Introduction
CRISPR-Cas technologies have transformed genome-editing of experimental organisms and have immense therapeutic potential. Despite significant advances in our understanding of the CRISPR-Cas9 system, concerns remain over the potential for off-target effects. Recent studies have addressed these concerns using whole-genome sequencing (WGS) of gene-edited embryos or animals to search for de novo mutations (DNMs), which may represent candidate changes induced by poor editing fidelity[1–⇓3]. Critically, these studies used strain-matched but not pedigree-matched controls and thus were unable to reliably distinguish generational or colony-related differences from true DNMs. Here we used a trio design and whole genome sequenced 8 parents and 19 embryos, where 10 of the embryos were mutagenised with well-characterised gRNAs targeting the coat colour Tyrosinase (Tyr) locus.
Results and Discussion
To perform our analysis, we chose two gRNAs targeting exon 2 of Tyr, Tyr2F and Tyr2R, that had typical off-target scores similar to that of the sgRNA used by Schaefer et al. [3](Supplementary Methods). Tyr is responsible for black coat colour and eye pigmentation in C57BL/6 mice[4], so its disruption should not be detrimental to embryonic development. The CRISPR-treated group was split to include five embryos treated with Tyr2F and five embryos treated with Tyr2R, while three untreated embryos from each of the three control groups (“Cas9 only”, “No injection” and “Sham injection”) were also collected (Fig. 1a). Microinjections were performed into the cytoplasm of 1-cell zygotes[5], which were then briefly cultured to assess viability and then transferred into 0.5 day post coital (d.p.c) pseudopregnant females. Embryos in the “Sham injection” group were microinjected with water only, and the “Cas9 only” embryos were microinjected with Cas9 protein only. All embryos were harvested at 12.5 d.p.c (Supplementary Table 1) and genomic DNA from both parents and embryos extracted.
(a) Experimental design: Four sets of C57BL/6N parents gave rise to 9 control embryos (3 “no injection”, 3 ‘‘sham injection” with water only and 3 “Cas9 only”), and 10 treated embryos (5 were injected with Cas9 and Tyr2F gRNA and 5 were injected with Cas9 and Tyr2R gRNA).
(b) Whole genome sequencing: All 27 mice / embryos were subjected to whole genome sequencing with median depth 39.5x and an average of 3.4% of bases with read depth less than 11x.
(c) Variant calling and filtering: starting from the joint variant call (bcftools mpileup + bcftools call), a sequence of filter steps were performed to detect only de novo mutations and remove likely false positives arising from low-level parental mosaicism and repeat regions. Parental-noise: alternate-allele reads present in either parent. Cross-noise: alternate reads from all other (non-parental) samples.
(d) Filtered SNV and Indel counts are not significantly different within control groups, within treatment groups, or between control and treatment groups.
Sequencing was performed on the Illumina X10 WGS platform yielding a median sequencing depth of 39.5x per genome (Fig. 1b, Supplementary Table 2). In parallel, targeted Illumina MiSeq sequencing to a mean depth of 10,800 reads of the Tyr target site was performed to comprehensively profile mosaicism, with these data analyzed using the CRISPResso software[6] (Supplementary Table 3). MiSeq analysis revealed a targeting efficiency of over 85% for Tyr2R and mosaicism, with a median of two variants per embryo (Supplementary Table 3). Importantly, to ensure our experiment was representative of the many thousands of CRISPR experiments performed worldwide, including those of the International Mouse Phenotyping Consortium, we compared these data to MiSeq data from 324 mice mutagenized using the Tyr2R gRNA, revealing good concordance of targeting efficiency and mosaicism (Supplementary Table 3).
We next performed variant calling on the WGS data using bcftools mpileup and bcftools call[7], configured to be sensitive to low allele-fraction indels (insertions/deletions)(Fig. 1c). Starting with a median of 324,561 variants per sample, we filtered the results to ensure adequate depth (10x) and variant quality (10) in all samples, resulting in a median of 225,671 variants per sample. Candidate DNMs (those not inherited from either parent) in the embryos were called using the TrioDeNovo software[8], which resulted in a median of 6,852 variants per embryo (489 SNVs / 6,450 indels) (Supplementary Table 4). Prior to further filtration of these candidate DNMs, we first looked for the presence of mutations within 10bp of any candidate CRISPR off-target site for Tyr2F or Tyr2R, as defined by the Cas-OFFinder software[9]; allowing up to 3 mismatches with a 1 nucleotide DNA/RNA bulge, and up to 4 mismatches without a DNA/RNA bulge. Importantly, we found no such coincident sites in any embryos from the CRISPR-untreated group and only on-target variants in the CRISPR-treated group (Supplementary Table 5), suggesting that if there are recurrent CRISPR-induced off-target alterations they are exceedingly rare. We next applied a validated filtration strategy to refine our candidate DNM calls[10], removing false positives arising from mosaic alleles in either parent, as well as those in proximity to repeats. Alignments for all SNVs and indel variants were then inspected visually for the presence of mosaic alleles (i.e. a second alternative allele at the same locus). In the same way, all indels were visually inspected to remove further false positives. This resulted in a median of 19 SNVs and 1 indel per embryo (Table 1), which is broadly consistent with prior work that aimed to define the de novo mutation rates in mice[10]. All variants were validated with targeted MiSeq sequencing to a depth of at least 10,000 reads per locus (Supplementary Table 8).
