In this Article we evaluate clinical lead guide RNAs (gRNAs) that, in combination with an ABE, achieve therapeutically desirable edits in PAH, ABCC6 or HPD in hepatocytes. To reduce off-target editing by these gRNAs, we explored a previously reported strategy, the in vitro or in cellulo use of hybrid gRNAs in which certain positions of the spacer sequence are substituted with DNA nucleotides11,12,13,14,15,16 -- a strategy compatible with the use of mRNA-lipid nanoparticles (LNPs) for in vivo delivery. In systematically assessing the effects of a series of hybrid gRNAs on the off-target profiles of clinical lead gRNAs, we unexpectedly observed that hybrid gRNAs have the potential to reduce bystander editing as well. Importantly, we found that hybrid gRNAs that reduced off-target editing by the lead PAH P281L-correcting and ABCC6 R1164X-correcting ABE/gRNA combinations in vivo, in PKU and PXE humanized mouse models, did so without compromising in vivo editing efficiency. On the contrary, the hybrid gRNAs <q>significantly increased the desired corrective editing in the liver while simultaneously reducing unwanted bystander editing.
In a previous study we characterized a clinical lead gRNA (hereon PAH1) targeting the human-specific protospacer sequence TCACAGTTCGGGGGTATACA, with the corresponding protospacer-adjacent motif (PAM) TGG, to correct the PAH P281L variant (Fig. 1a). The variant adenine base lies in position 5 of the protospacer sequence. A potential bystander adenine base lies nearby in position 3, and bystander A→G editing would result in a splice site disruption that has been reported to be pathogenic for PKU and thus would be undesirable. In combination with ABE8.8, the PAH1 gRNA results in efficient corrective editing of the P281L variant in cellulo (in HuH-7 hepatocytes homozygous for the P281L variant, derived from the HuH-7 human hepatoma cell line) and in vivo (in mouse liver), albeit with sizeable levels of bystander editing.
Most CRISPR-based off-target assay techniques are designed to detect double-strand breaks in DNA, whether in vitro or in cellulo. However, base editors such as ABE8.8 do not directly make double-strand breaks, and so conventional off-target assays such as GUIDE-seq do not accurately reflect off-target editing by base editors. We therefore utilized an ABE-tailored version of OligoNucleotide Enrichment and sequencing (ONE-seq) to nominate off-target sites, followed by genomic sequencing to verify these candidate sites (Fig. 1b) (Supplementary Note 1).
For the ABE8.8/PAH1 combination, we had previously performed a ONE-seq analysis followed by targeted amplicon sequencing of ~50 top-ranked sites that verified one genomic site of low-level off-target mutagenesis in HuH-7 hepatocytes. We now undertake a deeper analysis of the ONE-seq rank-ordered list, using hybrid capture sequencing to assess the entire set of 280 genomic sites with ONE-seq scores greater than 0.01 (normalized to the on-target site having a score of 1.0) (Supplementary Table 1). We verified an additional six sites of off-target mutagenesis in ABE8.8/PAH1 mRNA/gRNA-transfected P281L HuH-7 hepatocytes (Fig. 2a). None of the seven sites raises a concern of oncogenic risk (Supplementary Note 2). The third verified off-target site (PAH1_OT3) displayed the most off-target editing, with 1.3% editing.
To assess whether hybrid gRNAs could reduce off-target editing at the verified sites, we designed a series of 21 synthetic hybrid gRNAs (PAH1_hyb1-PAH1_hyb21) with single, double or triple DNA nucleotide substitutions ranging from positions 3 to 10 in the spacer sequence (Fig. 2b). We transfected ABE8.8 mRNA in combination with the PAH1 standard gRNA and each of the hybrid gRNAs into P281L HuH-7 hepatocytes and assessed for (1) on-target P281L corrective editing, (2) bystander editing at the on-target site and (3) PAH1_OT3 off-target editing in each of the samples (Fig. 2c). Although most of the PAH1 hybrid gRNAs had comparable on-target editing to the PAH1 standard gRNA (~90%), several had modestly decreased editing (as low as ~80%). Notably, and unexpectedly, most of the hybrid gRNAs reduced bystander editing (from 4.4% with the standard gRNA to as low as ~1%). The PAH1 hybrid gRNAs had highly varied effects on PAH1_OT3 off-target editing. Although triple DNA nucleotide substitutions generally reduced PAH1_OT3 off-target editing more than single or double substitutions, a couple of gRNAs with triple substitutions had greater PAH1_OT3 off-target editing than the standard gRNA. None of the 21 PAH1 hybrid gRNAs fully eliminated PAH1_OT3 off-target editing. As an orthogonal method of off-target profiling, we performed ONE-seq with ABE8.8 and each of the 21 PAH1 hybrid gRNAs (Fig. 2d). We used the number of genomic sites with ONE-seq scores greater than 0.01 as a metric of gRNA specificity -- the fewer the sites, the less potential for off-target editing. There was good agreement between the results of the site-specific PAH1_OT3 analysis and the more general ONE-seq off-target profiling for PAH1 hybrid gRNAs.
