fold change compared to control samples edited with the non-targeting guide.

PacBio analysis of on-target editing

A total of 150,000 iPSCs C9^(2/282), C9^(2/781), C9^(2/1063), and C9^(2/1243) were nucleofected (Lonza AAF-1003×, pulse code = DS138) with dual-gRNAs (100 μM) and spCas9 (40 μM). Two reactions were pooled into one well of 12-well plate. DNA was extracted 5 days post nucleofection using the Nanobind PanDNA Kit (Pacific Biosciences 103-260-000).⁶⁴ Single molecule sequencing was used to target the C9orf72 repeat expansion and surrounding region.⁶⁵ A total of 24 samples were multiplexed and sequenced on the Revio (Pacific Biosciences 102-090-600) for 24 h.⁶⁶ We processed the data using the nf-core/pacvar pipeline,⁶⁷ which aligns the sequences to the GRCh38 reference genome, followed by sorting and indexing. Phasing of the aligned reads was performed using a custom script³³ titled “PACBIO_analysis” that incorporated the pysam v.0.22.1⁶¹ package for sequence analysis as well as pandas v.2.2.2.⁶⁰ We phased C9^(2/781) based of an SNP of C-> A located on the WT allele at position chr9:27,572,257. We phased C9^(2/282) and C9^(2/1243) based on an SNP of C-> T on the mutant allele at position chr9:27,574,517 or the presence of the repeat expansion. And finally, we phased C9^(2/1063) based on an SNP of C-> T present on the WT allele at position chr9:27,574,805. We manually identified reads with inversions by visualizing the alignments in IGV (Integrative Genomics Viewer). We then used a custom script³³ to detect the presence of excision events in the remaining reads. An excision I was defined as a deletion that spanned the entire WT C9orf72 repeat region (chr9:27,573,529–27,573,547). If an excision overlapped with an exon or extended beyond the hexanucleotide repeat expansion (HRE) region into either exon 1A or exon 1B, then it was classified as an excision II (HRE + exon). We also quantified inversions. Data are presented in Figures 1E and 1F.

Change-seq

We used the published CHANGE-seq protocol⁶⁸ with modifications. First, we tagmented NA12878 human gDNA purchased from Coriell to an average of 400 bp using Tn5-transposase (MacroLab, UC Berkeley), and then purified gDNA with Solid Phase Reversible Immobilization (SPRI)-Guanidine beads. SPRI-Guanidine beads are made using Sera-Mag Magnetic Beads (Fisher Scientific 65152105050250). Beads (1 μL) were washed with Tris-EDTA (TE) buffer and added to 9g PEG8000 (Sigma-Aldrich P2139) dissolved in 10 mL of 10 M NaCl and brought to 50 mL with SPRI-guanidine binding buffer (4 M guanidine thiocyanate, 40 mM TRIS, 17.6 mM EDTA, pH 8.0, and TRIS 1 M pH 8). T4 Ligase (NEB M0202L) and HiFi HotStart Uracil+ Ready Mix (Kapa Biosystems KK2802) were used to repair gaps generated in the DNA, followed by additional purification using SPRI-Guanidine beads, and treatment with USER enzyme (NEB M5505L) and T4 Polynucleotide Kinase (NEB M0201L). We then performed intramolecular circularization of DNA with T4 DNA ligase and degraded unwanted linearized DNA with an exonuclease cocktail comprised of Lambda Exonuclease (NEB M0262L), exonuclease I (E. coli) (NEB M0293L) and Plasmid-Safe ATP-dependent DNase (Epicentre E3110K). Quantification with the Qubit dsDNA HS assay yielded circular DNA of 20–30 ng/μL. Next, we performed Cas9 cleavage reactions with 90 nM spCas9 and 270 nM sgRNA added to 125 ng of circular DNA in 1× NEB3.1 buffer (NEB B6003). Cleaved products were A-tailed using NEBNext dA-tailing Module (NEB E6053), adapter ligated, USER enzyme (NEB M5505L) treated, and then amplified using universal primers NEBNext Multiplex Oligos for Illumina (NEB E7600S). The products were purified with SPRI beads and eluted in TE buffer (TE pH 8.0). The libraries were quantified by qPCR using NEBNext Library Quant Kit for Illumina (NEB E7630L) and sequenced with (2 × 150) bp paired-end reads on the Illumina NextSeq 1000 sequencer.

