NmeCas9 is an intrinsically high-fidelity genome-editing platform

NmeCas9 protein accumulation in mammalian cells depends on co-expression of its sgRNA: full-length NmeCas9 is efficiently expressed when either of two sgRNAs is present but is far less abundant without one, whereas SpyCas9 accumulation is unaffected by the presence of its sgRNA. The type II-C ortholog CdiCas9 is dramatically stabilized against proteolysis in vitro by its cognate sgRNA, and similar protection could account for some or all of NmeCas9's sgRNA-dependent accumulation.

Wild-type NmeCas9 edits efficiently across human and mouse cells and tolerates several delivery routes. It edits all three of its natural bacterial protospacers in a split-GFP reporter in HEK293T cells at efficiency comparable to SpyCas9, and reprogrammed against 13 human chromosomal sites bearing N4GATT PAMs it produces indels at 12 of them, 5% to 47% by T7E1 assay and roughly 15% to 85% by targeted deep sequencing. Editing works by plasmid transfection, by ribonucleoprotein nucleofection in K562 cells and hTERT-immortalized human foreskin fibroblasts, and by lentiviral transduction of mouse embryonic stem cells and HEK293T cells.

NmeCas9's PAM requirement in mammalian cells is strict at the first position and flexible afterwards. Mutating the G1 residue of the 5′-N4GATT-3′ consensus to any other base drops editing to background, while four single-nucleotide variants (N4GCTT, N4GTTT, N4GACT, N4GATA) and two double variants (N4GTCT, N4GACA) support split-GFP editing at efficiencies approaching the consensus. N4GAGT, N4GATG and the double mutants N4GCCT, N4GCTA and N4GTTA give intermediate or low efficiency, and the remainder are at or near background.

At native chromosomal sites, every highly active NmeCas9 PAM variant tested except N4GACA supports editing, with indel efficiencies from roughly 8% to 60% by deep sequencing — so N4GCTT, N4GTTT, N4GACT, N4GATA and N4GTCT can be used alongside the N4GATT consensus to expand the targetable sequence space.

Sustained NmeCas9 expression is not genotoxic despite its RNA-dependent single-stranded DNA cleavage (DNase H) activity, which is independent of tracrRNA and PAM and could in principle degrade naturally occurring cellular R-loops. Mouse E14 embryonic stem cells stably expressing NmeCas9 or the DNase H-deficient dNmeCas9 show no increase in γH2AX foci over wild-type cells by immunofluorescence microscopy or flow cytometry, while UV exposure raises γH2AX as expected; HEK293T cells expressing NmeCas9 likewise show no difference from wild type.

Dual target sites bearing the composite PAM sequence NGGNGATT allow SpyCas9 and NmeCas9 to be compared without confounds, because both orthologs cleave the exact same internucleotide bond in the same sequence and chromatin context. When the target also carries a G residue at position −24 and another at −20 of the sgRNA-noncomplementary strand, the U6 promoter can express perfectly matched guides for both orthologs.

Neither NmeCas9 nor SpyCas9 is uniformly more efficient; which ortholog wins is site-specific. Across five shared chromosomal sites, SpyCas9 edits DTS1 and DTS8 more efficiently, NmeCas9 edits DTS3 and NTS7 more efficiently, and the two are approximately equal at DTS7. Adding an extra 5′-terminal G residue has little to no effect on either ortholog, while truncating NmeCas9 guides to 20 nucleotides has site-specific consequences: no reduction at DTS7, partial reduction at DTS3 and complete loss at DTS8.

At the same chromosomal sites, NmeCas9 and SpyCas9 leave similar but distinguishable repair signatures: small deletions predominate for both orthologs, but insertions are less frequent with NmeCas9 than with SpyCas9, and for both the vast majority of insertions are a single nucleotide. Deletion sizes vary considerably from site to site for both.

NmeCas9's longer guide and longer PAM make near-cognate sites in the human genome far rarer than for SpyCas9. Allowing permissive and semi-permissive PAMs, predicted off-target sites with six or fewer mismatches are two to three orders of magnitude less numerous for NmeCas9 than SpyCas9 — 21, 68 and 154 sites for three guides against 53,317, 31,144 and 60,706 for the SpyCas9 guides at the same loci. NmeCas9 sites with fewer than five mismatches are rare or absent, whereas the SpyCas9 guides have 49, 54 and 62 predicted sites with three or fewer mismatches.

