IntroductionCRISPR-Cas systems have revolutionized genetic engineering by enabling precise genome modifications. Derived from a bacterial adaptive immune mechanism, CRISPR-Cas9 allows targeted DNA edits in various organisms, enabling vast potential in medicine, agriculture, and biotechnology1,2. However, CRISPR-Cas9 faces challenges, notably off-target effects where unintended genomic regions similar to the target sequence can be edited3,4. This off-target activity is often due to sequence homology between the protospacer region of the gRNA and non-target regions of the genome. These off-target edits can lead to unwanted genetic changes and disrupt the function of essential genes leading to adverse effects on cellular function, viability, development of secondary diseases, or exacerbate existing conditions that pose risks in therapeutic applications.Recent advancements in gene editing have driven the parallel development of technologies to nominate off-target editing effects5,6,7,8,9,10,11,12. These technologies broadly fall into three categories: cell-based in cellulo/in situ, computational in silico, and biochemical in vitro assays each with distinct advantages and disadvantages. Cell-based assays like GUIDE-seq are known for high precision due to their ability to accurately mimic the cellular environment7,13,14. While cell-based methods provide nomination in a cell-type-specific and even donor-specific manner, these methods face challenges in representing diverse genome populations and typically require biologically variable processes such as DNA double-strand breaks (DSBs) and blunt integration of a dsDNA tag through non-homologous end joining (NHEJ) as a mechanism of action. Biochemical in vitro assays are highly sensitive and can represent patient genotypes with diverse editing modalities since they are executed on cell-free genomic DNA, but have low precision and suffer difficulties in replicating the cellular environment, which significantly impacts enzymatic activity6,7,13,14,15. Computational in silico approaches can rapidly predict all candidate off-target sites with homology to the gRNA in a patient genome/population-specific manner, but result in low precision due to insufficient data to inform predicted enzymatic and biological activity across diverse cell types11,12,13,14,16. All methods are setback by a lack empirical data and application of standards to characterize sensitivity against known truth, as it is challenging to ascertain a complete list of true positives for each gRNA in an unbiased manner. To better standardize off-target safety assessment, comprehensive information on assessment methods and guidance for their use is essential.Most methods for off-target nomination currently utilize next-generation sequencing (NGS) to detect off-targets. The efficiency of library preparation, library complexity (i.e., unique genome equivalents sequenced), read depth, and computational analysis methods are all known limitations affecting performance of NGS methods17. Some work has endeavored to improve upstream biological and library preparation inefficiencies of in-cellulo methods18,19. Other work has presented evidence of biological optimizations that proxy cell lines promiscuously editing off-targets may represent a superset of off-targets in HSPCs20. However, exploration of other NGS variables commonly benchmarked in parallel fields to drive uniformity in experimental procedures and reporting, are still largely lacking for gene editing off-target nomination17,21. The FDA has recently worked to address this need through draft guidelines for the cell and gene therapy industry, which currently recommends multiple orthogonal methods for off-target nomination22. This guideline for orthogonality ideally aims to mitigate known and unknown weaknesses in nomination, including variability in NGS method execution.Rapid innovations in gene editing have further complicated the off-target nomination process with the discovery of various editors possessing unique qualities beyond the targeted generation of DSBs with a Type II nuclease. This includes engineered systems generating staggered double-strand breaks (DSBs) in Type V CRISPR-Cas systems or leveraging single-strand DNA breaks (SSBs) coupled to engineered moieties for editing23,24,25. SSB engineered moieties like adenine base editors (ABE)26, cytosine base editors (CBE)27, and prime editors (PE)28 offer benefits such as decreased frequencies of unintended indels and even the ability to make non-indel DNA changes typically reserved for homology-directed repair (HDR). However, they lack consensus guidance regarding appropriate nomination assays for off-target assessment, and previous work has largely used DSB nomination methods as a proxy29,30. Due to the nature of these editors to generate both indels and SNPs, an orthogonal approach specializing on each event type may be needed to sensitively nominate off-targets.In this work, we develop an optimized end-to-end in cellulo nomination method and analysis pipeline, UNCOVERseq (Unbiased Nomination of CRISPR Off-target Variants using Enhanced RhPCR), leveraging NHEJ-based dsDNA tag integration for RNase H-dependent PCR (rhPCR)-based library preparation31. We demonstrate that using this method with a cell line under promiscuous editing conditions serves as a sensitive proxy for clinically relevant cell types. We establish experimental recommendations for library complexity, coverage depth, and end-to-end process controls that reliably lead to the successful nomination of risk-tiered edited sites with sub-0.01% editing (indel) frequencies observable empirically, and demonstrate that this optimized method outperforms other nomination methods through benchmarking comparisons. By screening 192 gRNAs using UNCOVERseq, we identify an experimental set of gRNAs spanning a broad range of specificities that are broadly representative of use-cases encountered in the field. We find that off-target effects of base editors are directly, rank-order correlated to DSB editing frequencies, demonstrating UNCOVERseq as an effective technology for nominating meaningful off-targets for both current and next-generation editing modalities. We conclude that, regardless of editor modality, the off-target burden of gRNAs quickly decreases to