DENT-seq for genome-wide strand-specific identification of DNA single-strand break sites with single-nucleotide resolution

Degenerate nucleotides provide a mutational signal demarking the locations of nicks

DENT-seq takes advantage of synthetic noncanonical deoxynucleotides capable of base-pairing with more than one of the four naturally occurring deoxynucleotides (Fig. 1). The bases 6H,8H-3,4-dihydropyrimido[4,5-c][1,2]oxazin-7-one (P) and N⁶-methoxy-2,6-diaminopurine (K) act as a universal pyrimidine and purine, respectively (Fig. 1A; Hill et al. 1998). dPTP and dKTP can be incorporated into a DNA molecule at the site of a single-strand break via nick translation (Fig. 1B) and generate readily detectable transition mutations when amplified by PCR and sequenced (Fig. 1C). Additionally, further extension of this nick translation product with biotin-tagged nucleotides and streptavidin pulldown enrich reads with mutational signal at nick sites (Fig. 1D).

View larger version:

• In this window
• In a new window

• Download as PowerPoint Slide

Figure 1.

DENT-seq overview and workflow. (A) Chemical structures of the two degenerate nucleotides, dPTP and dKTP, in both of their tautomeric forms enabling their respective activities as universal pyrimidine and purine. Black dotted lines represent hydrogen bonds between nucleotides. (B) During nick translation with only the two degenerate nucleotides, P residues (gold) are inserted across from A (blue) and G (purple) residues while K residues (cyan) are inserted across from C (green) and T (red) residues. (C) Sequencing DNA fragments without nicks or that had nick translation occur with only regular dNTPs will not show a mutational signal (top). Sequencing DNA fragments that underwent nick translation with dPTP and dKTP will show a mutational signal that extends a few bases 3′ of the nick's original location (bottom). (D) Workflow used to perform DENT-seq. Nick translations are performed consecutively with dPTP plus dKTP and then with regular dNTPs plus biotinylated dUTP. Of note, nick translation will move a nick from its original location but will not actually repair the nick. DNA is fragmented and streptavidin-based purification enriches for fragments around the original (pre-nick translation) sites of nicks. After PCR and sequencing, sequence coverage and mutational information allow for sensitive single-nucleotide-resolved and strand-specific identification of nick sites.

To characterize the mutational signatures that arise from P and K, we annealed custom oligonucleotides to obtain four DNA molecules, each with a nick present just 5′ and just 3′ of each of the four canonical deoxynucleotides (Fig. 2A). After PCR with Taq DNA polymerase and sequencing, we found evidence of P and K incorporation extending 4–8 bases downstream from the original nick site, with the ratio of C:T at sites where P was incorporated being ∼40:60 and the ratio of G:A at sites where K was incorporated being ∼15:85 in agreement with a previous study (Fig. 2B; Hill et al. 1998). PCR with KAPA HiFi DNA polymerase also results in 4–8 downstream bases displaying evidence of dPTP and dKTP incorporation, although with different distributions of C:T and G:A at such sites (Fig. 2C). Whereas PCR with Taq tends to result in higher signal at nicks, PCR with KAPA HiFi tends to result in lower background noise. Overall, although both polymerases are suitable for use in DENT-seq, KAPA HiFi is a better choice for complex samples like large genomes. We considered other degenerate nucleotides for use in DENT-seq, such as inosine (I) and ribavirin (R). I is recognized as G by DNA polymerases during PCR (Alseth et al. 2014), however, and as such would not allow for true single-nucleotide resolution of nicks flanked by G residues. We tested R in place of K as a universal purine (Sala et al. 1996) but found K to produce superior results (Supplemental Fig. S1).

View larger version:

• In this window
• In a new window

• Download as PowerPoint Slide

Figure 2.

Effect of degenerate nucleotides during PCR. (A) Schematic of the oligonucleotides used to test the mutational outcome when dPTP and dKTP are incorporated into DNA and amplified by PCR. Four different oligonucleotides were generated, each with a nick directly 5′ of a different one of the four native dNTPs. The strand with the nick also contained a 5′ biotin-TEG modification to allow for purification of just that strand before PCR. PCR primers were designed to specifically amplify the region of the oligonucleotides containing the nicks and were tailed with 5′ sequences compatible with secondary PCR using barcoded P5 and P7 sequencing primers. (B,C) Sequencing result of the four oligonucleotides following consecutive nick translations with dPTP plus dKTP and then with standard dNTPs, purification, and PCR with Taq DNA polymerase (B) or KAPA HiFi DNA polymerase (C). The black triangles represent the location of the nick for each sample.

