Reductive evolution of proteomes and protein structures

Structural Assignments.

A large data set of structures assigned to protein sequences using hidden Markov models (HMMs) of structural recognition in SUPERFAMILY (15) was analyzed in 745 genomes, including 215 Eukarya, 478 Bacteria, and 52 Archaea genomes (Table S1). Our analysis assumes that all genes are accurately predicted in the genomes. The averages tabulated describe complex populations of proteins including substantial minorities of multidomain proteins that make up 26% to 32% of the proteins in the three superkingdoms (Table 1). Here, the composite lengths of both domains and linkers are described so that, for all proteins, these lengths are the sums of multiple domains and multiple linkers when applicable. Fig. S1 shows whisker plots for average protein, domain, and linker lengths of representative sets of genomes, describing the typical dispersion of length expected in the genomes that were sampled. We note that the possible existence of unsolved structures (21, 22) and the limitations of the HMM method cannot guarantee that all structures in a sequence will be uncovered. This might introduce some uncertainties to length estimates. However, careful analysis of the overall distribution of domains and linkers segments, domains that overlap, and HMM performance within families of orthologues suggest that this is a negligible problem (SI Materials and Methods). We find that HMMs detect superfamilies equally well in the three superkingdoms, and they do not show preferences for homologues of the seed sequences of the models (Table S2) or for superkingdom-specific sequences. These results are consistent with previous observations (15) and suggest that domain boundaries of remote homologues of the seed sequence are reliable. Moreover, if we assume that cryptic domains in linkers follow the same length distribution of domains detected by current HMMs, the data show that few of the putative domains could be accommodated in linkers, even if they were overlapped (Figs. S2–S4). In fact, comparison of frequency distributions of domains and linkers reveals that, on average, cryptic domains match linkers with probabilities in the range 0.022 to 0.042 (nonparametric Mann–Whitney test, P < 0.0001). Given these estimates, the maximum cumulative influence of such cryptic domains on average linker length would be, at most, approximately 22% in each superkingdom (Table S3). In summary, the reliability of the protein length patterns recovered in the present study have been tested and found to be robust.

To determine if linker sequences are enriched in intrinsically disordered regions, we studied the amino acid composition of linkers and domains. We found that linkers and domains showed significantly distinct amino acid preferences (paired t test, P < 0.01) in the three superkingdoms (Fig. S5). These preferences included known disorder-promoting residues (e.g., S, R, P, Q, E) in linkers that are typical of intrinsically disordered regions and known order-promoting residues (e.g., C, I, F, V, L, H) in domains typical of 3D structure (13). As cryptic domains match linkers with low probability, it seems reasonable to infer that intrinsically disordered regions are widely present in linkers. In fact, we analyzed the entire set of confirmed intrinsically disordered proteins (n = 643 in total) and found that 63 entries contained domains and had fewer than 30% structurally undetermined regions. Within these, 88.6% of residues in linker regions were disordered (only 1.8% were ordered), suggesting that intrinsically disordered regions can be substantial in linkers.

Domain and Linker Lengths Among the Three Superkingdoms.

The data summarized in Table 1 and Fig. 1 show that the mean values of protein lengths in Archaea are very roughly 10% smaller for single-domain as well as for multiple-domain proteins compared with those of Bacteria. Much more pronounced differences are found in comparisons with eukaryotes. Here, the proteins of Archaea and Bacteria tend to be less than two thirds the lengths of proteins in Eukarya, which confirms earlier observations (1–4).

The immediate impression from the data summarized in Table 1 is that the most significant differences between overall lengths of proteins from Eukarya, and also those from Archaea and Bacteria, primarily result from systematic differences in the lengths of linker sequences. On average, the Eukarya have more than threefold greater linker lengths than Archaea and Bacteria (Table 1 and Fig. 1). The mean lengths of the linkers are 250, 86, and 73 aa for Eukarya, Bacteria, and Archaea, respectively.

In contrast, proteins in the three superkingdoms have similar domain lengths, with mean values of 281, 280, and 256 aa for Eukarya, Bacteria, and Archaea, respectively (Table 1). In addition, there is a characteristic scattering of domain and linker lengths from the three superkingdoms that is particularly evident in one- and two-domain proteins (Fig. 1). Here, the small cluster of archaeal domains and linkers overlaps the smaller end of the bacterial cluster. In contrast, the length of domains in multidomain proteins of Archaea overlaps but is generally smaller than in Bacteria (Fig. 1). Both Archaea and Bacteria are well segregated according to their shorter linker lengths from the Eukarya cluster. In general, the scattering of linker and domain lengths for Eukarya is greater than that for Bacteria and Archaea.

Multidomain Proteins.

An enhanced scattering of domain lengths and linker lengths in all multidomain categories parallels the progressive increase of their cumulative lengths with increasing domain number (Fig. 1). Furthermore, the relative numbers of sequences with three or more domains is approximately 30% greater in eukaryotes than in Archaea and Bacteria (Table 1).

