Computer Tools and Resources for CRISPR Gene Editing

The nucleotide composition of a sgRNA, particularly GC content, is essential to determine its efficiency and specificity. One of the most important applications of CRISPR/Cas tools is to perform whole-genome screening for gene functional analysis [31], which also provides important information for uncovering nucleotide preference of sgRNAs. Based on analyzing the data of 1841 sgRNAs designed for targeting endogenous mouse and human genes, Doench and colleagues developed a predictive model (named Rule Set 1, which is based on sgRNA sequence features) to clarify general rules for designing highly active sgRNAs [40]. After quantification of the sequence features correlated with the activities of sgRNAs, they found that the GC content of a sgRNA did not display a positive correlation with the sgRNA activity in genome editing; both high and low GC contents of sgRNAs led to less efficient genome editing [40]. A similar rule was also identified in performing genome-scale functional screens using human cells and zebrafish [31], [41]. Additionally, several large-scale datasets suggest that the type of nucleobase is important for sgRNA activity [40], [42]. The nucleotide at the position 20, located immediately upstream of PAM, is a key determinant. Guanine was highly favorable whereas cytosine was strongly unfavorable [31], [40], [41]. In contrast, the position 16, the last nucleotide of the seed region, preferred cytosine over guanine [40], [42]. Theoretically, the transcription of sgRNAs relies on RNA polymerase III that recognizes uracil-rich sequences for termination [43], [44]. The uracil-rich sequence structure might lead to early termination of sgRNAs and then impair expression [42]. Thus, sgRNA sequences with thymine-rich nucleobase are not favorable at their 3′ end region. Additionally, adenine is preferable in the middle of a sgRNA, whereas cytosine has negative effects at the position 3 [31], [40].

Zebrafish is an ideal model organism for performing large-scale analysis of sgRNA activity. To dissect the sgRNA molecular features affecting the efficacy of CRISPR/Cas9 in vivo, a sgRNA pool was constructed by introducing 1280 sgRNAs to target 128 genes in the zebrafish genome [45]. The researchers found that sgRNA stability in vivo plays a critical role in determining sgRNA activity. The formation of a guanine-quadruplex structure, which contains at least eight guanines, can significantly increase sgRNA stability. Additionally, several sequence features were identified by statistical analysis of the most efficient sgRNAs, such as guanine enrichment in the region of positions 1–14, cytosine enrichment between the position 15 and the position 18, and overall depletion of thymidine and adenine except the positions 9 and 10 [45]. Taken together, a linear regression-based predictive sgRNA-scoring algorithm, named CRISPRscan (http://CRISPRscan.org↗), was proposed for detecting the most active sgRNAs in vivo[45]. The CRISPRscan model is also implemented in other web-based sgRNA design tools, such as CHOPCHOP [46], [47] and CRISPOR [38].

Given the hypothesis that sgRNA activity could be influenced by several other features, such as the position-independent nucleotides, the location of the target sites in the gene, and the thermodynamic property of a sgRNA, the Rule Set 1 predictive model was further improved by integrating new prediction algorithms and generated “Rule Set 2”. It employs the improved algorithms for on- and off-target activity prediction, and the gradient-boosted regression tree model with the augmented feature set trained on the combined dataset, which is used not only for sgRNA libraries for general genome editing purposes (gene knockout and knockin) but also for CRISPR activation (CRISPRa) and CRISPR interference (CRISPRi) [37]. The Rule Sets 1 and 2 were widely implemented in many websites and computational tools for designing sgRNAs, including CHOPCHOP [46], [47], CRISPOR [38], GPP sgRNA Designer [48], and E-CRISP [49].

Some other factors also affect Cas nuclease binding and cleavage. It has been suggested that both sequence composition and locus accessibility are important to determine sgRNA activity, which subsequently influence the sgRNA design tools, such as sgRNAScorer [50], [51]. Additionally, chromatin accessibility [52], [53], [54], [55] and asymmetric sgRNA–DNA interactions also affect CRISPR/Cas cutting specificity [37], [56]. Currently, many groups have integrated these algorithms into their web-based applications, such as DeepCRISPR, CRISTA [56], [57], predictSGRNA [58], and uCRISPR [59]. GuidePro is a two-layer ensemble predictor for sgRNA efficiency prediction that enables the integration of multiple factors for the prioritization of sgRNAs for gene knockout [60].

