Sniper2L is a high-fidelity Cas9 variant with high activity

Genome editing with SpCas9 is limited by off-target effects and, for many high-fidelity variants, reduced on-target activity. Prior engineered SpCas9 variants (eSpCas9, Cas9-HF1, HypaCas9, R63A/Q768A, evoCas9, HiFi Cas9, Sniper-Cas9/Sniper1) improve specificity but often at the cost of general activity, creating a trade-off that constrains practical applications such as gene therapy and genetic screening. This study aims to develop and characterize a next-generation SpCas9 variant that overcomes this trade-off by preserving high on-target activity while enhancing specificity across many target sequences and delivery modalities. The authors used directed evolution of Sniper1 to create Sniper2L and assessed its activity/specificity at scale, investigated its biophysical mechanism of specificity, and built predictive models to facilitate guide selection.

- Directed evolution and mutagenesis: The authors used an E. coli-based ‘Sniper screen’ applying positive selection (Cas9 cleavage of a ccdB-containing plasmid) and negative selection (avoid cleavage at a mismatched genomic target) with the EMX1.6 sgRNA containing a PAM-distal mismatch at position 13 to bias for discriminating mismatches in the spacer middle region. Sniper1 libraries were generated via random mutagenesis (XL1-red, Genemorph II, Diversify PCR). After four screening rounds, sequencing revealed a hotspot at residue E1007. Site-saturation mutagenesis at E1007 yielded 20 substitutions; activities at matched/mismatched targets were profiled, identifying E1007L (Sniper2L) and E1007P (Sniper2P) for further study. - High-throughput cellular evaluations (lentiviral delivery): HEK293T cell lines individually expressing Sniper1, Sniper2L, or Sniper2P via lentivirus were generated with comparable protein expression. Pooled lentiviral libraries of sgRNA–target pairs were transduced and indels quantified by deep sequencing at days 4 and 7. Libraries: A (11,802 pairs; PAM and mismatch tolerance), B (23,679 pairs; NGG, NGH, non-NG targets for general activity), C (7,567 pairs; same targets as B but using tRNA-processed perfectly matched N20 sgRNAs). PAM compatibilities were assessed using NNNN-PAM targets. On-target activities were analyzed on 8,744 NGG targets (library B). Specificity was assessed using 30 sgRNAs each paired with 98 mismatched targets (one-, two-, or three-base mismatches) selected to normalize on-target activity across variants. - RNP-based high-throughput evaluation: To model therapeutically relevant delivery, SpCas9 variants (WT SpCas9, Sniper1, Sniper2L, HiFi Cas9, Cas9 R63A/Q768A) were electroporated as pre-assembled RNPs into a HEK293T library (library A). A guide-swapping strategy enabled variants to acquire library-expressed gRNAs. Co-delivery of an HPRT-targeting RNP allowed post-electroporation 6-thioguanine (6-TG) selection to enrich cells receiving RNPs, removing 65–80% of cells lacking RNP delivery. Post-selection, genomic DNA was sequenced to quantify indels at matched and mismatched targets and to validate PAM preferences under RNP conditions. - Single-molecule mechanism (smFRET): Cas9-gRNA RNP-induced DNA unwinding was measured by smFRET on surface-immobilized DNA targets with 0–4 consecutive PAM-distal mismatches or a single mismatch at position 10. The fraction of unwound molecules (funwound) and unwinding specificity (1 − funwound(mismatch)/funwound(perfect match)) were quantified across variants. - Deep learning models (DeepSniper): Separate CNN-based models were trained to predict on-target activities for Sniper1 and Sniper2L using (G/g)N19 vs tRNA-N20 sgRNA systems (30-nt context inputs), and off-target activities for mismatched targets with (G/g)N19 sgRNAs (inputs include guide, target, and mismatch identities). Data from libraries A–C were split into training and held-out test sets. Models used ReLU activations, dropout (0.3), mean absolute error loss, and Adam optimizer (lr=1e-4). Code provided on GitHub. - Data processing and statistics: Indel frequencies were background-corrected; targets with <100 reads or background indels >8% were excluded. Non-parametric tests (Kruskal–Wallis; Mann–Whitney U) were applied as indicated.