frequencies not measurable—that is, below the limit of detection - with current NGS methods (0.01% for indels and 0.5% for base editing) as gRNA specificity ratios rise. However, off-targets may still be observable at higher specificities in an editor/gRNA-dependent manner, emphasizing the need to consistently and empirically nominate and interrogate off-targets across the specificity spectrum.ResultsDevelopment and optimization of UNCOVERseqTo create our nomination method, we first started with the original GUIDE-seq protocol and designed an orthogonal dsDNA sequence with sufficient length to perform a modified rhPCR to multiplex primers in close proximity within a single reaction while avoiding primer-dimers (Supplementary Data 1). To streamline the process for preparing the nomination gDNA libraries we additionally converted from a mechanical to enzymatic DNA fragmentation (Fig. 1A). Upon analyzing data, we observed that freely adaptered dsDNA tag was allocated an average range of 37–67% of reads, varying across 4 gRNAs (Fig. 1B). This same artifact was also observed with the original GUIDE-seq protocol (Supplementary Fig. 1). To improve usable reads resulting from NGS, we introduced a blocking oligo into the PCR1 preparation designed to target the adapter:dsDNA junction (Fig. 1A). Introduction of this blocker reduced reads belonging to the adapter:dsDNA artifact to an average range of 0.3–0.5%, meaning >99% of reads were now belonging to gDNA:dsDNA junctions (Fig. 1B). Nomination frequencies were found to be conserved for all gRNAs (R2 = 0.99) with and without the blocking oligo (Supplementary Fig. 2).Fig. 1: Comparison of UNCOVERseq and the GUIDE-seq off-target nomination workflows.Full size imageA An overview of the UNCOVERseq workflow demonstrates that cells with a genomically integrated dsDNA tag have gDNA extracted and amplified with rhPCR in a single reaction, with dsDNA tag:adapter byproducts being blocked by a targeted oligo before being sequenced and analyzed using our described workflow. B A depiction is shown regarding what kind of events are targeted by the blocking oligo, with the usable reads (non-dsDNA:adapter reads) measured across nominated off-targets for 4 gRNAs in K562 (n = 3 per gRNA) with (light blue) or without (dark blue) the blocking oligo (mean ± s.d shown). C Tukey box plot comparison of different alignment methods used in publicly available nomination packages was performed to determine differences in Levenshtein distance 2 million reads per sample for maximum sensitivity (~100% analytical sensitivity in all Severity Bins) in assessing candidate off-target sites (Fig. 4H). It is possible that read depth requirements may vary with off-target number. However, by using a promiscuous gRNA (specificity ratio = 0.013) to determine this value, we propose that this represents the number of reads to successfully nominate sites using UNCOVERseq even with gRNAs with very poor specificity.Comparative nomination analysis of UNCOVERseq to other methodsA comparative analysis of UNCOVERseq to peer-reviewed accounts of other nominations methods was performed to better understand how the sensitivity and nomination frequencies of diverse methods compare to UNCOVERseq. Due to variable operational conditions, precision, and total nomination list sizes reported of different methods, we postulate that sensitivity is most appropriately measured using either off-targets with confirmed editing or off-targets from methods with high precision. Interrogation of the 60 LAG3 site 9 gRNA off-targets compared to CHANGE-seq7 and GUIDE-seq7 nominations demonstrated that both methods could nominate the most frequent group of confirmed off-targets (Severity Bin 1) with 91–100% sensitivity, but sensitivity rapidly decreased in the lower frequency off-target bins. CHANGE-seq was demonstrated to have a sensitivity between 66% and 75% for recovering Severity Bin 3–5 off-targets, while GUIDE-seq had a linear decrease from 16 to 0% for these same Severity Bins of confirmable off-targets (Supplementary Fig. 5). Previous analysis of the 60 LAG3 site 9 off-targets demonstrated that 100% of sites in Severity Bin 1–3 represented confirmed off-targets, demonstrating that published accounts of both GUIDE-seq and CHANGE-seq lack the ability to sensitively detect all off-targets (Fig. 4; Supplementary Fig. 5).Random sampling of the LAG3 site 9 dataset with 100% reproducibility showed UNCOVERseq-nominated sites had high precision, with confirmation frequencies correlating to average nomination frequencies down to the limit of detection (Fig. 4). Using this logic, we postulate the full 723 sites in this fully reproducible set are also likely to represent true positives. Investigation of sensitivity and frequencies of previous accounts of CHANGE-seq and GUIDE-seq for this gRNA yielded similar trends in sensitivity per Severity Bin, further supporting this (Fig. 5A). This provides additional evidence that reproducible UNCOVERseq sites have high likelihood of being true positives and may serve as an appropriate proxy for measuring sensitivity of different methods.Fig. 5: Comparative analysis of UNCOVERseq to other nomination technologies.Full size imageComparison of the sensitivity of other published accounts of nomination technologies to nominate 100% reproducible UNCOVERseq off-target sites using A LAG3 site 9 (n = 12 nomination replicates) B PCSK9 site 1 (n = 6 nomination replicates) C EMX1 (n = 6 nomination replicates) and D FANCF (n = 6 nomination replicates). E Total number of off-targets nominated per nomination method per gRNA with color gradient shifts at the number of off-targets easily interrogated using amplicon sequencing (250; gray/red). Legend displays the corresponding method to each column in the heatmaps along with the corresponding method type (in cellulo, in situ, in vitro). Rep. sites; Reproducible UNCOVERseq sites between all replicates. UNCOVERseq—All represents all Tier 1 to Tier 3 sites from the experiments.Using our previous finding that highly reproduced UNCOVERseq off-targets are likely indicative of true positives (Fig. 4B), we compared fully reproduced off-targets from 6 biological replicate UNCOVERseq experiments for three gRNAs (EMX1, FANCF, and PCSK9) to previous accounts of off-targets from