We used several approaches to identify CARF superfamily proteins. First, a CDD search was employed to identify all proteins in 2262 complete genomes (as of February 2013) that could be assigned to previously identified CARF families [namely COG1517, PF09455, PF09670, PF09659, PF09651, PF09623, PF09002, Csa3 (Lintner et al., 2010; Makarova et al., 2011b)]. Representatives of each family were used as queries for PSI-BLAST using the search strategy described in the Materials and Methods section (Altschul et al., 1997). Putative new members were validated using HHpred search (Soding et al., 2005). The same methods were used to identify other domains fused to CARF domains (Supplementary File 1). For further analysis incomplete protein sequences were discarded. The final data set included 1441 proteins (Supplementary File 1). This set was further clustered to generate a non-redundant subset (635 proteins) using BLASTCLUST (Wheeler and Bhagwat, 2007) with a length coverage cutoff of 0.8 and a score coverage threshold (bit score divided by alignment length) of 0.8. For this representative subset of 635 CARF domain-containing proteins, analysis of domain architecture and gene neighborhoods was performed as described under Materials and Methods. Because the extensive sequence divergence of the CARF domains results in saturation of substitutions and prevents building a high quality alignment for phylogenetic analysis, the relationships between families were determined approximately, on the basis of their similarity in HHpred searches (Figure 1 and Supplementary File 2).

Figure 1 shows the relationships between the CARF families, their domain organization and association (if any) with different types of CRISPR-Cas systems. The results of this analysis suggest that the CARF superfamily could be classified into at least 12 distinct major families with 10 or more representatives each and several minor families (Figures 1A,B, Supplementary File 1). In addition to the aforementioned CARF domain families, HHpred search using pfam09659 as the query identified significant sequence similarity between the CARF domain and an uncharacterized N-terminal domain of RtcR (Supplementary File 2), which is the regulator of the Rtc RNA repair system that consists of the 3′-terminal phosphate cyclase RtcA, and RNA ligase RtcB (Genschik et al., 1998; Chakravarty et al., 2012). Although this domain occurs in distinct protein architectural and genomic contexts (see below), it shares distinct sequence motifs with the CARF domains to the exclusion of other Rossmann fold domains. Hence we consider the predicted nucleotide-binding domain of RtcR a divergent member of the CARF superfamily.

The availability of five crystal structures of CARF domain proteins along with the above sequence analysis provides for a more detailed understanding of the conserved structural features of the superfamily and their functional implications. The core of the CARF domain is a six-stranded Rossmann-like fold with the core strand-5 and strand-6 forming a β-hairpin (Figure 2). The main regions of sequence conservation are associated with strand-1 and strand-4 of the core domain: the end of strand-1 is often characterized by a polar residue, typically with an alcoholic side chain (S/T), whereas immediately downstream of strand-4 is a highly conserved basic residue (K/R) often associated with [DN]X[ST]XXX[RK] signature. The position of these characteristic motifs is typical of the location of substrate-binding sites across a diverse range of Rossmann-like domains (Anantharaman and Aravind, 2006; Burroughs et al., 2006, 2009) with the implication that the ligand-binding capability is conserved throughout the CARF superfamily. Consistent with this prediction, probing the active site with a probe of 2 or more solvent radii shows the presence of a conserved pocket that is formed largely by the residues from the aforementioned motifs associated with strand-1 and strand-4 (Figure 2, Supplementary File 3). The conservation of K/R after strand-4 and its location in the pocket is consistent with the proposal of a nucleotide or nucleotide-derived molecule being the primary ligand of the CARF domains (Lintner et al., 2010). However, the RV2818 and RtcR families mostly lack the positively charged residue downstream of strand-4 suggesting that they might bind distinct ligands.

Examination of the structures also shows that the core fold of the CARF domain is prone to considerable divergence due to several distinct inserts (Figure 3). For example, in the group that consists of the SSO1393, sll7062, ST0035, and MA0186 families, there is an α-helical bundle inserted immediately after strand-1. Likewise, in the PF1127 family, a β-hairpin is inserted after strand-1 and multiple additional inserts are present after strand-2, strand-3 and in the β-hairpin formed by strand-5 and strand-6 (Figure 3). Based on the sequence alignments, we also detected smaller but comparable inserts after strand-1 in most members of the Aq_376 group and several members of the DET1451 group. These inserts typically are packed around the active site and form a “cap” that appears to shelter and augment the conserved ligand-binding site. The repeated emergence of inserts in similar locations in different families suggests that they might be determinants of ligand diversity across the CARF superfamily.

