Bacterial Protein Structures Reveal Phylum Dependent Divergence

Abstract

Protein sequence space is vast compared to protein fold space. This raises important questions about how structures adapt to evolutionary changes in protein sequences. A growing trend is to regard protein fold space as a continuum rather than a series of discrete structures. From this perspective, homologous protein structures within the same functional classification should reveal a constant rate of structural drift relative to sequence changes. The clusters of orthologous groups (COG) classification system was used to annotate homologous bacterial protein structures in the Protein Data Bank (PDB). The structures and sequences of proteins within each COG were compared against each other to establish their relatedness. As expected, the analysis demonstrates a sharp structural divergence between the bacterial phyla Firmicutes and Proteobacteria. Additionally, each COG had a distinct sequence/structure relationship, indicating that different evolutionary pressures affect the degree of structural divergence. However, our analysis also shows the relative drift rate between sequence identity and structure divergence remains constant.

1. Introduction

Quantifiable models of protein evolution are useful for developing robust tools to identify suitable drug-binding sites, to predict increases in susceptibility to a human genetic disease, and to predict and modify organismal niches. Some of the strongest arguments in favor of biological evolution draw from studies on protein evolution using sequence homology (Do and Katoh, 2008). Multiple sequence alignments are routinely used to create phylogenetic relationships (Chang, et al., 2008; Feng, 2007), which highlights sequence variability between organisms. The accepted view of protein evolution is that changes to the protein’s gene sequence are selected and modulated by a number of factors that includes structure (Pal, et al., 2006; Rocha, 2006).

What is the impact on protein structure as its sequence undergoes genetic drift? Maintaining the correct protein fold is fundamental to preserving its function (Forouhar, et al., 2007), but evolving the sequence would also be expected to result in structural changes (Chothia and Lesk, 1986a; Rost, 1999). The resulting observation is that sequence determines a protein’s structure, but the structure is relatively invariant over a large range of sequences. This is highlighted by the tremendous difference between the number of known protein structures versus protein folds (Sadreyev and Grishin, 2006). Even though the Protein Data Bank (PDB) (Berman, et al., 2000) contains 66,083 protein structures as of June 22, 2010, there are only 1,233 unique topologies and 1,195 unique folds in the CATH (Orengo, et al., 1997) and SCOP (Murzin, et al., 1995) structure classification databases, respectively. The significant reduction in the number of protein folds relative to the number of protein sequences implies a much stronger correlation between structure and function. Correspondingly, protein structures are generally viewed as more conserved relative to its sequence and recent studies have attempted to quantify this statement (Illergard, et al., 2009).

The explicit reason for the reduction in fold space remains unclear. However, some have suggested that protein fold space may be more appropriately described as a continuum instead of a collection of discreet folds (Kolodny, et al., 2006). In this manner, a protein fold should be considered as being plastic, where sequence changes are accommodated by local perturbations in the structure while maintaining the general characteristics of a particular fold (Illergard, et al., 2009; Panchenko, et al., 2005; Williams and Lovell, 2009). Correspondingly, the genetic drift in a protein’s sequence may imply a similar gradual divergence in structure instead of a sudden dramatic transition to a new fold. From this perspective, a comparative analysis of homologous proteins should identify correlated rates of structure and sequence divergence. Previous studies have looked at homologous structure similarity before but the datasets did not try to show structure divergence consequences on phylogenetic relationships (Illergard, et al., 2009; Panchenko, et al., 2005; Williams and Lovell, 2009). To help understand how protein plasticity affects organism divergence we compared 48 sets of homologous protein families annotated in the COG database for two bacterial phyla, Proteobacteria and Firmicutes.

2. Materials and methods

2.1. COG assignment of the protein data bank

Assignment of each bacterial protein in the PDB to a COG number in the clusters of orthologous groups (Tatusov, et al., 2003) database required downloading the complete sequence lists from both databases and running a pairwise Basic Local Alignment Search Tool (BLAST) comparison. The pairwise protein BLAST search was run using the Protein Mapping and Comparison Tool (PROMPT v0.9.2) (Schmidt and Frishman, 2006) that allowed for large pairwise BLAST searching and reported the best match between the two databases. The BLAST search was run using the BLOSUM62 matrix with a gap penalty of 11, gap extension penalty of 1, a word size of 5, and a BLAST expectation threshold (E-value) of 10⁻⁹. This E-value was used to unambiguously match genes in the COG database with proteins in the PDB. All PDB-to-COG matches were reported and stored in our PROFESS (Protein Function, Evolution, Structure, and Sequence) database (http://cse.unl.edu/~profess/).

After matching structures to their representative COG each PDB entry was matched with its source organism and phylum. The data set was then filtered according to the number of unique organisms. Specifically, only those COGs with structures from two or more different source organisms in both Proteobacteria and Firmicutes were analyzed further.

2.2. Pairwise structure comparison

The pairwise structure comparison program DaliLite v2.4.2 (Holm and Park, 2000) was installed on our 16-node Dual Athlon AMD 2.13 GHz with 1 GB of RAM Beowulf cluster running CentOS 4.4 Linux with a 2.25TB RAID array. A C-shell script matches the PDB files from each Proteobacteria-Proteobacteria comparison (−/−), Firmicutes-Firmicutes comparison (+/+) and Proteobacteria-Firmicutes comparison (−/+) and then submits the job to the program DaliLite. Each structural comparison took approximately 2–10 min, depending on the size and relative similarity of structures. The total time to run all 63,504 comparisons was approximately 7 weeks.

