Resources for the design of CRISPR gene editing experiments

Delivery of Cas9 and sgRNAs and Cas9 activity

The gene encoding S. pyogenes Cas9 (SpCas9) can be introduced by transfection or viral transduction with a Cas9 expression construct or by direct delivery of Cas9 protein [29–34]. Furthermore, a germ-line Cas9 mouse has been generated, providing a source of animals and primary cells in which Cas9 expression is already established [35, 36]. Delivery of Cas9 by transfection can be quite efficient in many cell types; frequently employed expression vectors include pX330-U6-Chimeric_BB-CBh-hSpCas9 and lentiCRISPRv2 [3, 37, 38] (available from AddGene). In hard-to-transfect cells, including many primary cell types, transduction with a viral vector provides an alternative, using, for example, lentiCRISPRv2. Furthermore, for pooled screening applications, each cell must receive only a single or small number of sgRNAs by treatment with a mixed sgRNA pool, and hence transduction is the only standard delivery option. Delivery of SpCas9, alone or together with an sgRNA, can be achieved with adeno-associated virus (AAV), retroviral or lentiviral vectors and is challenging owing to the generally poor viral packaging and titers of the 4-kb Cas9 gene. Whether employing transfection or transduction, Cas9 expression varies from cell to cell, and the levels also vary among cell lines. Transduced cells are typically obtained by selecting for a marker present on the Cas9 expression cassette. It is important to verify that the promoter construct employed is effective in the model of interest, and it can be helpful to grow clonal populations with empirically verified high Cas9 expression for subsequent experiments. By contrast, delivery of sgRNA oligonucleotides is relatively straightforward and can be achieved by transfection of plasmids or transduction with viral genomes driving sgRNA expression from the U6 promoter [2, 3]. Alternatively, sgRNAs can be delivered by transfection of in vitro transcribed sgRNA or chemically modified synthetic sgRNA [30].

It appears that most cell lines are amenable to CRISPR-based editing, but some cell types appear to exhibit low or no Cas9 activity even when Cas9 is expressed at high levels. In general, the factors that govern how uniformly the alleles in all the cells of a population receive the desired edit have yet to be fully teased apart and could include, for example, not only the Cas9 and sgRNA levels, but also Cas9 activity determinants such as localization, the kinetics of DSB formation, and the kinetics and fidelity of repair processes, all of which can vary across cells types. For the moment, the suitability of any particular model system of choice for CRISPR should be confirmed empirically.

A straightforward assay to assess CRISPR activity in a cell population involves transducing the cells with a cassette expressing both green fluorescent protein (GFP) and a validated high-efficacy GFP-targeting sgRNA [37] (available at AddGene). The cells are then analyzed by flow cytometry to determine the fraction of GFP-negative cells [37]. The parental line with no Cas9 should be uniformly GFP-positive, whereas a Cas9 line in which the cells are all active for CRISPR should be mostly GFP-negative. It should be noted that KO of a single GFP integrant can be considerably more efficient than targeting both alleles of an endogenous gene, so that this assay might represent a near-best-case scenario for KO rate. Additionally, the time required to achieve gene edits appears to depend on many factors, such as target gene, cell type, KO versus KI, and the levels of Cas9 and sgRNA. Generally, when feasible, it is necessary to wait a week or more following the introduction of Cas9 and sgRNA in order to accumulate edits in the targeted cells.

Target-site selection, sgRNA design

For CRISPR-based experiments, one must select a target site to achieve the desired modification. The Cas9 protein requires a PAM adjacent to the sgRNA homology region to achieve efficient Cas9 binding and DSBs. For gene KOs, there are typically many possible PAM sites from which to choose. Different sites can yield widely varying rates of gene KO, raising the question of how to predict activity in advance. Similarly, it is obviously desirable to predict which sgRNAs will be most specific to the intended target. Research is ongoing to determine criteria that predict sites favoring high activity and specificity. Here, we describe current criteria and tools for selecting sgRNAs.

Experimental evaluation of on- and off-target activity and clonal selection

As with all gene-perturbation technologies, various types of validation are needed to confirm the relationship between the perturbed gene and the phenotype, and to understand the observed phenotype and its mechanism. For CRISPR-based results, one useful validation experiment is to assess the genotype of the modified cells at the intended target site, and with respect to OT effects elsewhere in the genome. Numerous approaches have been employed (Fig. 3), and determining which ones to use and the degree of validation necessary can be challenging. Validation practices are currently far from standardized, but we will outline below some of the options and the key considerations in choosing a practical path to validating the link between the gene targeted for perturbation and the observed phenotypes.

