The C domain has two binding sites: one for the electrophilic donor substrate (the acyl group of the growing chain) and one for the nucleophilic acceptor substrate (the activated amino acid). The condensation reaction involves catalysis of a nucleophilic attack by the amino group of the aminoacyl adenylate bound to the downstream PCP on the acyl group of the growing peptide chain which is bound to the upstream PCP 211. The acceptor site of the C domain was shown to exhibit a strong stereoselectivity and significant side chain selectivity. The selectivity towards a specific side chain seems to be less pronounced at the donor site which, however, exhibits strong stereoselectivity 3.

In particular, C domains succeeding an E domain are expected to show specificity towards the configuration (L or D) of the C-terminal residue that is bound at the donor site because the preceding E domain does not specifically catalyze the epimerization from L to D but provides a mixture of configurations. It is the role of the C domain to select the correct enantiomer 11. Moreover, the C domain represents some kind of selectivity filter in that it supports the selection of the correct downstream nucleophile and prevents product mixtures 2.

C domains immediately downstream of E domains were shown to be D-specific for the upstream donor and L-specific for the downstream acceptor, thus catalyzing the condensation reaction between a D- and an L-residue. These C domains were termed ^DC_L-catalysts because of this behavior 12.

Accordingly, ^LC_L-catalysts promote the condensation of two L-amino acids. Both ^LC_L- and ^DC_L-catalysts possess a conserved His-motif in their active site. The consensus sequence of this motif is HHxxxDG where x denotes any residue (see Fig. 2, motif 3). The second His-residue seems to be essential for the catalytic function of the domain 2.

As a third type of C domain, so-called Dual Epimerization/Condensation (E/C) domains have recently been identified. This finding was based on the observation of NRPS which had products that contained D-residues although the NRPS itself did not show an E domain in the corresponding module. Biochemical experiments supported the hypothesis that Dual E/C domains exist which are ^DC_L-catalysts with epimerase activity 13. In the assembly line, a Dual E/C domain follows directly after a C-A-T module which activates and incorporates an L-amino acid. The module which contains the Dual domain also activates an L-amino acid. Then the Dual domain catalyzes the epimerization of the L-residue into D configuration and subsequently promotes the condensation of those two residues. In addition to the active site His-motif which is found in all C domains, Dual E/C domains exhibit a second His-motif, HH[I/L]xxxxGD, which is located close to the N-terminus of the domain 13 (It is partly located on motifs C1 & C2; see Fig. 2.)

C domains may be replaced by Heterocyclization (Cyc) domains which catalyze both peptide bond formation and subsequent cyclization of cysteine (Cys), serine (Ser), and threonine (Thr) residues. The five-membered heterocyclic rings which result from this reaction are important for chelating metals or interaction with proteins, DNA or RNA. Cyc domains are structurally related to C domains and are supposed to be evolutionary specialized C domains 2. In Cyc domains, however, the active site His motif is replaced by another conserved motif, DxxxxD. Keating et al. 14 found that the aspartate (Asp, D) residues are critical for both condensation and heterocyclization.

A total of 481 Condensation domains (including their homologues, Epimerization and Heterocyclization domains) were extracted from 182 (non-identical) NRPS and 31 NRPS/PKS hybrid sequences found in 62 bacterial genomes out of the 256 bacterial genomes screened, employing pHMMs as described in Section Methods (Note that only one genome was considered for our analysis if sequences of several strains of the same species were available, which reduced the number of NRPS or 'hybrid NRPS/PKS' containing genomes from 62 to 43). Altogether 108 C domains were obtained from 42 NRPS sequences from gene clusters downloaded from the UniProt database. After removing doublets, all 525 non-identical C domains and homologues obtained were multiply aligned and phylogenetic trees were built. The resulting tree topology was clearly dominated by the functional categories that are known for C domains (as described in the previous section), rather than species phylogeny or substrate specificity alone. The four main functions are: 1. condensation performed by ordinary C domains; 2. condensation and subsequent heterocyclization catalyzed by Heterocyclization (Cyc) domains; 3. epimerization followed by condensation which are both catalyzed by a Dual E/C domain; 4. Starter domains (see below) which are found on initiation (= first) modules and acylate the subsequent amino acid.