The shell script extracts all structural comparison information reported by DaliLite (comparison files, rmsd, %Sequence ID, Z-score) on a per chain basis. A single PDB file may contain multiple protein chains, where each chain may have a separate COG assignment. All structure information is stored in our PROFESS (PROtein Function, Evolution Sequence and Structure) database (Triplet, et al., 2010), which is parsed to find the largest Z-score for each pairwise structure comparison. The largest Z-score represents the best structure comparison for a pair of proteins and ensures that the correct PDB chains were used for the analysis and the correct COG assignments were made. All best matches from each COG were used to calculate the Fractional Structure Similarity score (FSS) described by eqn. 1.

$$\mathit{FSS}=\frac{Z_{AB}}{\mathit{max}(Z_{AA},Z_{BB})}$$ [1]

where Z_AB was the Z-score for comparing proteins A and B, Z_AA was the Z-score when protein A was compared to itself and Z_BB was the Z-score when protein B was compared to itself. Thus, Z_AA and Z_BB represent the Z-score that can be achieved for perfect similarity.

2.3. Manual filtering and data analysis

Manual refinement of the dataset included verification of each PDB assignment to a COG and filtering out redundantly solved structures from the same organism. When multiple structures were reported from the same organism (or organism with synonymous name), the structure that gave the largest Dali Z-score within the COG was kept while remaining structures were discarded from the analysis. This confirmed a single best PDB-COG match for each organism. Manual refinement was accomplished by opening all PDB IDs within a COG and checking biological information against the PDB (http://www.rcsb.org/pdb/home), COG (http://www.ncbi.nlm.nih.gov/COG/) and the NCBI (http://www.ncbi.nlm.nih.gov/) web servers. Consistency in functional and structural assignment within a COG coupled with low E-values between COG and PDB confirmed the best matches were functionally the same protein. Additionally, manual refinement was used to verify uniform sample conditions (i.e., the same ligand bound to all proteins within a COG or all proteins correspond to wild-type sequences) for cases of redundantly solved structures. The PDB to CATH linkage was obtained directly from the CATH v3.2 database. The CATH classification for each structural domain for the PDB files listed in Table 1 was manually verified using the CATH search engine. This was important because, even though 32 of the 48 COG structure families are single-domain proteins, the remaining 16 COG structure families have two or three domains.

Table 1.

COG Structure Families^a

COG	Function	Sgo_sim(COG)b	Phylogenetic Structure Treec	CATH	Which Domain?
28	Thiamine pyrophosphate requiring enzymes	0.59	Split	3.40.50.970	1st
				3.40.50.1220	2nd
				3.40.50.970	3rd
39	Malate/lactate dehydrogenases	0.80	Split	3.40.50.720	single domain
394	Protein-tyrosine-phosphatase	0.61	Split	3.40.50.270	single domain
446	Uncharacterized NAD (FAD) -dependent dehydrogenases	0.85	Split	3.50.50.60	1st
				3.50.50.60	2nd
				3.30.390.30	3rd
604	NADPH:quinone reductase and related Zn-dependent oxidoreductases	0.88	Split	3.40.50.720	single domain
605	Superoxide dismutase	0.76	Split	d	1st & 2nd
742	N6-adenine-specific methylase	0.73	Split	d	single domain
813	Purine-nucleoside phosphorylase	0.87	Split	3.40.50.1580	single domain
1012	NAD-dependent aldehyde dehydrogenases	0.58	Split	3.40.309.10	1st & 2nd
1057	Nicotinic acid mononucleotide adenylyltransferase	0.95	Split	3.40.50.620	single domain
1075	Predicted acetyltransferases and hydrolases with the alpha/beta hydrolase fold	0.70	Split	3.40.50.1820	single domain
1607	Acyl-CoA hydrolase	0.87	Split	d	single domain
1940	Transcriptional regulator/sugar kinase	0.31	Split	3.30.420.40	1st
				3.30.420.160	2nd
2124	Cytochrome P450	0.80	Split	1.10.630.10	single domain
2188	Transcriptional regulators	0.89	Split	3.40.1410.10	single domain
242	N-formylmethionyl-tRNA deformylase	0.87	Split with HGT	3.90.45.10	single domain
1052	Lactate dehydrogenase and related dehydrogenases	0.89	Split with HGT	3.40.50.720	1st & 2nd
2141	Coenzyme F420-dependent N5, N10-methylene tetrahydromethanopterin reductase and related flavin-dependent oxidoreductases	0.76	Split with HGT	3.20.20.30	single domain
3832	Uncharacterized conserved protein	1.00	Split with HGT	3.30.530.20	single domain
110	Acetyltransferase (isoleucine patch superfamily)	0.56	Starburst	2.160.10.10	single domain
171	NAD synthase	0.85	Starburst	3.40.50.620	single domain
251	Putative translation initiation inhibitor, yjgF family	0.00	Starburst	3.30.1330.40	single domain
346	Lactoylglutathione lyase and related lyases	0.11	Starburst	3.10.180.10	single domain
366	Glycosidases	0.51	Starburst	3.20.20.80	1st
				3.90.400.10	2nd
				2.60.40.1180	3rd
454	Histone acetyltransferase HPA2 and related acetyltransferases	0.83	Starburst	3.40.630.30	single domain
491	Zn-dependent hydrolases, including glyoxylases	0.50	Starburst	3.60.15.10	single domain
500	SAM-dependent methyltransferases	0.59	Starburst	3.40.1630.10	1st
				3.40.50.150	2nd
526	Thiol-disulfide isomerase and thioredoxins	0.96	Starburst	3.40.30.10	single domain
590	Cytosine/adenosine deaminases	0.70	Starburst	3.40.140.10	single domain
637	Predicted phosphatase/phosphohexomutase	0.52	Starburst	3.40.50.1000	1st
				1.10.150.240 or 1.10.164.10	2nd
664	cAMP-binding proteins	0.50	Starburst	2.60.120.10	1st
				1.10.10.10	2nd
745	Response regulators consisting of a CheY-like receiver domain and a winged-helix DNA-binding domain	0.73	Starburst	3.40.50.2300	single domain
753	Catalase	0.93	Starburst	d	single domain
778	Nitroreductase	0.64	Starburst	3.40.109.10	single domain
784	FOG: CheY-like receiver	0.48	Starburst	3.40.50.2300	single domain
796	Glutamate racemase	0.92	Starburst	3.40.50.1860	1st & 2nd
1028	Dehydrogenases with different specificities (related to short-chain alcohol dehydrogenases)	0.84	Starburst	3.40.50.720	single domain
1151	6Fe-6S prismane cluster-containing protein	0.71	Starburst	3.40.50.2030	1st
				3.40.50.2030	2nd
				1.20.1270.30	3rd
1309	Transcriptional regulator	0.80	Starburst	1.10.10.60	1st
				1.10.357.10	2nd
1396	Predicted transcriptional regulators	0.54	Starburst	1.10.260.40	1st
				2.60.120.10	2nd
1404	Subtilisin-like serine proteases	0.60	Starburst	3.40.50.200	single domain
1733	Predicted transcriptional regulators	1.00	Starburst	d	single domain
1846	Transcriptional regulators	0.85	Starburst	1.10.10.10	single domain
2159	Predicted metal-dependent hydrolase of the TIM-barrel fold	0.83	Starburst	3.20.20.140	single domain
2367	Beta-lactamase class A	0.93	Starburst	3.40.710.10	single domain
2730	Endoglucanase	0.88	Starburst	3.20.20.80	single domain
3693	Beta-1,4-xylanase	0.89	Starburst	3.20.20.80	single domain
4948	L-alanine-DL-glutamate epimerase and related enzymes of enolase superfamily	0.71	Starburst	3.30.390.10	1st
				3.20.20.120	2nd

