INTRODUCTION

Top Abstract Introduction Materials and Methods Results Discussion References

Escherichia coli, Vibrio cholerae, and Haemophilus influenzae are gram-negative -proteobacteria that can grow in human tissue and produce or contribute to disease. The principal habitat of E. coli is the human gut, V. cholerae is mainly a freshwater microbe, and H. influenzae is found in the human lung. On the other hand, Bacillus subtilis is a gram-positive, nonpathogenic soil bacterium. The minimal doubling time for these four bacteria in cultures is significantly less than 1 h. Fast growth implies many ribosomes, and these four bacteria have large numbers of rRNA operons per genome.

Predicted highly expressed (PHX) genes are characterized for the rapidly dividing bacteria E. coli, H. influenzae, V. cholerae, and B. subtilis using a method based on codon usage differences among gene classes (21). For complete lists of PHX genes, consult the website ftp://gnomic.stanford.edu/pub (see also Table 2).

	MATERIALS AND METHODS

Top Abstract Introduction Materials and Methods Results Discussion References

Assessments of gene or protein expression levels from codon usage were carried out as follows. High expression is predicted from codon usage as follows. Let G be a family of genes with average codon frequencies g(x, y, z) for the codon nucleotide triplet (x, y, z), normalized so that
where the sum extends over all codons (x, y, z) translated to amino acid a. Let f(x, y, z) indicate the average codon frequencies for the gene family F, normalized to 1 in each amino acid codon family. The codon usage difference of the gene family F (or a single gene) relative to the gene family G, termed the codon bias of F with respect to G, is calculated with the following formula:
where p_a(F) are the average amino acid frequencies of the genes of F (cf. references 19 and 20). Denoted by C is the collection of all genes, by RP the ribosomal protein genes, by CH chaperone/degradation protein genes, and by TF translation/transcription processing genes. Qualitatively, a gene g is deemed PHX if B(g|C) is appropriately high, whereas B(g|RP), B(g|CH), and B(g|TF) are suitably low. Predicted expression levels with respect to individual standards are based on the ratios
and the combined expression measure is
Other weighted combinations are also possible, but the results do not qualitatively change when different weights are used. We impose higher weight on the RP standard because the RP genes are generally the most PHX in all current completely sequenced genomes (21). The specification of the RP, CH, and TF gene classes as standards derives from the observation that these gene classes are consistently highly expressed in most genomes (21). Thus, these three gene classes (RP, CH, and TF) serve as representatives of highly expressed genes, and our method specifies genes with similar codon usages as PHX genes. These assignments are reasonable under fast growth conditions, where there is a need for many ribosomes, for proficient transcription and translation, and for many chaperone/degradation proteins needed to ensure correctly folded, modified, and translocated protein products.

A gene is predicted highly expressed (PHX) if the following two conditions are satisfied: at least two of the three expression values E_RP(g), E_CH(g), and E_TF(g) exceed 1.05, and the general expression level E(g) is 1.00. We sometimes refer to genes that do not unequivocally satisfy this definition but that have an E(g) of approximately 1.00 as marginally PHX.

	RESULTS

Top Abstract Introduction Materials and Methods Results Discussion References

To expose the significance of the PHX gene classes, we plotted B(g|C) versus B(g|RP) traversing all individual genes g (100 codons in length). The plots are given in Fig. 1 for each of the four rapidly growing bacteria. The distribution of points reveals two horns. The left horn effectively corresponds to the PHX genes. The right horn we refer to as putative alien genes. It consists of genes that significantly differ in their codon usages from the four classes C, RP, CH, and TF and will be discussed in a separate publication. If we replace the horizontal axis B(g|RP) with the coordinates of B(g|TF) or B(g|CH), the plots in Fig. 1 remain largely unchanged (data not shown).

View larger version (45K): [in this window] [in a new window]   FIG. 1.   Genes of 100 codons in the four fast-growing bacteria. Each gene is represented by a single point. Its position is determined by its bias relative to all genes B(g|C) and by its bias relative to the RP genes B(g|RP). PHX genes are indicated by red circles. Partial overlaps among the PHX and normal gene clusters are due to minor differences among B(g|RP), B(g|CH), and B(g|TF), which all contribute to the PHX predictions (see Materials and Methods). The upper right horn corresponds to putative alien genes (22, 28).

Top 20 PHX genes. The distribution of PHX genes among the four fast-growing bacteria is displayed in Table 1. The highest E(g) value exceeds 2 in all four genomes. Such high values are rare among the completely sequenced genomes (cf. reference 21). These four bacteria have a substantial number of PHX genes, ranging from 142 to 306. 