Biotinylated nick translation products can be enriched to focus sequencing effort at nick sites and enable confident strand determination

Nick translation with unmodified and biotin-tagged nucleotides after incorporation of the degenerate nucleotides allows for selective pulldown of DNA molecules containing P and K residues via a streptavidin-based purification. By using a low dNTP concentration (5000 times lower than typical PCR protocols), we limit incorporation of biotin-tagged nucleotides to within a few hundred bases of the original nick site. Increased sequence coverage around nick sites reduces the total amount of sequencing needed, provides independent evidence of nicks that can be analyzed in conjunction with the mutational signal, and reduces the number of bases that need to be tested for mutational signal (Fig. 1D). To test library enrichment, we digested a plasmid containing two recognition sites with the nicking endonuclease Nb.BsmI (Fig. 3A). After performing DENT-seq on this sample, we observed up to 1000-fold higher coverage of sequence reads at sites around the nicks (Fig. 3B). These coverage peaks can be identified by MACS2, a peak caller originally designed for use with ChIP-seq data (Zhang et al. 2008), and the mutational signal from P and K incorporation reveals the exact location of nicks within the peaks (Fig. 3C,D).

View larger version:

• In this window
• In a new window

• Download as PowerPoint Slide

Figure 3.

Single-base-pair resolution of nicks in plasmid DNA. (A) Schematic of the plasmid used to test DENT-seq in which red arrows represent the expected locations of nicks after treatment with the nicking endonuclease Nb.BsmI (both are expected on the reference strand). (B) Sequence coverage normalized to the location with maximum read depth for untreated plasmid (top) and plasmid with two nicks introduced by Nb.BsmI treatment (bottom) after performing DENT-seq. Black dashed lines represent the expected locations of nicks after Nb.BsmI treatment and gray dashed lines represent the bounds of MACS2 peaks called on data from the treated sample. (C,D) Mutational signal for all locations within the called MACS2 peaks (top) and close-up views where there is high mutational signal (bottom). Black dashed lines represent the expected locations of nicks after Nb.BsmI treatment. Data in gray represent the untreated plasmid, whereas data in blue and green represent the Nb.BsmI treated sample where a P or K residue would be inserted after a reference strand nick, respectively.

In enriched samples, we observe mutational signal from P and K in a large majority (∼90%) of unique sequence reads that contain the original nick site. These signals are robust to the enzyme used to incorporate the biotinylated nucleotides (Supplemental Fig. S2) and the enzyme used to conduct PCR (Supplemental Fig. S3). Furthermore, similar levels of enrichment are achieved when using nucleotides that are modified with desthiobiotin (Hirsch et al. 2002; Supplemental Fig. S4). Although we originally developed DENT-seq using a biotin-tagged uracil deoxynucleotide (biotin-dUTP), desthiobiotin-tagged nucleotides are a superior option. Use of a desthiobiotin-tagged adenosine deoxynucleotide (desthiobiotin-dATP) provides better compatibility with DNA polymerases and sequencing library construction protocols than biotin-dUTP. In addition, desthiobiotin-tagged nucleotides can be eluted from streptavidin-coated magnetic beads with free biotin, so that PCR can be carried out without potential interference from the presence of the beads. Peak width and the magnitude of enrichment are highly dependent on the concentration of nucleotides during the nick translation reaction. Concentrations higher than 40 nM per dNTP are not recommended due to an associated increase in background sequence coverage (Supplemental Fig. S5).