Prime editing is a new application of CRISPR/Cas technology in which a small-sized genetic sequence is altered without requiring a donor DNA template. In prime editing system, a prime editing guide RNA (pegRNA) is used to replace the traditional sgRNA, which contains a primer binding site (PBS) and a reverse transcriptase (RT) template sequence. After nCas9 cuts a target DNA sequence, the PBS sequence will be elongated and inserted into the original DNA sequence for DNA replacement [61]. Thus, prime editing can be used to repair any nucleotide error without a DNA template. Due to these advantages, prime editor has huge potentials for genome editing. However, evaluation of prime editing efficiency is time- and lab-intensive. To solve this problem, Kim and colleagues used deep learning to create a precise computational model for measuring the efficiency of pegRNAs based on high-throughput evaluation of 54,836 pegRNA–target pairs in human cells [62]. More importantly, this computational tool and resources can be found in their publicly available website http://deepcrispr.info/DeepPE/↗.

One of the main concerns about sgRNA design is off-target effects that are normally generated by unexpected cleavage at genomic sites similar to the target sequences [33], [63]. Thus, traditional short sequence alignment tools, such as Burrows-Wheeler Alignment Tool (BWA) and Bowtie [64], [65], [66], have been used to predict potential off-target sites [38], [49]. Given that BWA and Bowtie are originally designed for aligning short DNA reads to large reference genomes [64], [65], there are several innate defects for predicting off-target effects. For instance, CRISPR/Cas has been suggested to tolerate more mismatches than traditional BWA or Bowtie alignment allows [33], [67], [68]. Additionally, nucleotide positions are important for target specificity, and atypical PAM could be recognized by CRISPR/Cas9 as well [33], [37]. To overcome these problems, many improved off-target prediction tools have been reported. For example, CCTop can predict potential off-target sites with four mismatches differently distributed in the targeted genomic sites [39], and Cas-OFFinder is not limited by the number of mismatches and allows variations in PAM sequences [67].

To predict off-target sites more accurately, several computational models were built based on large amounts of experimental data. After evaluating more than 100 predicted genomic off-target loci in two human embryonic kidney cell lines [33], several rules were proposed to minimize off-target effects, including that (1) the potential off-target sequences should not be followed by a PAM with either a 5′-NGG or 5′-NAG sequence, and (2) the minimum mismatches between sgRNA and potential off-target sites should be limited to 3 nt and at least two mismatches are better in the proximal PAM region. These rules have been implemented in their specificity score tool, termed MIT, which has subsequently been implemented in web-accessible applications, such as CHOPCHOP [46], [47] and CRISPOR [38]. Another commonly used specificity score tool is Cutting Frequency Determination (CFD), proposed by Doench and colleagues [37]. In addition to mismatch position of sgRNA and atypical PAM effect, the identities of mismatched nucleotides and insertion and deletion (indel) variants can significantly affect sgRNA activity. CFD has been shown to predict most off-target sites and exhibit better performance than MIT and CCtop by using GUIDE-seq, an unbiased experimental method for detection of sgRNA off-target effects [69]. CFD has been implemented in CRISPOR, GPP sgRNA Designer, GUIDES, and other web-related tools.

Currently, there are many computational programs for designing sgRNAs and predicting their genome editing efficiency and specificity. To comprehensively benchmark these techniques and tools, several available on-target design tools, genome-wide off-target cleavage site (OTS) detection techniques, and in silico genome-wide OTS prediction tools have been systematically evaluated [70], [71]. A one-stop platform, named integrated Genome-Wide Off-target cleavage Search platform (iGWOS), was constructed by integrating these available OTS prediction algorithms and datasets [70], [71].