^aCOG Structure Families have two or more represented structures from among the Firmicutes and two or more from among the Proteobacteria

^bNormalized GO functional similarity score between each protein’s GO term set and the consensus GO term set for the COG (eqn. 2)

^c“Split” means the Firmicutes and Proteobacteria proteins were strongly separated from one another, “Starburst” means there was little to no evidence for a split according to phyla, and “Split with HGT” means there was strong evidence for a split according to phyla with the exception of one protein, which may indicate horizontal gene transfer. See Supplementary Table IS for a list of the PDB files associated with each COG.

^dThe protein structures in this COG family are in the CATH holding pen awaiting manual domain separation and/or final CATH assignment.

2.4. Structure based phylogenetic trees

In addition to pairwise alignment, all the protein structures from each COG were simultaneously aligned using the multiple structure alignment program MAMMOTH-multi (http://ub.cbm.uam.es/mammoth/mult/) (Lupyan, et al., 2005). The resulting aligned structures and the structure-based sequence alignment was used with in-house software to calculate an all-versus-all matrix of per-residue Cα distances. Standard boot-strapping techniques were then applied to the all-versus-all matrix of per-residue Cα distances to generate 100 distance-matrix tables. Columns of structure-based sequence alignments with the corresponding Cα distances were randomly selected until the total number of columns in the original sequence alignment was reached. The resulting set of Cα distances were then used to calculate a root mean square deviation (rmsd) between each pair of structures in the matrix. The 100 distance-matrix tables were imported into PHYLIP v3.68 (Felsenstein, 1989) to generate a consensus phylogenetic tree with bootstrap confidence levels (Efron, et al., 1996).

Each set of 100 bootstrapped distance matrices were analyzed by the Fitch-Margoliash method implemented in PHYLIP. Each matrix was jumbled with 100 replicates using a random number generator seed. This resulted in 10,000 unique and random distance matrices for each COG. The best tree was identified with the program Consense implemented in PHYLIP using the extended majority rule conservation. Since the bootstrapped trees do not show distance relationship, the original distance matrix generated by MAMMOTH-multi was used to generate a distance based phylogenetic tree. Each original distance matrix was jumbled with 100 replicates using a random number seed. The distance based phylogenetic tree was drawn using the program Drawtree implemented in PHYLIP.

Representative distance based phylogenetic trees are shown in (Fig. 4). Each tree was visually inspected and compared with the DaliLite analysis using the bootstrap values to determine if a tree fit the star, split or undetermined classification.

Figure 4. Protein structure based phylogenetic trees highlighting the split and starburst patterns.

The phylogenetic structure trees showed three different patterns: (top) strong split according to phyla; (bottom) starburst with no clear relationship to a common ancestor; and (Fig. 5) strong splits with the exception of one outlier. The Firmicutes protein structures are in blue and the Proteobacteria in black. The bootstrap values from 100 bootstrap replicates are indicated on branches and represent how often a branch appeared in the distance matrix. The two examples for the split pattern were from COG0028 (thiamine pyrophosphate requiring enzymes) and COG0446 (uncharacterized NAD(FAD)-dependent dehydrogenases). In the case of a strong split, the central branches were observed more than 95 times out of 100 replicate trials. The two examples for starburst pattern were from COG0491 (Zn-dependent hydrolases) and COG1309 (transcriptional regulator). For starburst patterns, very few branches were observed in more than two-thirds of the 100 replicate trials. The organism abbreviations are: A. hydrophila (Ahy); A. tumefaciens (Atu); A. viridians (Avi); B. cereus (Bce); B. japonicum (Bja); B. subtilis (Bsu); B. thuriagienes (Bth); E. carotovora (Eca); E. coli (Eco); E. faecalis (Efa); F. gormanii (Fgo); K. pneumonia (Kpn); L. lactis (Lla); L. sanfranciscens (Lsa); L. plantarum (Lpl); O. formigens (Ofo); P. aeruginosa (Pae); P. fluorescens (Pfl); P. pantotrophus (Ppa); P. putida (Ppu); P. species (Psp); S. aureus (Sau); S. marcescens (Sma); S. typhimurium (Sty); and X. maltophilia (Xma).