Comparison of predicted levels of expression in E. coli with 2D gel patterns. For many E. coli proteins, two-dimensional (2D) gel electrophoresis data for their abundances during growth in minimal medium are available. We compared the molar abundances of 96 proteins (with lengths of 100 amino acids [aa] [45, 46]) with the set of PHX genes (Table 4). Among the 20 most abundant of the 96 proteins, 17 were identified as PHX by our method. Among the 20 least abundant proteins of the 96, only 7 qualified as PHX. Of the remaining 56 proteins, which have intermediate molar abundances on 2D gels, 28 were identified as PHX. This agreement between high 2D gel abundances and high E(g) values supports naming the genes "highly expressed."

Classes of PHX genes. Tables 3 and 5 through 9 compare for the four fast-growing bacteria predicted levels of expression of all ribosomal protein genes, of the genes for the major transcription/translation processing factors, of the chaperone/degradation protein genes, and of the major energy metabolism genes. The extended repair gene repertoire of the four genomes and the vitamin biosynthesis genes of E. coli are evaluated in terms of E(g) levels (Tables 10 and 11). Each class is discussed in turn.

(i) Ribosomal protein genes (Table 3). Ribosomes of the four fast-growing bacteria have practically the same numbers of small- and large-subunit proteins. However, among all prokaryotic genomes, that number ranges from 50 to 65, while in eukaryotes, the number is constant at 79 (except in yeast, 78) (48, 50). This information suggests a greater range of variation in the patterns of protein synthesis among prokaryotes, consistent with the constrained phylogenetic origin of eukaryotic cells compared with the less constrained origin of prokaryotic species.

(ii) Genes for transcription/translation processing factors (Table 5). The majority of protein synthesis factors are PHX over all prokaryotic genomes. Expression levels correlate highly across species (Table 5, footnote a). As with the ribosomal proteins, the E(g) values cover a wide range. Elongation factor EF-G (fus) is distinctive, with an E(g) value exceeding 2 for each genome. The highest expression levels in E. coli occur for the RpoB and RpoC subunits of the core RNA polymerase. RpoA is PHX in B. subtilis but not in E. coli, V. cholerae, and H. influenzae. Why are the predicted expression levels for the RpoB and RpoC subunits higher than that for RpoA? Based on the RNA polymerase stoichiometry (one copy of RpoB, one copy of RpoC, but two RpoA units), should one expect elevated expression levels for RpoA compared to RpoB and RpoC? A possible explanation relates to the differences in protein sizes, RpoB and RpoC being larger proteins than RpoA. It has been observed for E. coli that codon choices in long genes tend to be more biased than those in short genes (10). Interestingly, Mycoplasma genitalium, its relative Ureaplasma urealyticum, and the spirochete Treponema pallidum feature PHX RpoA but not RpoB and RpoC.

(iii) Chaperone/degradation protein genes (Table 6). Among the top PHX genes in most eubacterial genomes are those for the major chaperone protein archetypes, DnaK and GroEL. These reach E(g) values exceeding 1.3 (>2 in E. coli). The gene for the multifunctional enzyme Pnp, fundamental in RNA processing and mRNA degradation, attains the highest predicted E(g) value, 2.66, among all E. coli genes. Pnp is PHX in many eubacterial genomes but not in B. subtilis.

(iv) Levels of expression of aminoacyl-tRNA synthetases (Table 7). There are 19 PHX tRNA synthetase polypeptides in E. coli, including two subunits of phenylalanyl-tRNA synthetase (PheS- and PheT-) and two subunits of glycyl-tRNA synthetase (GlyQ- and GlyS-). However, there are only eight in V. cholerae, seven in H. influenzae, and three in B. subtilis. IleS is missing from H. influenzae, and GlnS is missing from B. subtilis, which uses amidotransferase modifications to produce Gln-tRNA^Gln from Glu-tRNA^Glu synthetase. Actually, the GlnS gene is absent from most prokaryotic genomes (14).

