Background

Zinc finger (ZF) DNA binding proteins can be used to target functional protein domains to specific regions in complex genomes. For example, zinc finger nucleases (ZFNs) have tremendous potential for introducing site-specific gene knockouts or gene targeting events with high efficiency in various cell types including human[1–3]. A ZFN consists of two zinc finger proteins (ZFPs) each fused to a monomeric Fok I nuclease domain. When the ZFPs co-locate to adjacent sequences within the genome, the nuclease monomers are able to dimerize, generating an active nuclease that cleaves the double-stranded DNA at the target site. In the presence of exogenous donor DNA, genetic material may be exchanged through repair by homologous recombination; alternatively, the break may be repaired by non-homologous end joining, which is an error-prone mechanism that commonly results in knockout mutations [4, 5]. To date, ZFNs have been used to manipulate endogenous genes in several organisms, e.g., tobacco, maize, fruit fly, zebrafish, rats, and human [6–15], and are being evaluated in human clinical trials, including gene therapies to treat AIDS [16–18].

Zinc finger DNA binding domains, especially the C2H2 class of zinc fingers, have been exploited for performing targeted genome modification because they can be engineered to bind a wide range of desired DNA sequences. Each individual C2H2 zinc finger consists of an α-helix (the DNA "recognition helix") and a β-hairpin, stabilized by a single zinc ion coordinated through interactions with cysteine and histidine residues. Individual ZFs recognize and bind specific triplet DNA sequences through base-specific contacts within the major groove of double-stranded DNA[19]. Extended DNA sequences can be targeted by joining together several ZF domains [20, 21].

ZFPs engineered using the recently developed Oligomerized Pool ENgineering (OPEN) method have been reported to function with high success rates in vivo, particularly for zinc finger nucleases (ZFNs) [8, 9, 15, 20, 22]. For constructing ZFPs that recognize 9-bp targets, the OPEN method involves combinatorial assembly and subsequent selection of fingers from three pre-constructed pools, each of which contains up to 95 different engineered ZF recognition helix "solutions" for a chosen DNA triplet [8, 23]. Currently, pools are available for all 16 GNN triplets and several of the TNN triplets for each position in a three-finger array [8]. ZFNs generated by OPEN have been used to target genes in tobacco, zebrafish, and human cells with high efficiency [8–10].

Because using the OPEN procedure requires investment of time and effort and because there are often numerous potential targetable sites in any given gene, it is desirable to focus experiments on target sites that are most likely to yield functional ZFPs. For example, there are 315,186 OPEN ZFN sites in the protein encoding regions of the zebrafish genome (an average of 10.8 sites per transcript). While OPEN often generates ZFPs that function well in a bacterial two-hybrid (B2H) reporter system [8, 9], it does not have a 100% success rate. Thus, to reduce the experimental effort involved in applying the OPEN procedure, we sought to develop a computational approach to identify the "best" targets, i.e., those most likely to be successfully targeted by OPEN, from among the relatively large number of theoretically "targetable" ZFP sites that may exist for any chosen gene or genomic region of interest.

In this study, we demonstrate that sequence characteristics of ZFP target sites, when used as input to Naïve Bayes or Support Vector Machine (SVM) classifiers, can be used to reliably predict whether a specific DNA sequence will (or will not) be successfully targeted by OPEN. The performance of these classifiers on two experimentally validated datasets of ZF target sites suggests that their use could substantially reduce the experimental effort required to generate a functional ZFN using the OPEN method.

Results

Results from several groups [24–31] have suggested that ZFP recognition sites with a high purine nucleotide content, especially those containing several GNN-triplets, more frequently correspond to "active" targets for zinc finger proteins generated using modular assembly. To investigate whether such potential biases could be exploited to identify optimal sequences for ZFP targeting using OPEN, we analyzed sequence and base composition characteristics of sites targeted by this method.