GUIDE-seq5, INDUCE-seq10, Olitag-seq18, CHANGE-seq7, CIRCLE-seq6, AID-seq32, BLISS33, ONE-seq34, DIG-seq35, Digenome-seq34, and SITE-seq8 nomination methods (Supplementary Data 6). For PCSK9, only a combined set of ONE-seq and Digenome-seq sites was available, and nomination sensitivity was found to fluctuate from 0 to 83% through Severity Bins 2–5 (Fig. 5B). For EMX1, DIG-seq and BLISS sensitivity decreased to 29% and 57% after Severity Bin 1, respectively, and sharply decreased at lower frequency bins. Cell-based methods GUIDE-seq, OliTag-seq and INDUCE-seq sensitivity dropped after Bin 2 (11–53% sensitivity), ranging from 0 to 55% for off-target nomination frequencies observed after Severity Bin 2 (Fig. 5C). However, nomination frequencies for GUIDE-seq, OliTag-seq and INDUCE-seq correlated well with expected frequencies (R2 = 0.52–0.70), if identified (Supplementary Fig. 5). Remaining biochemical in vitro nomination methods had the highest sensitivity compared to UNCOVERseq, with approximate rank-order as follows: AID-seq/SITE-seq > CHANGE-seq >CIRCLE-seq (Fig. 5C). Sensitivity dropped for all these methods at Severity Bin 3, and ranged from 37 to 95% for Severity Bin 3–5 (Fig. 5C). For FANCF, GUIDE-seq and DIG-seq sensitivity quickly decreased below 50% after Severity Bin 2 to a range from 0 to 21% below Severity Bin 3 (Fig. 5D). AID-seq/SITE-seq achieved higher sensitivity compared to CIRCLE-seq/CHANGE-seq, similar to previous observations with EMX1 (Fig. 5C, D). CHANGE-seq and CIRCLE-seq sensitivity dropped to a range of 67—78% after Severity Bin 1 (Fig. 5D). SITE-seq and AID-seq sensitivity dropped after Severity Bin 2 and ranged from 75 to 100% (Fig. 5D).Each method nominated a widely variable number of sites per gRNA, which affects the calculation of analytical sensitivity and precision. The number of nominated sites ranged from 9 for FANCF using GUIDE-seq to 6,849 off-targets using LAG3 site 9 with CHANGE-seq (Fig. 5E). While biochemical in vitro methods like CHANGE-seq, AID-seq, and SITE-seq showed the greatest comparative sensitivity, they also had the largest nominated site list, ranging from 291 to 1328 sites for the FANCF and EMX1 gRNAs shared between most methods (Fig. 5E). Investigation of method normalized nomination frequencies demonstrate that in vitro methods have the poorest correlation to UNCOVERseq derived frequencies (R2 = 0.01–0.26), with AID-seq and SITE-seq having the lowest correlations (Supplementary Fig. 5). Nomination frequencies derived from cell-based in cellulo and in situ methods (GUIDE-seq, OliTag-seq, INDUCE-seq, BLISS) better correlate to UNCOVERseq nomination frequencies (R2 = 0.31–0.70), which previously were shown to correlate to observed confirmation frequencies (Supplementary Fig. 5; Fig. 4F; Fig. 5). These findings demonstrate that UNCOVERseq improves upon the sensitivity of existing in cellulo methods such as GUIDE-seq, in addition to subsequent improved methodologies such as OliTag-seq. Our findings also demonstrate that in vitro methods are not inherently more sensitive than in cellulo methods for discovering true off-targets, and UNCOVERseq nominates likely confirmable off-targets not detected in other methods.Comparative confirmation analysis of UNCOVERseq to other methodsTo characterize the sensitivity and precision of UNCOVERseq more comprehensively relative to other nomination technologies, we conducted targeted editing confirmation experiments in the promiscuous HEK293-Cas9 cell system using the two most widely implemented benchmarking gRNA loci in our set, EMX1 and FANCF (Fig. 6). This system was selected because stable Cas9 expression was previously shown to yield highly permissive, low specificity conditions, thereby maximizing the likelihood of correctly detecting confirmable off-target editing (i.e., true positives) that can be edited in biological contexts. To select nomination sites across technologies, nomination frequencies were first method normalized to the on-target value and these frequencies were used to sample sites across the full spectrum of predicted editing probabilities using the previously derived Severity Bins of nomination (Fig. 6A, B; Supplementary Fig. 7). Each nomination method had sites selected from all available Severity Bins per method while prioritizing top frequency sites across methods, for a total of 297 putative off-target loci for EMX1 and 181 putative off-target loci for FANCF (Supplementary Data 7). After intersecting and consolidating nominated sites common to different selections per method, the final confirmation panels comprised of 169 putative off-targets for EMX1 and 106 putative off-targets for FANCF, with variable numbers of sites per nomination methodology (Fig. 6A, B; Supplementary Data 7). Following consolidation of common off-targets across methods, all methods maintained relatively identical nomination frequency distributions to the original selection, with the exception of CIRCLE-seq, CHANGE-seq, SITE-seq, and AID-seq, which had a lower median off-target nomination frequency following consolidation due to contribution of sites nominated originally from other methods random selections (Supplementary Fig. 7). This indicates that sites interrogated in lower Severity Bins for these in vitro methods were enriched with sites predicted with more precise methods, which may ultimately inflate precision calculations for these methods (Fig. 6; Supplementary Fig. 7; Supplementary Fig. 8). Each nomination method had off-targets tested across all Severity Bins that the respective nomination technology produced targets within, with between 1 to 92 targets tested per nomination Severity Bin per method (Supplementary Fig. 8A, B).Fig. 6: Confirmation of editing at off-targets nominated by variable off-target nomination technologies in HEK293-Cas9.Full size imageA To compare the analytical sensitivity and precision of other published accounts of nomination technologies to UNCOVERseq off-target sites were selected across the dynamic range of each nomination technology for two gRNA with the total number of off-targets selected from each nomination methodology quantified for both EMX1 and B FANCF. C Confirmation of nominated off-targets was performed in HEK293–Cas9 (n = 3; paired treatment and control) to ensure true-positives were more easily