Ordinary C domains may further be classified into ^LC_L-catalysts and ^DC_L-catalysts according to the stereochemistry of their substrates. The existence of all these functional subtypes is reflected by the phylogeny. Fig. 3 shows a phylogenetic tree for subsets of each C domain subtype, as the whole tree of 525 taxa is far too large to be displayed here (see Additional files 1 and 2). The tree of all taxa showed a similar topology perfectly reflecting the functional categories.

For further analysis, the different subtypes were examined separately. While Cyc and Dual E/C domains could be identified by means of their characteristic sequence motifs (see Section Methods/Predicting of functional subtypes), ^LC_L- and ^DC_L-catalysts were either distinguished according to their domain structure or by their position in the phylogenetic tree. By this, 275 domains of all 525 C domains were classified as being ^LC_L-catalysts, 69 were ^DC_L-catalysts and 42 were Starter C domains (see next section).

When analyzing the Condensation (C) domain phylogeny, it became apparent that some domains did not cluster with the known C domain subtypes. A closer look at the location of these deviating C domains revealed that all of them were the very first C domain of the corresponding NRPS assembly line. The remaining C domains of these assembly lines appeared in other subtrees in the phylogeny.

Included in this set of starter C domains are those stemming from the biosynthesis clusters for the lipopeptides surfactin 15, lichenysin 16, fengycin 17 and arthrofactin 18. These lipopeptides are characterized by a β-hydroxyl fatty acid which is connected to the first amino acid of the peptide chain 19. The peptide synthetases involved in the production of these lipopeptides all have a C domain as their very first domain. This C domain is supposed to serve as an acceptor for a fatty acid which is transferred from an acyltransferase 19. This acylation process has also been observed for surfactin 20 and fengycin biosynthesis 21. Moreover, common to the Starter C domains of these biosynthesis clusters is their low sequence similarity to the remaining C domains of the same biosynthesis cluster 19.

The same has been observed for the synthesis of the acidic lipopeptide CDA in Streptomyces coelicolor 22 and the recently identified lipopeptide produced by protein NP_960354.1 of Mycobacterium avium 23. The Starter C domain of the pristinamycin cluster appears to diverge from this pattern at the first view. The C domain is the first domain of the polypeptide SnbC but the biosynthesis of pristinamycin is initiated by SnbA, which contains an A domain that activates 3-hydroxypicolinic acid (3-hydroxypyridine-2-carboxylic acid, "2-hydroxy-6-azabenzoate") but lacks an ACP 24. SnbA is homologous to EntE, which contains an A domain specific for 2,3-dihydroxybenzoate (DHB) and which is involved in the biosynthesis of enterobactin 25. A similar organization can be found in actinomycin biosynthesis. The process is initiated by AcmA, which activates 4-methyl-3-hydroxyanthranilic acid (MHA, 4-methyl-3-hydroxy-2-aminobenzoate) 26. In conclusion, what the C domains of SnbC, AcmB and EntF have in common is that they catalyze bond formation between a derivative of salicylic acid (2-hydroxy-benzoate) and an α-amino acid. Assured by the fact that these Starter C domains match significantly well to the profile HMM built from the Starter C domain sequences that process β-hydroxy fatty acids, we compared salicylic acid with β-hydroxy fatty acids. Because both are β-hydroxy-carboxylic acids with no amino-substituent at the α position, as a-amino acids would have, we assume that this is the structural characteristic recognized by the prototype of Starter C domains. The profile HMM built from all Starter C domains in our data set (together with the pHMMs of the other domains) presents a powerful instrument for exploring and understanding tricky NRPS domain-product relations.

Note that Formylation domains as found, for example at the N-terminus of linear gramicidin synthetase subunit A 27 are not C domains but belong to the Pfam "formyl transferase" domain family.

The different core motifs in Condensation domains have first been described by de Crécy-Lagard et al. 28 and recompiled by Marahiel et al. 29 but have never been updated since then. The core motifs of the C domain homologues, Epimerization and Heterocyclization domain are listed in the publication by Marahiel et al. 29 but the sequence motifs of the recently discovered ^DC_L domains 1230 as well as the Dual E/C 13 domains have never been comprehensively analyzed. Moreover the Starter C domain has not yet been recognized in the literature as a proper separate subtype.