CRISPOR provides multiple tools that include efficiency prediction, specificity prediction, and a primer design tool for vector construction as well as on-target and off-target detection. CRISPOR incorporates almost all empirical algorithms for predicting efficiency, such as Rule Set 2 [37], [40], CRISPRscan [45], Wang et al. [31], Chari et al. [51], and Xu and coworkers [35]. They also apply “deepCpf1” and “Najm et al.” to predict Cas12a and SaCas9 efficiencies [88], [89], [90], respectively. The predicted results are well visualized by these models. For specificity prediction, CRISPOR includes MIT and CFD that are two mainstream specificity prediction tools. CRISPOR also integrates two CRISPR/Cas outcome predictive models, out-of-frame score and frameshift ratio [84], [85], to further reduce cutting efficiency. In addition, several critical factors such as the GC content and the type and number of mismatches (0–4 nt) are labeled in the results. CRISPOR covers hundreds of organisms. Different nucleases and PAM types are also available for selection. These features allow the majority of researchers to use CRISPOR for designing different CRISPR/Cas genome editing experiments.

CHOPCHOP is also a comprehensive website for sgRNA design. Both CRISPR/Cas and transcription activator-like effector nuclease (TALEN) systems are supported by CHOPCHOP. Additionally, CHOPCHOP provides various targeting systems, such as knockout, knock-in, gene activation, and gene repression. Similar to CRISPOR, CHOPCHOP also provides multiple predictive models, and the user can choose one of them to predict cutting specificity and efficiency. In addition, CHOPCHOP has a “Custom PAM” option that is convenient for choosing different PAM sequences. It has been reported that cell types may affect the DSB repair pathway and then influence CRISPR/Cas genome editing outcomes [91], [92]. Several cell types, including mESC, U2OS, HEK293, HCT116, and K562, are optional in the CHOPCHOP website for accurate outcome prediction. It is also important that CHOPCHOP is compatible with more than 200 genomes. It allows researchers to design sgRNAs in a specific region of a gene, such as 5′ UTR, 3′ UTR, promoter, or the coding region.

CRISPR RGEN Tools is a CRISIPR/Cas library platform that contains multiple sgRNA design tools. For example, CRISPR RGEN Tools employs Cas-designer for conventional CRISPR/Cas nucleases, BE-Designer for CRISPR base editing, and PE-Designer for CRISPR prime editing [93]. In addition, PE-Designer only allows for SpCas9; both Cas-designer and BE-Designer have wide PAM compatibility. More than 100 organisms are well organized in those three tools. Microhomology-Predictor is an outcome-predictive tool that introduces out-of-frame score algorithm to evaluate potential in-frame deletions caused by the MMEJ repair approach [84]. In addition to CRISPR/Cas, this tool also supports other programmable nucleases, such as zinc finger nucleases (ZFNs) and TALENs, and an out-of-frame score over 66 is recommended. Thus, a user can utilize those tools to implement different experimental purposes; it is also helpful for designing sgRNAs with high accuracy.

CRISPR-GE is a web-based tool for designing sgRNAs in plants [94]. CRISPR-GE covers 41 plant genomes, including several agriculturally important crops, such as rice (Oryza sativa japonica), corn (Zea mays), and grape (Vitis vinifera). This tool also includes multiple Cas nucleases, such as SpCas9, FnCas12a, and AsCas12a, for helping the users to design sgRNAs for different CRISPR/Cas systems. Additionally, CRISPR-GE provides a “User defined” option that allows the users to customize PAM sequences (including 5′ and 3′ PAMs) and the length of target sites. CRISPR-GE provides warning notes for indicating “bad site”, such as very low or very high GC contents, poly-T site(s), and contiguous base-pairing with a sgRNA. CRISPR-GE implements CFD model to predict the specificity of a target site. CRISPR-GE also provides a primer design tool to assist vector construction and mutant detection.

CRISPR-P is another web-based tool for designing sgRNAs for plants [95], [96], which covers 75 plant genomes and the majority of them are important grain crops. Compared with CRISPR-GE, there are more CRISPR/Cas PAM types available in CRISPR-P, which include NGG (SpCas9), NNAGAAW (St1Cas9), N₄GMTT (NmCas9), NNGRRT (SaCas9), and NG (xCas9). Additionally, CRISPR-P allows the users to choose U3 or U6 sgRNA promoter-driven expression cassettes for designing sgRNAs. The users can input gene ID/name, position on scaffold and chromosome, and fasta format sequences for submitting data. CRISPR-P implements Rule Set 1/2 and CFD to predict on-target and off-target effects. The sgRNA predictive outputs are well visualized, which includes sgRNA GC content, restriction endonuclease site, secondary structure of sgRNA [97], and microhomology score [84].