2.5. Measuring functional similarity within a COG

Each protein in our dataset was annotated with the corresponding Gene Ontology (Ashburner, et al., 2000) identification number found in the PDB. By definition, a strong consensus requires each protein to share the same GO term. Instead, a weak consensus set of GO terms was generated for each COG, where only a majority of proteins are required to share the same GO term. A distance was measured between the weak consensus set and the set of GO terms assigned to each individual protein. An average, normalized distance is reported for each COG, where a score of 1 indicates an identical functional classification and a score of 0 indicates a lack of functional similarity. The normalized GO functional similarity score between each protein’s GO term set and the consensus GO term set for the COG was measured as follows:

$$S_{go\_\mathit{sim}}(\mathit{COG})=\sum _{p\in \mathit{COG}}\frac{\left|{GO}_{\mathit{cog}}(p)\cap {GO}_{\mathit{cog}\_wc}\right|}{\left|{GO}_{\mathit{cog}}(p)\cup {GO}_{\mathit{cog}\_wc}\right|}$$ [2]

where S_{go_sim}(COG) is the normalized GO functional similarity score, GO_{cog_wc} denotes the weak consensus set of GO terms for the COG, and GO_cog(p) denotes the set of GO terms set for each protein p in the COG.

3. Results

3.1 Creating the COG structure families

Current functional annotation tools available in the PDB include the Gene Ontology (GO) (Ashburner, et al., 2000) and Enzyme Classification (EC) (Schomburg, et al., 2004). Unfortunately, due to the potential for convergence of function, these annotation tools are not useful for the study of homologous structures. To accurately observe phylum dependent structure divergence of proteins, it is important to construct a dataset of functionally similar orthologs. Among the 20 resources for structural classification of proteins, the clusters of orthologous groups (COGs) scheme is the only one that attempts to identify orthology (Ouzounis, et al., 2003) while providing moderate functional information. Therefore, each sequence and structure in the PDB was annotated with one COG number. Additionally, each protein was annotated with GO numbers and the relative functional similarity for each COG was measured (Table 1). This was achieved by developing the PROFESS database (Triplet, et al., 2010) that contains the PDB to COG annotations among other biologically relevant information. This includes associating each structure with its phyla classification, which allowed for the structures from Firmicutes and Proteobacteria to be easily selected for further analysis (Table S1).

The most recent COG database was created by finding the genome-specific best-hit for each gene in 66 unicellular genomes (50 bacteria, 13 archaea, and 3 eukaryota). Specifically, the orthologs present in three or more genomes were detected automatically and then multidomain proteins were manually split into component domains to eliminate artifactual lumping. The online COG database contains 192,987 sequences distributed among 4,876 COGs, accounting for 75% of genes in these 66 genomes.

At the time of our COG-to-PDB annotation, the PDB included 45,368 protein structures, although many of them were composed of multiple subunits (and therefore associated with an even larger number of sequences). The two best-represented bacterial phyla, which accounts for nearly one-fourth of all structures in the PDB, were selected for annotation. The PDB contains 8,298 Proteobacteria protein structures and 3,416 Firmicutes structures. The sequences for each of these structures were compared to the COG reference sequences using the Basic Local Alignment Search Tool (BLAST) (Altschul, et al., 1990). An expectation cut-off of 1×10⁻⁹ was used to maximize the likelihood of matching each PDB with its correct COG. The BLAST comparison matched 82% of the Firmicutes and Proteobacteria sequences to specific COGs, resulting in the clustering of 2,728 Firmicutes structures and 6,881 Proteobacteria structures. Of these hits, 27% were 100% identical to the COG reference sequence and 97% matched with greater than 50% sequence identity. To carry out our comparative study, we selected only those COGs that contained a minimum of two Firmicutes organisms and two Proteobacteria organisms. This requirement gave 281 unique COGs with a total of 3,047 bacterial proteins (1,066 Firmicutes and 1,981 Proteobacteria). In addition to COG clustering, the eggNOG (http://eggnog.embl.de) (Muller, et al.) and OMA databases (Schneider, et al., 2007) (http://omabrowser.org) were also mined for generating orthologous sets of protein structures. However, there were no set of proteins that met our criteria of two structures per phylum per cluster.

To further support the COG-PDB clusters, the overall functional similarity for each COG was determined by measuring the average distance between the Gene Ontology annotations for each protein and a weak consensus list of GO annotations (Table 1). Overall each COG exhibited high functional similarity (0.72 ± 0.21) with 1 being functionally identical and 0 being functionally dissimilar. In addition to the high sequence and structure similarity within each COG, the GO functional similarity measure provides further support that the proteins have been properly annotated to the correct COG. Nevertheless, there are three apparent outliers; COG0251 (putative translation initiation inhibitor, yjgF family), COG0346 (lactoylglutathione lyase and related lyases) and COG1940 (transcriptional regulator/sugar kinase) have GO similarity scores of 0, 0.11 and 0.31, respectively. The low GO similarity scores for these COGs are driven by the inclusion of unannotated proteins in the dataset. All six single-domain proteins associated with COG0251 are classified as a conserved hypothetical protein and have no associated GO terms. Of the seventeen single-domain proteins associated with COG0346, nine lack GO term assignments and have no functional annotation. Additionally for COG1940, two of the five two-domain proteins have no GO terms assigned to the structure.

3.2 Pairwise structure similarity

The pairwise structure comparison tool DaliLite (Holm and Park, 2000) was used to perform 63,504 pairwise comparisons between all of the proteins in our dataset. In total, the backbone structure similarity corresponded to 31,542 Proteobacteria-Proteobacteria comparisons (−/−), 12,674 Firmicutes-Firmicutes comparisons (+/+), and 19,288 Proteobacteria-Firmicutes comparisons (−/+). All comparisons were manually filtered within their respective COG to remove all but one redundantly solved structure (the largest contributor to the size reduction of the dataset), multiple or non-functionally relevant conformations (mutant protein, non-native experimental conditions, inhibited ligand complex), and the shorter of two protein structures. The final dataset contained 48 COGs (Table I) with a total of 1,713 structural comparisons among 147 Firmicutes proteins from 58 unique organisms and 176 Proteobacteria proteins from 84 unique organisms.