(v) Levels of expression of major energy metabolism genes (Table 9). Enzymes of major catabolic pathways can be divided into four groups: glycolysis, pyruvate metabolism, the pentose phosphate pathway, and the TCA cycle. The glycolysis genes are predominantly PHX in all four fast-growing bacteria, with very high E(g) values, >2.00, for several of these genes in E. coli. Hexokinase and glucokinase are prominent glycolysis proteins in most eukaryotes, but the former is not found in most prokaryotes, including the four fast-growing bacteria under analysis in this study. Why? In glycolysis, hexokinase converts glucose to glucose-6-phosphate. However, glucose-6-phosphate arises from other hexoses and from glucose transported into the cell via the phosphotransferase system. Perhaps the multiplicity of sources means that glucokinase need not be PHX. Glucokinase occurs in many (but not all) eubacteria, normally at low to moderate E(g) values, 0.3 to 0.8.

Certain gene groups generally not PHX. Specific regulatory proteins or proteins responding to special demands and used few times, as in the highly specialized DNA repair processes, are not expected to be PHX. Also, specific transcription proteins and DNA replication proteins, because the cell assembles few replication machines, tend not to be PHX.

(i) Genomic repair proteins. Table 10 reports predicted expression levels for the main collection of repair proteins for the four genomes. Only two repair proteins of E. coli reach PHX levels: RecA and Ssb (single-stranded DNA binding protein) [E(g) for both, 1.48]. Two other repair proteins are borderline PHX: Dut (deoxyuridine 5'-triphosphate nucleotide hydrolase) and HepA [E(g) = 0.97 and 0.99, respectively]. Other repair proteins have low to moderate predicted expression levels, the E(g) values almost always in the range from 0.35 to 0.80. These evaluations parallel those for D. radiodurans, in which RecA [E(g), 2.04] has a dramatically high predicted expression level and MutT (gene no. DR2358) reaches an E(g) of 1.29, these being the only two proteins qualifying as PHX (22). The other repair proteins of D. radiodurans have E(g) values in the range 0.40 to 0.80.

(ii) Vitamin biosynthesis proteins (Table 11). Pathways to the synthesis of vitamins, of which only small amounts are needed to provide adequate cofactor function, have largely low predicted expression levels, with E(g) values of about 0.40 to 0.75. In E. coli, the genes acting in the synthesis of six vitamin cofactors, biotin, thiamine, riboflavin, lipoate, pyridoxal, and cobalamin, were examined. Only RibH, which participates in riboflavin biosynthesis, is PHX in E. coli. Although the enzymes of the biosynthetic pathways are poorly expressed, some of the enzymes that utilize the vitamins as cofactors are highly expressed, for example, biotin carboxylase (a subunit of E. coli acetyl-CoA carboxylase). In B. subtilis, RibE, which is not PHX, in the same pathway forms an oligomer complex with RibH in which the structural union (RibE-RibH) combines 3 units of RibE with 60 units of RibH (23). This anomalous stoichiometry makes it likely that RibH furnishes structural support and, for this reason, is PHX; in this guise, RibH may be used in other capacities. Paradoxically, RibH is not PHX in B. subtilis.

(iii) Genes of signal transduction pathways. In Table 8 of reference 21, the predicted expression levels for several two-component sensor genes (histidine kinases) of E. coli and B. subtilis are reported. In all of those examples, the predicted expression levels were low, the E(g) values ranging from 0.30 to 0.70.

(iv) Principal starvation genes of E. coli and their predicted levels of expression (Table 12). The genes shown in Table 12 are associated with starvation states, as discussed in the review (26). Three genes in this category are PHX: dps, also labeled pexB [E(g), 1.13], which provides protection from oxidative radicals; rpoH, which encodes ³² [E(g),1.46]; and the survival protein, SurA [E(g),1.10], a chaperone which is a member of the PPIase family. We expect these proteins, by virtue of their codon usage patterns, to be capable of high levels of expression, especially when induced by starvation. Other starvation proteins (Table 12) have low to moderate E(g) values. The ^E factor, which regulates the activity of other periplasmic proteins, is not PHX, and the same is true for ⁵⁴ and ³⁸, which respond to nitrogen and/or carbon starvation, respectively. However, ³² (rpoH), the principal chaperone sigma factor, pervasively registers as PHX, presumably to establish high levels of chaperone production.

Homologous PHX genes among the fast-growing bacteria. Table 13 compares the numbers of homologous PHX gene families among the four rapidly dividing bacteria. There are 60 gene families common to the four fast-growing bacteria, with each member PHX. Thirty-two of these are families of RP genes, eight are families of TF genes, and nine are families of genes essential for energy metabolism. Twenty-three gene families distinguish E. coli with PHX representatives, but these are not PHX in the other three fast growers, including five CH genes and five TF genes.