For this study, we first generated an experimentally validated dataset, ZFTS135, consisting of 135 nine bp target sites for which OPEN did or did not successfully yield ZFPs. ZFTS135 includes 53 ZF target sites from recently published OPEN experiments [8, 9] and 82 OPEN ZF target sites which we report here for the first time. Each target site in the dataset was assigned a class label of either "active" (79%) or "inactive" (21%). "Active" target sites were those yielding at least one ZFP that showed DNA-binding activity in a well-validated bacterial two-hybrid (B2H) reporter assay (defined as the ability to activate transcription by three-fold or more, a level previously shown to identify ZF arrays that possess high affinity and high specificity for their cognate DNA binding site [8, 23]). "Inactive" target sites were those that failed to yield a ZFP that showed activity in the B2H reporter assay. All 135 functionally validated ZFP target sites and their assigned labels are provided in Additional File 1 - Table S1.

Figure 1 presents analyses of the sequence and base composition characteristics of ZFP target sites in the ZFTS135 dataset. The average number of times each base occurs in active and inactive targets is shown in Figure 1A. On average, active sites contain more guanines and fewer thymines than inactive targets. Because OPEN ZF finger pools are available exclusively for GNN and TNN triplet subsites at present, total guanine and thymine counts are inflated, compared to adenine and cytosine counts. To account for this, as well as the fact that specific bases, when located in different positions within a triplet subsite, may preferentially contact different amino acids, the average base occurrences were calculated for each position within the triplets (Figure 1B). This analysis identified thymine frequency, at any position within a triplet, as the primary difference between active and inactive target sites. Guanine, adenine, and cytosine typically appear more frequently in active sites than in inactive sites, compensating for the decrease in thymine content.

Figure 1
figure 1

Base composition differs in active versus inactive ZFP target sites. A) Total base counts for active and inactive ZFP target sites (from ZFTS135, a dataset of 135 experimentally validated 9-bp target sites, see Additional File 1 - Table S1) reveal that variation in the average frequency of each base differentiates active versus inactive target sites. The total number of G and T residues relative to A and C is inflated because currently available OPEN pools are designed to target GNN and TNN triplets. B) Positional base counts, i.e., average base counts for each position within target site triplets (1st, 2nd, 3rd), suggest that thymine bases negatively impact ZFP binding at all three positions. C) An iceLogo [50] generated from ZFTS135 illustrates the difference in percentage composition of nucleotides at each position, from 1 - 9 (5' to 3'), between the positive class and the entire dataset. For example, 78% of all sites in ZFTS135 have a G in position 1, whereas 88% of all active sites have a G at position 1, resulting in a difference of 10%. Positive difference values indicate that, on average, the indicated bases are favored at those positions in active sites; negative difference values indicate that the indicated bases are disfavored. These position-specific differences in percentage composition also support the conclusion that thymine bases tend to occur in inactive targets (i.e., they have large negative propensities).

Full size image

Differences in base composition at each position within active 9-bp target sites were also analyzed. As shown in Figure 1C, thymine is generally disfavored in active target sites, with strong negative propensities in the 1st and 7th positions of active target sites. Other residues showed marginally positive propensities in most positions. Because available OPEN reagents are currently limited to those that target GNN and TNN triplets [8] (and one ANN triplet; M. Maeder & J.K. Joung, unpublished data), it is not possible to evaluate the significance of the relatively low percentage of adenine and cytosine residues in positions 1, 4 and 7.

Taken together, the results of these analyses suggested that base composition biases in active versus inactive ZFP target sites could be exploited by machine learning classifiers to predict whether a specific DNA sequence can be targeted successfully using the OPEN procedure. Machine learning classifiers that use a string of sequence identities as input have been successfully applied to a variety of problems, including protein functional site classification [32–35]. Because several different machine learning classifiers we tested gave comparable results (data not shown), here we present representative results obtained using two types of classifiers: Naïve Bayes and support vector machines (SVMs).

We compared classifiers trained using three different target site sequence encodings: i) sequence identity: 9 nucleotide identities corresponding directly to the target site sequence; ii) base counts: 4 numerical values representing the overall base counts of G,A,C,T in the target site; iii) positional base counts: 12 numerical values encoding the position-specific base composition of the target site (see Methods for details).