The resulting Dali Z-scores from the pairwise structure comparisons were plotted against sequence identity (Fig. 1) to reveal a saturating relationship as the percent identity rose to 100%. The lowest observed Z-score was 5.7 with a corresponding 16% sequence identity. This Z-score was well above the minimum cutoff of 2.0 (dashed line) for matches that were two standard deviations above a random match. This lowest Z-score came from the comparison of two Firmicutes proteins in COG0346 (lactoylglutathione lyase and related lyases): 2QH0 (Clostridium acetobutylicum); and 2QQZ (Bacillus anthracis). The average Z-score for all comparisons between these single-domain proteins was 27 ± 13, indicating that all structural matches were very significant even at sequence identities below 20%. All structure comparisons corresponding to 100% sequence identity in figure 1 result from a protein structure compared against itself. The inherent range in Z-scores at 100% sequence identity highlights the need to develop a normalized structure comparison score.

Figure 1. The relationship between structure similarity and sequence identity for 48 COGs.

Structure similarity is given as the raw Z-score, which increases as the protein length increases. The comparisons were for all proteins against all proteins, and include those for each protein against itself. The dashed line identifies a Dali Z-score of 2, which is the minimal limit for inferring structural similarity.

Since Z-scores increase as a function of the protein length, we normalized for this effect by calculating a Fractional Structure Similarity (FSS) scores (see eqn. 1). When the pairwise FSS scores were plotted against sequence identity (Fig. 2), a hyperbolic curve was obtained with all FSS values below an upper-limit at each percent identity. In fact, 20% sequence identity yielded a maximal FSS of 60%. This FSS limit was observed when all of the data were used (Fig. 2A), when only the pairwise comparisons within either phyla were used (Fig. 2B and C), or when only the pairwise comparisons between the two phyla were used (Fig. 2D). The pairwise comparison plot between the two phyla (Fig. 2D) showed an abrupt cutoff at 61% sequence identity and a 0.84 FSS score. This was not an artifact created by culling the dataset, since a similar plot prior to the manual filtering also demonstrated the same effect (supplemental Fig. 1).

Figure 2. The fractional structure similarity (FSS) and sequence identity for 48 COGs.

FSS was calculated using eqn. 1 to normalize the Dali Z-scores for their different sizes. The FSS values were plotted against sequence identity for (A) all the pairwise comparisons, (B) only Proteobacteria-Proteobacteria comparisons, (C) only Firmicutes-Firmicutes comparisons and (D) only Proteobacteria-Firmicutes comparisons.

The protein structures in COG0028 (thiamine pyrophosphate requiring enzymes) provides a useful example of the structural divergence that occurred after the Firmicutes and Proteobacteria phyla split. The overall fold is conserved between the phyla while there are discrete structural elements that are unique to each phylum. The two Firmicutes structures (Fig. 3A and 3B) yield a Z score of 59.6 and an FSS of 0.83, indicating very high structural conservation. There are more representative Proteobacteria structures that yield an average Z-score of 37.7 ± 1.6 and an average FSS of 0.58 ± 0.03. Again, the structures share a similar fold despite the slightly lower scores. Comparison of structures between the Firmicutes and Proteobacteria (Fig. 3C and D, respectively) phyla yield a lower Z-score of 34.8 ± 1.2 and a lower FSS of 0.49 ± 0.02 then the comparisons within each phylum. This suggests a divergence in structural details while conserving the overall fold. A detailed analysis reveals localized differences between the structures from the two phyla (see red highlights in Fig. 3C and D). In the Firmicutes representative structure, there is a continuous helix compared to helical breaks and loop insertions in the Proteobacteria structure. This is similar to the C-terminal domain of primase, where a long continuous helix found in the E. coli structure is broken by a loop region in B. stearothermophilus (Bailey, et al., 2007; Oakley, et al., 2005; Su, et al., 2006; Syson, et al., 2005).

Figure 3. Comparison of protein structures for COG0028 between two bacterial phyla.

The protein structures for COG0028 thiamine pyrophosphate requiring enzymes show (A) that the two Firmicutes structures have highly overlapping structures and (B) that the four Proteobacteria structures are very similar to each another. (see also the phylogenetic structure tree for COG0028 in Fig. 4). On the other hand, the major structural differences between the Firmicutes and Proteobacteria are highlighted in red on a representative Firmicutes (C) structure from L. plantarum (Lpl) (PDB ID: 1POW) (Muller et al. 1994) and the representative Proteobacteria structure (D) from P. fluorescens (Pfl) (PDB ID: 2AG0) (Mosbacher, Mueller, and Schulz 2005).

3.3 COG structure phylogenies

Structure based phylogenies were created from root-mean square differences (rmsd) in per residue Cα positions for optimally aligned protein structures using MAMMOTH-multi (Lupyan, et al., 2005). A separate phylogenetic tree was generated for each COG, where three distinct patterns were observed (Table I): 15 trees exhibited a strong split at the phylum level, 29 exhibited a starburst pattern suggesting little to no evidence for a split according to phyla, and 4 exhibited a strong split at the phylum level but with the exception of a single structure (split +1).

As shown in Table 1, the pattern of the structure based tree is not dependent on the relative GO functional similarity score for the proteins within each COG. All three tree patterns have a range of GO functional similarity scores with an average score of 0.75 ± 0.16, 0.88 ± 0.09 and 0.70 ± 0.24 for the split, split+1, and starburst tree pattern, respectively. Overall the high GO similarity scores within each COG are high, indicating conserved and consistent functional annotations for each COG.