AsCRISPR is a comprehensive web tool for designing sgRNAs for allele-specific genome elements, which can be used to discriminate between alleles. This tool is specifically designed for targeting dominant single nucleotide variants (SNVs) retrieved from ClinVar and OMIM databases [98]. In this publicly available web tool, several Cas enzymes, such as SpCas9, AsCas12a, and Cas12v, as well as CasX and their variants, can be selected. Currently, this web tool is only for targeting SNVs in the human and mouse genomes.

SNP-CRISPR is a web-based computational program for designing sgRNAs based on public variant datasets or user-identified variants [99]. It can be used for both model species and non-reference genomes as well as across varying genetic backgrounds, particularly for SNP-containing alleles. SNP-CRISPR also calculates the efficiency and specificity scores for sgRNA designs targeting both the variants and the reference.

PnB Designer is a web-based tool for designing sgRNAs for both prime and base editors, two newly developed CRISPR/Cas genome editors [100]. PnB Designer design sgRNAs for both single and multiple genome targets on several different plant and animal species.

Sequence scan for CRISPR (SSC; https://cistrome.org/SSC/↗) is one online web server for scanning sgRNA spacer [35]. It is not only for designing sgRNAs for CRISPR knockout but also for CRISPR inhibition or activation with sgRNA efficiency prediction.

In addition to academic-developed publicly available computational tools, certain CRISPR companies have also developed several useful computational tools and resources for the public. These design tools include, but are not limited to: IDT (https://www.idtdna.com/site/order/designtool/index/CRISPR_CUSTOM↗), Horizon (https://horizondiscovery.com/en/ordering-and-calculation-tools/crispr-design-tool↗), and Synthego (https://www.synthego.com/products/bioinformatics/crispr-design-tool↗).

To enable easy quantification of CRISPR/Cas9 genome editing products, several new methods have been developed by directly decoding Sanger sequencing data (Table 2). For example, tracking of indels by decomposition assesses (TIDE) is a decomposition algorithm that is able to precisely determine the indel spectrum and frequency of targeted mutations generated by CRISPR/Cas9 genome editing [103]. It is a very simple and effective method to assess the efficiency of well-performing sgRNAs. It only requires standard molecular biology reagents and involves three steps, including a standard PCR reaction, Sanger sequencing, and decoding raw sequencing data by the TIDE web tool. The algorithms accurately reconstruct the spectrum of indels from the sequence traces. The web tool reports the identity of the detected indels and their frequencies [104]. Moreover, it is highly effective to predict indels with all sizes in sample clones as well as tracing indels in heterozygotes [102]. TIDE has been further designed to decompose the sequence data produced by template-directed CRISPR/Cas genome editing experiments [105]. Since the majority of CRISPR/Cas-induced mutations in plants are biallelic (two distinct variations), homozygous (two identical mutations), and heterozygous (wild-type/single mutation) [106], Liu and colleagues established a web-based tool, termed DSDecode, to automatically decoding the superimposed sequencing chromatograms of CRISPR/Cas PCR products [107].

With rapid adaptation of genome editing technology, massively parallel sequencing methods have been employed for assessing CRISPR/Cas post-experimental data. Evaluation of targeted sequences by next-generation sequencing (NGS) strategies has been developed for deeper quantification of targeted amplicon sequences. The CRISPR Genome Analyzer (CRISPR-GA) evaluates the NGS dataset and quantifies and characterizes the indels and homologous recombination events [108]. NGS also provides information regarding the selected locus, including quantification of edited-sites and other mutations detected. After scanning the reads, locating indels, and computing the allelic replacements, CRISPR-GA provides a combined report-card to the user which includes all potential information about genome editing events. Similarly, CRISPResso2 and Cas-Analyzer also provide web-accessible tools for evaluating deep sequencing outcomes of CRISPR/Cas genome editing experiments [109], [110]; CRISPResso2 also provides specific optimizations on analyzing base editing outcomes [110].