Table 1 summarizes performance statistics for Naïve Bayes and SVM classifiers tested using the three different target site encodings and evaluated using leave-one-out cross-validation. In these experiments, classifiers were optimized for correlation coefficient, which is an indicator of how effectively a classifier identifies both positive (active) and negative (inactive) instances. All classifiers achieved correlation coefficients between 0.48 and 0.63, with accuracies ≥ 84%. For the practical application of identifying target sites for ZFPs that provide the greatest chance of success (for cases in which several potential target sites are available), it is appropriate to choose a classifier with a high specificity+ value, i.e., one that predicts a smaller number of "active" sites with higher confidence, rather than a high correlation coefficient per se.

Table 1 Performance of classifiers in predicting active OPEN target sites
Full size table

The receiver operating characteristic (ROC) curves in Figure 2 illustrate the tradeoffs between true positive rate (TPR), i.e., the percentage of active target sites correctly predicted as such, and false positive rate (FPR), i.e., the percentage of inactive sites incorrectly predicted to be active, for the different target sequence encodings. Using the base counts and positional base counts encodings, the Naïve Bayes and SVM classifiers gave similar results. Based on the Area Under the Curve (AUC) of the ROC curves, the best overall results were obtained using the sequence identity encoding with the Naïve Bayes classifier (AUC = 0.89), which slightly outperformed the best SVM classifier (AUC = 0.84). We designate the sequence-based Naïve Bayes classifier, ZiFOpT, for Zinc Finger OPEN Targeter.

Figure 2
figure 2

Receiver Operating Characteristic (ROC) curves for Naïve Bayes and SVM classifiers.

Full size image

To ensure that the performance of ZiFOpT on ZFTS135 was not over-estimated due to over-fitting, we generated a second completely independent data set of experimentally validated ZFP target sites. ZFTS140 consists of 140 9-bp target sites that were chosen by experts as ideal candidates for OPEN selection (see Additional File 2 - Table S2). Active ZFPs were found for 122 of the 140 sites tested. On this dataset, ZiFOpT performance was comparable in terms of overall accuracy (88%) and specificity+ (92%), but with reduced ROC AUC (0.77). To assist users in choosing the best ZFP target sites, therefore, we also provide a confidence score derived from the posterior probability returned by ZiFOpT (see Methods), which allows users to rank the predicted active target sites. As shown in Table 2, choosing potential targets with confidence scores ≥ 6 (as opposed to scores < 6) results in improved accuracy (90% vs. 67%), specificity+ (90% vs. 73%) and sensitivity+ (100% vs. 85%).

Table 2 Performance of ZiFOpT on an independent test set (ZFTS140)
Full size table

Due to the large number of potential OPEN target sites for most genomic targets of interest, it is desirable to identify a subset of target sites with the greatest chance of success. Currently, OPEN pools are available for 26 triplets in position 1, 21 triplets in position 2, and 23 triplets in position 3 of a 3-finger ZFP. Hence OPEN can, in theory, target 12,558 distinct sites. Because 415 of these sites are not targetable due to dam or dcm methylation, 12,143 distinct 9-bp ZFP target sites are currently targetable. The ZiFOpT classifier, when optimized for correlation coefficient, predicts that 8,412 (69%) of these sites will be active target sites. For ZF nuclease sites, which consist of two ZF array sites, OPEN can theoretically target a total 147,452,449 distinct nuclease sites (assuming a fixed number of nucleotides between the arrays). ZiFOpT predicts that only 70,761,744 (48%) of these nuclease sites will have two active sites.

An analysis of recently published OPEN ZFN sites in zebrafish [9] illustrates the value of ZiFOpT in reducing the experimental effort required to target a large number of genomic transcripts. In the previous study, at least one potential OPEN nuclease site was identified within the first three coding exons in ~86% of zebrafish transcripts [9]. As shown in Table 3, using a classification threshold that corresponds to a confidence score > 4 for the active sites (24% predicted FPR), ZiFOpT predicts that 15,565 (53%) of all zebrafish transcripts can be targeted successfully using OPEN. By restricting targets to those identified by ZiFOpT at a higher confidence score (> 8), the number of potential target sites for experimental testing could be reduced from 114,392 to 10,515, i.e., by ~ 90%. Thus, for functional genomic studies, ZiFOpT is a valuable tool for identifying sites most amenable to targeting by ZFNs. Indeed, we have used ZiFOpT to predict activity for all 315,186 OPEN ZFN targets previously identified in zebrafish [9]. These results are presented in Additional Files 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26 and 27.

