The development of recombinant DNA technology in the 1970s marked the beginning of a new era for biology. For the first time, molecular biologists gained the ability to manipulate DNA molecules, making it possible to study genes and harness them to develop novel medicine and biotechnology. Recent advances in genome engineering technologies are sparking a new revolution in biological research. Rather than studying DNA taken out of the context of the genome, researchers can now directly edit or modulate the function of DNA sequences in their endogenous context in virtually any organism of choice, enabling them to elucidate the functional organization of the genome at the systems level, as well as identify causal genetic variations.

Broadly speaking, genome engineering refers to the process of making targeted modifications to the genome, its contexts (e.g., epigenetic marks), or its outputs (e.g., transcripts). The ability to do so easily and efficiently in eukaryotic and especially mammalian cells holds immense promise to transform basic science, biotechnology, and medicine (Figure 1).

For life sciences research, technologies that can delete, insert, and modify the DNA sequences of cells or organisms enable dissecting the function of specific genes and regulatory elements. Multiplexed editing could further allow the interrogation of gene or protein networks at a larger scale. Similarly, manipulating transcriptional regulation or chromatin states at particular loci can reveal how genetic material is organized and utilized within a cell, illuminating relationships between the architecture of the genome and its functions. In biotechnology, precise manipulation of genetic building blocks and regulatory machinery also facilitates the reverse engineering or reconstruction of useful biological systems, for example, by enhancing biofuel production pathways in industrially relevant organisms or by creating infection-resistant crops. Additionally, genome engineering is stimulating a new generation of drug development processes and medical therapeutics. Perturbation of multiple genes simultaneously could model the additive effects that underlie complex polygenic disorders, leading to new drug targets, while genome editing could directly correct harmful mutations in the context of human gene therapy (Tebas et al., 2014).

Eukaryotic genomes contain billions of DNA bases and are difficult to manipulate. One of the breakthroughs in genome manipulation has been the development of gene targeting by homologous recombination (HR), which integrates exogenous repair templates that contain sequence homology to the donor site (Figure 2A) (Capecchi, 1989). HR-mediated targeting has facilitated the generation of knockin and knockout animal models via manipulation of germline competent stem cells, dramatically advancing many areas of biological research. However, although HR-mediated gene targeting produces highly precise alterations, the desired recombination events occur extremely infrequently (1 in 10⁶–10⁹ cells) (Capecchi, 1989), presenting enormous challenges for large-scale applications of gene-targeting experiments.

To overcome these challenges, a series of programmable nuclease-based genome editing technologies have been developed in recent years, enabling targeted and efficient modification of a variety of eukaryotic and particularly mammalian species. Of the current generation of genome editing technologies, the most rapidly developing is the class of RNA-guided endonucleases known as Cas9 from the microbial adaptive immune system CRISPR (clustered regularly interspaced short palindromic repeats), which can be easily targeted to virtually any genomic location of choice by a short RNA guide. Here, we review the development and applications of the CRISPR-associated endonuclease Cas9 as a platform technology for achieving targeted perturbation of endogenous genomic elements and also discuss challenges and future avenues for innovation.

A series of studies by Haber and Jasin (Rudin et al., 1989; Plessis et al., 1992; Rouet et al., 1994; Choulika et al., 1995; Bibikova et al., 2001; Bibikova et al., 2003) led to the realization that targeted DNA double-strand breaks (DSBs) could greatly stimulate genome editing through HR-mediated recombination events. Subsequently, Carroll and Chandrasegaran demonstrated the potential of designer nucleases based on zinc finger proteins for efficient, locus-specific HR (Bibikova et al., 2001, 2003). Moreover, it was shown in the absence of an exogenous homology repair template that localized DSBs can induce insertions or deletion mutations (indels) via the error-prone nonhomologous end-joining (NHEJ) repair pathway (Figure 2A) (Bibikova et al., 2002). These early genome editing studies established DSB-induced HR and NHEJ as powerful pathways for the versatile and precise modification of eukaryotic genomes.

To achieve effective genome editing via introduction of site-specific DNA DSBs, four major classes of customizable DNA-binding proteins have been engineered so far: meganucleases derived from microbial mobile genetic elements (Smith et al., 2006), zinc finger (ZF) nucleases based on eukaryotic transcription factors (Urnov et al., 2005; Miller et al., 2007), transcription activator-like effectors (TALEs) from Xanthomonas bacteria (Christian et al., 2010; Miller et al., 2011; Boch et al., 2009; Moscou and Bogdanove, 2009), and most recently the RNA-guided DNA endonuclease Cas9 from the type II bacterial adaptive immune system CRISPR (Cong et al., 2013; Mali et al., 2013a).