Current computation languages, such as Python and R, play a significant role in efficiency enhancement of several bioinformatic tools, which have been used to accurately detect modifications in the edited genomes by the NGS datasets. For example, “CRIS.py” is a simple and highly versatile program, which analyzes NGS data, and identifies knockout and multiple user-defined knock-in alterations from one and up to thousands of CRISPR/Cas9-edited samples [111]. CrispRVariants provides an R-based toolkit that is feasible to evaluate and visualize mutant allele types, locations, and frequency [112]. The repair outcomes of CRISPR/Cas9-generated DSBs were recently extensively studied in human primary T cells, in which Leenay and colleagues sequenced the repair outcomes at 1656 on-target genomic sites [113]; then, they used the sequencing data to develop and train a machine learning model, termed CRISPR Repair OUTcome (SPROUT). SPROUT incudes all the datasets generated from the 1656 CRISPR on-target sites and can be used to predict the length, probability, and sequences of indels generated by CRISPR/Cas9 [113]. In another study, Wang and colleagues collected 13 datasets obtained from previously reported different CRISPR/Cas genome editing experiments in six different species, including human, mouse, zebrafish, Drosophila, Ciona intestinalis, and C. elegans; after machine learning and featurization by eight different models, they developed an algorithm, called GNL-Scorer, for predicting CRISPR target activities [114]. GNL-Scorer, both GNL and GNL-Human, is a computational model based on the Bayesian Ridge Regression (BRR) model, which combines optimal datasets and features to address the cross-species problem. Both SPROUT and GNL-Scorer computational tools and resources will promote CRISPR sgRNA design and enhance the application of the CRISPR/Cas-based genome editing. BATCH-GE is another easy-to-use computational tool for identifying CRISPR/Cas-derived indel mutations and other precise genome editing events, including both on- and off-target impacts by analyzing huge data generated by deep sequencing technology [115], [116].

Given the size and diversity of data generated by pooled CRISPR/Cas9 screens, the majority of conventional methods are not sufficient to evaluate the huge datasets generated by pooled CRISPR/Cas9 screens. To this end, several algorithms have been specifically developed for interpreting raw sequencing outputs of CRISPR/Cas9 screens, such as Model-based Analysis of Genome-wide CRISPR/Cas9 Knockout (MAGeCK) [117], caRpools [118], HiTSelect [119], Screening Bayesian Evaluation and Analysis Method (ScreenBEAM) [120], Bayesian Analysis of Gene Essentiality (BAGEL), and Permutation Based Non-Parametric Analysis of CRISPR/Cas9 screen data (PBNPA) [121], [122]. Since these analysis methods were developed for persons skilled in bioinformatics, it is difficult for many biologists or researchers with less programming background to implement them. To simplify analysis procedure, web-based interfaces have been developed to enable the users to evaluate pooled CRISPR/Cas9 screening data. CRISPRAnalyzeR is the first end-to-end analysis pipeline that integrates eight different algorithms for identification of candidate genes. In addition, CRISPRAnalyzeR is constructed in R and can be easily installed locally [123]. PinAPL-Py workflow contains various statistical models, better sequence quality checks, automated sgRNA-seq extraction, precise sequence alignment, sgRNA enrichment or depletion analysis, and gene ranking facility [124]. Its workflow can deploy a variety of well-known sgRNA libraries as well as easily upload-able custom libraries. Importantly, it can analyze the multiple CRISPR/Cas-edited experiments. PinAPL-Py ranks both sgRNAs and genes, and it provides ready-to-publish plots. However, both CRISPRAnalyzeR and PinAPL-Py have several rate-limiting steps, such as long time for raw FASTQ file transfer and complicated parameter tuning for alignment. CRISPRCloud2 employs Amazon Web Service to decrease the covert time and satisfy data-privacy requirements. Additionally, an adaptive hash-mapping algorithm was introduced into CRISPRCloud2 to increase alignment speed and accuracy [125].