Table 3 Summary of zebrafish OPEN ZFN target sites, classified by ZiFOpT
Full size table

Discussion

Detailed analyses of available high resolution structures for DNA-protein complexes support the conclusion that there is no simple general code for DNA-protein recognition [36]. For certain classes of DNA binding proteins, including the C2H2 zinc finger proteins, it may be possible to decipher some of the rules that govern protein-DNA recognition by exploiting the increasing availability of data regarding sequence determinants of binding affinity and specificity. For example, Stormo's group has utilized contact propensities and weight matrices to predict which target sites a zinc finger motif is most likely to bind [27, 37]. Recently, Singh and colleagues utilized SVMs to predict whether a specific zinc finger protein will bind a specified target site [38]. Methods such as these utilize binding information for specific ZFPs interacting with a limited number of DNA target sites. In contrast, DNA microarray based experiments provide binding preferences of a transcription factor for thousands of potential sites [39–42]. These experiments should provide additional data for predicting and assessing transcription factor binding site models, including those for zinc finger proteins.

In the current study, we propose an approach for predicting whether a ZFP can be engineered to bind a specific DNA sequence without a priori knowledge of the ZFP amino acid sequence. We analyzed base composition features and position-specific base propensities in a dataset of 135 different DNA target sites for which the OPEN selection method had been experimentally attempted. Our goal was to use this information to develop a rapid and reliable machine learning classifier to identify DNA sequences most amenable to site-specific targeting by zinc finger arrays generated using the OPEN design procedure. Based on our results, we developed a server-based application, ZiFOpT, which implements a sequence identity-based Naïve Bayes classifier, and identifies active OPEN target sites with an estimated accuracy of 87% and ROC AUC of 0.89, when evaluated using cross-validation and optimized for correlation coefficient. ZiFOpT performance on an independent test set of 140 experimentally validated ZFP targets was lower in terms of AUC (0.77), as expected, due to the more challenging nature of this performance test. Importantly, confidence scores derived from posterior probabilities computed by ZiFOpT are provided for each predicted ZFP target site, allowing users to rank potential target sites and focus on those with the highest probability for success.

In our statistical analysis of active versus inactive target sites, we detected biases in position-specific base composition of ZF targets (Figure 1). Thus, we anticipated that classifiers in which we attempted to capture base count biases or position-specific base propensities in the sequence encoding might perform as well as those using sequence identity, particularly in light of the size of the dataset relative to the size of the feature space for the sequence identity representation. For the Naïve Bayes classifier, however, sequence identity outperformed positional base counts and gave the best overall performance, in terms of the AUC of the ROC curve (0.89). For the SVM classifier, using positional base counts as input did provide substantially better performance than sequence identity (0.84 vs. 0.76). Because the dataset used to train the SVM classifiers was smaller (to ensure a balanced number of positive and negative instances, see Methods), this difference in performance may be partly attributable to relatively sparse data for the sequence identity encoding.

Although the OPEN procedure tests only a small fraction of the total theoretical protein sequence space for the zinc finger recognition helix, it generates up to approximately 1 million ZFP combinations, clustered in what are expected to correspond to regions of optimal amino acid sequence space for the DNA target site of interest. Together with the results summarized in Figure 1, this suggests there are utilizable constraints on the DNA sequence space for 9-bp target sites that can be successfully targeted by ZFPs engineered by OPEN. For example, the results in Figures 1B and 1C indicate that increased thymine content in target sites, especially at positions 1 and 7, may preclude high affinity or high specificity binding. Previous studies have suggested that ZFP recognition sites with a relatively high purine nucleotide content are more often active targets for engineered zinc finger proteins [28, 29]. These earlier conclusions were based on analysis of target sites containing predominantly GNN-triplets and for ZFPs generated using modular assembly. The current analysis confirms and quantifies the contributions of high purine content as an important determinant of success for sequences targeted using OPEN. More specifically, our analyses indicate that for three-finger ZFPs, it is advisable to avoid target sites containing many thymine bases.