The 15 COG phylogenies with strong phylum-splitting patterns had two branches, one with closely related Firmicutes structures and the other with closely related Proteobacteria structures. Two examples are COG0028 (Thiamine pyrophosphate requiring enzymes) and COG0446 (Uncharacterized NAD/FAD-dependent dehydrogenases) (Fig. 4). The structures for both of these multi-domain COGs are classified in the CATH system as α/β 3-layer sandwiches, but differ in that COG0028 proteins have a Rossmann fold topology (Fig. 3) and COG0046 proteins have a FAD/NAD (P)-binding domain topology.

The 29 COGs with phylogenetic starburst patterns showed no evidence for the separation of structures according to phyla (Table 1). Two examples were COG0491 (Zn-dependent hydrolases) and COG1309 (Transcriptional regulator) (Fig. 4). The CATH classification for COG0491 Bacillus cereus Zinc-dependent β-lactamase (PDB ID: 1BC2) (Fabiane, et al., 1998) describes the protein as an α/β 4-layer sandwich with metallo-β-lactamase Chain A topology. The large category of β-lactamases constitutes a collection of enzymes that can be derived from any one of a group of proteins that bind, synthesize, or degrade peptidoglycans. The protein structures assigned to COG0491 gave FSS scores with large standard deviations, as is evident from the separated clusters within the Proteobacteria arm of the phylogenetic tree.

The two-domain COG1309 structural family falls into one of two CATH topologies represented by Arc Repressor Mutant subunit A and Tetracycline Repressor domain 2. Only those structures similar to the Arc Repressor Mutant (subunit A) topology were used for the pairwise comparison, since it was the dominant fold in this COG. The protein structures in the COG1309 structure family gave low FSS scores. However, even with a low overall FSS, the average absolute Z-score was 13 ± 2 indicating that it has significant overall structure similarity. The high FSS deviations of the COG0491 structural family and the low average FSS scores of the COG1309 structural family both indicate rapid structural divergence following the phyla split, consistent with the observed starburst phylogenetic patterns.

Four COG structure phylogenies showed a strong split pattern with a single outlier (Fig. 5). This result provides further evidence for the observation of phyla split based on protein structure similarity. The presence of the outlier in a clear split pattern suggests either a horizontally transferred gene (Table I) or a potential paralog. For all four families [COG0242 (N-formylmethionyl-tRNA deformylase) COG1052 (Lactate dehydrogenase and related dehydrogenases), COG2141 (Coenzyme F420-dependent N5, N10-methylene tetrahydromethanopterin reductase and related flavin-dependent oxidoreductases), and COG3832 (Uncharacterized conserved protein)] there was a large Dali Z-score and reliable BLAST Evalues, implying a correct match was made between COG and PDB. Additionally all four COGs exhibited high GO functional similarity scores suggesting a consistent functional assignment (Table 1). For COG0242, the Bacillus cereus gene def that encodes the N-formylmethionyl-tRNA deformylase protein (PDB ID: 1WS0) has been previously identified as a gene that has undergone horizontal gene transfer (Garcia-Vallve, et al., 2000).

Figure 5. Protein structure based phylogenetic trees highlighting the split+1 pattern.

Protein structure phylogenies of 4 COGs out of 48 had a strong split pattern with the exception of one outlier structure. The phylogenies were very reliable because the central branches were observed in 100 out of 100 replicate trials. When one Firmicutes or Proteobacteria protein structure clusters on a branch with the other phylum, its structure diverges from its closest relatives while resembling those of the other phyla. The COGs that fit this pattern are from COG0242 (N-formylmethionyl-tRNA deformylase), COG1052 (lactate dehydrogenase and related dehydrogenases), COG2141 (coenzyme F420-dependent N5, N10-methylene tetrahydromethanopterin reductase and related flavin-dependent oxidoreductases), and COG3832 (uncharacterized conserved protein). The organism abbreviations are: A. fermentans (Afe); A. tumefaciens (Atu); B. cereus (Bce); B. halodurans (Bha); B. stearothermophilus (Bst); B. subtilis (Bsu); C. violaceum (Cvi); E. coli (Eco); E. faecalis (Efa); H. methylovorum (Hme); H. pylori (Hpy); L. delbrueckii (Lde); L. helveticus (Lhe); M. species (Msp); N. europaea (Neu); P. aeruginosa (Pae); P. species (Psp), S. aureus (Sau); S. pneumoniae (Spn); and V. harveyi (Vha).

3.4. Structure divergence rates across phyla

As a way to quantify the relationship between structure and sequence differences, each phylogenetic tree was reduced to a single coordinate by calculating a structure similarity ratio (θ_FSS) and a sequence identity ratio (θ_SeqID). θ_FSS was determined for all 48 COGs by calculating an average FSS score for the Proteobacteria-Firmicutes structure comparisons, Avg(FSS_+/−), and dividing by the sum of the average Proteobacteria-Proteobacteria, Avg(FSS_−/−), and Firmicutes-Firmicutes, Avg(FSS_+/+), comparisons:

$${\theta }_{\mathit{FSS}}=\frac{\mathit{Avg}({\mathit{FSS}}_{+/-})}{\mathit{Avg}({\mathit{FSS}}_{+/+})/2+\mathit{Avg}({\mathit{FSS}}_{-/-})/2}$$ [3]

Similarly, a sequence identity ratio (θ_SeqID) was determined by calculating an average sequence identity for the Proteobacteria-Firmicutes structure comparisons, Avg(SeqID_+/−), and dividing by the sum of the average Proteobacteria-Proteobacteria, Avg(SeqID_−/−), and Firmicutes-Firmicutes, Avg(SeqID_+/+), comparisons:

$${\theta }_{\mathit{SeqID}}=\frac{\mathit{Avg}({\mathit{SeqID}}_{+/-})}{\mathit{Avg}({\mathit{SeqID}}_{+/+})/2+\mathit{Avg}({\mathit{SeqID}}_{-/-})/2}$$ [4]

In general, most starburst phylogenies (see representative COG0491 and COG1309 in Fig. 4) had a branch length between members of different phyla that was much shorter than the branch lengths between members within the same phyla. That is, a starburst phylogeny was expected to have θ_FSS and θ_SeqID values greater than unity. Likewise, most split phylogenies had longer branches between phyla than within each phyla (see representative COG0028 and COG0446 in (Fig. 4) and were expected to yield θ_FSS and θ_SeqID of less than unity.