Another striking feature revealed by the comparison of the available structures is the diversity of spatial positions of the C-terminal wHTH and effector domains (Figure 3) vis-à-vis the CARF domain. This diversity of spatial positions is in sharp contrast to the strong positional polarity that is typical of prokaryotic one-component transcription factors with respect to their upstream ligand-binding domains (Aravind et al., 2010). Instead, it appears likely that the spatial organization of the C-terminal domains reflects optimization for transmitting the signal generated by the bound ligand to different C-terminal effector domains. This observation is compatible with the proposal that in most members of the CARF superfamily ligand binding is not directly linked to transcription but rather affects other DNA-associated activities (See discussion below).

The majority of the families contain a wHTH domain downstream of the CARF domain (Figures 1A,B, Supplementary File 2). In the PF09659 and PF09670 related families, we were unable to identify a HTH domain; instead, proteins in both these families contain a distinct, conserved alpha-helical region (6H domain) (Supplementary File 3). In the largest family (PF09455), the wHTH domain cannot be identified by sequence similarity searches (Kim et al., 2013) but an α-helical domain of uncertain provenance, potentially derived from a wHTH is present at the C-terminus, and harbors a partly disordered insertion that contains a highly modified remnant of the HEPN domain. In addition to the previously described fusions to DNases and RNases, several new domain architectures were identified in this analysis, namely (1) fusion of two CARF domains, (2) a membrane-associated CARF, (3) fusion with a HD phosphoesterase domain, (4) fusion to a TIM barrel adenosine deaminase Ada domain the enzyme that catalyzes deamination of adenosine to inosine in the purine salvage pathway (Nygaard, 1977; Holm and Sander, 1997). Notably, fusion of the CARF domain with nuclease domains of the same family might have occurred independently on several occasions. In particular, we detected at least four distinct CARF families associated with the HEPN domain and two families associated with the PIN domain (Figures 1A,B). Overall, most of the C-terminal catalytic domains of the CARF superfamily proteins are predicted to be nucleases or other enzymes targeting nucleic acids (Makarova et al., 2013).

A small family consists of large multidomain proteins in which a Zn ribbon, a serine/threonine/tyrosine protein kinase and a distinctive AAA+ ATPase domain with an arginine finger within the P-loop motif are fused upstream of the CARF and wHTH domains (Figure 1B). In this case, the CARF domain might function as part of a signal transduction pathway mediated by the kinase. The RtcR proteins in addition to the divergent CARF domain contain a NtrC-like AAA+ ATPase and HTH domains. Furthermore, BLAST search initiated with the CARF-like domain of RtcR detects high similarity with a family of proteins that, similar to RtcR, contain NtrC-like AAA+ ATPase and HTH domains but are not linked to Rtc system. Instead these proteins are often associated with restriction-modification (R-M) systems (Supplementary File 4). One of the close homologs of these proteins, PspF, which contains AAA ATPase and HTH domains only, has been shown to be involved in sigma-54 dependent activation of membrane-associated phage shock protein (PSP) system in response to phage infection and other stress factors (Model et al., 1997; Joly et al., 2009, 2010). Thus, these systems are likely to function as sigma-54 dependent activators of their respective downstream genes, with the NtrC-like AAA+ domain binding the sigma factor. In these proteins, CARF domains might sense ligands generated during or after phage infection, such as RNA with 2′–3′ cyclic phosphate ends or a phage-specific nucleotide to regulate either RNA repair or DNA restriction. Thus, the central functional theme for the majority of CARF superfamily domains, whether associated with CRISPR-Cas systems or not, seems to be antivirus defense and stress response.

To further characterize potential functional partners of the CARF proteins, we analyzed their genomic context by examining both known and new proteins families in the respective genomic neighborhoods. All gene products from these neighborhoods were collected, clustered using BLASTCLUST and analyzed using PSI-BLAST to further expand the respective families. The most common families associated with the CARF-domain proteins are shown in Figure 1C.

The WYL (named for three conserved amino acids found in a subset of domains of this superfamily) domain proteins are most abundant. Recently, it has been shown that a WYL domain protein (sll7009) is a negative regulator of the I-D CRISPR-Cas system in Synechocystis sp. (Hein et al., 2013). Further analysis of the WYL domain showed that the domain boundaries, as currently defined in the Pfam database (PF13280), are inaccurate because they encompass both a copy of the domain WYL domain (Supplementary File 5) and an additional C-terminal extension which is found primarily in the subset of WYL proteins with wHTH domains. HHpred searches revealed similarity of the refined WYL domain with SH3 β-barrel fold related to Sm domains (Supplementary File 2). Additionally, these searches showed that the uncharacterized Pfam DUF2693 family and the YolD family encoded in SOS DNA repair-associated operons (Permina et al., 2002; Aravind et al., 2013) are also members of the WYL domain superfamily (Supplementary File 2). Although the WYL domain was originally named for the 3 eponymous amino acids, examination of the refined and expanded alignment generated in the course of this work showed that these residues are not strongly conserved throughout the family. Rather, the conservation pattern includes four basic residues and a position often occupied by a cysteine (Supplementary File 5), which are predicted to line a ligand-binding groove typical of the Sm-like SH3 β-barrels (Gutierrez et al., 2007). Given that WYL domains often occurs in two copies in the same polypeptide or are encoded alongside other genes encoding multi-WYL proteins, it is conceivable that they form torroidal multimeric assemblies similar to other Sm-like SH3 β-barrels with a central ligand-binding channel (Schumacher et al., 2002).