Functional knockout of individual genes

The knockout of protein function for individual genes has been a powerful tool to determine the functional role of a gene in cell-based or in vivo models [9, 74, 75]. In this approach, a cell, tissue or animal model is assayed for phenotypic changes following the selective knockout of one or more genes. CRISPR has arguably become the go-to gene-perturbation technology to evaluate gene function, and CRISPR-based gene phenotyping has become an accepted standard for confirming gene function hypotheses. Before CRISPR technology, a workhorse for mammalian loss-of-function experiments was RNA interference (RNAi), but CRISPR approaches are now favored over RNAi for many or most applications, mainly owing to its dramatically improved target specificity. In addition to its improved specificity, CRISPR can provide complete functional knockout, which has the potential to generate stronger and more-uniform phenotypes than might arise from the varying degrees of incomplete loss of function achieved by RNAi. It should be noted that RNAi represents a fundamentally different type of gene perturbation than genomic DNA modifications, and this might, in some cases, offer important advantages (e.g., if reductions in the levels of transcripts more accurately model the biology of interest), but, for many experiments, CRISPR has supplanted RNAi approaches.

For small-scale gene KO experiments, the basic three issues of (i) reagent delivery and CRISPR activity in the cells of interest, (ii) efficiency of the desired edit(s), and (iii) specificity are all important. As in current CRISPR implementations the per-cell rate of CRISPR KO typically ranges from 30–60 %, it is not possible to produce genetically uniform cells without a step of single-cell cloning to isolate and identify lines that have been modified in the desired manner. As single-cell cloning is unavoidable to obtain uniformly edited cells and requires considerable effort, it is highly desirable to achieve high CRISPR efficiencies in order to minimize the number of clones needed to obtain the desired target-site modifications. How does one contend with OT effects? Again here, cell-to-cell heterogeneity is a problem, and the cost and effort associated with fully characterizing all possible OT modifications in every cell clone are impractical. A standard strategy has been to produce multiple distinct cell clones employing several distinct sgRNAs and target sites for each gene of interest. If these clonal lines all exhibit a concordant phenotype, one can assume it is due to the common on-target perturbation rather than OT effects that would generally differ among clones, especially if different sgRNAs were employed. A minimum of three effective sgRNAs per gene is recommended. The benefits of obtaining multiple good clones places an even higher premium on good design to minimize the clone picking required. When targeting a single or very small number of genes, it is practical to curate the sgRNA selection process manually and to factor in gene-specific knowledge for each gene to optimize the on-target site selection. This permits more flexibility than for larger-scale CRISPR applications for which computational tools must be fully automated and fast enough to evaluate hundreds or thousands of genes.

Large-scale KO screens

An increasingly common application of CRISPR-Cas9 technology is to functionally evaluate hundreds, thousands or all genes in the genome by a high-throughput screening approach. Genome-wide and genome-scale pooled screens have been successfully executed [10, 11, 76–80]. Particularly exciting with respect to these screens is the frequency of ‘multiple-hit’ genes for which most or all of the sgRNAs score strongly. In analogous RNAi screens, much lower concordance is observed among short hairpin RNAs (shRNAs) or small interfering siRNAs (siRNAs) targeting the same gene [10]. Furthermore, the validation rate of hits from these early CRISPR screens appears to be generally quite high (albeit with relatively few examples thus far), supporting the notion that these reagents will generally yield far more accurate hit-lists than RNAi.

Pooled screens require that cells with the hit phenotype can be enriched or depleted within the screened cell population. This is feasible for phenotypes that are distinguishable using FACS or by proliferation–viability (‘selections’). To perform such screens, a cell population is treated with a pooled viral library carrying many different sgRNAs. The cell population is transduced at low titer such that each cell receives a single sgRNA to knock out a different gene in each cell. At the end of the screen, genomic DNA is harvested from the hit-enriched cell population (e.g., a population that has been subjected to FACS for the hit phenotype), and PCR-sequencing is used to determine which sgRNAs were enriched among the hit cells and, therefore, by inference the list of genes whose KO produces the phenotype. CRISPR pooled-screen publications provide detailed descriptions of the methods employed [10, 11, 76]. Here we highlight several key considerations for design of sgRNA pooled screens.