- Variant identification: E1007L and E1007P substitutions in Sniper1 (Sniper2L and Sniper2P) uniquely combined high on-target activity with low off-target activity across multiple test sites; other E1007 substitutions did not. - PAM compatibility: Sniper1, Sniper2L, and Sniper2P exhibited similar PAM preferences with highest activities at NGG; noncanonical PAM activities were low when delivered as RNPs. - Large-scale on-target activity (lentiviral): On 8,744 NGG targets, Sniper2L had significantly higher indel efficiencies than Sniper1; Sniper2P was lowest (n=7,702 in Fig. 2a; p≈1.3×10^-35 for Sniper1 vs Sniper2L). General activity comparisons across thousands of targets showed Sniper2L maintained activity comparable to WT SpCas9. - Specificity (lentiviral, mismatches): Using 30 sgRNAs with matched on-target activities, Sniper2L showed significantly lower activity at single-mismatch targets than Sniper1 and Sniper2P (n≈1,732; Fig. 2b) and thus significantly higher specificity (Fig. 2c). Specificity was elevated across mismatch positions, with greater discrimination in PAM-proximal regions for all variants (Fig. 2d). Multi-base transversion mismatches (2–3 bp) dramatically reduced relative indel frequencies (Fig. 2e). Wobble mismatches were better tolerated than transversions (Extended Data Fig. 5). - Guide format effects: With (G/g)N19 sgRNAs from U6 promoters (5′G enforced, possibly mismatched at position 1) vs perfectly matched tRNA-N20 sgRNAs, Sniper2L/Sniper2P generally showed slightly higher activities with (G/g)N19 overall, but for targets starting with 5′C or T, tRNA-N20 improved Sniper2L activity (Fig. 2f). - Trade-off outlier: Across eight benchmark sgRNAs, Sniper2L simultaneously increased specificity and maintained high on-target activity, representing an outlier to the common specificity–activity trade-off observed in high-fidelity variants (Fig. 2g). - RNP delivery (high-throughput): Under RNP delivery with 6-TG enrichment, variants retained NGG preference. On 105–113 perfectly matched NGG targets per variant, general on-target activities were similar (no significant difference; Fig. 3b). At mismatched targets, Sniper2L was markedly less active than other variants (one-, two-, three-base mismatches), indicating superior specificity (Fig. 3c–d). Targeted tests at six loci showed Sniper2L had higher on-target and lower off-target editing than WT SpCas9 (Extended Data Fig. 6). - Mechanism (smFRET): The number of PAM-distal mismatches needed for >2-fold reduction in funwound was lowest for Sniper2L (np≥1) versus Sniper1/Sniper2P (np≥2) and WT SpCas9 (np≥3), making Sniper2L most stringent. Unwinding specificity with a single PAM-distal mismatch and at position 10 was highest for Sniper2L (e.g., p10 mismatch specificity 0.83 for Sniper2L vs 0.33 for WT; Fig. 4b–d). - Predictive modeling (DeepSniper): Deep learning models accurately predicted Sniper1/Sniper2L activities on held-out targets: matched targets r=0.96 (R=0.94) and mismatched targets r=0.92 (R=0.90) (Fig. 5b–c). Web tool provided. - Practical screening note: 6-TG selection removed ~65–80% of cells, enriching for RNP-delivered cells and substantially increasing HPRT indel frequencies (Extended Data Fig. 7).

The study addresses the longstanding trade-off between specificity and on-target activity in SpCas9 engineering. By directed evolution focused on mismatch discrimination in the spacer middle region and hotspot saturation at E1007, the authors identified Sniper2L, which increases specificity while retaining high activity across thousands of targets and under both lentiviral and therapeutically relevant RNP delivery. High-throughput analyses demonstrated that Sniper2L is broadly active on NGG targets with significantly reduced activity at mismatched sites. Single-molecule smFRET experiments showed Sniper2L avoids unwinding even single-mismatch targets more effectively than other variants, providing a mechanistic basis for its enhanced fidelity. The DeepSniper models enable practical guide selection by predicting activities on matched and mismatched targets, further facilitating adoption. Collectively, these results show that rationally guided directed evolution at strategic sites can produce Cas9 variants that defy the typical specificity–activity trade-off, with immediate relevance to precise and efficient genome editing applications, including ex vivo therapeutic contexts where RNP delivery is preferred.

The authors developed Sniper2L, a next-generation high-fidelity SpCas9 variant (E1007L in Sniper1) that uniquely combines high on-target activity with enhanced specificity, overcoming the common trade-off. They validated its performance at scale under both lentiviral and RNP delivery, elucidated a mechanistic basis for specificity via smFRET (stringent DNA unwinding discrimination), and released DeepSniper models to predict variant activity. These contributions enable efficient and specific genome editing, streamlining sgRNA selection and potentially improving outcomes in research and therapeutic settings. Future work could include directed evolution starting from Sniper2L with alternative mismatch positions or sgRNAs to discover additional fidelity-enhancing mutations, evaluation across diverse cell types and in vivo systems, integration with editors (base/prime editors) to assess effects on specificity profiles, and structural/biophysical studies to fully resolve how E1007 substitutions modulate conformational activation and mismatch proofreading.

- Cell-type scope: Experiments were conducted in HEK293T cells; while prior work suggests relative variant performance is consistent across cell types, absolute activities can vary. - Library coverage under RNP: The 6-TG enrichment reduced library coverage by ~3–5-fold, necessitating read-count filtering and potentially excluding low-coverage targets. - smFRET correlation: Single-molecule unwinding for perfect matches was lower for Sniper2L than for WT/Sniper1 despite high cellular editing, indicating this assay does not fully capture on-target editing dynamics. - PAM/target scope: Most detailed analyses focused on NGG PAMs; noncanonical PAM activities were low under RNP conditions due to shorter exposure, and broader PAM compatibility under varied conditions was not extensively profiled here. - Generalizability to other editors: Effects of E1007L/P on the fidelity of other CRISPR modalities (e.g., base editors) remain to be determined.