In terms of domain architectures, WYL domains are most frequently associated with different predicted DNA-binding N-terminal wHTH domains. However, similar to the CARF domains, WYL domains also show fusions to several enzymatic domains (Supplementary File 6). In some of the type I CRISPR-Cas systems, a WYL domain is fused to the Cas3 protein which consist of a HD phosphoesterase domain and Superfamily-II helicase module. Additionally, WYL domains combine with 3′→5′ exoRNase, Mrr-like REase, HNH endonuclease, Superfamily-I helicase, AbiGII-like nucleotidyltransferase (DUF1814), BRCT, and TerB domains (Anantharaman et al., 2012). These fusions, the relationship between the WYL domain and the Sm-like domains, and the sequence conservation pattern of the WYL domain together seem to suggest that this is another ligand-sensing domain that could bind negatively charged ligands, such as nucleotides or nucleic acid fragments, to regulate CRISPR-Cas and other defense systems such as the abortive infection AbiG system (O'Connor et al., 1996; Makarova et al., 2013).

Several cas genes are enriched in the gene neighborhoods of the CARF superfamily (Figure 1C, Supplementary File 7). One of these, csx19, is always associated with CRISPR-Cas systems, and is predicted to represent a diverged version of the RAMP domain (RRM-like fold) that is found in many Cas proteins (Makarova et al., 2011a). Thus, colocalization of the csx19 genes with the genes encoding CARF domain proteins might simply reflect their shared association with the CRISPR-Cas systems rather than a direct functional link. In addition, cas genes of another, less common family, Csx15, are fused to the genes coding for CARF domain proteins on several occasions (Figure 1B). The Csx15 proteins show no significant similarity to any known domains, and their functions remain obscure. However, the presence of several highly conserved residues, namely two histidines, glutamate, and arginine are reminiscent of active site residues of metal-independent RNases (Zhang et al., 2012) and could be potentially involved in catalysis (Supplementary File 7). This together with the CARF domain fusions (Figure 1B), suggest that Csx15 might be a novel nuclease.

The association of CARF domain-containing proteins with CRISPR-Cas systems, especially those of type III, has been noted previously (Makarova et al., 2011a,b; Anantharaman et al., 2013; Koonin and Makarova, 2013). Here we sought to identify specific associations with CRISPR-Cas systems for each major family of CARF-domain proteins separately. The assessment was based on the proximity of the respective genes to CRISPR-Cas loci. Most of the 12 major CARF families are indeed typically found in vicinity of other cas genes (Figure 1A, Supplementary File 8), with the exception of DET1451, MA0186, and the divergent RtcR-like family. Those families of CARF-domain proteins that are associated with CRISPR-Cas systems most often are contained within type III CRISPR-Cas systems, and some show specific preference for type III-A or III-B. All these CARF domain protein families possess a third domain, a nuclease, which is predicted to function as a toxin that targets non-self or self-nucleic acids (Koonin and Makarova, 2013). The only CARF family (Csa3) that displays clear affinity to type I systems, and subtype I-A in particular, lacks a C-terminal catalytic effector domain. However, these associations notwithstanding, there are genes in each CARF family that are not linked to CRISPR-Cas and thus might not be functionally involved in the CRISPR-Cas-mediated defense. Some of the CARF genes that are not linked to CRISPR-Cas (e.g., Daci_4198 from Delftia acidovorans) of the VC1899 family (PF9002) are embedded within a novel Type-VII secretion system gene cluster predicted to function as a DNA-transfer agent and additionally encompassing multiple Ter genes that have been implicated in phage restriction (Anantharaman et al., 2012).

CARF domain-containing proteins are present in 145 genomes (among the representative set of 659 complete archaeal and bacterial genomes) of which only 9 genomes possess neither cas1 nor cas10 (the signature protein families of CRISPR-Cas systems), suggesting a strong link of these proteins to CRISPR-Cas (Supplementary File 9). Type III CRISPR-Cas systems often co-occur with type I system, so it was of interest to clarify whether a specific link existed between CARF domain and type III systems and whether or not this linkage depended on the presence of a C-terminal catalytic effector domain in the CARF-domain proteins. To address this question, we compared the co-occurrence of at least one CARF-domain protein containing a (predicted) effector domain with type I, type II, and type III CRISPR-Cas systems (Supplementary File 9). The data presented in Figure 1D clearly demonstrate a strong, specific link between CARF proteins containing a C-terminal catalytic effector domains and type III systems. This association suggests that CARF-domain proteins with this type of architecture play important roles in the majority of type III systems.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.