To deliver Cas9 activity to the cell population to be screened, a Cas9 stable cell line can be established first, and the sgRNA pooled virus added later, or the Cas9 can be delivered simultaneously with the sgRNA. As Cas9 packages poorly and yields low viral titers, there is a practical advantage to making and expanding a stable cell line first, whereas combining the Cas9 and sgRNA in a single vector reduces the titer of the library pool, but does have the advantage of permitting single-step perturbation of the cells. As indicated above, Cas9 activity in the cells to be screened should be experimentally confirmed. The design of the sgRNA library is another key factor in screen performance. As always, sgRNAs should be designed to maximize activity and specificity. High sgRNA activity is particularly important for screens because, unlike small-scale experiments, it is not possible to select single-cell clones with the desired mutations before evaluating phenotype. The entire population of cells receiving any particular sgRNA must represent, in bulk, the phenotypic effect of that sgRNA. Thus, cells that receive an sgRNA but do not fully lose function of the target gene will dilute the apparent effect of that sgRNA in the screen. Tools for sgRNA selection for large-scale libraries must be capable of fully automated design for every gene. Multiple sgRNAs per gene are advised: first, to provide more chances for efficacy and, second, so that consistency of sgRNAs per gene can be used as a gauge of gene specificity. Popular sgRNA libraries include approximately half a dozen sgRNAs per gene (i.e., 120,000 sgRNAs for a whole genome of 20,0000 genes). Improved designs yielding higher proportions of highly active sgRNAs could reduce the number of guides employed without sacrificing the power of the library to identify hit genes. Reducing the size of the library reduces the scale and cost of the screen, permitting more cells or conditions to be tested. In cases where the cells are difficult to obtain or the screen is particularly difficult or prohibitively expensive, reducing screen scale can be not only helpful but necessary. A few publically available software tools permit the high-throughput sgRNA design and scoring required for large libraries, but those that do are generally computationally intensive and must be installed and run locally (SSFinder [46], CRISPRseek [48], sgRNAcas9 [45]).

As an sgRNA can produce heterogeneous phenotypic results for both technical (non-uniform gene modifications) and biological reasons (inherent cell-to-cell variability and stochasticity of responses), a screen must employ enough cells to ensure that each sgRNA is tested in many cells. Experience with shRNA and sgRNA screens suggests that approximately 1000–2000 cells per sgRNA (combining across all replicates) is typically sufficient, assuming that the library pool is evenly represented, with all sgRNAs present in similar abundance. In practice, for each screen, the actual number of cells needed to converge to reproducible results depends on many variables, and the scale required should be validated for each screen by comparison of independent replicates to determine whether the hit-list has converged. Several scoring schemes have been proposed for RNAi pooled screens that similarly apply to sgRNA screens. None has become standard, and they are not reviewed here. Such schemes combine the phenotypic enrichment scores from the multiple sgRNAs targeting each gene, and vary mostly in the degree to which they emphasize the magnitude of scoring (of the best sgRNA) versus constituency among the multiple sgRNAs per gene. In any case, detailed experimental validation of findings from large-scale screens is essential to confirm gene effects.

There are several contexts in which in vivo mouse pooled screens are feasible using either RNAi or CRISPR. One uses tumor xenograft models in which cancer cells are library-perturbed ex vivo and then implanted into the animal subcutaneously, orthotopically or into the blood [81]. More complex in vivo screens involve library transduction of mouse hematopoietic stem cells (HSCs) or immune cells ex vivo and then reconstituting them into the mouse through bone marrow transplant or adoptive transfer or by injection of virus into the tissue of interest for in vivo transduction [82–84]. To date, these approaches have been performed at sub-genome scale on focused sets of 20–2000 genes. For pooled screens, either in vitro or in vivo, inducible Cas9 systems for delayed gene perturbations can provide additional possibilities in screen design. Inducible systems optimized to both avoid leakiness and provide rapid efficient gene editing upon induction are in development by many groups.

Pooled screens for gene activation or inhibition are performed in a similar manner, but the library designs for such CRISPRa or CRISPRi systems differ as described above. Few such screens have been published to date, and these systems are not reviewed here, but given the advantages of modulating the endogenous gene in context versus expressing the CDS from an artificial promoter, CRISPR transcriptional modifications promise to be a popular screening approach [13, 22, 24, 26].