Off-target impact is one of the major challenges for CRISPR/Cas application in gene therapy and crop improvement as well as other areas, such as gene function studies. To reduce potential off-target impacts, many strategies have been developed, which include but are not limited to selecting high-affinity Cas enzymes, designing better sgRNAs, and using the right CRISPR/Cas reagent delivery system. However, identifying all potential off-targets is still a challenge. Identifying and quantifying unexpected genome targeting events are essential to assess the fidelity of genome editing tools as well as to guarantee the safety of gene therapeutic applications. Currently, NGS has been proved as a reliable technology to identify all potential off-target impacts as well as targeted and cleaved genome sites. However, NGS generates a vast number of reading sequences that require special computational programs to identify off-target sequences. To solve this problem, in the past several years, several research laboratories have developed computational tools that can highlight off-target activities besides the edited DNA sequences in the genome by using NGS (Table 2). Crosetto and colleagues presented a method called “direct in situ breaks-labeling enrichment on streptavidin and next-generation sequencing (BLESS)” that scans the DSBs at the whole-genome level by using Instant-seq software for Illumina sequencing data [126]. The efficiency of BLESS was tested in human and mouse cells by using various DSB-inducing reagents and sequencing platforms. The aforementioned method can identify telomere ends, Sce endonuclease-induced DSBs, and complex genome-wide DSBs. In human cells, the identified mutations (> 2000) were in the form of un-evenly distributed aphidicolin-sensitive-regions (ASRs) that was the principal proof of utilization of BLESS at the whole-genome level. Genome-wide unbiased identification of DSBs enabled by sequencing (GUIDE-seq) is an experimental approach for global detection of DNA DSBs for identifying off-target cleavage generated by Cas nucleases and potentially other nucleases, such as TALENs [69]. During identifying off-target sequences by GUIDE-seq, the authors customized a bin-consensus variant-calling algorithm based on molecular index and SAMtools; this computational program distinguishes off-target sequences from the reference sequences. This method can be used to detect off-target cleavage activities that previous computational methods or chromatin immunoprecipitation sequencing (ChIP-seq) could not detect. GUIDE-seq also detects Cas-independent genomic DSB hotspots. Giving that linear double-stranded integrase-defective lentiviral vectors (IDLVs) possesses the propensity of integrating preferentially into nuclease-induced DSBs by NHEJ repairing pathway, it has been employed to detect CRISPR/Cas-induced off-target cleavages with a very low frequency of 1% [127]. IDLVs also shows that Cas9 protein induces frequent off-target cleavages at 1-bp bulge or up to 13-bp mismatches between the sgRNA and its genomic DNA target, which may help in refining sgRNA design [127]. Circularization for in vitro reporting of cleavage effects by sequencing (CIRCLE-seq) identities off-targets at the genome-wide level by mapping the paired-end read sequences for searching off-target sites using bwa mem and samtools mpileup. This NGS and computational approach can be used not only for organisms with reference genome sequences but also for organisms without reference genomes [128]. However, off-target discovery methods using purified genomic DNA/specific cellular models are not capable of direct-in-vivo detection. To overcome this issue, a recently developed universally applicable approach called “discovery of in situ Cas off-targets and verification by sequencing (DISCOVER-Seq)” can be used to detect off-target effects in vivo[129]. This unbiased off-target identification approach recruits the DNA repair factors both in cells and organisms. By tracking these factors as “MRE11” [a subunit of the MRE11–RAD50–NBS1 (MRN) complex, which is tightly distributed around the Cas9 cut site], this program can detect off-target activities with single-base resolution. Moreover, DISCOVER-Seq works with several sgRNA formats and different types of Cas proteins that enable the characterization of new genome editing tools. Based on large-scale data analysis and a machine learning model, Lazzarotto and colleagues developed a “circularization for high-throughput analysis of nuclease genome-wide effects by sequencing (CHANGE-seq)” method for measuring the genome-wide activity of Cas9 in vitro, which includes both genetic and epigenetic impacts as well as off-target effects. Using this method, the authors identified 201,934 off-target sites from 110 sgRNA targets across 13 therapeutically relevant loci in human primary T cells [130]. From this study, they also observed that CRISPR/Cas9-induced off-target impacts were more likely to occur near active promoters, enhancers, and transcribed regions. With the rapid development of these NGS-based off-target detection approaches, more data can be produced from living therapeutic cells, which will boost the evolution of machine learning models and enhance alignment algorithms for identifying off-target impacts of CRISPR/Cas at the whole-genome level.