An online bioinformatics tool predicts zinc finger and TALE nuclease off-target cleavage

Abstract

Although engineered nucleases can efficiently cleave intracellular DNA at desired target sites, major concerns remain on potential 'off-target' cleavage that may occur throughout the genome. We developed an online tool: predicted report of genome-wide nuclease off-target sites (PROGNOS) that effectively identifies off-target sites. The initial bioinformatics algorithms in PROGNOS were validated by predicting 44 of 65 previously confirmed off-target sites, and by uncovering a new off-target site for the extensively studied zinc finger nucleases (ZFNs) targeting C-C chemokine receptor type 5. Using PROGNOS, we rapidly interrogated 128 potential off-target sites for newly designed transcription activator-like effector nucleases containing either Asn-Asn (NN) or Asn-Lys (NK) repeat variable di-residues (RVDs) and 3- and 4-finger ZFNs, and validated 13 bona fide off-target sites for these nucleases by DNA sequencing. The PROGNOS algorithms were further refined by incorporating additional features of nuclease-DNA interactions and the newly confirmed off-target sites into the training set, which increased the percentage of bona fide off-target sites found within the top PROGNOS rankings. By identifying potential off-target sites in silico, PROGNOS allows the selection of more specific target sites and aids the identification of bona fide off-target sites, significantly facilitating the design of engineered nucleases for genome editing applications.

Figure 1.

PROGNOS search interface and comparison to previous prediction methods. (A) The PROGNOS online interface allows users to enter the target site of their nuclease pair and specify search parameters and primer design considerations. (B) A comparison of PROGNOS predictions to previously reported methods identifying off-target sites for different ZFNs (6,9). The Homology and Conserved G’s algorithms were used to determine what percentage of the sites with previously identified off-target activity fell within the top fractions of PROGNOS rankings. The ‘1X’ top fraction corresponds to searching the same number of top PROGNOS sites as were investigated in the original paper and ‘3X’ corresponds to searching three times as many PROGNOS sites as were investigated in the original manuscript. (C) A comparison of the PROGNOS search algorithms to previously reported methods identifying off-target sites for TALENs (5,8). The top PROGNOS rankings using the Homology-5TC and RVD-5TC algorithms were searched to determine what percentage of off-target sites found to have activity fell within the top fractions of PROGNOS rankings. (D) Venn diagram displaying the 13 known off-target sites identified for the heterodimeric CCR5 ZFNs during development and testing of the original PROGNOS algorithms (9,10). The sites ranked at the top of the PROGNOS Homology and Conserved G’s in silico algorithms [allowing 3X the number of sites searched by Pattanayak et al. (9)] are compared to the 12 sites identified previously and one site uncovered in this study.

Figure 2.

Using PROGNOS to identify nuclease off-target sites. (A) Outline of the procedure to identify nuclease off-target activity. (B) Sample outputs of the PROGNOS online software showing all sites found and what types of genomic regions they are located in as well as rankings of the top potential off-target sites. The rankings include the closest gene, the number of mismatches, the size of PCR product from the automatically designed primers, and other helpful information. (C) Comparison of the success of the automatically designed PROGNOS primers used in high-throughput full-plate PCR of off-target sites to primers designed in other off-target publications. (D) Sequencing reads of an off-target location for the 3F ZFN pair that show evidence of NHEJ. In the wild-type (WT) sequence, the ZFN binding sites are highlighted in yellow and mismatches to the intended target sequence are lowercase red. In the sequencing reads, inserted bases are lowercase and highlighted in blue. The size of the indel is displayed to the right of the sequence, along with the number of times that mutation was observed.

Figure 3.

Using SMRT Sequencing to analyze nuclease activity. (A) SMRT sequencing produced very similar results to standard TOPO sequencing over a range of mutation rates from ∼20% to ∼76%. Error bars are 90% confidence intervals. S2/S5 NK and S2/S5 NN are the TALENs targeting beta-globin compared in this study. S116/S120 and J7/J8 are NK-TALENs targeting beta-globin and CDH1, respectively (30). (B) Comparison of the range and frequency of different sizes of indels observed in cells treated with TALENs or ZFNs. The observed frequencies of the different sizes are normalized to the frequency of the most common indel size for each nuclease type.

Figure 4.

Improved performance of the refined PROGNOS algorithms. (A) The performance of the two initial ZFN algorithms and the refined ‘ZFN v2.0’ algorithm are compared for their ability to predict off-target sites for all the ZFNs in the training and validation sets. Percentages of off-target sites located were calculated according to 3X limits for previous studies and within the number of sites interrogated for PROGNOS-based studies (typically the top 24 ranked sites). Error bars represent SD. (B) The expanded landscape of 38 total heterodimeric off-target sites for the CCR5 ZFNs found by four different experiment-based prediction methods and the refined ‘ZFN v2.0’ PROGNOS algorithm. The PROGNOS sites are drawn from the top rankings spanning 3X the number of predictions by the Bayesian abstraction of the in vitro cleavage profile. (_**) Note that only six of the sites found using ChIP-Seq were provided by Sander et al. (11), so the full degree of overlap of all ChIP-Seq sites with sites found by other methods is unclear. (C) The performance of the four original TALEN algorithms and the refined ‘TALEN v2.0’ algorithm are compared for their ability to predict off-target sites for all TALENs in the training and validation sets.

Figure 5.

Sensitivity and specificity analysis of PROGNOS algorithms. (A) Average false positive ratios are shown for the PROGNOS investigation of novel nucleases using the initial algorithms, and for previous experimental prediction methods. Ratios are also shown for individual nucleases in the three different categories of nuclease that have been investigated previously by experimental prediction methods. (B) The false negative rates of the different PROGNOS algorithms and previous experimental prediction methods are shown. These were determined by each method’s ability to identify the 38 known hetero-dimeric off-target sites of the CCR5 ZFNs in their top ranking predictions.