detected. Editing quantification at assays with >1000× coverage. Each dot represents a gRNA on/off-target with the average raw frequency of indels of the control (x-axis) and treatment (y-axis) plotted. Blue dots indicate sites with no statistical significance while orange dots indicate significant sites (p adj1000×) ranged from 92 to 100% per panel, with a range of 0–15 targets not being at sufficient coverage per panel (Supplementary Fig. 12B).To determine the frequency of UNCOVERseq HEK293-Cas9 nominations that convert to empirically edited sites in variable DSB editing contexts, we compared this frequency for S.p. Cas9 both in HEK293-Cas9 and HSPCs (HiFi Cas9). For HEK293-Cas9, a range of 54.5% of nominated off-targets all the way to 100% of off-targets had confirmed editing ranging from to 0.02 to 95% indel editing, demonstrating the true positive rate for UNCOVERseq-nominated sites remains high even with only a single replicate in the appropriately paired confirmation context (Fig. 7F; Supplementary Fig. 13A–F). For HSPCs, a range of 2.4–34.5% of nominated targets were successfully confirmed per gRNA, with confirmed indel editing ranging from 0.06 to 88% (Fig. 7F; Supplementary Fig. 14A–F). Nomination: confirmation frequencies trended to increase as gRNA specificity decreased, suggesting that the method is still successfully nominating relevant off-targets, but that these sites likely no longer exceed detectable frequencies or are no longer edited in the higher genome editing specificity context of HSPCs delivered a HiFi Cas9 mRNA editor (Fig. 7F). Furthermore, at higher gRNA specificities the only nominated target being confirmed is the on-target site for HiFi Cas9 in HSPCs (Supplementary Fig. 14A–F).Off-targets that were confirmed were compared to the list of those that would have been dropped given a different previously evaluated alignment method (Regex method; Fig. 1). A range of 3–23 bona fide off-targets per gRNA in HEK293-Cas9 were successfully nominated using the UNCOVERseq alignment method that were missed using GUIDE-seq analysis Regex method, with a range of observed indel editing from 0.02 to 68% (Supplementary Fig. 15A). In HSPCs, 1 bona fide HiFi Cas9 off-target was identified with a frequency of 0.01% with our alignment method that was missed using the Regex alignment method (Supplementary Fig. 15B). This demonstrates that DSB off-target sites that were nominated due to differences in alignment criteria can result in bona fide off-target indel editing in both HEK293-Cas9 and HSPCs, further emphasizing the importance of the improvements implemented in UNCOVERseq.Comparative analysis of non-DSB editors in HSPCs (off-target)Off-targets were simultaneously confirmed for both indel and base editing in the non-DSB treatments for HSPCs (ABE and CBE). Similarly, we interrogated the frequency that UNCOVERseq HEK293-Cas9 nominations convert to empirically edited sites in HSPCs being delivered a base editor. For ABE treatments, a range of 2.4–29% of nominated targets had confirmed editing ranging from 0.53 to 75.9% cumulative ABE editing (Fig. 7G–H; Supplementary Fig. 16A–F). For CBE treatments, a range of 2.4% of nominated targets to 14.5% of targets had confirmed editing ranging from 0.51 to 32.9% CBE editing (Fig. 7G; Fig. 7I; Supplementary Fig. 17A–F). Significant indel editing was observed for all ABE and CBE editing treatments, with largely only the on-target gRNA containing indels at higher specificity gRNAs (Fig. 7C; Supplementary Fig. 18A–F; Supplementary Fig. 19A–F). Off-target indel frequencies for ABE treatments ranged from 0.02 to 0.88% indels across different gRNAs (Supplementary Fig. 18A–F). Interestingly, three off-target sites were found to generate indel events at the higher specificity TRAC7 gRNA under ABE treatment conditions, which lacked any significant off-targets in paired DSB Cas9 treatment (Supplementary Fig. 18D). Off-target indel frequencies for CBE treatments ranged from 0.08 to 0.66% indels across different gRNAs (Supplementary Fig. 19A–F). Generally, it was observed that indel and base editing frequencies were lower in CBE-treated samples in comparison to ABE-treated samples (Fig. 7I, J; Supplementary Fig. 16A–F; Supplementary Fig. 17A–F), although this could be a result of lower overall activity of the base editor instead of off-target propensity.When comparing the list of confirmed ABE/CBE off-targets to those that would have been excluded given a different alignment method during nomination we found 1 bona fide off-target of the PDCD1 gRNA that was identified for both ABE and CBE treatments with a frequency range of 0.5–3.1% base editing that was missed using the Regex alignment method (Supplementary Fig. 15B). This demonstrates that off-target sites that were nominated due to differences in alignment criteria can also result in bona fide off-target base editing activity for both ABE and CBE editors in HSPCs.To investigate relationships between DSB indels, SSB indels, and base editing, we binned confirmed base editing off-targets based on their presence of indels in either DSB or SSB systems. Base editing with the highest frequencies (median 20.6% and 2.3% for ABE and CBE, respectively), were found to coincide with indel editing for both DSB and SSB systems (Fig. 7H). Interestingly, only ABE treatments were found to have an increased frequency of base editing at SSB only sites, with eight detected SSB-only off-targets with a median 13.1% cumulative base editing compared to zero sites for CBE (Fig. 7). This may coincide to activity differences, as both ABE editing/indel activity was generally higher than CBE editing/indel activity across the different sites (Supplementary Figs. 16−19). DSB-only and sites with no evidence of significant indel editing were present in confirmed sites for both ABE and CBE treatments, albeit with lower median cumulative base editing frequencies (Fig. 7). The on-target indel and base editing activity of the different gRNAs were rank-order correlated (r = 0.66–0.89), suggesting that indel editing frequencies may be predictive of base editing frequencies (Supplementary Fig. 20). Similarly, off-target DSB indel editing frequencies from HiFi Cas9 demonstrated rank-order correlation with off-target base editing frequencies (r = 0.77–0.78) at sites that had