Meganuclease, ZF, and TALE proteins all recognize specific DNA sequences through protein-DNA interactions. Although meganucleases integrate its nuclease and DNA-binding domains, ZF and TALE proteins consist of individual modules targeting 3 or 1 nucleotides (nt) of DNA, respectively (Figure 2B). ZFs and TALEs can be assembled in desired combinations and attached to the nuclease domain of FokI to direct nucleolytic activity toward specific genomic loci. Each of these platforms, however, has unique limitations.

Meganucleases have not been widely adopted as a genome engineering platform due to lack of clear correspondence between meganuclease protein residues and their target DNA sequence specificity. ZF domains, on the other hand, exhibit context-dependent binding preference dueto crosstalk between adjacent modules when assembled into a larger array (Maeder et al., 2008). Although multiple strategies have been developed to account for these limitations (Gonzaelz et al., 2010; Sander et al., 2011), assembly of functional ZFPs with the desired DNA binding specificity remains a major challenge that requires an extensive screening process. Similarly, although TALE DNA-binding monomers are for the most part modular, they can still suffer from context-dependent specificity (Juillerat et al., 2014), and their repetitive sequences render construction of novel TALE arrays labor intensive and costly.

Given the challenges associated with engineering of modular DNA-binding proteins, new modes of recognition would significantly simplify the development of custom nucleases. The CRISPR nuclease Cas9 is targeted by a short guide RNA that recognizes the target DNA via Watson-Crick base pairing (Figure 2C). The guide sequence within these CRISPR RNAs typically corresponds to phage sequences, constituting the natural mechanism for CRISPR antiviral defense, but can be easily replaced by a sequence of interest to retarget the Cas9 nuclease. Multiplexed targeting by Cas9 can now be achieved at unprecedented scale by introducing a battery of short guide RNAs rather than a library of large, bulky proteins. The ease of Cas9 targeting, its high efficiency as a site-specific nuclease, and the possibility for highly multiplexed modifications have opened up a broad range of biological applications across basic research to biotechnology and medicine.

The utility of customizable DNA-binding domains extends far beyond genome editing with site-specific endonucleases. Fusing them to modular, sequence-agnostic functional effector domains allows flexible recruitment of desired perturbations, such as transcriptional activation, to a locus of interest (Xu and Bestor, 1997; Beerli et al., 2000a; Konermann et al., 2013; Maeder et al., 2013a; Mendenhall et al., 2013). In fact, any modular enzymatic component can, in principle, be substituted, allowing facile additions to the genome engineering toolbox. Integration of genome- and epigenome-modifying enzymes with inducible protein regulation further allows precise temporal control of dynamic processes (Beerli et al., 2000b; Konermann et al., 2013).

The recent development of the Cas9 endonuclease for genome editing draws upon more than a decade of basic research into understanding the biological function of the mysterious repetitive elements now known as CRISPR (Figure 3), which are found throughout the bacterial and archaeal diversity. CRISPR loci typically consist of a clustered set of CRISPR-associated (Cas) genes and the signature CRISPR array—a series of repeat sequences (direct repeats) interspaced by variable sequences (spacers) corresponding to sequences within foreign genetic elements (protospacers) (Figure 4). Whereas Cas genes are translated into proteins, most CRISPR arrays are first transcribed as a single RNA before subsequent processing into shorter CRISPR RNAs (crRNAs), which direct the nucleolytic activity of certain Cas enzymes to degrade target nucleic acids.

The CRISPR story began in 1987. While studying the iap enzyme involved in isozyme conversion of alkaline phosphatase in E. coli, Nakata and colleagues reported a curious set of 29 nt repeats downstream of the iap gene (Ishino et al., 1987). Unlike most repetitive elements, which typically take the form of tandem repeats like TALE repeat monomers, these 29 nt repeats were interspaced by five intervening 32 nt nonrepetitive sequences. Over the next 10 years, as more microbial genomes were sequenced, additional repeat elements were reported from genomes of different bacterial and archaeal strains. Mojica and colleagues eventually classified interspaced repeat sequences as a unique family of clustered repeat elements present in >40% of sequenced bacteria and 90% of archaea (Mojica et al., 2000).

These early findings began to stimulate interest in such microbial repeat elements. By 2002, Jansen and Mojica coined the acronym CRISPR to unify the description of microbial genomic loci consisting of an interspaced repeat array (Jansen et al., 2002; Barrangou and van der Oost, 2013). At the same time, several clusters of signature CRISPR-associated (cas) genes were identified to be well conserved and typically adjacent to the repeat elements (Jansen et al., 2002), serving as a basis for the eventual classification of three different types of CRISPR systems (types I–III) (Haft et al., 2005; Makarova et al., 2011b). Types I and III CRISPR loci contain multiple Cas proteins, now known to form complexes with crRNA (CASCADE complex for type I; Cmr or Csm RAMP complexes for type III) to facilitate the recognition and destruction of target nucleic acids (Brouns et al., 2008; Hale et al., 2009) (Figure 4). In contrast, the type II system has a significantly reduced number of Cas proteins. However, despite increasingly detailed mapping and annotation of CRISPR loci across many microbial species, their biological significance remained elusive.