When θ_FSS and θ_SeqID for all 48 COGs were plotted versus one another (Fig. 6), 79% of the starburst phylogenies were equal to or greater than unity for both structure and sequence whereas 84% of the split phylogenies were below a θ_FSS of 0.9 for structure and 73% of split phylogenies were below a θ_SeqID of 0.80 for sequence. This indicated that split phylogenies occur when the structure differences are less than their sequence differences. In addition, the plot of θ_FSS versus θ_SeqID conformed to a linear relationship regardless of the shape of the phylogenetic tree indicating that all homologous protein structure differences are constant with respect to homologous protein sequence differences (θ_FSS = 0.55θ_SeqID + 0.45; R² = 0.7). Thus, this curve represents the relative structural drift rate for each COG structural family between the two phyla. The slope indicates that structure branch lengths change approximately half as fast as sequence branch lengths.

Figure 6. Constant rate of structural drift.

The relationship between structure and sequence change was constant regardless of the phylogenetic starburst (x) or split (■) pattern. Structure changes measured using a structure similarity ratio (θ_FSS), where the average FSS between members of the two phyla (Firmicutes versus Proteobacteria) was divided by the average FSS between members of the same phyla (see eqn. 3). Sequence change was calculated similarly (see eqn. 4). The best-fit line, θ_FSS=0.55θ_SeqID + 0.45, yielded an R² of 0.70.

3.5. Fold dependency on structure similarity

A plot of FSS versus sequence identity for the two most populated CATH families in our dataset (Fig. 7) was used to investigate if particular protein architectures are more amenable to structural changes. Thirty-one of 66 total domains (47%), the largest portion of our data set, are classified as CATH 3.40 (α/β, 3-layer (αβα) sandwich). The CATH 3.40 classification is more often associated with the split phylogenetic tree pattern (12 out of 22 total domains or 55%) then the starburst pattern (17 out of 39 total domains or 44%).

Figure 7. Fold dependency on fractional structure similarity (FSS) and sequence comparisons.

The FSS between two CATH families, CATH 1.10 (●) CATH 3.40 (⋄). CATH 1.10 (mainly α, orthogonal bundle) family is apparently limited to approximately 40% sequence identity and 0.6 FSS while CATH 3.40 (α/β, 3-Layer (αβα) sandwich) fills in the complete curve. 87.5% of the COGs (7 of 8) represented by CATH 1.10 give a starburst structure similarity tree. Contrastingly, only 50% (12 of 24) of the COGs represented by CATH 3.40 give a starburst structure similarity tree. The remaining 12 COGs formed either split (11 of 12) or split+1 (1 of 12).

The second most populous CATH family is CATH 1.10 (mainly α, orthogonal bundle) with 11% of our COGs belonging to this CATH family. Most (85.7%) of the COGs (6 of 7) in the CATH 1.10 family are represented by the starburst phylogenetic tree pattern with only one COG represented by a split pattern. There appears to be a limit in structure similarity at approximately 0.6 FSS and a corresponding sequence identity limit at 40% for CATH 1.10 (Fig. 7, solid circles). This limit is not observed in the CATH 3.40 family (Fig. 7, open diamonds). The sequence and structure similarity limit for CATH 1.10 combined with a larger percentage of COGs assigned to the starburst family suggests that CATH 1.10 is more susceptible to mutations that affect the protein structure.

4. Discussion

There is an inherent challenge in obtaining an accurate functional annotation for a large set of proteins from a relatively small number of experimentally determined functions (Andrade, 2003; Frishman, 2007; Karp, et al., 2001; Rentzsch and Orengo, 2009; Valencia, 2005). The available functional information is incomplete, ambiguous and error-prone (Benitez-Paez, 2009; Schnoes, et al., 2009) and requires multiple sources (Rentzsch and Orengo, 2009) to improve the accuracy in the annotation of a protein. There is also the complicating factor of correctly distinguishing between orthologs and paralogs, where it has been previously noted that the COG database does include some paralog members (Dessimoz, et al., 2006; Tatusov, et al., 2003). Thus, the accuracy of our analysis of structural divergence is fundamentally dependent on a reliable functional assignment for each protein structure. Given these challenges, the independent and separate utilization of both COG and GO terms provides a reasonable and robust approach to identify clusters of functionally similar proteins. The overall high sequence (E-value ≤ 10⁻⁹, sequence identity ≥ 16%), structure (Z-score > 5.7) and GO term similarity (0.72 ± 0.21) within each COG supports this conclusion. The lack of identity for the GO term similarity scores should not be interpreted as evidence for functional divergence. GO terms are assigned based on a validated source. So, a missing GO term for a protein is more likely attributed to the fact that the protein has not been explicitly tested for the specified activity. Similarly, a protein being assigned a GO term does not provide definitive evidence that the function is relevant in vivo (Canevascini, et al., 1996; Lindorff-Larsen, et al., 2001; Otsuka, et al., 2002; West, et al., 2004).

