Gene-editing nucleases, such as zinc finger nucleases (ZFNs), transcription activator-like effector nucleases (TALENs) and clustered regularly interspaced short palindromic repeats (CRISPR)/CRISPR-associated (Cas) nucleases, can be used to create targeted sequence alterations with high efficiencies in numerous cell types and organisms (1–7). Repair of nuclease-induced double-stranded breaks can be exploited to introduce either insertion/deletion (indel) mutations via non-homologous end-joining (NHEJ) or specific sequence alterations from a donor template via homology-directed repair (1,2). Comprehensive delineation of unintended off-target mutations is important for customized nucleases in many biological applications and will be essential for developing therapeutic strategies based on these proteins.

Two different methods have recently been described for characterizing the genome-wide specificities of ZFNs (8,9), but neither study comprehensively identified the full-spectrum of possible off-target mutations. One method, previously developed by Liu and colleagues, used an in vitro cleavage site selection assay to identify sequences from a large partially degenerate library (based on the intended target DNA site) that can be cleaved by ZFNs. In vitro selections with a CCR5-targeted ZFN pair identified 36 potential off-target cleavage sites that occur in the sequence of the human genome; analysis of these sites in human cells in which CCR5-targeted ZFNs had been expressed revealed nine bona fide off-target sites (8). Another approach, described by von Kalle and colleagues, exploited the incorporation of integrase-deficient lentivirus (IDLV) DNAs into nuclease-induced double-stranded breaks to map ZFN cleavage sites in human cells (9). Application of this approach to the same CCR5-targeted ZFNs characterized with the in vitro selection approach identified four off-target genomic sites. However, the substantial lack of overlap between off-target sites identified in these two studies (only one site was common to both sets) strongly suggested that neither identified all possible off-target sites. In addition, these results also suggested that a broader range of potential off-target sites might exist beyond the sets identified by these two methods.

Here, we show that in silico abstraction of ZFN cleavage profiles generated by the selection method of Liu and colleagues provides an improved approach to screen the human genome for potential ZFN off-target sites. This enhanced strategy identifies both previously described as well as dozens of additional off-target sites for a ZFN pair targeted to CCR5 gene. We also show that this improved method works effectively for another ZFN pair targeted to the VEGFA gene. Our results demonstrate that the potential landscape of off-target mutagenesis effects for ZFNs may be broader than delineated in previous studies.

We sought to improve the original strategy of Liu and colleagues by addressing its inability to interrogate cleavage site libraries in vitro to a depth sufficient to identify all possible off-target sites present in the human genome. To do this, we added a machine-learning-based step that uses cleavage site preferences from the in vitro selection experiments to predict what sequences in the human genome are most likely to be cleaved (Figure 1). We used standard machine-learning techniques to construct Naïve Bayes classifiers that quantify how the nucleotide identity at each position within a DNA site differs between members of a partially degenerate library that were cleaved efficiently in vitro and those that were not (‘Materials and Methods’ section). The scores generated by each classifier range from 0 to 1, with lower scores representing a higher probability that any given site will be cleaved (‘Materials and Methods’ section). Figure 1.

Schematic illustrating the original method by Pattanayak et al. (8) (blue arrows) and the enhanced approach that incorporates addition of a classifier-based step (green arrows).

Our results show in silico abstraction of in vitro cleavage data provides a strategy that more broadly identifies all potential off-target sites of ZFN activity in human cells than previously described methods. Our classifier-based approach not only successfully re-identified all previously known off-target sites for two different ZFNs but also enabled the identification of many additional novel off-target sites, including some that differ from the target sequence by as many as 8 (of 24) or 6 (of 18) bp for the CCR5- or VEGFA-targeted ZFNs, respectively (Tables 3 and 4) (8–10). Because sequences harboring these numbers of mismatches will occur frequently in the human genome, identifying these off-targets would have been previously intractable by simple mismatch counting approaches; indeed, using such strategies would require screening hundreds of thousands of potential sites (Supplementary Tables S7 and S8). Table 3.