The multidomain proteins also reveal distinctive size distributions (Fig. S4). The fractions of proteins with increasing numbers of domains decrease very rapidly in all three superkingdoms, whereby proteins with one or two domains constitute approximately 90% of all of sequences (Table 1). There is a marked tendency for the domains to shrink in size as their numbers per protein increase; this is most pronounced for Eukarya. Consequently, the aggregate domain lengths of the largest multidomain proteins among Eukarya may be as much as 20% shorter than their counterparts among Archaea and Bacteria (e.g., average domain lengths of 604 and 731 aa residues in five-domain proteins for Eukarya and Bacteria, respectively). Likewise, the cumulative linker sequences of the eukaryotic multidomain proteins do not increase at the same rate as the cumulative domain lengths do with increasing domain numbers. The archaeal multiple domain proteins, as a whole, are marginally shorter than their bacterial counterparts, which are significantly shorter than those of Eukarya. Nevertheless, the most striking discrepancy is that Eukarya proteins in all categories are equipped with linkers that are approximately two to three times longer than those of Archaea and Bacteria.

Continuous Variation of Linker Lengths.

We compared in greater detail the linker lengths for proteins in cells from the three superkingdoms and some phyla. Briefly, N-terminal linkers (79 and 25 residues for Eukarya and Archaea, respectively) are generally slightly shorter than C-terminal linkers (87 and 28 residues for Eukarya and Archaea, respectively) in Archaea and Eukarya, but not in Bacteria, in which there is the opposite inequality. In addition, the N-terminal linkers in almost all animals and fungi are shorter than C-terminal linkers with a small number of exceptions. In most Bacteria and Archaea, the lengths of their internal linkers are shorter than N-terminal or C-terminal linkers. Average lengths from animals, fungi, plants, and protists suggest that plants have the shortest linkers and that fungal linkers are the most narrowly distributed, whereas protists have the longest and most dispersed linker sequences. All this detailed variability needs further study. At this point, these data suggest that different phylogenetic groups of organisms have evolved linker lengths in idiosyncratic ways.

The present data contradict the expectation (3) of discrete differences between eukaryote linker lengths and prokaryote linker lengths. Fig. 2 is a 3D plot that highlights the aggregate linker lengths of all proteins compounded from their N-terminal, C-terminal, and, if present, internal linker sequences. We observe a continuous variation of the aggregate linker length vector with all the proteins of the three superkingdoms. Thus, linker lengths for all proteins of the three superkingdoms can be sorted in a continuous series rather than in discontinuous blocks characterized by discrete sizes.

Linker and Domain Length Correlations with Proteome Diversity.

We recall that it is primarily linker lengths that vary between proteomes and not domain lengths (as detailed earlier). Accordingly, we might speculate that, if diversity of proteomes is approximately proportional to genome size, this might be reflected in plots of average linker lengths against genome sizes for cells of the three superkingdoms (4–7, 23). Although we observed a gradual increase of prokaryote linker lengths along with the corresponding genome lengths, eukaryote linker lengths were much longer than expected and approximately randomly distributed around a higher average value that did not correlate with genome length. The interpretation of these results was complicated by the potential influence of uncorrelated amounts of noncoding repetitive genome sequences.

To circumvent an influence of repetitive noncoding sequences on the analysis, we plotted average linker lengths against total CDS lengths. The CDS is the sum of codons in ORFs in a genome and is accordingly a measure of the total coding capacity for amino acid sequences of a genome, i.e., the diversity of the corresponding proteome. The CDS is also a measure of DNA and mRNA sequence heterogeneity for the corresponding genome. When the CDS values are plotted against domain diversity represented as the numbers of different domains at the fold superfamily (FSF) level of structural classification, significant linear correlations (R² = 0.609–0.874; P < 0.001) are obtained (Fig. S6). In brief, the CDS is a useful measure of the potential proteome diversity of a particular cell, but it does not take into account variations of expression levels. When we plot average linker lengths against total CDS for individual genomes, similar results to those obtained with uncorrected genome lengths are obtained (Fig. 3).

Fig. 3 describes the domain length and linker length distributions as a function of aggregate CDS for the genomes analyzed. The interpretation of this plot is based on the simplistic assumption that the combined coding capacity of a cell, as represented by its CDS value, is a valid measure of the diversity of that cell's proteome. Two convincing correlations surface in Fig. 3. First, the domain lengths of proteins from all three superkingdoms are relatively stable throughout the range of variation of CDS; the net increase over that range is less than 5%. In contrast, the lengths of linkers increase approximately twofold over the same range of CDS values for Archaea and Bacteria. In other words, the average lengths of the corresponding linker sequences are correlated with the molecular diversity encoded in the prokaryote genomes. This correlation suggests that reductive pressures are constraining the variation of linker lengths in the archaeal and bacterial populations (4–7, 23).

Here, one additional factor that may generate the observed scatter of these points is the dependence of the selective pressure for minimal protein lengths on the expression levels of individual proteins (4). The individual expression levels of proteins vary enormously within and between different cellular proteomes.

The results for the eukaryote cohort are much more chaotic. Thus, the plots of linker lengths versus CDSs for Eukarya reveal substantial random variations of linker lengths around a relatively constant average length (Fig. 3). This average length is greater than would be expected from an extrapolation of minimal lengths expected for eukaryotes based on the prokaryote relationship between linker length and CDS in Fig. 3. In other words, the best-fit simulation constructed for the Eukarya suggests that the average length of linkers within that cluster is not responding systematically to substantial increases in the molecular diversity (i.e., CDS) of the corresponding genomes (Fig. 3). The correlation coefficient of that best-fit second-order polynomial curve is a respectable 0.96. It would seem that the greater eukaryotic linker lengths are indifferent to the reductive pressures correlated with increasing CDS in prokaryotes. Similar conclusions could be drawn from plots that use domain diversity instead of CDS (Fig. S7).