Based on the results reported here, ZiFOpT will be valuable for guiding investigators using OPEN to ZFN target sites with the greatest opportunities for success. The calculations shown in Table 3 illustrate the potential reduction in experimental effort that could be achieved by using ZiFOpT to identify ZFP target sites for every protein encoded by the zebrafish genome. Also, ZiFOpT should be valuable for selecting targets among the 695,819 total OPEN nuclease targets identified in protein-encoding transcripts of the human genome (Ensemble V51.1) [D. Reyon and J. Sander, unpublished], and could assist investigators who wish to apply OPEN technology to target specific genes or genomic regions of interest in other organisms. ZiFOpT classifies potential target sites for OPEN-generated ZFPs as "active" or "inactive" and provides a confidence score for the prediction. ZiFOpT is freely available and incorporated in the Zinc Finger Targeter web server (http://bindr.gdcb.iastate.edu/ZiFiT) [43, 44]. ZiFiT can scan a given DNA sequence of interest and identify every potential DNA site targetable by OPEN. With the integration of ZiFOpT, users will be able to evaluate the expected success rate of OPEN for target sites identified by ZiFiT.

Conclusion

In this study, we developed machine learning classifiers that reliably identify DNA sites highly amenable to targeting by the OPEN zinc finger protein engineering method. Analysis of a dataset of 135 experimentally validated ZFP binding sites identified high thymine content as a significant barrier to effective targeting by OPEN. In addition, comparison of results obtained using three different target sequence encodings as input for Naïve Bayes and SVM classifiers suggested that positional context plays a significant role in ZFP target site recognition. Importantly, however, a simple encoding based on sequence identity is sufficient to identify the most promising ZFP target sites, with ~87% accuracy. As more ZFP functional data become available and we learn more about the sequence composition of fingers in OPEN pools, our predictions should improve. At present, the ZiFOpT classifier presented here is expected to reduce the experimental effort required to identify an active ZFP-target site pair by ~75%, compared with selection of target sites without classification. By restricting experimental targets to "active" OPEN sites predicted with highest confidence, experimental success rates should be significantly enhanced. This in turn should accelerate the application of zinc finger proteins as tools for precise genetic manipulation in basic genomics research as well as in gene therapy.

Methods

Definition of active and inactive ZFP target sites based on B2H assays

An active target site is a 9-bp DNA sequence for which the OPEN procedure has been used successfully to obtain at least one ZFP capable of binding the site with sufficient affinity and specificity to provide three-fold activation in a bacterial 2-hybrid (B2H) assay, i.e., to induce production of β-galactosidase by at least three-fold above the basal level of induction obtained using control constructs that lack the cognate ZFP target site [8, 23, 29]. An inactive target site is a 9-bp DNA sequence for which none of the corresponding OPEN-generated ZFPs tested were capable of producing a three-fold activation in the B2H assay.

Datasets of experimentally validated ZFP-target sites

ZFTS135 (cross-validation dataset)

A zinc finger target site dataset generated from a group of 135 potential 9-bp zinc finger target sites (ZFTSs) that have been experimentally targeted using OPEN. For each ZFTS in the dataset, ZFPs have been selected using OPEN [8] and evaluated for DNA-binding activity in vivo using the B2H assay [10, 23, 29]. The sequences of all 135 ZFTS, together with their experimentally determined functional activity labels (active or inactive) are provided in Additional File 1 - Table S1. For 82 target sites in ZFTS135, functional activity labels, based on B2H assays, are reported here for the first time. The remaining 53 target sites, denoted by asterisks (*) were characterized previously [8, 23, 29] and experimental activity data were extracted from the Zinc Finger Database, ZiFDB (http://bindr.gdcb.iastate.edu/ZiFDB) [45].

ZFTS140 (independent test set)

This dataset is an independent group of 140 potential 9-bp ZFN target sites (none of which overlap with those in ZFTS135), which have been experimentally targeted using OPEN. These sites were chosen by experts in the field in order to generate a test set for rigorous evaluation of ZiFOpT performance. 122 (87%) of these sites were determined to be 'active' based on B2H assay results, as described above. The sequences of all 140 target sites, along with classification and confidence scores, are provided in Additional File 2 - Table S2.