In vivo validated off-targets for ZFNs targeted to CCR5

Importantly, because we only assessed a small sampling of the top scoring potential off-target sites in cells, we believe that the full range of potential off-target sites for the two ZFN pairs we examined is likely more expansive than just those identified in this study. This expectation is supported by another experimental screen (data not shown) that identified six additional bona fide off-target sites for the CCR5-targeted ZFNs (Supplementary Table S9 and Supplementary Discussion). Collectively, these results clearly demonstrate that ZFN off-target sites may occur at low rates much more widely on a genome-wide scale than suggested by data from previously described reports.

Although our data clearly demonstrate that sites with low classifier scores are highly enriched for bona fide off-target sites, our results also show that bona fide off-targets are present (albeit at a much lower frequency) among loci with higher classifier scores. This suggests that comprehensive identification of off-target sites will require interrogation of a large number of loci by deep sequencing. We expect that decreases in the price per base and increases in the number of bases that can be sequenced should increase the number of potential sites with low classifier scores that can be examined, thereby enabling the identification of a greater number of bona fide off-target sites. However, until such reductions in sequencing costs become reality, an alternative approach might be to look at off-targets with the best scores or to pre-screen off-targets bioinformatically for sites that fall in regions of high priority such as promoters, exons and non-coding RNAs.

We note that the number of off-target sites identified by our approach may be larger or smaller depending on the cell type examined as well as the level and duration of ZFN expression. Not all of the sites with low classifier scores we examined showed evidence of mutagenesis. Potential reasons for this might include DNA methylation of the target site or chromatin status of the gene. These parameters will be cell-type specific and would not be accounted for by in vitro selections or in silico classifiers. As large-scale efforts such as ENCODE and the NIH Roadmap Epigenomics Mapping Project define these variables in multiple different cell types, it may be possible to use such information to prioritize sites with low classifier scores and thereby to increase the yield of bona fide off-target sites identified by deep sequencing. In addition, we expressed ZFNs from a strong constitutive CMV promoter using transiently transfected plasmids and harvested genomic DNA from cells 5 days post-transfection. Lower levels and shorter durations of ZFN expression might be expected to induce fewer off-target mutations, whereas higher levels and longer durations might induce an even greater number of such mutations.

More broadly, the combined strategy of using in vitro cleavage site selection data together with machine-learning-based classifiers might also be extended to specificity information from other sources (e.g.–SELEX or bacterial selection) and to define the specificities of nucleases built on other platforms (e.g.—TALENs or CRISPR-Cas RNA-guided nucleases). The use of machine learning to improve the predictive power of data derived from in vitro selection experiments could be particularly useful for ZFNs composed of greater numbers of fingers in each monomer and for TALENs. These nucleases target longer sites, making it challenging to adequately sample all potential off-targets even in an in vitro system. Continuing to better define off-target effects of targeted nucleases will provide important information to guide refinement of the genome-wide specificities of these reagents. These improvements will be critically important, as these targeted nucleases are more widely applied for both research and therapeutic approaches.

Supplementary Data are available at NAR Online.

National Institutes of Health (NIH) Director’s Pioneer Award [DP1 GM105378], [NIH P50 HG005550] and Defense Advanced Research Projects Agency (DARPA) [W911NF-11-2-0056 to J.K.J.]; The Jim and Ann Orr Massachusetts General Hospital (MGH) Research Scholar Award (to J.K.J.); [NIH T32 CA009216 to J.D.S.]; National Science Foundation Graduate Research Fellowship and a Ford Foundation Predoctoral Fellowship (to C.L.R.); National Science Foundation (NSF) ward [DBI-0923827 to D.R.]; [DARPA HR0011-11-2-0003, DARPA N66001-12-C-4207] and the Howard Hughes Medical Institute (to V.P. and D.R.L.); Award number [T32GM007753 to V.P.] from the National Institute of General Medical Sciences. National Human Genome Research Institute Grant [U54 HG004570 to N.S., M.K., B.E.B.]. Funding for open access charge: NIH [DP1 GM105378].

Conflict of interest statement: J.K.J. has a financial interest in Transposagen Biopharmaceuticals. J.K.J.’s interests were reviewed and are managed by Massachusetts General Hospital and Partners HealthCare in accordance with their conflict of interest policies.