A key turning point came in 2005, when systematic analysis of the spacer sequences separating the individual direct repeats suggested their extrachromosomal and phage-associated origins (Mojica et al., 2005; Pourcel et al., 2005; Bolotin et al., 2005). This insight was tremendously exciting, especially given previous studies showing that CRISPR loci are transcribed (Tang et al., 2002) and that viruses are unable to infect archaeal cells carrying spacers corresponding to their own genomes (Mojica et al., 2005). Together, these findings led to the speculation that CRISPR arrays serve as an immune memory and defense mechanism, and individual spacers facilitate defense against bacteriophage infection by exploiting Watson-Crick base-pairing between nucleic acids (Mojica et al., 2005; Pourcel et al., 2005). Despite these compelling realizations that CRISPR loci might be involved in microbial immunity, the specific mechanism of how the spacers act to mediate viral defense remained a challenging puzzle. Several hypotheses were raised, including thoughts that CRISPR spacers act as small RNA guides to degrade viral transcripts in a RNAi-like mechanism (Makarova et al., 2006) or that CRISPR spacers direct Cas enzymes to cleave viral DNA at spacer-matching regions (Bolotin et al., 2005).

Working with the dairy production bacterial strain Streptococcus thermophilus at the food ingredient company Danisco, Horvath and colleagues uncovered the first experimental evidence for the natural role of a type II CRISPR system as an adaptive immunity system, demonstrating a nucleic-acid-based immune system in which CRISPR spacers dictate target specificity while Cas enzymes control spacer acquisition and phage defense (Barrangou et al., 2007). A rapid series of studies illuminating the mechanisms of CRISPR defense followed shortly and helped to establish the mechanism as well as function of all three types of CRISPR loci inadaptive immunity. By studying the type I CRISPR locus of Escherichia coli, van der Oost and colleagues showed that CRISPR arrays are transcribed and converted into small crRNAs containing individual spacers to guide Cas nuclease activity (Brouns et al., 2008). In the same year, CRISPR-mediated defense by a type III-A CRISPR system from Staphylococcus epidermidis was demonstrated to block plasmid conjugation, establishing the target of Cas enzyme activity as DNA rather than RNA (Marraffini and Sontheimer, 2008), although later investigation of a different type III-B system from Pyrococcus furiosus also revealed crRNA-directed RNA cleavage activity (Hale et al., 2009, 2012).

As the pace of CRISPR research accelerated, researchers quickly unraveled many details of each type of CRISPR system (Figure 4). Building on an earlier speculation that protospacer adjacent motifs (PAMs) may direct the type II Cas9 nuclease to cleave DNA (Bolotin et al., 2005), Moineau and colleagues highlighted the importance of PAM sequences by demonstrating that PAM mutations in phage genomes circumvented CRISPR interference (Deveau et al., 2008). Additionally, for types I and II, the lack of PAM within the direct repeat sequence within the CRISPR array prevents self-targeting by the CRISPR system. In type III systems, however, mismatches between the 5′ end of the crRNA and the DNA target are required for plasmid interference (Marraffini and Sontheimer, 2010).

By 2010, just 3 years after the first experimental evidence for CRISPR in bacterial immunity, the basic function and mechanisms of CRISPR systems were becoming clear. A variety of groups had begun to harness the natural CRISPR system for various biotechnological applications, including the generation of phage-resistant dairy cultures (Quiberoni et al., 2010) and phylogenetic classification of bacterial strains (Horvath et al., 2008, 2009). However, genome editing applications had not yet been explored.

Around this time, two studies characterizing the functional mechanisms of the native type II CRISPR system elucidated the basic components that proved vital for engineering a simple RNA-programmable DNA endonuclease for genome editing. First, Moineau and colleagues used genetic studies in Streptococcus thermophilus to reveal that Cas9 (formerly called Cas5, Csn1, or Csx12) is the only enzyme within the cas gene cluster that mediates target DNA cleavage (Garneau et al., 2010). Next, Charpentier and colleagues revealed a key component in the biogenesis and processing of crRNA in type II CRISPR systems—a noncoding trans-activating crRNA (tracrRNA) that hybridizes with crRNA to facilitate RNA-guided targeting of Cas9 (Deltcheva et al., 2011). This dual RNA hybrid, together with Cas9 and endogenous RNase III, is required for processing the CRISPR array transcript into mature crRNAs (Deltcheva et al., 2011). These two studies suggested that there are at least three components (Cas9, the mature crRNA, and tracrRNA) that are essential for reconstituting the type II CRISPR nuclease system. Given the increasing importance of programmable site-specific nucleases based on ZFs and TALEs for enhancing eukaryotic genome editing, it was tantalizing to think that perhaps Cas9 could be developed into an RNA-guided genome editing system. From this point, the race to harness Cas9 for genome editing was on.