significant DSB indel editing frequencies in HSPCs (Fig. 7J). This provides evidence that DSB editing may be indicative of base editing activity, meaning that DSB-nominated sites are meaningful for interrogation in the context of both indel and base editing off-target assessment for both ABE and CBE modalities.Comparative translocation analysis and overall editing burden across editing modalitiesTo investigate differential frequencies of editor modalities to generate large structural variants (>0.1% frequencies) in HSPCs, we investigated the previously described six sites for on-target:off-target and off-target:off-target translocations using amplicon sequencing. Of the six interrogated gRNAs, only the PDCD1 gRNA had detectable translocations, with two out of three of the translocations being shared between the S.p. Cas9 and ABE conditions (Supplementary Fig. 21). Shared translocations included a fusion of OTE132 to OTE94 and OTE160 to OTE158, with comparable average frequencies ranging between 1.0–1.7% and 0.3–0.5%, respectively (Supplementary Fig. 21). The on-target site was not a translocation partner in any significant events, which is likely due to the low frequency of on-target indels for the PDCD1 gRNA relative to off-target indel activity (Fig. 7B, C; Supplementary Fig. 14A; Supplementary Fig. 18A). The overall estimated translocation burden for this gRNA was estimated to be an average of 1.4% translocations for S.p. Cas9 and an average of 2.2% translocations for ABE conditions (Supplementary Fig. 21). This suggests that translocations are either below 0.1% or not occurring in healthy HSPC donors across higher gRNA specificities. However, it is noteworthy that they are still occurring for both SSB and DSB modalities, as has been previously observed39.When calculating the normalized risk of cumulative off-target frequencies (indels, base edits, and translocations) across editor modalities throughout our six gRNAs displaying a range of specificities, off-target ratios were observed for the PDCD1 gRNA (specificity ratio = 0.001) over a range 9.4–89.0 off-target events per 1 on-target event (Supplementary Fig. 21). Off-target ratios for the CYP2C18 gRNA (specificity ratio = 0.290) ranged from 0.07 to 0.76 off-target events per 1 on-target event (Supplementary Fig. 21). Trends consistently showed that overall off-target burden of high-fidelity DSB editors was actually decreased in comparison to SSB base editors for the cumulative frequency of event types monitored using this strategy (Supplementary Fig. 21). Even though the B2M gRNA was considered higher-specificity (specificity ratio = 0.785), a single significant ABE off-target was observed for this treatment contributing to a higher ratio (Supplementary Fig. 21). This highlights that even higher-specificity gRNAs may generate observable off-targets in clinically relevant cell types, and that the off-target burden of high-fidelity DSB enzymes can actually be lower than that of current generation SSB base editing modalities in translational contexts.DiscussionOur study presents an improved, end-to-end characterized in cellulo method for the nomination of off-target sites in CRISPR experiments that we collectively refer to as UNCOVERseq (v1.0). This method leverages several technological improvements to collectively streamline the in cellulo nomination process, improve NGS data quality, and increase the number of high-confidence nominated sites through computational analysis improvements compared to other previously published methods. By demonstrating recommended operational conditions that can allow the experiments to be performed independent of cell context with controls grounded in empirical data, we provide a framework to ensure translation to different treatment modalities with quantifiable levels of performance from experiment to experiment. Furthermore, we demonstrate the workflow is capable of nominating relevant unique and shared off-targets for both DSB-based and SSB-based CRISPR editing systems and demonstrate correlations between DSB formation and the frequency of a site to be edited by ABE/CBE editors. Finally, we perform a thorough assessment of off-targets nominated by different methods to ground both UNCOVERseq and other methods with measured analytical sensitivity and precision to demonstrate that UNCOVERseq outperforms other available nomination methods.To ensure all relevant off-targets are assessed, high analytical sensitivity is a critical off-target nomination metric. However, accurate calculations of false-negative rates from nomination methods have been challenged by technical difficulties in obtaining an empirically defined gold-standard of all true-positive off-targets. In this work, we create a starting point for a gold-standard dataset for measuring sensitivity by selecting common gRNAs to multiple nomination methods and performing multiplexed ultra-deep NGS sequencing to confirm off-targets nominated across diverse methods in a promiscuous editing system. This leads us to find that UNCOVERseq is the most sensitive gene editing off-target nomination method currently available (>95% sensitivity) and demonstrate that many other existing methods have sensitivity gaps leading to missed nomination of ~20% to ~80% of off-targets that can exhibit significant editing in cells. Previous work has led to a perception that in vitro biochemical methods are inherently more sensitive than in cellulo methods as evidenced by: (1) true positive sites captured by in vitro methods like CHANGE-seq that are missed with GUIDE-seq and (2) multiple accounts of in cellulo methods being largely a subset of in vitro results6,7. While we demonstrate that previously published accounts of in cellulo methods were not as sensitive as UNCOVERseq, it is not clear whether this is due to insufficient operational conditions (read depth, library complexity, etc.) to maximize capabilities of many assays as opposed to the technical improvements that confer enhanced nomination capabilities to UNCOVERseq. To reduce risk of false negatives, we demonstrate a generalizable strategy for in cellulo off-target nomination where high gDNA input and promiscuous editing conditions are used to greatly amplify nomination signal to reproducibly detect sub-0.05% editing events in UNCOVERseq data