In Table 1, we show the percentage of proteomes that may be assigned folds for five minimal organisms and for E. coli, an example of a well-studied organism that is not "minimal." For the minimal organisms considered in this study, nearly all proteins annotated as soluble and globular may be assigned to a known fold. The aphid symbiont B. floridanus has the highest coverage, at 96% of soluble, globular proteins (431 of 451 proteins). 58 of the remaining proteins in the proteome (10% of the proteome) have unknown structure, but are predicted to have at least one transmembrane helix. 3 additional proteins have unknown structure and no predicted transmembrane helices, but 20% or more of their residues are in predicted low complexity or coiled coil regions, and thus not easily tractable in experimental structural studies. Overall, the folds of 502 of 583 B. floridanus proteins (86%) may be annotated by sequence similarity to a protein of known structure. Other minimal organisms also have high structural coverage: 95% of soluble, globular W. glossinidia proteins, 94% of soluble, globular B. aphidicola proteins, 87% of soluble, globular M. genitalium proteins, and 87% of soluble, globular T. whipplei proteins can reliably be assigned folds. In contrast, only 78% of soluble, globular E. coli proteins can reliably be assigned folds. The low numbers of predicted transmembrane proteins in several of the minimal organisms (e.g., only 87 of 572 B. aphidicola proteins) is also notable; previous analyses suggest that some transmembrane proteins (e.g., proteins with a role in cell defense or transporters of diverse nutritional sources) are less important to intracellular symbiotes than to free-living bacteria 27.

For the proteins that could be reliably assigned folds, we examined their structural classification in the SCOP database 38. SCOP is a widely used, manually curated database in which protein structures are divided into domains, which are classified in a hierarchy indicating different types of structural and evolutionary relationships between the domains. Domains classified together in a single "family" or "superfamily" are hypothesized to have a common evolutionary origin on the basis of sequence or structural evidence. Superfamilies that share similar secondary structural features and topology, but for which there is little or no evidence to suggest a common evolutionary origin, are classified together at the "fold" level. SCOP folds are grouped together in seven major "classes" (all-α, all-β, α/β, α+β, multi-domain, membrane, and small), based on common physical characteristics such as the predominant type of secondary structure or the order of connection of the different secondary structures (Figure 1). Note that the SCOP "multi-domain" class encompasses folds that are comprised of multiple domains that individually would belong to different classes; individual domains from multi-domain proteins are not classified in the "multi-domain" class. Although we use the term "fold" to refer to a protein's overall 3D backbone structure, we use the term "SCOP fold" to refer to a specific fold classification within the SCOP database.

The fraction of proteins found in each organism belonging to each of these SCOP classes is shown in Figure 2. Those proteins that could not reliably be assigned folds, and those that were assigned a fold based on homology to a protein not yet classified in SCOP, are described as "Unsolved" and "Unclassified," respectively. For all organisms, the highest proportion of SCOP folds are in the α/β class, and those in the α/β and α+β classes together comprise over half of the assigned SCOP folds. This reflects the observation that the α/β class contains some of the most functionally diverse "superfolds" that act as scaffolds for a wide array of molecular or chemical functions 39.

In order to analyze how the annotated cellular function of each protein correlates with its structure, we examined the "functional role" annotation for each protein as provided in the TIGR database 40. We found that the distribution of proteins among SCOP fold classes was highly conserved within some roles and showed much more variability in others.

Figure 3 shows the fold class distribution of proteins in the "Protein Synthesis" functional category across all 6 proteomes. The fraction of these proteins in each structural class shows little variability, with no more than a 4% difference between proteomes. Furthermore, the proteins in this functional category comprise a relatively large fraction of the proteins in each proteome (99 proteins on average, or 8% of the proteome). The extremely low variability is consistent with the idea that these proteins have been fundamental part of cellular biochemistry since early evolution, and are thus essential to any organism regardless of its environment.

In contrast, Figure 4 shows the fold class distribution of proteins in the "Cell Envelope" functional category across all 6 proteomes. This functional category is also highly represented in each proteome (73.8 proteins on average), but the proteins show a much higher degree of variation in fold usage. This category contains the highest proportion of unassigned folds, as well as a diverse array of assigned SCOP folds: for example, 6% and 4% of domains from W. glossinidia and E. coli cell envelope proteins belong to the all-α structural class, while cell envelope proteins from the other proteomes contain few or no all-α structures. E. coli also contains a number of solved transmembrane structures, while other proteomes contain significant numbers of proteins predicted to be transmembrane proteins not detectably homologous to any protein with a known 3D structure. M. genitalium and T. whipplei contain the largest fractions of cell envelope proteins that could not be reliably assigned a fold at this time, although most of these M. genitalium proteins are expected to be soluble and globular, while the majority of such proteins from T. whipplei are predicted to contain at least one transmembrane helix. The high amount of variability suggests that proteins in the "Cell Envelope" category evolve rapidly in response to specific pressures in an organism's environment, and different sets of these proteins remain after reductive evolution in the different environments occupied by the different species of minimal organisms.