To further reduce off-target editing, we combined triple and double substitutions within a single hybrid gRNA. We chose the triple substitutions of PAH_hyb17 (positions 4, 5 and 6) or PAH_hyb16 (positions 3, 4 and 5), which maximally reduced PAH1_OT3 off-target editing and maximally reduced bystander editing while preserving on-target editing, and added the double substitutions of PAH_hyb15 (positions 9 and 10), which substantially reduced PAH1_OT3 off-target editing and bystander editing, yielding PAH1_hyb22 and PAH1_hyb23. We added an additional substitution in position 11 to PAH1_hyb23, yielding PAH1_hyb24. We compared these three hybrid gRNAs directly against the PAH1 standard gRNA and negative controls via mRNA/gRNA transfection in P281L HuH-7 hepatocytes, and all three significantly reduced PAH1_OT3 off-target editing and bystander editing (Fig. 2e). ONE-seq confirmed the improved off-target profiles of the three new hybrid gRNAs relative to the 21 original hybrid gRNAs; indeed, PAH1_hyb24 had zero off-target sites with ONE-seq scores greater than 0.01 (Fig. 2f). At all seven verified PAH1 off-target sites, OT1-OT7, PAH1_hyb22 reduced the detectable off-target editing to the background levels of the negative controls (Fig. 2g).
All standard off-target assessment techniques share a critical limitation: each is tied to the specific individual genome represented by the cells or by the in vitro genomic DNA sample used for analysis. For this reason, most off-target analyses have simply ignored the potential for naturally occurring human genetic variation to create novel off-target editing sites in some patients. A cautionary tale is provided by the finding that the gRNA used in the recently approved CRISPR-based therapy for sickle cell disease, exa-cel, has substantial off-target editing in hematopoietic stem cells at a genomic site created by a genetic variant present in ~10% of individuals of African ancestry, resulting in both indel mutations and chromosomal rearrangements. The ONE-seq technique is uniquely capable of performing a variant-aware off-target analysis, accommodating naturally occurring human genetic variation by using oligonucleotide libraries designed not just using the reference human genome but also incorporating data from the 1000 Genomes Project, the Human Genome Diversity Project, and so on. We used the bioinformatic tool CRISPRme to design a variant-aware ONE-seq library for the PAH1 protospacer with ~7,000 non-reference-genome oligonucleotides. In performing a variant-aware ONE-seq experiment with the ABE8.8/PAH1 combination, we identified 40 variant sites with ONE-seq scores greater than 0.01 (Fig. 2h). One of the top variant sites, with a ONE-seq score of 0.19, is a rare singleton 1-bp deletion that alters a reference genomic site that has seven mismatches with the on-target protospacer, with an NTG protospacer-adjacent motif (PAM), into a site that has three mismatches, with an NGG PAM (Extended Data Fig. 1). Another top variant site, with a ONE-seq score of 0.11, is common enough to have been previously catalogued as rs76813758, a single-nucleotide variant that alters a reference genomic site that has four mismatches with the on-target protospacer, with an NGG PAM, into a site that has three mismatches, with an NGG PAM. This variant is present in ~8% of individuals of African ancestry and ~3% of individuals of European ancestry (Extended Data Fig. 1).
One limitation of variant-aware off-target analysis is that, upon identifying candidate variant sites, it can be challenging to verify whether editing actually occurs at any of those sites in the therapeutically relevant cells (for example, hepatocytes) if there is no way to obtain such cells from individuals with those variants. We reasoned that the use of hybrid gRNAs should reduce the potential for off-target editing not only at reference genomic sites, but also at variant sites, mitigating the need to evaluate variant sites in cellulo. Accordingly, we performed variant-aware ONE-seq experiments for the ABE8.8/PAH1_hyb22 and ABE8.8/PAH1_hyb24 combinations (Fig. 2h and Extended Data Fig. 1). With PAH1_hyb22, there were just three variant sites with ONE-seq scores greater than 0.01, and with PAH1_hyb24, there were no variant sites with ONE-seq scores greater than 0.01. For rs76813758, the ONE-seq score dropped from 0.11 to 0.004 with PAH1_hyb22 and to 0.0001 with PAH1_hyb24 (1/10,000th of the in vitro editing activity of the on-target site, within the background of the assay).