Library enrichment also plays a crucial role in determining the strandedness of a nick. As DNA polymerases extend DNA in the 5′→3′ direction and the nick translation steps are carried out consecutively, the biotinylated nucleotides are always 3′ of the degenerate nucleotides and the position with maximum enrichment is expected 3′ of the mutational signal. Sequence data are typically viewed with respect to the reference strand of double-stranded DNA, so a nick that occurred on the reference strand will result in the coverage peak appearing 3′ of the mutational signal. A nick that occurred on the non-reference strand, however, will result in the coverage peak appearing 5′ of the mutational signal because it is on the 3′ side with respect to the non-reference strand (Supplemental Fig. S6). Determining the strandedness of a nick is, in turn, important for single-nucleotide resolution because multiple P and K residues can be incorporated at a nick, resulting in a few consecutive bases with mutational signal that enhances the mutation signal's detectability, but does not unambiguously identify the nick site. Stranded determination unambiguously identifies the nick site, which is immediately 5′ of the mutational signal on the nicked strand.

Finally, this enrichment is essential to DENT-seq's high sensitivity. Not all DNA molecules in a sample will necessarily be nicked at the same locus, but by including this enrichment we can selectively sequence a larger fraction of DNA fragments originating from nicked molecules versus those originating from non-nicked molecules. In this way, the mutational signal arising from dPTP/dKTP can remain high even as the fraction of molecules containing nicks at a given locus decreases. Even so, DENT-seq has finite sensitivity and specificity. To explore these limits in different scenarios, we carried out a series of experiments and analyses.

By serially diluting a nicked plasmid in a non-nicked plasmid, we showed that DENT-seq identifies nicks present in as few as 1 in 1000 DNA molecules (Supplemental Fig. S7). We then performed bootstrap resampling of reads in each of these nick dilution data sets to estimate DENT-seq's false-negative rate. Similarly, we performed bootstrap resampling of a data set obtained by performing DENT-seq on non-nicked DNA to estimate the false-positive rate (false-positive and false-negative rate estimates are provided in Supplemental Table S2). False-negative rate is highly dependent on nick penetrance; whereas we estimate rates as high as 12.18% when nicks are rare at a locus (1 nicked molecule out of 1000 copies at a single locus), we estimate false-negative rates <0.1% for more penetrant nicks (≥1 nicked molecule out of 100). Of note, the false-negative rate for low penetrance nicks is based on a conservative mutation rate cutoff for nick calling and lowering this cutoff could reduce the false-negative rate with little effect on false-positive rate. False-positive rate is highly dependent on the class of DNA being tested. Based on our results, we estimate about one false positive occurring across every four experiments performed on human genomic DNA (i.e., less than one false positive call in 10 billion genomic bases) and lower rates for less complex samples like microbial genomes and synthetic DNAs.

Nicking endonuclease Nb.BsmI has detectable off-target activity

Nb.BsmI has been reported to display off-target activity in reactions with too high an enzyme concentration, that proceed for too long, or that have modified buffer conditions. Utilizing an increased concentration of Nb.BsmI (cf. Fig. 3) causes off-target activity that is detectable by DENT-seq (Fig. 4A). Enrichment of the two on-target sites still occurs, but to a lesser extent than when off-target nicks are absent. Because read coverage is relatively high throughout the plasmid, we did not restrict our nick calls to MACS2 peaks in this instance. In addition to the on-target sites, there are 12 sites with mutational signal above background level (Fig. 4B; Supplemental Table S3). Each of these sites has a single-nucleotide mismatch within the 6-nt consensus Nb.BsmI recognition sequence. Given a total of 57 such sites in the plasmid, observing 12 sites with high mutational signal occurring at such near-cognate Nb.BsmI sites by chance is highly unlikely (P < 10⁻²⁵, hypergeometric test), suggesting these mutational signals indeed reflect the presence of nicks resulting from true off-target Nb.BsmI activity. Because these off-target nicks are not associated with highly enriched MACS2 peaks, their strandedness cannot be determined from the mutation signal alone. Here, even without the presence of enrichment peaks, the orientation of the off-target Nb.BsmI recognition sequence in the plasmid reference can be used as an alternative method to determine the nicked strand. Although the identified off-targets do not occur in enrichment peaks, the enrichment procedure significantly enhances their detection (Supplemental Fig. S8). This likely results from such off-target nicks having low penetrance and the enrichment selecting for DNA molecules that underwent the nick translation with dPTP and dKTP.

View larger version:

• In this window
• In a new window

• Download as PowerPoint Slide

Figure 4.