Previous comparative sequence genomic analyses of symbionts have shown that the number of proteins in most cellular function categories varies little between symbiont proteomes, and that many of the most highly conserved proteins have cellular functions related to information storage and processing, particularly translation and ribosomal structure 31. We calculated the coefficient of variation (CV) in the number of proteins in each functional role category (N₁ for the first species, N₂ for the second species, etc.), as shown in Equation 1.

C V s e q u e n c e = S t d e v ( N 1 … N 6 ) M e a n ( N 1 … N 6 )       ( 1 ) MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqWGdbWqcqWGwbGvdaWgaaWcbaGaem4CamNaemyzauMaemyCaeNaemyDauNaemyzauMaemOBa4Maem4yamMaemyzaugabeaakiabg2da9maalaaabaGaem4uamLaemiDaqNaemizaqMaemyzauMaemODayNaeiikaGIaemOta40aaSbaaSqaaiabigdaXaqabaGccqWIMaYscqWGobGtdaWgaaWcbaGaeGOnaydabeaakiabcMcaPaqaaiabd2eanjabdwgaLjabdggaHjabd6gaUjabcIcaOiabd6eaonaaBaaaleaacqaIXaqmaeqaaOGaeSOjGSKaemOta40aaSbaaSqaaiabiAda2aqabaGccqGGPaqkaaGaaCzcaiaaxMaadaqadaqaaiabigdaXaGaayjkaiaawMcaaaaa@59BA@

Results are shown in Table 2. As expected, the category with the lowest variation in the number of proteins is "Protein synthesis," and the top three categories are all closely related to transcription or translation.

We also calculated the coefficient of variation in the number of protein domains assigned to each SCOP class (N_1,all-α for the first species in the all-α class, N_2,all-α for the second species in the all-α class, etc), then averaged that data across all 7 structural classes, as shown in Equation 2.

C V s t r u c t u r e = ∑ c l a s s = 1 7 s t d e v ( N 1 , c l a s s … N 6 , c l a s s ) M e a n ( N 1 , c l a s s … N 6 , c l a s s ) 7       ( 2 ) MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=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@877C@

CV_structure was calculated separately for each functional role category, and these data are shown in Table 2 and Figure 5A. The functional category with the lowest variation in the number of domains in each structural class is "Protein Synthesis," as would be expected from Figure 3. However, there are some interesting differences between the rankings based only on the CV_sequence, and the rankings based on CV_structure. For example, fold usage of proteins involved in biosynthesis of cofactors, carriers, and prosthetic groups varies to a higher degree than the variation in total numbers of these proteins in each proteome. This implies that the repertoire of specific functions in this broad category is specialized to the particular needs of each organism, even though the overall number of such proteins varies little. As expected, the distribution of structures in "catch-all" classes such as hypothetical and unclassified proteins are more varied than the distribution of structures found in more well-defined functional categories.

We also analyzed the degree of variation using data from only the five near-complete minimal organisms, excluding data from E. coli. Results are shown in Table 3 and Figure 5B. As before, fold usage of proteins in the "protein synthesis" category shows the least variance of all functional categories. The total genome size also slows relatively little variation among minimal organisms, as has been observed previously 31. However, some functional categories show relatively more variation between minimal organisms than between minimal organisms and E. coli. For example, the cellular function categories "Cell envelope," "Central intermediary metabolism," and "Amino Acid Biosynthesis" all drop in rank (the relative degree of conservation in fold usage among functional categories) by 7 positions relative to Table 2, indicating higher diversity of folds in these functional categories among minimal organisms. In contrast, fold usage of proteins in the "Regulatory functions" category shows relatively less variation among minimal organisms than between minimal organisms and E. coli. This suggests that although the minimal organisms have lost many of the regulatory pathways unnecessary for survival in their relatively unchanging environments, they maintain a relatively conserved set of proteins responsible for common regulatory functions. A more thorough phylogenetic analysis of these proteins would be necessary to test this hypothesis.

We examined the most common protein folds (as defined in SCOP 1.67) in minimal organisms. Results are shown in Table 4. Four of the eleven most common SCOP folds (TIM barrel, nucleoside triphosphate hydrolase, flavodoxin-like, and ferredoxin-like) are among the nine superfolds originally described by Orengo and colleagues as scaffolds that can support a wide array of molecular functions 39. However, all have fewer copies in minimal organisms than are found in E. coli.

Table 5 shows SCOP folds that are found in both minimal organisms and in E. coli, which are represented at equal or greater levels in the minimal organisms. Proteins with these folds are presumably important for the survival of the organisms, and were not eliminated during reductive evolution. Five SCOP folds are present in slightly greater numbers in minimal organisms than in E. coli. For example, the DNA primase core fold (e.13) has 3 representatives in M. genitalium: the DNA primase protein itself (dnaE) and two conserved hypothetical proteins (NP_072670 and NP_072719). All five folds are involved in the critical functions of transcription, translation, or DNA replication. Forty-two other SCOP folds are present in the same numbers in each minimal genome as in E. coli. The five with the largest number of copies per genome are shown in Table 5. Some appear to be key metabolic enzymes, while others are involved in transcription, translation, or DNA replication.