Sequence analysis was preformed using a modified version of CHANGE-seq analysis software.⁶⁹ Briefly, raw FASTQ files were trimmed to remove transposon and Illumina adapter sequences using Cutadapt v.3.4.^70,71 Trimmed reads were then aligned to the human reference genome GRhg38.p14 using BWA v.0.7.17.⁷² Aligned reads were classified as edited in both treated and untreated control samples according to the following criteria (1) paired-end reads had outward-oriented alignments, (2) the distance between break points was under 3 bp, (3) alignments were of stringent quality threshold mapping quality MAPQ > 40, and (4) the overlapping reads, including ±30 bp of flanking region, contained a sequence matching the gRNA and NGG PAM below a Levenshtein distance of 7. Lastly, to account for variability between replicates, raw read counts of identified sites were normalized using median normalization. Final sites were reported if at least one replicate contained a site with at least 6 reads.

Whole genome sequencing for off-targets

A total of 150,000 iPSCs C9^(2/282) were nucleofected with dual-gRNAs (100 μM) and spCas9 (40 μM) or nucleofected without Cas9 or gRNA. Genomic DNA was collected 48 h post-nucleofection and was submitted to Novogene (Sacramento, CA, USA) for WGS. Each sample consisted of ≥200 ng of high-quality gDNA, with an OD260/280 ratio between 1.8 and 2.0, free from degradation and contaminants. DNA quantity was determined using the Qubit fluorometric method, and integrity was confirmed prior to library preparation. WGS libraries were constructed for each sample using NEBNext Ultra II DNA Library Prep Kit for Illumina (NEB E7645L). For samples requiring deep (∼300×) sequencing, libraries were generated across three independent rounds of preparation and sequencing, each targeting 300 Gb of raw data per sample per round. Libraries were sequenced on the Illumina NovaSeq X Plus platform using 150 bp paired-end (PE150) reads. The sequencing run was configured to yield a minimum of 900 Gb of raw data per sample, achieved by merging data from the three sequencing rounds. A Q30 score of ≥85% was required to ensure high base quality across reads. To account for the inherent genetic variation in the patient line, we generated a consensus genome representing the control sample. The control sample FASTQ file was aligned to the human reference genome (hg38) using bwa-mem2 v.2.2.1^73,74 and the resulting alignment was sorted and indexed using SAM tools v.1.21^62,75 to produce a binary alignment map (BAM) file. Variants were then called from the BAM file using bcftools v1.21^62,75 and incorporated into the hg38 reference genome using bcftools v1.21^62,63 consensus to generate the consensus genome.⁶² A chain file was also produced with bcftools v1.21^62,63 to enable coordinate conversions between the generated consensus genome and the hg38 reference genome. FASTQ files from all the samples were then aligned to this generated consensus genome using bwa-mem2 v.2.2.1,^73,74 and the resulting BAM files were sorted and indexed with SAMtools v.1.21.^62,75 We then extracted subsets of these BAM files corresponding to the off-target coordinates nominated by Cas-OFFinder and CHANGE-seq, doing so for each replicate of the four edited samples and the control. To quantify editing at each nominated site, we developed a custom Python script titled “WGS_Analysis” that measured the rates of insertions and deletions that overlapped or were contained within the nominated window, capturing both the length and precise coordinates of these modifications in the sequencing reads.³³ The software libraries used for the analysis included pysam v.0.22.1⁶¹ for sequencing read parsing, and numpy v.1.26.4,^76,77 pandas v.2.2.2,⁶⁰ and scipy v.1.14.0^78,79 for data analysis. We then compared each edited sample to the control sample to identify regions where the editing rate was significantly higher than in the control. Statistical significance was evaluated using Fisher’s exact test for each replicate, and a final p value was calculated using Fisher’s method to combine replicates. Benjamini-Hochberg correction was applied to control the false discovery rate.