The sequence motifs represented in Fig. 2 improve the C domain core motif consensus sequences published by Marahiel et al. 29 which, at that time, were based on much fewer sequences and did not differentiate between the C domain subtypes. The motifs are represented as sequence logos 31 which make it easier to identify variably conserved positions compared to simple consensus sequences. We adhere to the core motifs identified by Marahiel et al. 29, and also show the surrounding "landscape" if there are highly conserved positions nearby, especially if they are important for distinguishing between the C domain subtypes. The motifs were built on the basis of 40 verified and 198 predicted ^LC_L sequences, in which "predicted" means that they were classified based purely on their position in the phylogenetic tree while "verified" sequences were checked individually taking into account their position in the succession of neighboring NRPS domains, the presence of discriminative unique motifs (see Methods Section) and/or literature information. For the ^DC_L motifs, 23 verified and 46 predicted sequences were used, 7 verified and 35 predicted for the Starter domains, and domains 9 verified and 47 predicted for the Dual E/C domains.

Based on three publications, four residues are likely to be essential for the catalytic activity of the C domain. The most important residue is the 2nd His of the active site His-motif 32.

Furthermore, six residues have been identified as being structurally important or as playing a role in correct folding of the domain. In the following, these residues are presented, grouped by their role (the numbering is according to their linear occurrence on the peptide; see Fig. 2). This information is also presented in Table 1 where the sites are sorted by their relative position in the domain.

While not being conserved at residues Nb. 5, Nb. 7, Nb. 9, and Nb. 10, all remaining 6 functionally important residues are highly conserved throughout the putative Starter domains. When comparing ^LC_L and Starter domains, 18 discriminative positions were found by SDPpred and 5 more were found in the motifs by FRpred. Those positions are highlighted in Fig. 2. Common to these residues is the fact that they seem to be highly conserved among extender (= ^LC_L) domains but show no conservation among Starter C domains. When we compare C domain sequence motifs, it is apparent that motifs C2 and C4, despite being well conserved in ^LC_L, are unconserved in Starter domains, which presumably can be explained by the much broader structural range of substrates processed by Starter domains.

The reconstructed phylogeny of C domain subtypes reveals that ^LC_L and Starter C domains are more closely related to each other than to other subtypes (see Fig. 3). Comparing sequence motifs confirms this observation, though pronounced differences in some segments of the protein (especially in motifs C2 and C3, as can be seen in Fig. 2) account for the unequal donor substrates (amino vs. β-hydroxy-carboxylic acid). Furthermore the phylogenetic tree shows that Dual E/C and ^DC_L domains share a common ancestor. We tested the reliability of the phylogenies depicted in Fig. 3 and Fig. 4 by repeating the reconstruction on biased profile alignments. These biased alignments were generated by producing MUSCLE profile-profile alignments in a step-wise manner, assuming evolutionary relationships of the different domain subtypes that are contradictory to what the original trees suggest. The topology of the resulting trees supports the shared ancestry of ^LC_L and Starter C domains as well as of Dual E/C and ^DC_L domains. In addition, we generated an alignment using DIALIGN 37, which is a non-progressive alignment method, and subsequently reconstructed a PHYML-tree based on this alignment. Here also, the Dual E/C and ^DC_L domains are grouped together as are ^LC_L and Starter C domains.

Especially in motif C5, Dual E/C and ^DC_L domains are very similar to each other and dissimilar to ^LC_L and Starter domains. This observation of the relationship between the four subtypes is consistent with the stereochemistry of the substrates, bearing in mind that Dual E/C domains function as ^DC_L because the substrate L-amino acid is first epimerized by the intrinsic epimerization activity of the domain 13.

Within the subtrees of ^DC_L and ^LC_L domains, the tree topology reflects the species phylogeny of the bacteria rather than substrate specificity of any kind. We analyzed this by reconstructing phylogenies for ^DC_L domains and ^LC_L domains separately to be able to see the topology within these subtypes in more detail (data not shown). The reconstructed phylogenies did not give any evidence that would support the hypothesis that C domains cluster according to their specificity towards the condensated amino acids. This analysis, however, is based on the complete C domain sequence. A strategy to investigate whether C domains exhibit substrate specificity would involve predicting putative specificity determining positions using entropy and/or conservation based approaches (e.g. SDPpred, FRpred), or inferring of putative active site residues by homology with the VibH structure (as done by Rausch et al. 38 for the adenylation domain).