Summary table of initial variant counts, de novo variant counts, and filtered SNVs and Indels. Treatment Groups: Cas9 only (without gRNA), no injection (uninjected embryos), sham injection (water only), Tyr2R treated (Tyr2R gRNA + Cas9), Tyr2F treated (Tyr2F gRNA + Cas9). Variants passing basic depth / quality filters: bcftools joint call variant count per animal passing joint-depth and genotype quality filters (see Supplementary Methods). Candidate de novo mutations: all candidates produced by TrioDeNovo caller. Final SNVs / Final Indels: all SNVs/Indels remaining after filtering for false positives arising from low-level mosaicism, known C57BL/6NJ & C57BL/6N variants, proximity to UCSC repeat regions and further visual inspection.
A comparison of the expected variants at the Tyr locus detected by WGS and targeted MiSeq sequencing (Supplementary Table 6) shows that of the 20 indels detected by the MiSeq pipeline, 18 were also detected by the WGS pipeline, with the missing indels having low allele frequencies (7% and 7.5%, as defined by MiSeq sequencing). We are therefore confident that our WGS pipeline will detect genome-wide off-target damage with a range of allele fractions and mosaicism similar to on-target variants. Further, given our median depth (39.5x) and our minimum required de novo allele frequency (10%, Supplementary Methods), our power to detect a DNM occurring in the single-cell or two-cell stage of the zygote is at least 99%. Our power to detect a DNM with on-target variant allele fraction 0.17 (the median allele fraction seen in our experiments, Supplementary Table 6) is 85%.
Using the final counts of filtered SNVs and indels for each embryo (Table 1), we conducted a Kruskal-Wallis Rank test, detecting no significant difference in DNM counts between the “no injection”, “sham” and “cas9 only” untreated embryo groups (p=0.30 and p=0.37 for SNVs and indels, respectively). Similarly, a Wilcoxon Rank Sum test failed to detect significantly different SNV- or indel-DNM counts between the “Tyr2F” and “Tyr2R” CRISPR-treated groups (p=0.25 and p=0.43 for SNVs and indels, respectively). Based on these analyses (Fig. 1d), we combined variant calls from embryos in the two CRISPR-treated groups and in the same way combined data from the three untreated groups. Notably, using these data a Wilcoxon Rank Sum test failed to detect a significant difference in SNV or indel counts between the CRISPR-treated and untreated groups; p=0.30 and p=0.45, respectively (Supplementary Table 7).
Finally, we measured the impact of using unrelated parents on the false-positive DNM rate by deliberately choosing the parents of the Cas9-only embryos when analyzing all embryos in the study; the male parent (CBLT8902) was up to 7 generations removed from all other male parents (Supplementary Figure 1). Performing a comparable subset of filtrations and comparing variant counts by sample to the correctly analysed embryos at the same filtration point showed a median increase of 66 false variants per embryo (Supplementary Figure 1, Supplementary Table 9), highlighting the importance of using trios of mice when studying potential off-target rates.
(a) Network of heredity between the matings of mice from which mice parents were drawn. All female parents were drawn from mating CBLT9125 (red). Male parents were drawn from matings CBLT8712, CBLT8762 and CBLT8902 (dark blue) that is 7 generations distant.
(b) Scheme of the filtration pipelines for both the “correct” calling approach and the “incorrect parent” calling approach, generating variant counts that can be compared.
(c) Graphs of variants counts from “correct parent” pipeline and “incorrect parent” pipeline showing the effect of choosing distantly related parents on de novo variant calls: an average increase of 60 variants.
Conclusion
We conclude that if CRISPR mutagenesis were causing SNV or indel off-target mutations in treated embryos, then the number of these mutations is not statistically distinguishable from the background rate of DNMs occurring due to other processes. This work should support further efforts to develop CRISPR-Cas9 as a therapeutic tool.