Although our rational exploration of hybrid gRNAs was successful in rendering editing at the seven verified PAH1 off-target sites undetectable in cellulo, while preserving on-target editing efficiency in cellulo, it remained to be answered whether hybrid gRNAs could function as effectively in vivo. We formulated LNPs with ABE8.8 mRNA and PAH1 standard gRNA, PAH1_hyb22 gRNA, PAH1_hyb23 or PAH1_hyb24, exactly paralleling LNPs we used in a recently published study. We administered the LNPs to homozygous humanized P281L PKU mice at a dose of 2.5 mg kg. We observed that only the PAH1_hyb23 and PAH1_hyb24 gRNAs resulted in complete normalization of blood Phe levels (mean <125 µmol l) by 48 h after treatment, significantly outperforming the PAH1 standard gRNA (Extended Data Fig. 2). With ABE8.8/PAH1 LNPs, there was 40% (mean) whole-liver P281L corrective editing, and with each of the hybrid gRNAs, there was mean 50-60% editing, establishing that the use of the hybrid gRNAs did not compromise, and in fact significantly improved, on-target editing (Fig. 2i). The increased on-target editing was accompanied by significantly reduced bystander editing (mean 0.8% editing with standard gRNA compared to mean 0.2-0.3% editing with hybrid gRNAs).
We performed ONE-seq for the ABE8.8/PAH1 combination against the reference mouse genome (Supplementary Table 2). In interrogating the candidate sites with the top ONE-seq scores in liver samples from the LNP-treated mice, we verified a site (PAH1_mOT3) with mean 2.1% off-target editing with the PAH1 standard gRNA (Fig. 2i). PAH1_hyb22 gRNA, PAH1_hyb23 and PAH1_hyb24 all significantly reduced the off-target editing at this site (<0.2%).
Of the recurrent ABCC6 variants causative of PXE, we focused on the R1164X variant because of features suggesting it would be amenable to adenine base editing (Figs. 1a and 3a), namely the positioning vis-à-vis a TGG PAM compatible with a gRNA (designated PXE1) with protospacer sequence GTCACGGGAAACTGATCCTC, the variant adenine base lying in position 4, and potential bystander adenine bases lying in positions 9, 10 and 11, outside the reported editing window of ABE8.8. We used prime editing to generate a homozygous R1164X HuH-7 cell line (Supplementary Note 3). Transfection of the cell line with ABE8.8 mRNA and the PXE1 gRNA achieved substantial corrective editing of the R1164X variant (32%), albeit with a moderate amount of unwanted bystander editing (6.3%) (Fig. 3b). For the ABE8.8/PXE1 combination, we performed a ONE-seq analysis followed by targeted amplicon sequencing of top-ranked sites that verified six genomic sites of off-target mutagenesis in HuH-7 cells (Supplementary Table 3). None of the six sites raises a concern of oncogenic risk (Supplementary Note 2). The first and second verified off-target sites (PXE1_OT1 and PXE1_OT2) displayed the most off-target editing.
As we had done with the PAH1 gRNA, we designed a series of 21 hybrid gRNAs for PXE1 (PXE1_hyb1-PXE1_hyb21) with single, double or triple DNA nucleotide substitutions ranging from positions 3 to 10 in the spacer sequence (Fig. 3a), matching the designs of the PAH1 hybrid gRNAs. Following transfection of ABE8.8 mRNA and each of the gRNAs into R1164X HuH-7 hepatocytes, we found that all of the hybrid gRNAs reduced off-target editing at the OT1 and OT2 sites, some of the hybrid gRNAs increased on-target editing (to as high as ~50%), and some of the hybrid gRNAs reduced bystander editing (to as low as ~3%) (Fig. 3b). We performed ONE-seq for each of the hybrid gRNAs and, judging by the number of genomic sites with ONE-seq scores greater than 0.01, improvement of off-target editing was concordant with the effects on the individual OT1 and OT2 sites (Fig. 3c). We further evaluated PXE1_hyb18 and found that it significantly increased on-target editing and significantly, though not entirely, reduced off-target editing at the OT1 and OT2 sites (Fig. 3d). It is possible that additional DNA nucleotide substitutions would yield hybrid gRNAs that further reduce off-target editing.