The comparison of homologous protein structures with the same function provides quantitative evidence that protein structures diverged following the speciation events that created the modern bacterial phyla of Firmicutes and Proteobacteria. The abrupt cutoff at 61% sequence identity and 0.84 fractional structure similarity observed between Firmicutes and Proteobacteria proteins was mirrored by an approximate 60% protein sequence identity between these two phyla observed by 16S rRNA sequence similarity (Konstantinidis and Tiedje, 2005a; Konstantinidis and Tiedje, 2005b). Thus, this maximum observed sequence identity imparts limits to the maximum possible structure similarity between homologous proteins from these two phyla. This is consistent with prior observations that sequence identity ≤ 40–50% sometimes results in significant structural and functional differences (Chothia and Lesk, 1986a; Rost, 1999; Rost, 2002). Furthermore, the results imply an inherent allowable structural plasticity that does not perturb function. Additionally, the random drift after speciation inexorably leads to non-identical structures despite maintenance of function. There are a number of cases where FSS was below 0.20 indicating a significant structural change. Proteins with completely different folds but the same function are extreme examples of the plasticity of the structure-function relationship and include such proteins as peptidyl-tRNA hydrolases (COG1990) (Powers, et al., 2005), pantothenate kinase (KOG2201) (Yang, et al., 2006), polypeptide release factors (Kisselev, 2002) and lysyl-tRNA synthetases (COG1190) (Ibba, et al., 1997), these proteins are not in our dataset.

Forty percent of the COGs we examined have evolved slowly enough that it was possible to generate phylogenetic trees consistent with this ancient split. The other COGs have either evolved too rapidly or are otherwise subject to few evolutionary constraints to provide evidence for this split. This distinction between the COGs is clearly apparent from the comparison of θ_FSS and θ_SeqID in (Fig. 6). The slope of (Fig. 6) indicates a fixed relative structure drift rate, where structure changes half as fast as sequence across phyla. This correlation in the divergence of protein sequences and protein structures has additional ramifications beyond bacterial evolution. Our analysis implies a continuum of protein folds that adapt to large sequence changes by incurring local structural modifications (Illergard, et al., 2009; Kolodny, et al., 2006; Panchenko, et al., 2005; Williams and Lovell, 2009). This continuum of protein folds makes it challenging to apply protein structural classification to identify function, as has been previously noted (Hadley and Jones, 1999; Pascual-Garcia, et al., 2009).

Does the nature of the protein’s three-dimensional structure play a role in protein structure divergence? Our analysis demonstrates that some proteins evolve slowly and maintain high sequence identity (>80%) and structure similarity (> 0.80 FSS) while other proteins exhibit rapid evolution rates where sequence identity is ≤ 20% and FSS ≤ 0.40. This implies that the underlying architecture of a particular protein may be more or less amenable to amino-acid substitutions in order to maintain functional activity. A specific protein fold may have a higher intrinsic plasticity that enables it to readily accommodate sequence changes through local conformational changes without a detrimental impact on activity. This is exactly what was observed.

Structural variations were localized to specific regions as illustrated by the comparison of the COG0028 protein structures see (Fig. 3). This is consistent with the observation that there are different structure divergence rates within a protein (Chirpich, 1975; Lin, et al., 2007). Regions of the protein that do not impact biological activity are expected to yield a higher divergence rate and incur larger local structural changes (Chothia and Lesk, 1986b; Lesk and Chothia, 1980). As a result, a fold with a relatively high plasticity would experience an elevated structural diversity between phyla, where the rate of change may closely parallel the mutation rate (Illergard, et al., 2009). Conversely, another fold may be extremely sensitive to amino-acid substitutions, where minor sequence perturbations may result in a decrease in structural integrity and a corresponding loss of activity. This analysis is consistent with the known range of protein thermodynamic stabilities (Robertson and Murphy, 1997), and the general observation that most mutations destabilize protein structures (Sanchez, et al., 2006).

For instance, CATH 1.10 was the second most abundant protein architecture observed in our study, comprising 11% of the total domains. It was very strongly associated with the fastest evolving protein structure and corresponds to an orthogonal α-helical bundle. Conversely, the highly populated CATH 3.40 is a 3-layer (αβα) sandwich with a slower evolution rate compared to CATH 1.10 structures. β-sheets are strongly influenced by long-range interactions and, on average, have a higher hydrophobic environment compared to α-helices (Gromiha and Ponnuswamy, 1995). Effectively, the protein environment is an important factor in β-sheet folding (Parisien and Major, 2007). Since the stability and structure of a protein is strongly dependent on the integrity of the hydrophobic core (Vlassi, et al., 1999), which is formed by the β-sheet in the 3-layer (αβα) sandwich, mutations in the β-sheet are probably less tolerated.

Our study illustrates the inherent value in solving structures for functionally identical proteins from multiple organisms. A major challenge in creating our COG-to-PDB dataset was the fundamental requirement to have structures from at least two Firmicutes organisms and two Proteobacteria organisms. Only 48 (~1%) of the 4,876 COGs meet this stringent requirement. The limited number of multiple homologous structures has partly occurred because structural biology efforts are focused on obtaining single representative structures for each functional class or protein fold (Chandonia and Brenner, 2005) and understandably biased toward therapeutically relevant proteins (Mestres, 2005). If we are to achieve a more accurate understanding of the relationship between the evolution of protein fold, protein sequence, and the organisms in which they function, the fields of bioinformatics and structural biology must expand their focus to include efforts to obtain a more diverse set of homologous protein structures.

ABBREVIATIONS

FSS

Fractional Structure Similarity

COG

Cluster of Orthologous Groups

PDB

Protein Data Bank

Split

Clusters showing strong phylogenetic split pattern based on structure

Split+1

Clusters showing strong phylogenetic split pattern with one outlier based on structure

Starburst

Clusters with variable phylogenetic patterns based on structure

Z_AA and Z_BB

Dali Z-scores for self comparisons

Z_AB

Dali Z-scores for pairwise comparisons

θ_FSS

Structure similarity ratio

θ_SeqID

Sequence similarity ratio