while still retaining sites derived from higher fidelity modalities and primary cell lines. This finding is in agreement with previous comparative work of in cellulo tag-based methods in HSPCs20, meaning it likely extends to additional cell types as well. We also find that in vitro biochemical assays do not currently sensitively cover the full range of true positive off-target sites as evidenced by (1) missed nomination of ~20–60% of significantly edited EMX1 and FANCF off-targets and (2) missed sites that were generated from highly reproduced UNCOVERseq nomination. A likely explanation is that in vitro systems may lack or modify key features of biological systems that have been shown or hypothesized to have impacts on genome editing, including genomic accessibility15, local DNA topology40,41, macromolecular crowding42, and non-physiological buffer systems43,44. Together, these factors in addition to our findings indicate that in vitro assays may currently underestimate certain biologically relevant off‑target events due to sub-optimal operating conditions, reinforcing that assay sensitivity is not an intrinsic property of assay class. Overall, our findings ground the sensitivity of multiple nomination methods in empirical data and highlight the importance of establishing empirically defined, modality‑appropriate conditions for each platform. Future work should focus on expanding the size of a gold-standard dataset with paired nomination and confirmation, in addition to further quantifying and decreasing the limit of detection for confirming off-target editing from targeted sequencing to limit potential false negatives.Precision is another important metric for off-target nomination methods to appropriately select sites for downstream confirmation. Some off-target nomination methods produce extensive, low‑precision site lists that are costly to confirm, and provide limited actionable correlation with cellular editing outcomes, which complicates both regulatory evaluation and the development of effective risk mitigation strategies. Using UNCOVERseq with sufficient replication, we show that nominated sites can achieve very high precision, enabling highly sensitive nomination while producing approximately an order of magnitude less nominated sites than low precision in vitro methods. We also find that sites nominated by UNCOVERseq have nomination frequencies that strongly correlate with observed editing frequencies—an association that in vitro biochemical methods generally fail to match when compared with in cellulo and in situ approaches. This type of actionable correlation enables acceleration of translational programs with integrated, risk‑based decision making that incorporates both the likelihood of editing and its potential biological impact. As a result, UNCOVERseq substantially improves method‑specific site prioritization, helping to filter or rank large off‑target lists based on biologically relevant metrics. The importance of this step of the process has already been realized and exemplified in the de-risking of translational in vivo correctional gene therapies, further demonstrating clinical impact45. Thus, by providing a more accurate and biologically relevant map of off‑target activity, UNCOVERseq enables therapeutic developers to better quantify and characterize the risk profiles of candidate guides and editing systems. Additional work should further explore whether the output of UNCOVERseq can be used to better refine the sensitivity and precision of faster turn-around time in silico nomination methods to further bolster the translational process in an orthogonal manner.Using a simple prioritization method based on frequency, replication, and high-level indicators of risk (exonic regions vs. intergenic, etc.), we demonstrate UNCOVERseq nominations with recommended experimental structures can result in manageable sequencing panel sizes for downstream confirmation (e.g., 25% of putative targets being missed without use of biological replicates. However, to better understand risk after off-target nomination and confirmation across methods, a more standardized scoring system to prioritize off-targets is needed in the future. The fields of oncology and heritable diseases have encountered similar issues and derived guidelines including tiered scoring systems from the American College of Medical Genetics (ACMG) and Association of Molecular Pathology (AMP) and modifications leveraging these criteria21,46,47. Gene editing may be able to leverage some of these learnings but will face unique challenges in categorization of off-target risk since even off-targets in intergenic space during the nomination phase can be at risk for known structural variations derived from DSBs and SSBs. This includes events such as translocations48, loss of heterozygosity49 (LoH), aneuploidy50, and other large variants like multi-kilobase deletions51. Given this, it seems likely that probability, frequency, and even proximity to other coding regions will be important criteria for triaging off-targets for assessment. While our work shows that UNCOVERseq can provide predictive values for probability and frequency of off-target events, future work should target determining generalized annotation prioritization and risk-tiering scoring schemas that can be broadly applied across technologies.Benchmarked nomination and confirmation methodologies are needed for both DSB and SSB-based editing modalities. By selecting a variable range of gRNA specificities, we demonstrate that even in popular ex vivo models like HSPCs with mRNA delivery, high specificity gRNAs are still sensitive to both SSB indels and base editing off-target effects at frequencies >0.01% and 0.5%, respectively. Furthermore, we demonstrate that indel editing and base editing are rank-order correlated across 34 base editing on/off-targets, supporting the idea that DSB-based nomination methods are effective tools for nominating both indel and base editing activity. Base editing-specific nomination methods, such as CHANGE-seq-BE, have been developed to target base editing events, while demonstrating unique off-target confirmation findings in comparison to published methods52. It is our belief that orthogonal methods for nominating both base editing events and indel events are likely