We used CRISPR-Cas9 targeting in mouse embryos to generate a humanized PXE model, in the C57BL/6J background, in which we replaced a small portion of the endogenous mouse Abcc6 exon 24 with the orthologous human sequence spanning the PXE1 protospacer/PAM sequences and containing the R1164X variant (Supplementary Note 4). Consistent with previous studies of Abcc6 knockout mice, homozygous R1164X mice had reduced blood pyrophosphate levels compared to wild-type littermates (Extended Data Fig. 3).
We formulated LNPs with ABE8.8 mRNA and either PXE1 standard gRNA or PXE1_hyb18 gRNA. We administered the LNPs to homozygous R1164X mice at a dose of 2.5 mg kg. With ABE8.8/PXE1 LNPs, there was mean 24% whole-liver P281L corrective editing in the absence of bystander editing, and mean 3.5% corrective editing with bystander editing; with the hybrid gRNA, there was 29% and 2.1% editing, respectively (Fig. 3e). The PXE1 hybrid gRNA significantly increased the on-target editing and reduced bystander editing in vivo, as was observed with PAH1 hybrid gRNAs in vivo. LNP treatment also largely normalized the blood pyrophosphate levels in the homozygous R1164X mice (Extended Data Fig. 3a).
We performed ONE-seq for the ABE8.8/PXE1 combination against the reference mouse genome (Supplementary Table 4). In interrogating the candidate sites with the top ONE-seq scores in liver samples from the LNP-treated mice, we verified a site (PXE1_mOT24) with mean 8.4% off-target editing with the PXE1 standard gRNA (Fig. 3e). The PXE1_hyb18 gRNA significantly reduced off-target editing at this site, with mean 0.6% editing. We also assessed blood alanine aminotransferase (ALT) and cytokine/chemokine levels in LNP-treated mice, and we observed similar safety profiles with the standard and hybrid gRNAs (Extended Data Fig. 3b and Supplementary Table 5).
Lacking any previous data regarding adenine base editing to inactivate the HPD gene (our previous study assessed only cytosine base editing), we screened for precise disruption of a splice donor or acceptor site in the human HPD gene by adenine base editing (Supplementary Note 5). We observed the highest levels of editing with the gRNAs designated 'HPD20', 'HPD3', 'HPD4' and 'HPD5', targeting either the splice acceptor of exon 7 or the splice donor of exon 8 (Fig. 1a and Extended Data Fig. 4a). To evaluate for off-target editing, we used two ONE-seq libraries against the reference human genome: a combined library for the cluster of closely spaced HPD3, HPD4 and HPD5 protospacers (consecutive protospacers shifted by one nucleotide), and a library for the HPD20 protospacer. Targeted amplicon sequencing of top-ranked sites in ABE8.8/gRNA-treated wild-type HuH-7 hepatocytes verified one genomic site of off-target mutagenesis each for HPD3 and HPD4, and none for HPD5 and HPD20 (Extended Data Fig. 4b and Supplementary Tables 6-9).
Because of its more favourable on-target and off-target editing profiles, we prioritized HPD20 for further study as a clinical lead gRNA. We assessed a full series of HPD20 hybrid gRNAs (HPD20_hyb1-HPD20_hyb21) (Fig. 4a), matching the designs of the PAH1 hybrid gRNAs, via ABE8.8 mRNA/gRNA transfection in wild-type HuH-7 hepatocytes, for both editing of the target splice site adenine (in position 6 of the protospacer sequence, TCTGCAGAAAGCACGGGAAC, with a GGG PAM) and bystander editing of positions 8, 9 and/or 10 (Fig. 4b). Because we had not verified any individual genomic sites of off-target editing with the HPD20 standard gRNA, we used ONE-seq to assess the off-target editing profiles of the HPD20 hybrid gRNAs, with many showing decreased off-target potential (that is, number of sites with ONE-seq scores greater than 0.01), but a few showing increased off-target potential (Fig. 4c). Two hybrid gRNAs, HPD20_hyb22 and HPD20_hyb23, with five DNA nucleotide substitutions each, maintained on-target editing, substantially reduced bystander editing, and had the fewest sites with ONE-seq scores greater than 0.01 (Fig. 4a-c).