Codon usages along the gene and expression levels. For relatively long genes (600 codons long), we determined expression levels with the gene length divided into three equal parts (5', middle, and 3' parts). The pairwise correlations among the three parts of the E. coli genes are high, 0.86, 0.85, and 0.88, respectively, indicating that expression levels calculated from codon biases are effectively the same for the three parts of genes.

PHX ORFs shared by the four fast-growing genomes. Genes are considered homologous if their SSPA (significant segment pair alignment) score (percent similarity; see reference 5) is 40%. Examples include three ORFs (yaaH, yajC, and yeeX) common to E. coli and V. cholerae, three similar ORFs (yfiD, yjjK, and yebC) present in the genomes of E. coli, V. cholerae, and H. influenzae, respectively, and one ORF (ybaB) common to E. coli and B. subtilis. These PHX genes of unknown function offer attractive candidates for mutagenesis and knockout studies to determine their functions.

Distributions of PHX genes over the chromosomes. Clusters of PHX genes are displayed in Table 15. Statistical significance was assessed using the r-scan analysis protocol described elsewhere (18).

	DISCUSSION

Top Abstract Introduction Materials and Methods Results Discussion References

Gene expression can be evaluated in several ways. One currently popular way centers on DNA microarrays (DNA chips) aiming to dissect gene expression under varied physiological, clinical, and environmental conditions. These DNA chips have been applied to the monitoring of genes in different situations for the discovery of genes associated with diseases; for assessment of gene expression under inducements from drugs, chemicals, or toxins; for ascertainment of genes compensatory for knockout mutations; and for profiling of gene expression patterns in temporal and tissue-specific localizations. The current microarray methodology is restricted to discriminating transcription levels and not levels of translation or protein abundances (33, 42). Also, DNA chip hybridizations are generally unable to detect unambiguously low-abundance gene transcripts. Experimental evaluations of protein abundances under different cellular conditions can be assayed by 2D gel electrophoresis (reviewed in reference 46) supplemented by mass spectrometry (51), by antibody associations, and by biochemical tests. Also, correlations of 2D gel proteomes and microarray assessments of transcriptomes generally appear to be weak (13). However, Futcher and coworkers (11) reexamined these correlations in yeast and found generally good agreement.

Codon choice is presumably influenced by protein structure via evolutionary selection for the most accurately translated sequences at structurally important locations. Codon choices may be different at the beginning of a gene than at the central part of the gene (6). It has been suggested that translation pause sites, especially early in the coding sequence, can slow translation initiation (16). Accordingly, there appear to be conflicting selection pressures imposed by constraints on ribosomal binding for the rate of initiation, rate of elongation, and overall translation fidelities. In rapidly growing cells, where ribosomes are limiting for protein synthesis, a ribosome stalled at a rare codon is unavailable for the synthesis of other proteins, and the higher the molar abundance of the stalled protein, the greater the disruption of cellular growth (52). Protein structure may be correlated with codon usage (e.g., see references 30 and 44). Thanaraj and Argos (44) argue the rare-codon hypothesis for domains and secondary structures, in which repetition of rare codons reduces translation rates and introduces translation pauses, allowing time for protein domains and secondary structures to fold into native structural conformations.

Codon usage offers another way to evaluate gene expression with a different set of limitations. Our sequence methods are effectively complementary to the experimental procedures of 2D gel electrophoresis and DNA microarray analysis in assessing gene expression levels. By our methods, genes similar in codon frequencies to RP, TF, and CH genes but strongly deviant in codon usage from the average gene are identified as PHX. Our analyses and data support the hypothesis that each genome has evolved codon usage patterns indicating "optimal" gene expression levels for most situations of its habitat, energy sources, and lifestyle. The three protein familiesribosomal proteins, major translation/transcription processing factors, and chaperone/degradation proteinsare fundamental at many stages of the cell life in promoting growth and stability. Generally, PHX genes exploit favorable codon usages, tend to possess strong Shine-Dalgarno sequences, and putatively possess strong promoter sequences (cf. reference 21). Some limitations of our method result from an implicit assumption that the codon usage of a gene is not affected by its location in the genome, e.g., G+C-rich versus A+T-rich regions. The high variance of G+C composition (isochores) along mammalian genomes may be prohibitive with respect to predicting gene expression levels from codon usages. However, the nucleotide compositions of bacterial genomes are largely homogeneous. Some genes that deviate in G+C content (e.g., those for transposases or specialized pathogenicity islands) tend to be detected as "putative alien" genes (22, 28).