INDUCE-seq

INDUCE-seq libraries to measure double-strand breaks were prepared essentially as published.³⁶ A total of 400,000 iPSCs C9^(2/282) were nucleofected (Lonza AAF-1003×, pulse code = DS138) with sgRNAs (100 μM) targeting the C9orf72 locus and spCas9 (40 μM) or electroporated with no gRNA or spCas9. Two reactions were pooled into one well of a 12-well plate and incubated at 37°C for 2 h. There were three biological replicates (i.e., 3 wells) per condition. The cells were lifted with Accutase (STEMCELL Technologies 07922) and 200,000 cells in 100 μL of media were dispensed per well of a 96-well plate. The 96-well plate was previously coated with 100 μL CellAdhere laminin-521 (STEMCELL Technologies 77003) (diluted to 10 μg/mL in cold DPBS [Ca²⁺, Mg++]) overnight at 4°C. The 96-well plate was centrifuged at 300 g for 2 min and incubated at room temperature for 20 min. A total of 100 μL of 8% paraformaldehyde (PFA) was slowly added to the side of each well and the plate was incubated at room temperature for an additional 20 min. Supernatant was removed and wells were washed with 100 μL of PBS twice. Following fixation, cells were permeabilized with Buffer 1 (10 mM Tris-HCl, 10 mM NaCl, 1 mM EDTA, 0.2% Triton X-100; 1 h, 4°C) and Buffer 2 (10 mM Tris-HCl, 150 mM NaCl, 1 mM EDTA, 0.3% SDS; 1 h 37°C) with three PBS washes in between. Cells were washed with CutsmartTM buffer (NEB B6004S) then DNA breaks end repaired (Quick Blunting Kit [NEB E1201L plus 100 μg/mL bovine serum albumin (BSA); 1 h, room temperature]), washed with 3× CutsmartTM buffer washes and A-tailed (NEBNext(R) dA-Tailing Module (NEB E6053L); 30 min, 37°C). A custom P5 adapter³⁶ was ligated with T4 DNA ligase (NEB M0202M; 16–20 h 16°C) and excess adapter removed with ten washes of wash buffer (10 mM Tris-HCl, 2 M NaCl, 2 mM EDTA, 0.5% Triton X-100). Cells were lysed and cross-linking reversed in DNA extraction buffer (1 mg/mL Proteinase K, 10 mM Tris-HCl, 100 mM NaCl, 50 mM EDTA, 1% SDS; 10 min, 37°C, 1 h 65°C). Genomic DNA was extracted (ZR-96 Genomic Clean & Concentrator-5 Zymo) to give 500–700 ng DNA (Qubit dsDNA HS, Invitrogen). Genomic DNA was fragmented to 300–500 bp (Covaris ME220 and size selected using CleanNGS beads [GC Biotech CNGS-0050, confirmed with High Sensitivity D1000 Tapestation, Agilent]). Fragmented DNA was end-repaired (NEBNext UltraII End Repair/dA-Tailing Module, NEB E7546L; 30 min 20°C, 30 min 65°C) before a custom P7 adapter³⁶ was ligated (NEBNext Ultra II Ligation Module, NEB E7595L). Libraries were purified with CleanNGS beads (GC Biotech CNGS-0050) and 50 μL of all libraries pooled and sequenced across two high output 75 cycles NextSeq 550 (Illumina) sequencing runs.