needed for the foreseeable future until more complete cross-modality benchmarking can be done, especially given some of our findings that some UNCOVERseq-nominated sites generate confirmable indels only in conditions using the SSB base editing modalities in translational cellular contexts.We envision UNCOVERseq coupled with promiscuous conditions being a powerful tool to help sensitively identify CRISPR-Cas off-targets for interrogation during pre-clinical development phases. By enhancing the quality of off-target nomination and grounding NGS operational conditions in empirical data, we believe that UNCOVERseq improves informed risk assessment of gene editing in translational systems. In a clinical development pipeline, this improved off‑target resolution facilitates more rational guide RNA selection, more targeted mitigation approaches (e.g., guide redesign, high‑fidelity nucleases, delivery modulation), and more efficient toxicology study design. Ultimately, the ability to characterize off‑target effects with high precision strengthens the regulatory framework for genome editing therapeutics, accelerates candidate advancement, and reduces uncertainty in long‑term safety projections. As gene therapy continues to shift toward in vivo editing and systemic delivery modalities, the need for such robust, context‑relevant off‑target analysis will only grow—positioning UNCOVERseq as a powerful enabling technology in the path toward safe, effective, and durable genomic medicines.MethodsHuman cell culture and transfection (K562 and HEK293-Cas9)K562 (ATCC; CCL‑243) and HEK293-Cas9 (ATCC; CRL‑1573) cells were cultured in Iscove’s Modified Dulbecco’s Medium (IMDM; ATCC) and Eagle’s Minimum Essential Medium (EMEM; ATCC) supplemented with 10% FBS at 37 °C with 5% CO2. Ribonucleoprotein (RNP) complexes were formed by mixing Alt-RTMS.p. Cas9 Nuclease V3 (IDT) or SpyFiTM Cas9 Nuclease (Aldevron) and Alt-R CRISPR-Cas9 sgRNA (gRNA, IDT) and incubating for 20 min at room temperature (Molar Ratio: 1:1.2, Cas9:gRNA). For each transfection, 8.0 × 105 cells were washed with 1X phosphate-buffered saline, resuspended in 20 µL of solution SF (Lonza). For K562 cells, RNP complexes at 4 µM were combined with 4 µM of the dsODN (Supplementary Data 1) into the SF solution, while for the HEK293-Cas9 cells, 5 µM gRNA and 0.5 µM dsODN were added to the SF solution. This mixture was transferred into 1 well of a 96-well Nucleocuvette plate (Lonza) and electroporated using program FF-120 (K562) or DS-150 (HEK293-Cas9). Two nucleofections per replicate were performed and each treatment done in triplicate. Following electroporation, cells were transferred to a 6-well plate preheated with either IMDM or EMEM and were incubated at 37 °C with 5% CO2 for 72 h. After incubation, genomic DNA (gDNA) was extracted using either the PurelinkTM Pro 96 Genomic DNA Purification kit or the MonarchTM Spin gDNA Extraction Kit (New England Biolabs) according to the manufacturer’s instructions, eluted in low-EDTA TE buffer (IDT, 11-05-01-05), and quantified using a NanoDrop 8000 UV-Vis Spectrophotometer (ND-8000-GL).Primary T-cell culture and transfectionFrozen human primary pan-T cells (STEMCELL Technologies; 70024) from 2 unique human donors were thawed in ImmunoCult-XF T Cell Expansion Medium including 300IU IL-2 (Cytiva) and activated with 10 μL/mL TransAct, human, T cell activator (Miltenyi Biotec) for 48 h. To prepare for transfection using Lonza 96-well plate 4-D Nucleofector system, cells were counted, pelleted using centrifugation (300 × g, 10 min at room temperature), and washed gently with 10 mL 1X phosphate-buffered saline. Cells were again pelleted and resuspended in Lonza Nucleofection Solution P3 at 2.5 × 106 cells/mL. For each electroporation, 5 μL of RNP complex and 3 μL dsODN was added to 20 µL of cells in P3 (5 × 105 cells/nucleofection) for a final concentration of 4 μM RNP (1:1.2 ratio of Cas9 to gRNA) and 1–4 μM dsODN. Where tag was not included, 3 μL of IDT Alt-R Cas9 Electroporation Enhancer was added for 3 µM final concentration to achieve a fixed final nucleofection reaction volume of 28 μL. Each reaction was mixed by pipetting and 25 µL was transferred to an electroporation cuvette plate. The cells were electroporated according to the manufacturer’s protocol using the Amaxa 96-well Shuttle and nucleofection protocol 96-EH-140. After electroporation, the cells were resuspended in 75 μL pre-warmed IL-2 culture media in the electroporation cuvette. Triplicate aliquots of 25 μL of recovered cells were further cultured in 175 μL pre-warmed IL-2 media with TransAct. Cells were incubated for 72 h, after which gDNA was isolated and quantified.iPSC culture and transfectioniPSCs from fibroblasts (Coriell Institute, GM23338) were cultured in mTeSR™ Plus media (Stemcell Technologies) at 37 °C with 5% CO2. RNPs were formed as described above. For transfection using Lonza 96-well plate 4-D Nucleofector system, cells were detached using ReLeSR™ (Stemcell Technologies) and washed with 1X phosphate-buffered saline. Cells were resuspended in P3 buffer at 2 × 105 cells/nucleofection. CRISPR reagents at required final concentrations (4 µM RNP; 0.5 µM dsODN) were added to the mix to make a final volume of 25 µL, and of which 20 µL was transferred to the nucleocuvette for electroporation. The nucleovette plate was electroporated using code CA-137. After the nucleofection, cells were recovered and plated in complete mTeSR Plus medium with 1X CloneR™ 2 supplement (Stemcell Technologies). Recovery media was added to the electroporated cells to achieve a final volume of 100 µL, and 25 µL of this was added to 175 µL media per replicate well for final plating in a vitronectin-coated 96-well plate. During recovery and growth at 37 °C with 5% CO2 for up to 96–120 h, media changes were performed as desired and/or following manufacturer’s protocols for media and CloneR 2 supplement. gDNA extraction and quantification was performed as described above.Off-target nomination with UNCOVERseq500 ng of purified gDNA was enzymatically fragmented and adapter-ligated using the xGenTM DNA Library Prep EZ UNI kit along with the xGen Deceleration Module (IDT, xGen DNA Library Prep EZ UNI 96 rxn, 10009822; xGen Deceleration Module 96 rxn, 10009823) according to the manufacturer’s instructions