In 2011, Siksnys and colleagues first demonstrated that the type II CRISPR system is transferrable, in that transplantation of the type II CRISPR locus from Streptococcus thermophilus into Escherichia coli is able to reconstitute CRISPR interference in a different bacterial strain (Sapranauskas et al., 2011). By 2012, biochemical characterizations by the groups of Charpentier, Doudna, and Siksnys showed that purified Cas9 from Streptococcus thermophilus or Streptococcus pyogenes can be guided by crRNAs to cleave target DNA in vitro (Jinek et al., 2012; Gasiunas et al., 2012), in agreement with previous bacterial studies (Garneau et al., 2010; Deltcheva et al., 2011; Sapranauskas et al., 2011). Furthermore, a single guide RNA (sgRNA) can be constructed by fusing a crRNA containing the targeting guide sequence to a tracrRNA that facilitates DNA cleavage by Cas9 in vitro (Jinek et al., 2012).

In 2013, a pair of studies simultaneously showed how to successfully engineer type II CRISPR systems from Streptococcus thermophilus (Cong et al., 2013) and Streptococcus pyogenes (Cong et al., 2013; Mali et al., 2013a) to accomplish genome editing in mammalian cells. Heterologous expression of mature crRNA-tracrRNA hybrids (Cong et al., 2013) as well as sgRNAs (Cong et al., 2013; Mali et al., 2013a) directs Cas9 cleavage within the mammalian cellular genome to stimulate NHEJ or HDR-mediated genome editing. Multiple guide RNAs can also be used to target several genes at once. Since these initial studies, Cas9 has been used by thousands of laboratories for genome editing applications in a variety of experimental model systems (Sander and Joung, 2014). The rapid adoption of the Cas9 technology was also greatly accelerated through a combination of open-source distributors such as Addgene, as well as a number of online user forums such as http://www.genome-engineering.org and http://www.egenome.org.

The family of Cas9 proteins is characterized by two signature nuclease domains, RuvC and HNH, each named based on homology to known nuclease domain structures (Figure 2C). Though HNH is a single nuclease domain, the full RuvC domain is divided into three subdomains across the linear protein sequence, with RuvC I near the N-terminal region of Cas9 and RuvC II/III flanking the HNH domain near the middle of the protein. Recently, a pair of structural studies shed light on the structural mechanism of RNA-guided DNA cleavage by Cas9.

First, single-particle EM reconstructions of the Streptococcus pyogenes Cas9 (SpCas9) revealed a large structural rearrangement between apo-Cas9 unbound to nucleic acid and Cas9 in complex with crRNA and tracrRNA, forming a central channel to accommodate the RNA-DNA heteroduplex (Jinek et al., 2014). Second, a high-resolution structure of SpCas9 in complex with sgRNA and the complementary strand of target DNA further revealed the domain organization to comprise of an α-helical recognition (REC) lobe and a nuclease (NUC) lobe consisting of the HNH domain, assembled RuvC subdomains, and a PAM-interacting (PI) C-terminal region (Nishimasu et al., 2014) (Figure 5A and Movie S1).

Together, these two studies support the model that SpCas9 unbound to target DNA or guide RNA exhibits an autoinhibited conformation in which the HNH domain active site is blocked by the RuvC domain and is positioned away from the REC lobe (Jinek et al., 2014). Binding of the RNA-DNA heteroduplex would additionally be sterically inhibited by the orientation of the C-terminal domain. Asaresult, apo-Cas9 likely cannot bind nor cleave target DNA. Like many ribonucleoprotein complexes, the guide RNA serves as a scaffold around which Cas9 can fold and organize its various domains (Nishimasu et al., 2014).

The crystal structure of SpCas9 incomplex with an sgRNA and target DNA also revealed how the REC lobe facilitates target binding. An arginine-rich bridge helix (BH) within the REC lobe is responsible for contacting the 3′ 8–12 nt of the RNA-DNA heteroduplex (Nishimasu et al., 2014), which correspond with the seed sequence identified through guide sequence mutation experiments (Jinek et al., 2012; Cong et al., 2013; Fu et al., 2013; Hsu et al., 2013; Pattanayak et al., 2013; Mali et al., 2013b).

The SpCas9 structure also provides a useful scaffold for engineering or refactoring of Cas9 and sgRNA. Because the REC2 domain of SpCas9 is poorly conserved in shorter orthologs, domain recombination or truncation is a promising approach for minimizing Cas9 size. SpCas9 mutants lacking REC2 retain roughly 50% of wild-type cleavage activity, which could be partly attributed to their weaker expression levels (Nishimasu et al., 2014). Introducing combinations of orthologous domain recombination, truncation, and peptide linkers could facilitate the generation of a suite of Cas9 mutant variants optimized for different parameters such as DNA binding, DNA cleavage, or overall protein size.