Detection of off-target nicks in plasmid DNA. (A) Schematic of a plasmid used to test DENT-seq, in which red arrows represent on-target activity of Nb.BsmI and purple arrows represent detected off-target activity of Nb.BsmI. (B) Normalized sequence coverage (top) and base transition rate (bottom) for the plasmid in A. Black dashed lines represent on-target locations of Nb.BsmI activity and gray dashed lines represent the bounds of called MACS2 peaks. Purple and cyan dashed lines represent detected off-target locations of Nb.BsmI activity on the reference and non-reference strands, respectively. (C) Schematic of a plasmid used to test DENT-seq in which the red arrow represents on-target activity of D10A spCas9 nickase and the purple arrow represents detected off-target activity (both are on the non-reference strand). (D) Normalized sequence coverage (top) and base transition rate (bottom) for the plasmid in C. The black dashed line represents on-target activity of D10A spCas9 nickase, whereas the cyan dashed line represents off-target activity. The gray dashed lines represent the bounds of called MACS2 peaks.

Cas9 nickase has detectable off-target activity in plasmid DNA

Cas9 off-target activity has been observed at locations with significant, but not full, complementarity to the guide RNA (Kim et al. 2015; Tsai et al. 2015, 2017; Slaymaker et al. 2016; Yan et al. 2017). We were able to observe this phenomenon directly in D10A spCas9 nickase activity via DENT-seq in plasmid DNA. We used a guide RNA with one fully complementary on-target site and one predicted off-target site where proper base-pairing would occur for 15 of the 20 nucleotides of the guide RNA (Fig. 4C). Significant enrichment and a strong mutational signal occurred at both sites, with slightly lower enrichment observed at the off-target site (Fig. 4D; Supplemental Table S4).

Nick detection at genome scale

After demonstrating the capabilities of DENT-seq on plasmid DNA, our next goal was to detect nicks in a genome. We used D10A spCas9 nickase to incorporate nicks in an Escherichia coli genome of 4.62 Mbp by carrying out a reaction with two guide RNAs wherein we predicted nine nicks would occur at fully cognate target sites (Fig. 5A). After performing DENT-seq, all nine nicks were identified with no false positive identifications. Eight nicks, all originating from the same guide RNA, resulted in highly significant sequence coverage peaks (Fig. 5B), and single-nucleotide resolution was achieved using the mutational signal (Fig. 5C; Supplemental Fig. S9A). The remaining nick, originating from the other guide RNA, was located within a MACS2 called peak but with lower enrichment. In fact, this peak could not be distinguished from background peaks based on sequence coverage or MACS2 statistics alone (Fig. 5B). The mutational signal present in this peak allowed the nick to be detected and localized (Fig. 5D). We hypothesize that the difference in sequence coverage between the peaks containing nicks is owing to different guide RNA activity levels that result in differential nick penetrance. Other sequence coverage peaks either lacked mutational signal altogether, contained singleton mutational signal (not at multiple consecutive or near-consecutive nucleotides), or contained mutation types other than the transition mutations that dPTP and dKTP are known to cause (Supplemental Fig. S9B). These signals likely arose from sources other than P and K incorporation, such as misincorporation of canonical dNTPs during PCR or sequencing errors, and as such were straightforwardly filtered out to achieve high specificity. We did not expect to observe off-target activity from either guide RNA in this experiment. One guide showed low on-target activity, and there were no sites in the genome with significant homology to the other, high activity guide that lacked mismatches at the 3′ end of the target sequence and were also adjacent to a Cas9 PAM site.

View larger version:

• In this window
• In a new window

• Download as PowerPoint Slide

Figure 5.

Single-base-pair resolution detection of nicks in genomic DNA. (A) Schematic of an E. coli genome used to test DENT-seq, in which red arrows represent the activity of Cas9 nickase paired with one guide RNA and the purple arrow represents the activity of Cas9 nickase paired with a different guide RNA. Circularized plot data (all colors) show normalized sequence coverage. Green signals represent the locations of called MACS2 peaks with quality score greater than 0.4. Red signals represent locations of such peaks wherein a nick is identified with the exact location of that nick represented by a black dot. (B) Cumulative number of nicks contained within MACS2 peaks (green) and cumulative number of MACS2 peaks containing no nick (red) as a function of P-value threshold. The left gray dashed line represents the P-value where the final nick-containing peak is identified, and the right gray dashed line represents the P-value where the first peak not containing a nick is identified. This panel uses only the coverage signal and does not incorporate mutational data. (C) Normalized sequence coverage (top) and base transition rate (bottom) for four MACS2 peaks that contain nicks resulting from Cas9 nickase in conjunction with one of the guide RNAs used (corresponding to the red arrows in A). The second set of plots from the left represent a single MACS2 peak, which encompassed two nicks owing to their close proximity. (D) Normalized sequence coverage (top) and base transition rate (bottom) for the MACS2 peak that contains the nick resulting from Cas9 nickase in conjunction with the other guide RNA used (corresponding to the purple arrow in A). Coverage is lower for this guide, but there are still nucleotide positions proximal to the Cas9 nickase target site with high mutational signal that enable nick identification.