Sequencing reads were processed essentially as described with modifications to enable alignment to a custom variant-aware reference genome.³⁶ The custom reference genome was generated by applying all known variants to the hg38 reference genome using “bcftools consensus v.1.17.”^62,63 The “-c” option was specified to generate a chain file for conversion of coordinates from the standard hg38 reference genome to the custom variant-aware reference genome. This chain file was used to modify the ENCODE blacklist and any other relevant accessory files.⁸⁰ Optical duplicate removal was performed using Clumpify,⁸¹ and adapter sequences were removed using Trim Galore v.0.6.10.⁸² Reads were aligned to the custom variant-aware reference genome using BWA-mem v.0.7.18,^72,83 followed by read filtering with SAMtools v.1.18.^62,75 Region-level filtering to exclude regions from the ENCODE blacklist,⁸⁰ was performed with BEDtools v.2.31.1.⁸⁴ AWK v.5.2.18⁸⁵ was used to assign double-stranded break positions as the first 5′ nucleotide upstream of the read relative to strand orientation using the following command: awk “BEGIN {OFS = “\t”} {if ($6 = = “+”) $3 = $2 + 1; else if ($6 = = “-”) $2 = $3–1; print $0}.” Quantification of single-nucleotide resolution breaks was performed using BEDtools.⁸⁴ Specifically, BEDtools merge⁸⁴ with a distance of 0 bp was used to define break sites in a bed format. Break recurrency describes the propensity of a genomic region to break. Recurrency curves were then generated by grouping break sites into recurrency classes based on break scores (x axis), with the y axis representing the frequency of a given recurrency class. Statistical significance was assessed using two methods, using Wald’s test after modeling the counts according to a negative binomial distribution using DESeq2³⁷ and Fisher’s exact test, comparing the signal at each genomic site—excluding those with only a single detected break—between treated samples and the electroplated control. p values were adjusted for multiple comparisons using the Benjamini-Hochberg correction. Mismatch plots were generated by intersecting breaks with Cas-OFFinder predicted cut-sites using a modified version of GUIDE-seq Visualization.⁸⁶

Data and code availability

• •Data used in Figure 2G were obtained from the ANSWER ALS Data Portal (AALS-01184) and the Target ALS postmortem tissue core.
• •Sequencing data have been deposited in NIH Bioproject: PRJNA1283233.
• •All custom code is archived at and directly available on github.³³
• •All other data generated and presented in this study are included in the article or supplemental files.

Acknowledgments

C.D.C is supported by Target ALS NI-2023-NAI-S1, NIH/NINDS K08-NS112330KO8, U19NS132303, U01NS134062, CIRM DISC2-16738, Alzheimer's Association NIAP24-1272942, Carol and Gene Ludwig Foundation, Larry H. Hillblom Fellowship, Rainwater Charitable Trust, Packard Foundation, Weill Neurohub, and the UCSF Memory & Aging Center. P.G. is supported by NIH/NIAID U01AI176469. Graphical abstract was made, in part, with BioRender content per appropriate licensing agreements (https://BioRender.com/ednkut2).

Author contributions

C.D.C. conceived the study and designed experiments with O.A.A., K.J.K.L., and K.A.B.; O.A.A. performed the allele-specific screen and validation of dual gRNAs and the screen for enhancing excision efficiency; K.J.K.L. performed the bi-allelic dual gRNA screen and validation of dual gRNAs; K.A.B. performed poly-GA experiments, PacBio sequencing, and prepared samples for off-target analysis; T.J. developed code for processed sequencing data and performed statistical analysis on PacBio sequencing, whole genome sequencing, and INDUCE-seq data; P.O.I.-L. developed code for and analyzed population data for SNP frequencies; C.H. performed the RNP dose-response experiment; T.R.-H. and K.B. performed CHANGE-seq assay and analysis, supervised by P.G.; C.A.G. performed PCR cleanup for NGS; N.K. performed NGS; S.K.W. analyzed NGS; R.H.C.W. performed INDUCE-seq and K.O. and P.v.E. analyzed INDUCE-seq data, supervised by S.H.R.; O.A.A., K.J.K.L., K.A.B., T.J., and C.D.C. wrote the paper.

Declaration of interests

C.D.C. is a founder with equity in Ciznor Co., a gene therapy company. P.G. is a co-founder and director of Need Inc., and a consultant for IMDPath and ResearchDx. S.H.R. and P.v.E. are co-founders of Broken String Biosciences and employed by Broken String Biosciences. S.H.R. received institutional research support from Broken String Biosciences, which also supported R.H.C.W. and K.O.’s positions.