Glycopeptide antibiotics are a subgroup of nonribosomal peptide antibiotics of which the best known representatives are probably vancomycin and teicoplanin. To date, all identified glycopeptide antibiotics are produced by actinomycetes. They interrupt cell wall formation of gram-positive bacteria by binding to the D-Ala-D-Ala termini of the growing peptidoglycan, thereby inhibiting the transpeptidation reaction. All glycopeptide antibiotics consist of a heptapeptide backbone which is synthesized by NRPS.

Modification reactions involve extensive cross-linking of the aromatic side chains to rigidify the molecule 3940. The modular organization of some NRPS which were identified in glycopeptide-producing actinomycetes are depicted in Fig. 5.

All these NRPSs comprise seven modules. They show an identical domain composition, with the exceptions of module M3 in the A47934 (sta) and M3 and M6 in complestatin (com) clusters which contain an E domain not present in the other clusters. The M3-E domain, however, is assumed to be inactive 41, while the presence of an E domain in com M6 has not been reported elsewhere so far. We were able to detect it with an hmmpfam scan using the specific E domain pHMM. All six NRPSs contain a domain X* of unknown function. Until now, it has been characterized as an atypical C or E domain but its role in glycopeptide synthesis remains to be clarified. In general, it is assumed that the stereochemistry of a NRPS product can be predicted from its domain structure. In the case of the known glycopeptides, the domain organization implies the stereochemistry NH₂-L-D-L-D-D-L-L-COOH, provided that the E in module M3 is inactive and that the X* domain does not function as an E domain. This stereochemistry is inconsistent with the chemically determined structure of the products: NH₂-D-D-L-D-D-L-L-COOH 41. The assumption is that the A domain of the first module activates a D-amino acid. For the cep cluster, however, Trauger and Walsh 42 show that the A domain of M1 prefers L-Leu over D-Leu in a 6:1 ratio; but on the other hand, they could not show which stereoisomer is processed further. This suggests the existence of an unknown E domain that acts on the L-Leu activated by M1. With the discovery of Dual E/C domains, a new possible strategy arises for the incorporation of a D-residue by the first module. However, no Dual E/C domain could be detected in all glyco-NRPS. Alternatively, one could imagine an external racemase as is found in the cyclosporin cluster 43, which provides a D-Leu that can be incorporated directly.

Having gained knowledge about the differences between ^LC_L, Starter and ^DC_L domains as described above, we examined all glyco-NRPSs. When we reconstructed the phylogeny of C domains including all homologous domains from glyco-NRPSs, it was staggering to find that all C domains were clustered in the ^DC_L subtree and the X* domain clustered in the ^LC_L subtree (see Fig. 4). This finding could be confirmed by analyzing all instances of the C domain motifs found in these domains. How could this be interpreted, given the fact that M4 and M7 C domains clearly act as ^LC_L domains, as we can tell by the stereochemistry of the products? Our hypothesis is that those C domains are former ^DC_L domains that have developed ^LC_L activity by convergent evolution. Accumulating supportive evidence is possible: When we look at the phylogeny of the C domains, the sequences of the com cluster from Streptomyces lavendulae are always most distant from the others and more closely related to the hypothetical common ancestor, implying that they can serve as a model for the archetype of glyco-C domains. It is likely that in the archetype, all C domains were true ^DC_L catalysts, supposing that the E domains which are still present in com modules M4 and M7 were still active.

In a similar way, we can trace back the origin of the X* domain: in the com cluster (and only there) it is followed by remnants of an adenylation domain (which has several larger insertions and deletions; see Additional file 4). This tells us that the X* domain used to be the first domain of a new module followed by an adenylation domain.