We also assessed the HPD4 gRNA with protospacer sequence CACTCACAGTTTAGGAAGTA and a GGG PAM, with the target adenine base in position 6 and a bystander adenine base in position 8. Having verified a site of off-target editing with the HPD4 standard gRNA (HPD4_OT13) with mean 4.9% editing, we assessed a full series of HPD4 hybrid gRNAs (HPD4_hyb1-HPD4_hyb21) (Fig. 4d), matching the designs of the PAH1 hybrid gRNAs, via ABE8.8 mRNA/gRNA transfection in wild-type HuH-7 hepatocytes (Fig. 4e). All the hybrid gRNAs reduced OT13 off-target editing, some close to the background level, and many hybrid gRNAs reduced bystander editing while maintaining on-target editing.
In a previous study, we characterized a gRNA (herein termed 'PAH4') that corrects the PAH R408W variant, the most frequent variant in PKU, in cellulo and in vivo. Because there are no NGG PAMs situated appropriately for adenine base editing of the PAH R408W variant, we screened PAM-altered and PAM-relaxed versions of ABEs and identified an optimal combination of SpRY-ABE8.8 and protospacer sequence GGCCAAGGTATTGTGGCAGC with an AAA PAM (Fig. 1a). The variant adenine base lies in position 5 of the protospacer sequence. Potential bystander adenine bases lie in positions 6 and 10; bystander editing at position 6 would result in a benign synonymous variant, whereas editing at position 10 would result in a nonsynonymous variant that has been reported to be a pathogenic variant for PKU.
Editors based on the SpRY variant of Streptococcus pyogenes Cas9 are near-PAMless in their ability to engage genomic sites, greatly expanding their targeting range, but also incurring an increased potential for off-target editing. To evaluate the SpRY-ABE8.8/PAH4 combination, which corrects the PAH R408W variant, we performed ONE-seq against the reference human genome. Rather than interrogate candidate off-target sites nominated by ONE-seq with the PAH4 standard gRNA, we immediately tested a limited series of hybrid gRNAs (double/triple substitutions, PAH4_hyb9-PAH4_hyb21, matching the designs of the analogous PAH1 hybrid gRNAs) with the goal of reducing the potential of SpRY-ABE8.8 for off-target editing, and prospectively winnowing the list of candidate off-target sites for verification (Fig. 5a). Several of the PAH4 hybrid gRNAs displayed increased on-target editing and/or decreased bystander editing at protospacer position 10 (nonsynonymous variant) (Fig. 5b). ONE-seq demonstrated reduced off-target potential with some of these favourable hybrid gRNAs, despite the use of a PAM-relaxed ABE (Fig. 5c).
In comparing our results across the five studied loci, our findings suggest that there will not be a single set of hybrid gRNA modifications that will fully optimize off-target and on-target editing across all genomic loci, although suggestive patterns are evident. The 'hyb16' and 'hyb17' designs (DNA nucleotide substitutions in protospacer positions 4, 5 and 6, or in positions 5, 6 and 7) substantially reduced off-target editing potential while maintaining on-target editing across all loci, so these designs could be a reasonable starting point for future hybrid gRNA screening campaigns. To test this proposition, we compared standard, hyb16 and hyb17 gRNAs for corrective editing of variants at four additional loci (Extended Data Fig. 5). The CPS1 Q335X variant, the focus of our recent report of a personalized N-of-1 base-editing therapy ('k-abe') administered to an infant with carbamoyl phosphate synthetase (CPS1) deficiency, can be corrected with NGC-ABE8e-V106W (also termed SpCas9-LWKYQS-ABE8e-V106W) and 'CPS1-1' gRNA. The CPS1 R780H variant, another ultra-rare cause of CPS1 deficiency, can be corrected with NGC-ABE8e-V106W and 'CPS1-2' gRNA. The ABCC6 R1141X variant is the most frequent variant in PXE and can be corrected with SpRY-ABE8.8 and 'PXE2' gRNA. The PAH c.1066-11G>A variant is the second most frequent variant in PKU and can be corrected with SpRY-ABE8.8 and 'PAH5' gRNA. In all four cases, one or both hybrid gRNAs maintained or improved the desired corrective editing and reduced bystander editing in cellulo, while substantially improving ONE-seq off-target profiles (Extended Data Fig. 5). We note that the six-month development time of the 'k-abe' therapy did not allow for extensive screening of candidate hybrid gRNAs, but in future time-limited N-of-1 development efforts, quick testing and validation of the hyb16 design or hyb17 design would permit it to be incorporated into a therapy administered to a patient.