and cleaned with AMPure XP beads (Beckman). Following fragmentation and adapter ligation, rhPCR was performed using rhAmpSeqTM Library Mix 1 (IDT) to amplify the DNA in a single tube using a forward primer specific to the P5 adapter, a reverse primer specific for top and bottom strand of the integrated dsODN tag, and an adapter-blocking oligo corresponding to each strand of the dsODN31. Following PCR, samples were diluted 1:40 with nuclease-free water and used in a second PCR with rhAmpSeqTM Library Mix 2 (IDT) that added a unique P7 adapter to each library. Libraries were then cleaned with AMPure XP beads and run on an Agilent Fragment Analyzer for library quality assessment. All libraries were quantified with the Qubit 1X dsDNA HS Assay kit (Invitrogen) and pooled in equimolar amounts. All libraries were sequenced using an Illumina MiSeq or NextSeq2000 instrument with 150-bp paired-end reads. All oligonucleotides used in preparation of UNCOVERseq libraries are listed in Supplementary Data 1. All oligonucleotides used in this study (Supplementary Data 1) were synthesized by Integrated DNA Technologies.Vector construct and in vitro transcription of modified mRNAGenes encoding HiFi Cas953, ABE8e37, AncBE4max36, and PE228 were each cloned into a plasmid vector with a dT7 promotor followed by a 5’UTR, Kozak sequence, ORF, and 3’UTR for subsequent in vitro transcription. Chemically modified Cas9, ABE8e, AncBE4max and PE2 mRNA was transcribed in vitro from the PCR templates with full substitution of uridine by N1-methylpseudouridine. Co-transcriptional capping was achieved using CleanCap AG analog (TriLink), yielding a 5′ Cap 1 structure. Transcription reactions were carried out with the HiScribe T7 High Yield RNA Synthesis Kit (New England Biolabs) in 0.5× transcription buffer supplemented with 4 mM CleanCap AG. Following synthesis, mRNAs were purified using the Monarch Spin RNA Cleanup Kit (New England Biolabs). PCR templates incorporated mammalian-optimized UTRs (TriLink) and a 120-nucleotide poly(A) tail.HSPC culture and transfectionCD34⁺ HSPCs from a single human donor (Fred Hutch Cancer Center, RO04089) were cultured at 1 × 10⁵ cells/mL in StemSpan SFEM II medium supplemented with 100 ng/mL SCF, 100 ng/mL TPO, 100 ng/mL FLT3L, 100 ng/mL IL-6, 20 µg/mL streptomycin, and 20 U/mL penicillin. Cultures were maintained at 37 °C in a humidified incubator with 5% CO₂. For electroporation, Cas9 mRNA and gRNA were mixed at a 1:1 weight ratio (3 µg each per reaction). ABE, CBE, and PE2 mRNAs were used at equimolar amounts to Cas9 (3.4 µg ABE, 4 µg CBE, and 6 µg PE2), and pegRNA was added at the same molar amount as gRNA (4.2 µg). HSPCs were resuspended in 20 µL Lonza P3 buffer and electroporated using a Lonza 4D-Nucleofector (program DZ−100). After electroporation, cells were plated at 1 × 10⁵ cells/mL in HSPC medium.Computational analysis—nominationFollowing NGS, Illumina adapters and UMIs were identified and annotated using Picard MarkIlluminaAdapters. Tag sequences were identified and trimmed using Cutadapt v4.254. Sequencing reads were aligned to hg38 (GRCh38.p12) reference genome using BWA mem55 v0.7.15 and UMI consensus reads were generated based on consensus from a single-strand (minimum UMI consensus size = 1) using fgbio v0.7.0 (https://github.com/fulcrumgenomics/fgbio). Nomination of candidate off-target sites began by using mapped UMI consensus reads to create a flanked search space (±40 bp) to perform alignment between the guide and empirical target region using a glocal implementation (+2/−1/−10/−1 match/mismatch/gap open/gap extension; penalize_end_gaps = False) of the Needleman-Wunsch alignment56,57. After a candidate match to the gRNA spacer region was identified in the sequencing data, nominated off-target sites were identified using a hypergeometric test with multiple testing correction (Benjamini and Hochberg; FDR T (CBE) transversions in the relevant base-editing window.For identifying indels, the window for event quantification was centered on the canonical cut site and events quantified utilizing the default window size for Cas9 (8 bp). To determine whether indels found in the sequencing data could result from bona fide off-target cleavage, indels were grouped by location relative to the cut site (prioritizing minimum distance to cut site) followed by fitting counts of events to a negative binominal model with a Wald test for significance in each location bin per off-target using the DESeq2 package35 within IDT’s OTEasy tool (Schmaljohn et al., Manuscript in Preparation). For classification of indel off-target editing, the tool requires: (1) sufficient read coverage for the site (>1000×) in all replicates, (2) significant edits to occur at or adjacent to the cut site after optimal alignment, (3) the classified cumulative significant edits to exceed 0.01%, (4) the comparison of treatment/control samples at the site to have a significant adjusted p-value (p 10 bp in size from the CRISPAltRations output to the expected dsDNA tag using the biopython implementation of the Needleman-Wunsch aligner56. Alignments with an alignment score greater than 50 were quantified as a tag integration event, and events were considered imperfect if any base was mutated or missing from the expected dsDNA sequence alignment.For identifying base editing-generated off-target effects, the window for event quantification was centered in the middle of canonical base-editing window between position +5/+6 of the spacer (5’ to 3’) with a 5 bp window for quantification. To determine significant base-editing transitions resulting in off-target editing, all individual events that contained an ABE (A>G or T>C) or CBE (C>T or G>A) transition were grouped according to unique base editing events in the window and fitting counts of events to a negative binominal model with a Wald test for significance in each location bin per off-target using the DESeq2 package60 within IDT’s OTEasy tool (Schmaljohn et al., Manuscript in Preparation). For classification of adenine base editing at off-targets, the tool requires: (1) sufficient read coverage for the site (>1000×) in all replicates, (2) the classified cumulative significant edits to exceed 0.5%, (3) the comparison of treatment/control samples at the site to have a significant adjusted p-value (p