The assumption that the diverged C domains of modules M4 and M7 would have adopted mutations at positions that we have previously determined as "specificity determining positions" was disproved. Probably, a few spontaneous mutations in the ^DC_L domains relaxed the stereo-selectivity; supposing that this altered stereochemistry of the product resulted in a highly selective advantage (arising from a vancomycin-like product), the loss of the functional E domains in M3 and M6 would have been a selective gain. Comparing all M4 and/or all M7 C domains with all ^DC_L domains using SDPpred did not reveal any significant positions; comparing them against the other glyco-C domains gave thirty positions. As all glyco-C domains are very closely related and differences between them might also reflect substrate selectivity (not only stereo-selectivity) or different inter-domain interacting residues, we cannot decide which of them confer the altered stereo-selectivity. One point to notice however, is a (positively charged) His in all M4 glyco-C domains at position 6 in the extended motif C2 where an (uncharged polar) Gln is highly conserved in other ^DC_L domains. This position has also been selected by FRpred as a significant (= subtyping) position. The other positions do not represent mutations in highly conserved residues (data not shown). It would be necessary to check their significance experimentally with mutation studies. It would also be helpful to compare the peculiar sequences with more glyco-C domains, but others are -unfortunately – not publicly available.

However, although we could not discover which altered positions are responsible for the functional shift from ^DC_L to ^LC_L in glyco-C domains, interesting experimental questions can be formulated based on our findings. For example, one could think of mutational studies with the goal of altering the stereo-selectivity of a ^DC_L domain and to determine the relevant residues experimentally. A starting point could be, for example the M6 C domain of any glyco-NRPS.

The second His of the His-motif in motif C3 which is important for catalysis is replaced by Arg (R). Also, the Gly of the His-motif is not present but replaced by Arg in all but one X* domain. Note, however, that while the second active site His is invariant in C domains, Gly138 is not.

SDPpred predicted 13 specificity determining residues when comparing M7-X* to ^LC_L-domains of Streptomyces species. Only three of these coincide with residues of functional importance: His126, Arg278 and Asn335. Furthermore, a C terminal region could be detected in which M7-X* and ^LC_L differ strikingly. The concordance of M7-X* with the most highly conserved residues of Streptomycete ^LC_L domains supports the phylogenetically based suggestion that M7-X* is an inactive ^LC_L domain.

The protein sequences and GenBank entries for all completely sequenced bacterial genomes available to date were obtained from the NCBI FTP site ftp://ftp.ncbi.nlm.nih.gov/genomes/Bacteria/. In total, the genomes of 256 bacterial species were downloaded and screened for NRPS protein sequences (including NRPS/PKS hybrids). Additional protein sequences of PKS and NRPS which are part of known secondary metabolite biosynthesis clusters were obtained from the UniProt database 45. NRPSs were retrieved from 14 known biosynthesis clusters, of which 13 came from Actinomycetes and one from Pseudomonas (see Additional file 8).

A common strategy for the identification of a specific type of domain is to use Profile Hidden Markov Models (pHMMs), which are statistical models extracted from multiple sequence alignments. In contrast to simple sequence motifs of fixed length, i.e. position specific scoring matrices, pHMMs are suited for identifying motifs that are interrupted by segments of variable length, and are used to characterize position-specific sequence similarities within a family of proteins. A collection of pHMMs for a wide array of domains and domain families is availabe from the database Pfam 46 and TIGRFAMs 47. The pHMM implementation HMMER 4849 and self-written Perl 50 scripts and BioPerl 51 scripts were used to search for NRPS in the genome sequences and biosynthesis clusters and to extract single domains from a given protein sequence. To identify a protein sequence as an NRPS, the occurrence of at least one complete NRPS module with one C domain, one A domain and T domain was required (Pfam accession numbers PF00668, PF00501 and PF00550), with an E-value threshold of 0.1 (thus we accepted to miss freestanding starter modules containing only A and T domains, or had to add them manually, as in the case of the biosynthesis clusters).