Off-target Cas9 nickase activity can be detected across the human genome

We performed DENT-seq on human genomic DNA samples treated with D10A spCas9 nickase and guide RNAs targeting either the AAVS1 gene on Chromosome 19 (the same guide used in the plasmid experiment described above) or the VEGFA gene on Chromosome 6. With the AAVS1 guide, this allowed for the identification of six nicks without using a priori knowledge of expected Cas9 on- or off-target activity to inform nick calling. Of these six nicks, the one with the highest enrichment is located at the on-target site (Fig. 6A), and the other five are located at off-target sites each with three or four mismatches to the guide RNA (Fig. 6B). Comparing maximum read depth around the on-target nick site to average read depth across the whole genome reveals ∼28-fold higher sequencing effort would have been required to obtain similar read depth around the on-target site without the streptavidin pulldown. Of note, this calculation underestimates the benefit of the pulldown because it only considers read depth. Mutational signal is also enhanced by the pulldown, particularly for nick sites where the nick is not fully penetrant as reads from the nick translation products are enriched relative to unmodified sequences from non-nicked molecules at the same locus. In particular, nuclease off-target sites often have lower nick penetrance than the on-target site, and some may not have any apparent mutational signal above background if the pulldown was not performed. Nonetheless, in the absence of a priori target information, we only identified the on-target nicking site in the VEGFA nicking experiment (Supplemental Fig. S10). Although prior work using the same VEGFA guide, wild-type spCas9 nuclease, and genome-wide DSB detection did identify off-target DSB sites (Kim et al. 2015; Tsai et al. 2015; Slaymaker et al. 2016; Yan et al. 2017), we wanted to explore off-target nick sites in a minimally biased fashion, and for this reason focused our further analyses on data from experiments using the AAVS1 guide.

View larger version:

• In this window
• In a new window

• Download as PowerPoint Slide

Figure 6.

Detection and analysis of D10A spCas9 nickase on- and off-target activity in human genomic DNA. (A) D10A spCas9 nickase single-site on-target activity is represented by the most prominent MACS2 peak, denoted by the gray dashed lines, which also contains strong mutation signal. The reverse complement of the guide sequence is shown in blue with the PAM in green at bottom. (B) One of the five off-target nicks that were identified without using a priori knowledge of expected off-target sites. Gray lines mark the MACS2 peak bounds. At bottom, the genomic sequence where the guide RNA binds is shown in blue with mismatches in red. (C) Ideogram showing the site of on-target activity in red and candidate off-target sites with DENT-seq mutational signal in purple. (D) Comparison of the frequency of mismatch locations in the guide RNA at the sites of detected off-target activity to all potential sites proposed by Cas-OFFinder. (E) Sequence motifs created from the sites with evidence of D10A spCas9 nickase off-target activity and the top hits predicted by multiple Cas9 off-target predictors (left). The motif for the sites identified by DENT-seq is compared to the motifs for each of the software predictions using DiffLogo (right). The bases composing the PAM sequence are shown with a light gray background. Dispersion metrics for the prediction software sequence logos were generated by performing 10,000 bootstrap resamples of the entire list of potential off-target sites proposed by each algorithm and calculating the position probability matrix for the highest rated 52 sites in each resample. Ninety percent confidence intervals for the frequency of each base at each position are displayed under the assumption that the information content at each position remains unchanged.