Machine learning classifiers

Naïve Bayes is a probabilistic classifier that assumes the independence of each attribute and generates models that are amenable to user interpretation, usually without compromising performance [46]. We used the implementation available in the WEKA package version 3.5.7 [47]. For each instance, the classifier returns a classification of either "active" or "inactive" based on the posterior probability (Bayes' rule). The value of the classification threshold (θ) can be selected based on the desired trade-off between sensitivity and specificity. We evaluated several classification performance measures (see below), using a standard leave-one-out cross validation procedure.

Support Vector Machines (SVMs) find a hyperplane in high-dimensional space that maximizes the distance between the different classes of data in that space. We implemented the SVM classifier using the wrapper class available for LIBSVM [48]. We tested several different kernel functions. Best results were obtained using the radial basis function (RBF) kernel. Optimal cost and gamma parameters were determined using a grid search algorithm. Because SVM classifiers are sensitive to the number of positive and negative instances in the training set, and because our dataset is unbalanced (106 positive and 29 negative instances), we used a variation of the standard leave-one-out cross validation technique. For each test case, we removed that instance and generated 10 randomized balanced training sets. The probability assigned to each test case was an average of the probability estimate generated from 10 randomized balanced training sets.

We also tested several other types of classifiers, including Decision Trees, and obtained results that were either comparable to or significantly worse than those obtained using ZiFOpT. Among the several Decision Tree algorithms we tested, the Logistic Model Tree (LMT) classifier performed the best with an AUC of ROC of 0.86.

Target site sequence encoding

For each classifier, three different input sequence encodings were evaluated. The sequence identity input window consists of a target site represented as a 9 nucleotide DNA sequence, reading in the 5' to 3' direction on one strand (e.g., GTTGACGGC). The base counts input window consists of four single-digit values that represent the number of occurrences of each of the four DNA bases (G, A, C, T) within a target site (e.g., 4,1,2,2 for the target site in the preceding example). The positional base counts input window consists of a string of 12 values (3 sets of 4 digits), ranging from 0 to 3 and representing the number of times each base occurs in the first, second, and third positions within a triplet (e.g., 3,0,0,0;1,1,0,1;0,0,2,1, for the target site in the preceding example, in which G occurs in the first position of a triplet 3 times, once in the second and 0 times in the third.).

Classification performance measures

We used several standard performance measures: accuracy, correlation coefficient (CC), specificity+ , and sensitivity+, and the AUC for standard ROC curves as described by Baldi et al. [49]. Here True Positives (TP) is the number of validated targets correctly predicted to be "active" target sites, i.e., sites that have been targeted successfully by an OPEN-generated ZFP to produce > 3-fold activation in the B2H assay; False Positives (FP) is the number of "inactive" target sites incorrectly predicted to be "active" sites; True Negative s (TN) is the number of "inactive" target sites correctly predicted as such; False Negatives (FN) is the number of "active" target sites incorrectly predicted to be "inactive" sites.

Accuracy = T P + T N T P + F P + T N + F N
CC = T P * T N − F P * F N ( T P + F N ) ( T P + F P ) ( T N + F P ) ( T N + F N )
Specificity + = T P T P + F P
Sensitivity + = T P T P + F N
False Positive Rate  ( FPR ) = F P F P + T N
True Positive Rate  ( TPR ) = T P T P + F N

A Receiver Operating Characteristic (ROC) curve displays the tradeoff between the true positive rate (hit rate) and the false positive rate (false alarm rate) for different discrimination thresholds [49]. The Area Under the Curve (AUC) of the ROC plot is valuable for comparing performance of different classifiers because it portrays the tradeoff between the false positive rate and the true positive over the range of classification threshold values.

Confidence Score

The posterior probability returned by ZiFOpT for classifying each target site was used to generate a confidence score. Target sites with posterior probability were classified 'active' if they had posterior probability ≥ 0.5 and 'inactive' otherwise. For the 'active' class, the posterior probability was transformed to a scale from 0 to 9 by incrementing the confidence score by 1 as the posterior probability increased by 0.05 above 0.5. Therefore, a posterior probability of 0.75 corresponds to an 'active' classification with a confidence score of 5. For the 'inactive' class, the confidence score was incremented by 1 as the posterior probability decreased by 0.05 below 0.5. Therefore, a posterior probability of 0.25 corresponds to an 'inactive' classification with a confidence of 5.