The Pfam pHMM Condensation (PF00668) recognizes both the Condensation (C) and Epimerization (E) domain of NRPS. The intention, however, is to be able to distinguish between these two domain types. Therefore C domain and E domain specific pHMMs were generated from a multiple sequence alignment (MSA) of Epimerization domains and non-Epimerization domains, both of which were recognized by the Pfam C pHMM. To obtain a set of Epimerization domains, all NRPS sequences with complete modules were extracted from all bacterial protein sequences in the Uniprot database 45 as described above. Whenever two consecutive C domains followed by an A domain were detected with Pfam pHMMs, the "first C" domain was extracted. That way, we obtained a set consisting mainly of E domains (151 of 237 sequences). By phylogenetic subtyping (as described below) we determined the E domain sequences from the phylogenetic tree of the "first C" domains, which were forming a distinct subtree. The E and non-E sequences were aligned with MUSCLE 5253, and specific pHMMs were build for them with hmmbuild and hmmcalibrate from the HMMER package (As a control, it was not possible to detect E domains in the 771 "second C" domains). The domain sequence covered by our own pHMMs for C and E domains is identical with that of the Pfam Condensation pHMM; in other words it extends from four positions before our extended C1 motif to the fourth position after the extended C5 motif (these motifs were first revealed by de Crécy-Lagard et al. 28 and reviewed by Marahiel et al. 29). Phylogenetic reconstruction is always based on this part of the C domain (see Fig. 2). To extract the complete N-terminal part of the C domains, we followed the dissections applied by Roche and Walsh 33 and checked the secondary structure with Quick2D of the MPI Bioinformatics Toolkit 5455.

The quality of a reconstructed phylogenetic tree crucially depends on the underlying multiple sequence alignment. All sequence alignments in our study were generated using MUSCLE 5253. The alignment algorithm can be divided into three stages. First, a progressive alignment is built based on a UPGMA guide-tree. In the second stage, the underlying guide-tree is iteratively improved, yielding a new progressive alignment. The third stage involves refinement of the tree: Based on the tree, bipartitions of the dataset are produced; their profiles are extracted and realigned to each other. Thus, the finally generated alignment is not solely based on a single guide-tree, which is why we can rule out that the phylogenies reconstructed on the basis of these alignments merely reflect the guide-tree used in the first step of the algorithm.

C domains catalyze the condensation of two amino acids, thus, they have two binding sites: the acceptor and the donor site. To be able to investigate whether the substrate specificity of one of these sites influences the phylogeny of the domain, the specificity of the preceding and succeeding A domain in the assembly line was predicted with the NRPSpredictor 38 and stored for each C domain.

Functional subtypes may be distinguished on the basis of sequence features, domain architecture or clustering behavior during tree reconstruction. Condensation and Heterocyclization domains may be discriminated by the sequence motif they exhibit at their active site. The occurrence of a sequence motif within a longer sequence can be detected with the help of a position specific score matrix (PSSM) 48. PSSMs were generated and applied for the detection of the active site His-motif of the C domain and the DxxxxD-motif of the Heterocyclization domain. These were used to discriminate between the two subtypes. The His-motif was built from 86 sequences and the Cyc motif from 15 sequences. The PSSMs were only applied to a region of 100 residues which was expected to contain the active site. In addition, a PSSM was generated for the N-terminal His-motif found in Dual E/C domains. It was constructed from 55 sequences which had been identified as Dual E/C domains by their clustering behavior in the phylogeny and by additional visual inspection of the alignment. The PSSM was applied for validation purposes to make sure that this N-terminal His-motif is unique to Dual E/C domains and cannot be found in any other C domain subtype. Predicting whether a C domain is a ^LC_L- or a ^DC_L-catalyst was established according to the observed domain organization of the modules in an NRPS sequence (^DC_L-catalysts were first described by Luo et al. 30). It is assumed that the role of a module with the domain structure C-A-T-E is the activation and epimerization of a residue that is in the L stereo configuration with the intention of incorporating a D residue into the final product. Alongside this, a C domain directly following an E domain is expected to be selective for residues in D-configuration, which is why it was assigned to the ^DC_L-type. All other C domains were assumed to be ^LC_L-catalysts. Classification as a ^DC_L-catalyst is supposed to be fairly reliable. A false positive should only occur if the preceding epimerase turns out to be nonfunctional. The ^LC_L classification, however, is prone to errors when the respective C domain is the very first (N-terminal) domain in the protein. In this case, the type of the condensation reaction can only be assigned if the order in which the proteins act in the assembly line is known. To overcome this problem, we checked all assignments with the classification suggested by the phylogeny.