Nick calling can be enhanced further by considering only prospective off-target locations. To this end, we used Cas-OFFinder (Bae et al. 2014) to identify all 714,765 PAM-adjacent sites in the genome with up to eight mismatches to the AAVS1 guide RNA or with up to two mismatches and a 2-nt DNA or RNA bulge. Analyzing these locations for sequence reads and mutational signal enabled the identification of 53 sites with strong evidence for off-target activity (Fig. 6C; Supplemental Table S5). These off-target sites were generally concordant with previously reported characteristics of wild-type Cas9 off-target DSB activity (Tsai et al. 2015). Mismatches closer to the 5′ end of the guide RNA, distal from the PAM, are better tolerated than mismatches at the 3′ end, proximal to the PAM (Fig. 6D). Further, the PAM sequence 5′-NGG is the most prominent but there is observable off-target nicking activity at the alternative PAM 5′-NAG.

The guide RNA used here was previously tested for off-target activity in the human genome with wild-type spCas9 (Wang et al. 2014; Tan et al. 2015). The methods previously used for DSB analysis with this particular guide were neither unbiased nor genome-wide, but rather relied on PCR and deep sequencing of select locations in the genome predicted as likely off-target sites (99 unique sites were tested between the two studies), which hinders direct comparison to our findings. However, DENT-seq provides evidence of D10A spCas9 nickase activity at five off-target sites where the previous studies described no wild-type spCas9 DSB off-target activity using the same guide RNA. Three of these five sites were detectable as nicks without the use of a priori knowledge of expected off-target sites and showed highly significant peak enrichment as measured by MACS2 (Supplemental Table S5). Furthermore, the previous studies identified eight off-target sites with wild-type spCas9 DSB activity where DENT-seq did not show D10A spCas9 nickase activity, suggesting that wild-type Cas9 and the D10A nickase mutant could have different off-target spectra.

To summarize the off-target activity of the D10A spCas9 nickase variant, we generated a sequence motif from the 52 sequences where nickase activity was identified and base-pairing between the guide RNA and target DNA did not result in a DNA or RNA bulge. To compare the D10A spCas9 nickase off-targets detected by DENT-seq with predicted wild-type spCas9 off-targets, we used several DSB off-target prediction tools. CCTop (Stemmer et al. 2015), CRISPOR (Concordet and Haeussler 2018), DeepCRISPR (Chuai et al. 2018), and CROP-IT (Singh et al. 2015) ranked all genome-wide potential off-targets and we generated sequence motifs for the top 52 sites proposed by each algorithm (for direct comparison with the 52 DENT-seq identified sites). Two motifs were created based on CRISPOR's output as it reports both MIT off-target scores (score calculation is based on data from Hsu et al. 2013) and CFD off-target scores (Fig. 6E, left column; Doench et al. 2016). To compare the motifs created from our results and from each of the prediction tools, we used DiffLogo (Nettling et al. 2015). Each position in the DiffLogo motif summarizes the base magnitude and polarity of difference between the two motifs being compared and enables the calculation of a scalar total difference score between the motifs. We created DiffLogo motifs to perform pairwise comparisons between the DENT-seq identified off-targets and the proposed off-targets from each prediction tool, including the background distribution of all possible off-targets identified by Cas-OFFinder (Fig. 6E, right column). From the DiffLogo scores, the motif generated by the CCTop predictions most closely resembles the motif generated by the DENT-seq results. The biggest discrepancy between these two motifs is at the base just next to the PAM region where T is not as frequent at off-target sites as CCTop predicts. This is in contrast to CROP-IT and, to a lesser extent, DeepCRISPR, where too little emphasis is placed on this position. For the CFD and MIT scores, it appears as though the bases at the 5′ end of the target sequence are being weighed too heavily and the 5′-NAG alternative PAM is nearly ignored.

All of the prediction tools used here are meant to predict wild-type spCas9 DSB off-target activity. We expect there to be some differences in the off-target strand breaks created by wild-type spCas9 and its D10A nickase variant, so our analysis here is not a test of how well these tools perform at the task they were designed for. Rather, it reveals which of the currently available methods for a related task performs best at predicting D10A spCas9 nickase off-target activity. As additional D10A spCas9 nickase off-target sites are characterized using DENT-seq, optimal parameters for each algorithm can be identified for predicting D10A spCas9 nickase activity and the prediction tools could have the option to search for expected wild-type/DSB or D10A nickase off-target activity.