Interestingly, all 47 SCOP folds present in equal or greater numbers in all minimal organisms as in E. coli are also folds for which only a single superfamily is characterized in SCOP; i.e., all proteins sharing the fold are also annotated as evolutionarily related to each other. The case of multiple superfamilies sharing one fold may arise from two alternative causes: convergent evolution of two or more families to one fold, or a single family that has diverged enough that homology between different branches of the family are no longer evident even from structure (in this case, each branch would be classified as a different superfamily in SCOP). These data imply that proteins that play sufficiently important roles to avoid elimination during reductive evolution have also not diverged as much as other protein families due to this same evolutionary pressure.

An additional set of SCOP folds found only in minimal organisms and not in E. coli is given in Table 6. None of these folds are found in all five minimal organisms, and the proteins are not generally related to essential cellular functions such as transcription, translation, or replication. Some are presumably adaptations to the specific environment of the organism, and several (e.g., viral coat and capsid proteins, and the MHC antigen-recognition domain) are not typically found in bacteria. These may represent lateral gene transfers or erroneous annotations.

Our database of known protein structures, knownstr, was created on 22 Feb 2005. This database contained sequences of every protein chain released by the PDB 45, including those of obsolete entries, sequences of proteins deposited in the PDB and made available while the structures were still on hold, and sequences from TargetDB 46, for which a structure had been solved by a participating structural genomics center.

Pfam 47 classification of known structures was evaluated using Pfam version 16.0. The HMMER tool (version 2.3.2) 48 was used to compare the Pfam_ls library of hidden Markov models to the knownstr database, using the family-specific "trusted cutoff" score as a cutoff for assigning significance.

INTEGR8 version 12 2 was used for sequence data. The Integr8 database contains data for 238 complete proteomes, including 19 eukaryotes. The proteome for each organism is composed of proteins curated from the Swiss-Prot and TrEMBL databases. All proteins were annotated with hidden Markov models 4849 from the InterPro 50 database. Since InterPro includes models from Pfam, we used the supplied InterPro annotations to map Pfam domains onto each protein. The version of InterPro used to annotate Integr8 version 12 includes Pfam 16.0

SUPERFAMILY 51 version 1.67 contains hidden Markov models based on superfamilies from the SCOP database 3852, also version 1.67. Recent versions of SUPERFAMILY 53 provide pre-calculated annotations of genomes downloaded from NCBI with all the superfamily models. We used these precalculated annotations to assign SCOP domains to sequences from minimal organisms and E. coli, as described below. The false positive rate for SUPERFAMILY annotations is estimated to be less than 1% 54.

The Comprehensive Microbial Resource 40 contains annotations of TIGR role categories in its OMNIOME database. We obtained TIGR role annotations from the version of OMNIOME downloaded on 12 May 2005. Of 19 TIGR role categories, two ("signal transduction" and "other categories") were found in low average abundance in the proteomes we analyzed (averaging 0.7 and 9.0 proteins per proteome, respectively), and these categories were excluded from our analysis. The remaining 17 categories are listed in Table 2.

To use annotations from the SUPERFAMILY and OMNIOME databases, we mapped proteins from the Integr8 database onto corresponding proteins in the NCBI and CMR Locus databases, respectively. In most cases, this was done by mapping identical sequences from the corresponding genome. However, in some cases, the gene or ORF annotations of the same genomes varied between the databases, resulting in different protein sequences. In these cases, we used BLAST 55 version 2.2.9 to map each Integr8 sequence to the most similar sequence in the other databases. We mapped each protein in Integr8 that could not be mapped by direct sequence match to the most significant BLAST hit in the other database, provided the BLAST E-value of the hit at least as significant as an empirically chosen threshold of 10^-10. An average of 16.3 proteins in each proteome could not be mapped to any of the functional categories in OMNIOME, and were not included in this analysis.

We identified all proteins with a predicted transmembrane helix, or with 20% or more residues in low complexity regions, or with 20% or more residues in coiled coil regions, as likely to be intractable in high-throughput experiments. Other proteins were annotated as soluble, globular proteins. The 20% threshold were used in more recent target selection rounds at the Berkeley Structural Genomics Center 25. Similar thresholds have also been justified by recent comprehensive crystallization trials on the Thermotoga maritima proteome 56.

The "seg" program 57 (version dated 5/24/2000) was run on all sequences in Integr8 to identify putative low complexity regions. The "ccp" program 58 (version dated 6/14/1998) was used to predict coiled coil regions in all sequences, and TMHMM 2.0a 59 was used to predict the locations of transmembrane helices. TMHMM can distinguish between soluble and membrane proteins with both specificity and sensitivity greater than 99%, but frequently produces false positive predictions when signal peptides are present. Default options were used for all programs.