If the order of the subunits is unknown, temporarily incorrect assignments can only be revised later in the analysis.

In a set of homologous enzymes, we may find subsets that each contain sequences with one distinct substrate specificity. These subsets of common function are called subtypes and often vary at certain positions, whereas the same positions may be conserved within a given subtype. Li et al. 56 call these specificity-determining residues (SDR); Kalinina et al. 35 refer to them as specificity determining positions (SDP). To determine SDPs from an alignment, calculating each column's mutual information is a possible way, as described by Li et al. 56 and Kalinina et al. 35. For this paper, SDPs were determined using the freely accessible SDPpred server 35. Here, the mutual information is based on so-called smoothed frequencies, which allow substitution of residues with similar physico-chemical properties. In addition to that, the significance of the mutual information of each position is estimated by calculating Z-scores and evaluating their significance. Predictions by SDPpred were compared with the highest scoring positions predicted by FRpred 3657 which combines a mutual information term with a conservation score.

Several methods were applied for reconstructing phylogenetic trees from the multiple sequence alignments that were generated for each domain type. Trees presented in this article were reconstructed using protein sequences, as amino acid sequences are preferred to nucleotide sequences because they are more conserved and are not influenced by compositional bias like G+C content and codon usage. In addition, the mathematical model for the evolutionary change of amino acid sequences is much simpler than that of nucleotide sequences, which reduces the risk that the phylogeny is based on wrong evolutionary assumptions, since just a suitable substitution matrix has to be selected 58. The amino acid substitution matrix employed in this study was the Jones-Taylor-Thornton (JTT) matrix 59.

In some cases, the rate of amino acid substitution may be assumed to be the same for all positions in the alignment. In general, however, this does not reflect reality since the substitution rate is usually higher at positions of lower functional importance. A more realistic model is achieved if the substitution rate is taken to vary among sites according to the gamma distribution 60.

Apart from PHYLIP 61, all methods used in this study offer an estimation of parameter α which determines the shape of the Γ distribution as an option. Whenever a gamma distributed rate variation was assumed, four gamma-rate categories were used to approximate the distribution. Several tree reconstruction methods were applied to each dataset to determine whether different methods yield different topologies, which in turn would indicate that the proposed topologies are unreliable. As a distance-based method, the Neighbor-Joining (NJ) method 62 was applied. The distances were calculated with the program protdist and NJ was performed with neighbor, both available from the PHYLIP package. For NJ, only uniform substitution rates were used. As a maximum likelihood method, the programs IQPNNI 63 and PHYML 64 were applied.

Bootstrapping 65 was performed to test the reliability of the topologies.

In general, a topology is taken as reliable if tree reconstruction results in the same topology for at least 95% of the datasets generated by bootstrapping. This is a quite strict approach and it has been shown that subtrees of a tree may be accepted as being significant if they are supported by only 70% of the trees 66. Using the PHYLIP package, bootstrap datasets were generated with seqboot and used as input data for neighbor. PHYML also offers an option that allows a bootstrap analysis of the original data. This results in a set of trees which can be visualized as a consensus network using SplitsTree4 67. The specification of a cutoff value allows a clearer view of the bootstrap tree/network where only those edges which are supported by boostrap values higher than the cutoff are included.

The program meme 6869 was used to detect the sequence motifs in C domains. Meme discovers one or more motifs in a collection of unaligned DNA or protein sequences. The C domain subtypes were aligned using MUSCLE 5253, the multiple alignments were visualized using JalView 70 and the motifs found by meme were extracted (cut out). It was ascertained that the C domain motifs described by Sieber and Marahiel 2 were included as well as remarkable sequence positions in the proximity of the motifs, such as single conserved residues or positions which were important for discerning the subtypes. The dissected motif sequences were used to create pHHMs with HMMER and also to create sequence logos using seqlogo by Crooks et al. 31. Sequence logos were prefered over consensus sequences, as they provide a more precise description of sequence similarity and reveal significant features of the alignment which are otherwise difficult to perceive. For sequence logos, positions with > 10% gaps were removed. Sequence logos of all C domain motifs created with seqlogo are available online as Additional file 9.