Cas9 is exclusively associated with the type II CRISPR locus and serves as the signature type II gene. Based on the diversity of associated Cas genes, type IICRISPR loci are further subdivided into three subtypes (IIA–IIC) (Figure 5B) (Makarova et al., 2011a; Chylinski et al., 2013). Type II CRISPR loci mostly consist of the cas9, cas1, and cas2 genes, as well as a CRISPR array and tracrRNA. Type IIC CRISPR systems contain only this minimal set of cas genes, whereas types IIA and IIB have an additional signature csn2 or cas4 gene, respectively (Chylinski et al., 2013).

Subtype classification of type II CRISPR loci is based on the architecture and organization of each CRISPR locus. For example, type IIA and IIB loci usually consist of four cas genes, whereas type IIC loci only contain three cas genes. However, this classification does not reflect the structural diversity of Cas9 proteins, which exhibit sequence homology and length variability irrespective of the subtype classification of their parental CRISPR locus. Of >1,000 Cas9 nucleases identified from sequence databases (UniProt) based on homology, protein length israther heterogeneous, roughly ranging from 900 to 1600 amino acids (Figure 5C). The length distribution of most Cas9 proteins can be divided into two populations centered around 1,100 and 1,350 amino acids in length. It is worth noting that a third population of large Cas9 proteins belonging to subtype IIA, formerly called Csx12, typically contain around 1500 amino acids.

Despite the apparent diversity of protein length, all Cas9 proteins share similar domain architecture (Makarova et al., 2011a; Chylinski et al., 2013, 2014; Fonfara et al., 2014), consisting of the RuvC and HNH nuclease domains and the REC domain, an α-helix-rich region with an Arg-rich bridge helix. Unlike type I and III CRISPR systems, which are found in both bacteria and archaea, type II CRISPRs have so far only been found in bacterial strains (Chylinski et al., 2013). The majority of Cas9 orthologs in fact belong to the phyla of Bacteroidetes, Proteobacteria, and Firmicutes (Figure 5D).

The length difference among Cas9 proteins largely results from variable conservation of the REC domain (Figure 5E), which associates with the sgRNA and target DNA. For example, the type IIC Actinomyces naeslundii Cas9, which is more compact than its Streptococcus pyogenes ortholog, has a much smaller REC lobe with substantially different orientation (Jinek et al., 2014).

A critical feature of the Cas9 system is the protospacer-adjacent motif (PAM), which flanks the 3′ end of the DNA target site (Figure 2C) and dictates the DNA target search mechanism of Cas9. In addition to facilitating self versus non-self discrimination by Cas9 (Shah et al., 2013), because direct repeats do not contain PAM sites, biochemical and structural characterization of SpCas9 suggested that PAM recognition is involved in triggering the transition between Cas9 target binding and cleavage conformations (Sternberg et al., 2014; Jinek et al., 2014; Nishimasu et al., 2014).

Single-molecule imaging indicated that Cas9-crRNA-tracrRNA complexes first associate with PAM sequences throughout the genome, allowing Cas9 to initiate DNA strand separation via unknown mechanisms (Sternberg et al., 2014). DNA competitor cleavage assays additionally suggested that formation of the RNA-DNA heteroduplex is initiated at the PAM site before proceeding PAM distally by interrogating the target site upstream of the PAM for guide sequence complementarity (Sternberg et al., 2014). Binding of the PAM and a matching target then triggers Cas9 nuclease activity by activating the HNH and RuvC domains, supported by the observation of HNH domain flexibility within the Cas9-sgRNA-DNA ternary complex (Nishimasu et al., 2014).

The complexity of the PAM sequences also determines the overall DNA targeting space of Cas9. For example, the 5′-NGG of SpCas9 allows it to target, on average, every 8 bp within the human genome (Cong et al., 2013; Hsu et al., 2013). Additionally, SpCas9 can target sites flanked by 5′-NAG PAMs (Jiang et al., 2013; Hsu et al., 2013), albeit at a lower efficiency, further expanding its editing versatility. The PAM is specific to each Cas9 ortholog, even within the same species, such as 5′-NNA GAAW for Streptococcus thermophilus CRISPR1 (Deveau et al., 2008) and 5′-NGGNG for Streptococcus thermophilus CRISPR3 (Horvath et al., 2008). Another Cas9 from Neisseria meningitidis with a 5′-NNNNGATT PAM requirement (Zhang et al., 2013) was recently applied in human pluripotent stem cells (Hou et al., 2013).

Computational (Chylinski et al., 2013, 2014; Fonfara et al., 2014) or metagenomic analysis of bacteria and archaea containing CRISPR loci could lead to the discovery of Cas9 nucleases with additional PAMs to expand the targeting range of the Cas9 toolkit. Delivery of multiple Cas9 proteins with different PAM requirements facilitates orthogonal genome engineering, in which independent but simultaneous functions are applied at different loci within the same cell or cell population. NmCas9 and SpCas9, for example, can be employed for independent transcriptional repression and nuclease activity (Esvelt et al., 2013).