Author Contributions
KB, BD, ER performed the experiments. VI and MT performed the analysis. DJA oversaw the project. All authors contributed to the writing of the manuscript.
Competing Financial Interests
The authors declare no competing financial interests.
Supplementary tables
Supplementary tables provided in excel format (see zip file).
Supplementary Table 1: List of zygote injections.
Counts of embryos microinjected, transferred and harvested at 12.5 d.p.c by treatment group. Supplementary Table 2: Illumina X10 read coverage.
Worksheet: Total Depth
Total numbers of bases sequenced by Illumina X10 and average depth by mouse sample.
Worksheet: Depth by Coverage Bin
For each sample the columns 1+, 11+ etc show the fraction of all sequenced bases at coverage greater >= 1x, >=11x, etc.
Supplementary Table 3: Summary of mosaicism for determining cutting efficiency.
Cutting efficiency and mosaicism obtained from Tyr2R, using different Cas9 sources (mRNA, protein) and different gRNA sources (in vitro transcribed or synthetic). Historical cutting efficiency is compared to the current experiment.
Worksheet: Efficiency and Mosacism
Shows cutting efficiency and mosaicism rates from both historical mouse data (Rows “Cas9 Protein” and “Tyr2R RNP”) and the present experiment (Rows “parent”, “No Injection”, “Sham injection”, “Cas9 only”, “Tyr2F”, “Tyr2R” describe each treatment condition). Historical data shows broad similarity to current data conditions “Tyr2R” and “Tyr2F” for targeting efficiency (column E) and average variants per embryo (column G).
Columns:
Targeting efficiency: the number of animals displaying a mutant genotype with allele frequency > 5%. Percentage is stated of the number of pups analysed.
Number with Mosaic genotype: The number of animals or embryos displaying multiple alleles at the same target site. Percentage is stated of the number of pups analysed.
Average variants per embryo: the average number of distinct alleles per embryo or mouse in each historical or current treatment group.
Worksheet: TestsOfDifferentConditions
Historical data. Shows the expected cutting efficiency and mosaicism of Tyr2R under different Cas9 sources and gRNA sources. Synthetic gRNA was most efficient with 57 out of 82 pups born showing indels (70%). On average, 53% of mice showed one allele, 33% two alleles, 12% three alleles, 2% four alleles and 0.5% five alleles. As the synthetic gRNA showed the highest cutting efficiency, we used these for further experiments.
Supplementary Table 4: Total variant counts and filtered counts by sample.
This is a detailed version of Table 1, with further details on the reduction of variant counts as successive filters are applied. (see Supplementary Methods). Samples are grouped into parent / offspring groupings.
Columns:
Sample: sample label of each parent or embryo
Treatment Group:
Cas9 only - gRNA omitted in injection
No injection - uninjected embryos
Sham injection - injected with water
Tyr2R treated - treated with Tyr2R gRNA + Cas9
Tyr2F treated- treated with Tyr2F gRNA + Cas9
Quality passed: Variants per sample passing basic genotype quality > 10
Depth and quality passed: variant count per animal passing joint-depth (>10) and genotype quality filters
TrioDeNovo, depth, quality: all candidates produced by TrioDeNovo caller passing the previous two thresholds.
TrioDeNovo, depth, quality SNVs / Indels: breakdown of TrioDeNovo column by variant class.
On target: Indel variants produced by TrioDeNovo lying in the expected Tyr mutation regions
Not on-target, vaf and parent noise filtered: all SNVs/Indels not in the Tyr locus, remaining after filtering for false positives arising from parental mosaicism and minimum Variant Allele Fraction (>=0.1)
Not on-target, vaf, parent noise and repeat filtered = Variants passing prior filters which are not in or 1bp adjacent a UCSC repeat region.
Not on-target, vaf, parent noise, repeat and BL6NJ/N filtered = Variants passing prior filters which are not known BL6NJ or BL6N variants.
Not on-target, vaf, parent /cross noise, repeat and BL6NJ/N filtered: Variants passing prior filters which are not present at more than 2% in any other samples.
… Without shared DNVs: Variants passing prior filters, which are not shared between any two embryos:
Final SNVs: SNV variants passing all previous filters (and manual re-addition of any mosaic variant at same locus, if it exists).
Filtered Indels: Indel variants passing all previous filters (and manual re-addition of any mosaic variant at same locus, if it exists).
Final indels: indel variants passing all previous filters, visually inspected for correctness.
Supplementary Table 5: Off-target locations with adjacent de novo mutations.
Worksheets: RGEN_Tyr2F_Offtarget_Sites, RGEN_Tyr2R_Offtarget_Sites: All expected CRISPR off target locations for Tyr2F and Tyr2R.