What does the expression level E(g) for a gene g reflect? Gene expression in prokaryotes is regulated at initiation, elongation, and termination of transcription and of translation, by different rates of transcription and translation, by differential mRNA stabilities, by segmental stability differences in polycistronic messages, by codon preferences, and by interactions with chaperone and other proteins. Expression is also influenced by lifestyle, habitat, and energy sources. The classes (RP, TF, and CH) of proteins that we have chosen to represent highly expressed genes are needed in high molar abundances when a high rate of protein synthesis is essential.

Multifunctional proteins and PHX levels A protein that belongs to a PHX class and that performs several functions might be expected to register higher E(g) values than the average PHX gene. We offer several examples.

Contrasts in PHX levels among genes involved in energy metabolism in E. coli and B. subtilis. As indicated earlier, certain genes of energy metabolism are predominantly PHX in all four fast-growing bacteria and have high expression levels [E(g), often >2.00]. This is manifestly valid (Table 9) for glycolysis genes and for genes of pyruvate oxidation. Why should most of the TCA genes of E. coli be PHX but not those of B. subtilis? We suggest four possible contributing causes. (i) Perhaps B. subtilis makes less use of the TCA cycle for ATP production than E. coli. The principal glycolysis genes of B. subtilis, unlike those of E. coli (dispersed all over the E. coli genome), are encoded from a single cluster (gap, pgk, tpi, pgm, and eno); see our earlier discussion of PHX clusters. (ii) The TCA cycle has at least two main tasks: the first, aerobic energy (ATP) production, and the second, synthesis of carbon chain precursors to various essential metabolites, such as amino acids. Can many of these precursors be more easily acquired by other means in B. subtilis? B. subtilis, in marked contrast to E. coli, has four PHX flagellin genes (flagellin [hap], flagellar hook protein [flgE], flagellar hook basal body [fliE], and flagellin homolog [yvzB]), whereas a single flagellin gene of E. coli is PHX (21). Moreover, flagellar genes are strictly regulated and inducible in E. coli but constitutive in B. subtilis (40). Assuming that soil is the primary B. subtilis habitat and that the human gut is the primary habitat for E. coli, different metabolic patterns may be appropriate. The swimming movements of B. subtilis mediated by its PHX flagellar proteins may facilitate the acquisition of nutrients, such as amino acids, from an assortment of soil sources. B. subtilis also excretes many digestive enzymes in gathering macromolecular nutrients for possible predatory objectives (1). (iii) There are also differences between E. coli and B. subtilis in energy pathways, which can influence expression levels. For example, E. coli uses succinyl-CoA as a precursor in the biosynthesis of lysine and methionine, whereas B. subtilis uses acetyl-CoA for this objective. E. coli possesses isocitrate lyase (AceA) in competition with isocitrate dehydrogenase, the first enzyme of the glyoxylate shunt pathway, which is very effective for acquiring a net carbon gain in the metabolism of fatty acids, whereas B. subtilis lacks AceA. The early genes in the TCA cycle of B. subtilis, those for aconitase and isocitrate dehydrogenase, are PHX, whereas the remaining genes are only predicted moderately expressed. Apparently, the order of TCA genes can be important. (iv) B. subtilis and E. coli are both facultative aerobic organisms (29). For anaerobic respiration, B. subtilis relies exclusively on nitrate or nitrite as its terminal electron acceptor, whereas E. coli has many alternative acceptors.

Highly expressed genes under varying conditions. Can our methods be applied in conjunction with microarray analysis? We cannot change the codon usage of a given gene, but we can change the gene class standards for discerning expression levels relative to these gene classes (see Materials and Methods). Here, the gene class standards are RP, TF, and CH. It is hypothesized that similarity of codon usages, as characterized in Materials and Methods, for two or more natural gene classes may identify new genes with similar properties, as in the defining gene classes. Effectively, codon usage patterns provide a means to correlate genes and functional categories (20). By using several gene classes as standards, a figure corresponding to Fig. 1 but in multiple dimensions, when coupled to a suitable clustering analysis, may discriminate additional genes highly expressed relative to the different gene class standards. For example, when we compare codon usages of genes with respect to the B. subtilis sporulation genes versus the class of all genes, the two coordinates plot a straight line. In another example, yeast mitochondrial genes feature a melange of PHX genes, putative alien genes, and average genes, and the genes for the ribosomal proteins functioning in the mitochondrion tend to show codon usages akin to average genes.