PAM specificity can also be modified. For instance, orthologous replacement of the PAM-interacting (PI) domain from the Streptococcus thermophilus CRISPR3 Cas9 with the corresponding domain from Streptococcus pyogenes Cas9 successfully altered PAM recognition from 5′-NGGNG to 5′-NGG (Nishimasu et al., 2014). PAM engineering strategies could also be exploited to generate short Cas9 orthologs with flexible 5′-NGG or 5′-NG PAM domains.

To date, the Streptococcus pyogenes Cas9 (SpCas9) has been used broadly to achieve efficient genome editing in a variety of species and cell types, including human cell lines, bacteria, zebrafish, yeast, mouse, fruit fly, roundworm, rat, common crops, pig, and monkey (see Sander and Joung [2014] for a detailed list). SpCas9 is also dramatically expanding the catalog of genetically tractable model organisms, for example, by enabling the introduction of multiplex mutations in cynomolgus monkeys (Niu et al., 2014).

SpCas9 can be targeted either with a pair of crRNA and tracrRNA (Cong et al., 2013) or with a chimeric sgRNA (Cong et al., 2013; Mali et al., 2013a; Cho et al., 2013; Jinek et al., 2013), as the crRNA or sgRNA contains a 20 nt guide sequence that directly matches the target site. The only requirement for the selection of Cas9 target sites is the presence of a protospacer-adjacent motif (PAM) immediately downstream of the target site.

An early discrepancy in the use of SpCas9 editing of the human genome was the drastically higher levels of NHEJ-induced indels given the same target site, when using the engineered dual guide RNA system (Cong et al., 2013) compared to the engineered sgRNA(+48) scaffold, which only contained up to the 48th base of tracrRNA. Although sgRNA(+48) is sufficient for cleaving DNA in vitro (Jinek et al., 2012), extension of the 3′ tracrRNA sequence preserved several hairpin structures (sgRNA(+72) and sgRNA(+84)) that were critical for effective sgRNA-mediated genome editing in vivo (Mali et al., 2013a; Hsu et al., 2013). The additional stem loops enhance the stability of the sgRNA (Hsu et al., 2013) and are important for proper Cas9-sgRNA-DNA ternary complex formation (Nishimasu et al., 2014). These analyses of the sgRNA structure and function indicate that careful sgRNA design is critical for optimal Cas9 activity, especially for testing novel Cas9 candidates derived from metagenomic analysis.

One hallmark of the natural CRISPR-Cas9 system is its inherent ability to efficiently cleave multiple distinct target sequences in parallel (Barrangou et al., 2007; Garneau et al., 2010) by converting a pre-crRNA transcript containing many spacers into individual guide RNAs duplexes (mature crRNA and tracrRNA) through hybridization with tracrRNA (Deltcheva et al., 2011). Harnessing this unique aspect of CRISPR interference would enable highly scalable multiplex perturbations. Indeed, coexpression of a CRISPR array containing spacers targeting different genes (Cong et al., 2013) or a battery of several sgRNAs (Mali et al., 2013a; Wang et al., 2013) together with SpCas9 has led to efficient multiplex editing in mammalian cells. Surprisingly, CRISPR arrays containing direct repeats interspaced by designer spacers were processed into mature guide RNAs without the introduction of bacterial RNase III. Because RNase III is required for crRNA maturation in prokaryotic cells (Deltcheva et al., 2011), it is likely that endogenous mammalian RNases play compensatory roles (Cong et al., 2013).

Because genome editing leads to permanent modifications within the genome, the targeting specificity of Cas9 nucleases is of particular concern, especially for clinical applications and gene therapy. A combination of in vitro and in vivo assays has been typically used to characterize the specificity of ZFNs and TALENs (Gabriel et al., 2011), but systematic analysis has remained challenging due to difficulties in synthesizing large libraries of proteins with varying sequence specificity. However, Cas9 target recognition is dictated by the Watson-Crick base-pairing interactions of an RNA guide with its DNA target, enabling experimentally tractable and systematic evaluation of the effect of guide RNA-target DNA mismatches on Cas9 activity.

Streptococcus pyogenes Cas9 specificity has been extensively characterized by multiple groups using mismatched guide RNA libraries, in vitro selection, and reporter assays (Fu et al., 2013; Hsu et al., 2013; Mali et al., 2013b; Pattanayak et al., 2013). In contrast to previous studies that suggested a seed sequence model for Cas9 specificity, wherein the first 8–12 PAM-proximal guide sequence bases determine specificity (Jinek et al., 2012; Cong et al., 2013), these studies collectively demonstrate that Cas9 tolerates mismatches throughout the guide sequence in a manner that is sensitive to the number, position, and distribution of the mismatches (Fu et al., 2013; Hsu et al., 2013; Mali et al., 2013b; Pattanayak et al., 2013). Although the PAM-distal bases of the guide sequence are less important for specificity, meaning that mismatches at those positions often do not abolish Cas9 activity, all positions within the guide contribute to overall specificity. Importantly, off-target sites followed by the 5′-NAG PAM can also lead to off-target cleavage, demonstrating the importance of considering both NGG and NAG PAMs in off-target analysis.