Gene editing

Another mainstay application of CRISPR-Cas9 technology is to produce precise gene edits — for example, to introduce specific alleles that correlate to, and might have a causal role in, a disease phenotype. In contrast to the low-throughput and high-throughput strategies for producing gene KOs described above, this method relies on the introduction of a repair template, such that new sequence is substituted at the site of the DSB. Using these HDR-mediated edits — KI alterations — any desired sequence can be inserted to produce, for example, loss of function, gain of function or altered (neomorphic) function or to investigate variants of unknown functional status. One could engineer coding variants to model a human disease or to introduce reporter genes or epitope tags into endogenous loci [15, 17]. It is clearly advantageous to obtain specifically chosen gene edits, versus the ‘take-what-you-get’ modifications resulting from NHEJ, but it comes at the cost of reduced editing efficiency. Use of HDR currently necessitates single-cell cloning to isolate the small percentage of cells with the desired modification. As with the production of high-value KO cell lines, it is strongly advisable to produce multiple correctly modified clones generated by multiple sgRNAs to enable the consistent on-target effects to be discerned from OT effects that might be exhibited by individual clones. The required single-cell cloning and analysis make KI strategies strictly low-throughput processes, but ongoing efforts to make isolation and identification of the desired clones more efficient [85, 86], or to avoid it entirely by dramatically increasing the efficiency of the HDR process [21, 87–90], could make larger scales more feasible.

When designing KI strategies, the first consideration is the location of the DNA break. For small mutations such as single-nucleotide replacements, a DSB in close proximity to the desired site of mutation can be efficiently repaired with a short single-stranded DNA oligo encoding the desired mutation and an approximately 50-nucleotide flanking sequence on both sides [91, 92]. Introduction of large insertions such as GFP reporters can be achieved by using a longer repair template such as a targeting plasmid with 400- to 1000-bp homology arms on either side of the mutation site [15, 17, 18]. In some instances, a suitable PAM might not occur within 20 bp of the mutation site, or the sgRNA in closest proximity might have excessive OT liabilities. It is preferable to select a more specific sgRNA, even if it is over 100 bp away from the mutation site, and to use a targeting plasmid with 400- to 1000-bp homology arms to improve HDR efficiency. When using either short single-stranded DNA repair templates and longer dsDNA plasmids or PCR products, mutating the targeted PAM site is advised to prevent subsequent cleavage of modified or repaired alleles [93]. In some cases, it might be desirable to introduce several silent mutations in the repair template at the sgRNA-binding site, so as to create a distinct primer-binding site in repaired alleles to facilitate genotyping. Alternatively, introduction of silent mutations that generate a new restriction enzyme recognition sequence can be leveraged for genotyping strategies. It is important, however, that any introduced mutations in the PAM or elsewhere be silent or not disrupt splicing. Thus, it is advisable to evaluate gene expression from the modified locus, and to verify on-target integration of the repair template. Several approaches are available for detecting spurious integration of repair templates and other OT indels [94].

Mouse models

Soon after CRISPR-mediated genome engineering was demonstrated in cultured cells, it was adapted to the generation of mutant mice [14, 15, 17, 95, 96]. Many of the same considerations for in vitro genome engineering in cells apply in vivo as well, such as the selection of target sites to maximize target efficacy and specificity against OT liabilities. To generate mice, Cas9 and sgRNA can be delivered into embryonic stem (ES) cells or injected directly into zygotes. Injection of in vitro transcribed sgRNA and Cas9 mRNA into zygotes and subsequent implantation into pseudopregnant foster mothers has produced efficient generation of KO alleles [14]. When targeting a single gene, indel mutations can be detected in a majority of the resulting mice, and two out-of-frame alleles can be observed in up to 35–40 % of mice, provided that loss of function does not compromise viability [93]. Although founder mice tend to exhibit mosaicism [97], germline transmission of modified alleles is quite efficient, suggesting that the majority of indels occur early during blastocyst development. OT mutations presumably will also be transmitted efficiently to subsequent generations [98]. By sequencing predicted OT sites in CRISPR-modified mice, investigators have documented variable OT effects depending on the sgRNA selected, but in vivo CRISPR can be quite selective, consistent with in vitro observations [14]. To try to reduce OT effects, dCas9 nickase has been employed with paired sgRNAs in vivo as well as in vitro, but unfortunately on-target efficiency is also reduced with this strategy. Nevertheless, it is possible to obtain up to 20 % of mice with homozygous loss-of-function alleles [93]. Even with highly selective sgRNAs, OT effects cannot be discounted when generating mice. Evaluating undesired mutations by sequencing of predicted OT sites is fairly straightforward; however, as noted, prediction of OT sites is relatively poor in both directions — it can generate an overly long list of candidate sites of which few are actually found to be modified, and yet still miss many actual OT sites. Thus, many researchers might wish to maintain breeder colonies of CRISPR-modified mice by backcrossing to wild-type mice [15, 17, 93]. There are many potential applications for such in vivo modifications, such as creation of disease models, engineering of reporter mice for in vivo assays, and even in vivo screening using pooled sgRNAs delivered, for example, to the lung or immune cells [82, 99].