GUIDE-seq at three chromosomal sites shared between the orthologs shows a difference in empirical accuracy of a different order. SpyCas9 induced 271 off-target sites with one guide and 54 with another, 88% and 77% of them at canonical NGG PAMs, and at five off-target sites the read counts exceeded those at the on-target site; a third guide gave only a single off-target with three reads. NmeCas9 at the same three sites yielded no reads at any site with six or fewer guide mismatches for two of the guides, and for the third only one site with four mismatches and 12 reads, roughly 100-fold below its 1222 on-target reads.

Repeating GUIDE-seq with a different transfection reagent separates genuine off-targets from experimental noise: 29 of 30 SpyCas9 off-target sites with up to five mismatches were recovered under both conditions, whereas the few NmeCas9 candidate sites showed no overlap between conditions — evidence that NmeCas9's sparse off-target reads reflect background rather than real editing.

Across a wider panel, NmeCas9 off-targeting is the exception rather than the rule. GUIDE-seq at 16 additional target sites — 10 with canonical N4GATT PAMs and six with variant functional PAMs — detects on-target editing at every site but no off-target reads for 14 of the 16 guides; the two exceptions have two and one off-target sites respectively. Of 19 guides tested in total, only two produced confirmed cellular off-target editing, and of the three bona fide off-target sites found, only one was edited at substantial frequency (11.6%), the other two at 0.3% or lower.

Both PAM quality and the position of guide-target mismatches govern which NmeCas9 near-cognate sites get edited. The efficiently edited NTS1C-OT1 carries two wobble pairs and one mismatch, all confined to the nine PAM-distal nucleotides, together with a canonical N4GATT PAM. The weakly edited NTS1C-OT2 has only a single mismatch, at the 11th nucleotide, but a non-canonical N4GGTT PAM. NTS25-OT1 is edited roughly 1000-fold less efficiently than its on-target site even though it carries an optimal N4GATT PAM while the on-target site carries the variant N4GATA.

SITE-Seq, a biochemical off-target assay independent of cellular DNA repair, corroborates NmeCas9's accuracy at the level of nuclease cleavage. At the three shared target sites, SpyCas9 ribonucleoproteins recovered hundreds of biochemical off-targets at 4 nM and thousands at 256 nM, while NmeCas9 recovered only its on-target sites at 4 nM and at most 29 off-targets at 256 nM. Of 12 further NmeCas9 target sites, eight recovered only on-targets at 4 nM and six of those no more than nine off-targets at 256 nM. Mismatches at recovered NmeCas9 off-targets are enriched toward the 5′ end of the target sequence.

Cellular validation of biochemically identified sites indicates the NmeCas9 off-target census is essentially complete. Across all NmeCas9 ribonucleoproteins only three cellular off-targets were found, all belonging to one guide and two of them already detected by GUIDE-seq. Every high-cleavage-sensitivity SITE-Seq site — recovered at all concentrations tested — showed editing in cells, while low-sensitivity sites recovered only at 64 nM or 256 nM were rarely edited (2 of 93). SpyCas9 by contrast gave 14 cellular off-targets across its three guides (8 of 70, 6 of 83 and 0 of 79 sites tested).

NmeCas9 tolerates guides from 19 to 26 nucleotides, with the lower bound depending on the target. In a split-GFP reporter, 20 to 23 nucleotides of complementarity match the natural 24-nucleotide guide while 18 and 19 give lower activity. At a chromosomal N4GATT site, 20 to 22 nucleotides of complementarity perform comparably to the natural length and 1 to 2 unpaired 5′ G residues cause no harm, but guides of 19 nucleotides or shorter carrying a mismatch in the first or second position give no detectable editing while a perfectly matched 19-nucleotide guide works. At an N4GCTT site, only 21 to 24 nucleotides of complementarity support editing, so 20-nucleotide guides work at some sites but not all.

Truncated sgRNAs suppress the residual off-targeting of NmeCas9 without creating new off-targets. GUIDE-seq with truncated guides at two chromosomal sites revealed no novel off-target sites, and at the most efficiently edited off-target found in this work, off-targeting persisted with 24- through 21-nucleotide guides but was abolished with a 20-nucleotide guide at no cost to on-target editing; a 19-nucleotide guide also abolished it but reduced on-target efficiency.

The 5′-terminal G residue needed for U6-driven sgRNA transcription can defeat truncation-based off-target suppression, so truncated guide design must account for it. Off-targeting at NTS1C-OT1 reappeared with an 18-nucleotide-complementarity guide because its 5′-terminal G pairs with a C at the off-target site, whereas in the 19-nucleotide guide both that G and the adjacent C are mismatched — making the more heavily truncated guide the better match to the off-target. sgRNAs produced by chemical synthesis, which need no 5′ G, avoid this problem.