Interestingly, Cas9 requires extensive homology between the guide RNA and target DNA in order to cleave but can remain semi-transiently bound with only a short stretch of complementary sequence between the guide RNA and target DNA. These observations suggest that Cas9 has many off-target binding sites but cleaves only a small fraction of them (Wu et al., 2014). Thus, concerns about off-target activity could vary widely given a desired application that exploits Cas9 for its DNA binding or cleavage capabilities.

Enzymatic concentration is also an important factor in determining Cas9 off-target mutagenesis. This is particularly important because Cas9 can tolerate even five mismatches within the target site (Fu et al., 2013). Mismatches appear to be better tolerated when Cas9 is present at high concentrations (Hsu et al., 2013; Pattanayak et al., 2013), leading to higher off-target activity; decreasing Cas9 concentration significantly improves the on- to off-target ratio at the expense of the efficiency of on-target cleavage (Hsu et al., 2013). The duration of Cas9 expression is likely an additional factor that tunes off-target activity, though its contributions remain to be carefully investigated.

While potential off-target sites have typically been computationally determined by searching for genomic sequences with high sequence similarity to the desired target locus, wholegenome sequencing or other unbiased ways of labeling DNA DSBs genome-wide may illuminate off-target sites that are not predictable by first-order sequence comparison. Unbiased genome-wide characterizations have been previously used to characterize ZFN off-target mutagenesis (Gabriel et al., 2011) and could easily be adapted for Cas9 nuclease activity. Such data, perhaps in combination with thermodynamic characterization of guide RNA and target DNA hybridization, will likely provide a quantitative framework for assessing and predicting the off-target activity of Cas9. Multiple groups now provide Cas9 target selection tools (e.g., http://tools.genome-engineering. org, http://zifit.partners.org, and http://www.e-crisp.org).

Cas9 nucleases cleave DNA through the activity of their RuvC and HNH nuclease domains, each of which nicks a strand of DNA to generate a blunt-ended DSB (Figure 2C). SpCas9 can be converted into a DNA “nickase” that creates a single-stranded break (SSB) by catalytically inactivating the RuvC or HNH nuclease domains (Gasiunas et al., 2012; Jinek et al., 2012; Sapranauskas et al., 2011) via point mutations (Figure 6A). Because DNA single-strand breaks are repaired via the high-fidelity base excision repair (BER) pathway (Dianov and Hübscher, 2013), Cas9 nickases can be exploited for more specific NHEJ as well as HR.

To improve on-target DSB specificity, a double-nicking approach analogous to dimeric ZFNs or TALENs can be used to increase the overall number of bases that are specifically recognized in the target DNA. Using pairs of guide RNAs (Mali et al., 2013b; Ran et al.,2013)and an SpCas9 HNH⁺/RuvC⁻ nickase mutant (D10A), properly spaced cooperative nicks can mimic DSBs and mediate efficient indel formation (Figure 6B). Because off-target nick sites are precisely repaired, this multiplexed nicking strategy can improve specificity by up to 1,500× relative to the wild-type Cas9 (Ran et al., 2013).

Because Cas9 nuclease or multiplex nicking activity both stimulate NHEJ, a population of cells cotargeted with a homology donor would eventually possess a mix of indel mutants and donor integrants. Single DNA nicks, however, are also able to mediate donor recombination, albeit at a lower level than with DSBs (Hsu et al., 2013). Cas9 nickases with single sgRNAs can thus be used to mediate HR rather than NHEJ. Furthermore, off-target integration is highly unlikely due to long homology arms flanking the donor cassette.

In addition to the double-nicking strategy, sgRNAs truncated by 2 or 3 nt have been reported to significantly increase targeting specificity of SpCas9, potentially due to greater mismatch sensitivity (Fu et al., 2014). These truncated sgRNAs can be combined with multiplex nicking to further reduce off-target mutagenesis (Fu et al., 2014). Future structure-function analyses and Cas9 and protein engineering via rational design or directed evolution may lead to further improvements in Cas9 specificity.

Cas9 can be used to facilitate a wide variety of targeted genome engineering applications. The wild-type Cas9 nuclease has enabled efficient and targeted genome modification in many species that have been intractable using traditional genetic manipulation techniques. The ease of retargeting Cas9 by simply designing a short RNA sequence also enables large-scale unbiased genome perturbation experiments to probe gene function or elucidate causal genetic variants. In addition to facilitating co-valent genome modifications, the wild-type Cas9 nuclease can also be converted into a generic RNA-guided homing device (dCas9) by inactivating the catalytic domains. The use of effector fusions can greatly expand the repertoire of genome engineering modalities achievable using Cas9. For example, a variety of proteins or RNAs can be tethered to Cas9 or sgRNA to alter transcription states of specific genomic loci, monitor chromatin states, or even rearrange the three-dimensional organization of the genome.