Columns state Chromosome, Position and Direction of expected off-target site, as well as the extent of the mismatch (number of basepair mismatch and whether there is a “bulge” - i.e. a 1 bp insertion or deletion, and whether the “bulge” is a DNA or RNA bulge).
Worksheet: RGEN_DNM_10bp_overlap_by_sample
The intersection (with a 10bp window) of unfiltered de novo variant calls with all Tyr2F and Tyr2R gRNA candidate off-target locations. This shows that no untreated animals have any overlap within 10bp of a candidate Tyr2R or Tyr2F off-target site, and that only the on-target indels overlap the expected on-target site in the treated animals.
Supplementary Table 6: WGS and MiSeq on-target alleles for comparison.
Shows a comparison of on-target variants revealed by high-depth MiSeq sequencing and CRISPResso analysis (cutoff allele frequency 5%), compared to the variants at the same locations revealed by whole-genome X10 sequencing and bcftools analysis. We note that CRISPResso found no on-target variants in the untreated samples (MD5624a-MD5632a), so these samples are not presented.
CRISPResso found 20 variants in the Tyr-treated variants (MD5633a-MD5642a), with variant allele fractions ranging from 0.1 to 0.33, and with a typical mosaicism of 2 alleles per embryo. This confirms that the Tyr gRNAs are active in the treated samples, and that the activity at the on-target locations measured by mosaicism and allele fraction are within expected ranges. Deep-sequencing of the on-target location by MiSeq therefore shows mosaic indels in treated animals, and no mutation in controls.
X10 sequencing confirms both location and mosaicism in treated animals, with the exception of three indels, two of which are at low allele fraction (7% and 7.5%), and the third which is a second mosaic allele at exactly the same location as a called allele. Our pipeline manually adds in such mosaic alleles.
Supplementary Table 7:Results for KrusKal-Wallis and Wilcoxon Rank Sum tests
There are three control groups, containing three samples each: a “no injection” group, a “sham” injection group (injected with water) and a “Cas9-only” group. Due to the small number of samples in each group we elected to perform a Kruskal-Wallis Rank Sum test to detect any differences between these three groups. This test failed to detect any difference in SNV counts between groups - i.e. to reject the null hypothesis (chi-squared = 2.4, df = 2, p-value = 0.3012) and failed to detect any difference in indel counts between groups (chi-squared = 1.9874, df = 2, p-value = 0.3702).
The Tyrosinase-treated groups were either Tyr2R-treated (5 samples) or Tyr2F-treated (5 samples). We performed a Wilcoxon Rank Sum test and again failed to detect any difference between the groups for SNV counts (W = 6.5, p-value = 0.2492) or indel counts (W = 8.5, p- value = 0.4338).
Finally, based on these results, we merged all three untreated groups together and both Tyr- treated groups together. Using the Wilcoxon Rank Sum test, we were unable to detect any difference in SNV counts between the Tyr-untreated and Tyr-treated groups (W = 58, p-value = 0.3059) or indel counts (W = 35.5, p-value = 0.4471).
Supplementary Table 8: Filtered variant positions for experimental validation.
This lists the genomic location of all final DNM variants (SNVs and Indels) passing all filters and subsequently sent for validation, as well as PCR primer pairs used for validation, and the validation test results.
Supplementary Table 9: Filtered variant counts for alternative de novo variant analysis.
Lists filtered variant counts for an alternative de novo variant analysis in which the parents of the Cas9-only trio were deliberately set to be the parents for all embryos. De novo variant calling and filtration was carried out as before, with filtration stage stopped after the removal of parental mosaicism, proximity to repeats and the overlap with known BL6NJ/N variants. (This is the last sensible filter stage suitable for a comparison of this approach with the original data.) Column G lists the combined SNV and indel counts at this filtration point. As expected, the “Cas9 only” group - which was analysed using the correct parents - has significantly lower values of variants than the other groups. Column G is directly comparable with Supplementary Table 4 Column L (“not on-target, vaf, parent noise, repeat and BL6NJ/N filtered”).
Acknowledgements
We thank William C. Skarnes and Allan Bradley for their scientific advice. All mouse work was undertaken by the Sanger Research Support Facility (RSF), with assistance from Evelyn Grau, Joanne Doran, Ellen Brown, Mike Woods and Catherine Tudor. Sequencing was performed by the Sanger DNA sequencing pipeline, with additional analysis by the Cancer Genome Project Informatics team. This work was supported by Wellcome funding.
Footnotes
↵* Contact information: Vivek Iyer, vvi{at}sanger.ac.uk
References