Although Cas9 has already been widely used as a research tool, a particularly exciting future direction is the development of Cas9 as a therapeutic technology for treating genetic disorders. For a monogenic recessive disorder due to loss-of-function mutations (such as cystic fibrosis, sickle-cell anemia, or Duchenne muscular dystrophy), Cas9 may be used to correct the causative mutation. This has many advantages over traditional methods of gene augmentation that deliver functional genetic copies via viral vector-mediated overexpression—particularly that the newly functional gene is expressed in its natural context. For dominant-negative disorders in which the affected gene is haplosufficient (such as transthyretin-related hereditary amyloidosis or dominant forms of retinitis pigmentosum), it may also be possible to use NHEJ to inactivate the mutated allele to achieve therapeutic benefit. For allele-specific targeting, one could design guide RNAs capable of distinguishing between single-nucleotide polymorphism (SNP) variations in the target gene, such as when the SNP falls within the PAM sequence.

Some monogenic diseases also result from duplication of genomic sequences. For these diseases, the multiplexing capability of Cas9 may be exploited for deletion of the duplicated elements. For example, trinucleotide repeat disorders could be treated using two simultaneous DSBs to excise the repeat region. The success of this strategy will likely be higher for diseases such as Friedreich's ataxia, in which duplications occur in noncoding regions of the target gene, because NHEJ-mediated repair may lead to imperfect or frameshifted repair junctions.

In addition to repairing mutations underlying inherited disorders, Cas9-mediated genome editing might be used to introduce protective mutations in somatic tissues to combat nongenetic or complex diseases. For example, NHEJ-mediated inactivation of the CCR5 receptor in lymphocytes (Lombardo et al., 2007) may be a viable strategy for circumventing HIV infection, whereas deletion of PCSK9 (Cohenet al., 2005) orangiopoietin (Musunuru et al., 2010) may provide therapeutic effects against statin-resistant hypercholesterolemia or hyperlipidemia. Although these targets may be also addressed using siRNA-mediated protein knockdown, a unique advantage of NHEJ-mediated gene inactivation is the ability to achieve permanent therapeutic benefit without the need for continuing treatment. As with all gene therapies, it will of course be important to establish that each proposed therapeutic use has a favorable benefit-risk ratio.

Cas9 could be used beyond the direct genome modification of somatic tissue, such as for engineering therapeutic cells. Chimeric antigen receptor (CAR) T cells can be modified ex vivo and reinfused into a patient to specifically target certain cancers (Couzin-Frankel, 2013). The ease of design and testing of Cas9 may also facilitate the treatment of highly rare genetic variants through personalized medicine. Supporting these tremendous possibilities are a number of animal model studies as well as clinical trials using programmable nucleases that already provide important insights into the future development of Cas9-based therapeutics.

Recently, hydrodynamic delivery of plasmid DNA encoding Cas9 and sgRNA along with a repair template into the liver of an adult mouse model of tyrosinemia was shown to be able to correct the mutant Fah gene and rescue expression of the wild-type Fah protein in ∼1 out of 250 cells (Yin et al., 2014). In addition, clinical trials successfully used ZF nucleases to combat HIV infection by ex vivo knockout of the CCR5 receptor. In all patients, HIV DNA levels decreased, and in one out of four patients, HIV RNA became undetectable (Tebas et al., 2014). Both of these results demonstrate the promise of programmable nucleases as a new therapeutic platform.

However, numerous challenges still lie ahead. Most importantly, successful clinical translation will depend on appropriate and efficacious delivery systems to target specific disease tissues. To achieve high levels of therapeutic efficacy and simultaneously address a broad spectrum of genetic disorders, homologous recombination efficiency will need to be significantly improved. Although permanent genome modification has advantages over monoclonal antibody or siRNA treatments, which require repeated administration of the therapeutic molecule, the long-term implications remain unclear. As researchers further develop and test Cas9 toward clinical translation, it will be paramount to thoroughly characterize the safety as well as physiological effects of Cas9 using a variety of preclinical models.

The story of how a mysterious prokaryotic viral defense system became one of the most powerful and versatile platforms for engineering biology highlights the importance of basic science research. Just as recombinant DNA technology benefited from basic investigation of the restriction enzymes that are central to warfare between phage and bacteria, the latest generation of Cas9-based genome engineering tools are also based on components from the microbial antiphage defense system. It is highly likely that the future solutions for efficient and precise gene modification will be found in as of yet unexplored corners of the rich biological diversity of nature.