The two types of CAI were calculated for all genes in 318 bacterial strains and fungal genomes, and the correlations between the derived tCAI and dCAI values are illustrated for eight different bacterial phyla, with any remaining bacterial species grouped into 'Other bacteria', and fungi depicted separately (Figure ). For most groups, the correlation between the two CAI measures is high (median above 0.5). Only for chlamydiae and spirochaetes are the median correlations below 0.5, indicating that the dominating codon biases are not translational for most of the species included in these groups. However, it is not surprising that there appears to be little selection for strong tCAI bias in these genomes because most of the bacteria in both of these phyla have slow replication times. Presumably, fast-replicating bacteria have optimized their replication machinery as opposed to slow-replicating bacteria, for which other factors might be more important [7
]. Consequently, we were able to confirm a significant relationship between the level of translational codon adaptation and replication time across the entire range of genomes (Spearman's rank correlation, rho about 0.46) using the number of 16S rRNAs as an indirect measure of doubling time, as previously suggested [13
], since the number of 16S rRNAs indirectly influence replication times [14
Figure 1 Box plot summarizing correlations between tCAI and dCAI for eight major bacterial phyla and fungi. The group 'Other bacteria' comprises a number of minor bacterial phyla (Aquificae, Chloroflexi, Fusobacteria, Planctomycetes, Acidobacteria, and Thermotogae) (more ...)
Next, the codon preferences, which are measurable by the relative adaptiveness of each codon (wij
), were compared between tCAI and dCAI and the difference (wij
for tCAI minus wij
for dCAI) was used for cluster analysis of all 318 bacterial strains and the five fungal genomes (Figure ; also see Additional data file: 1, additionally available at our website [15
]). Figure shows a clear separation into several clusters with AT-rich bacteria towards the left and GC-rich bacteria towards the right, whereas bacteria with intermediate base composition are in the middle. This is also reflected in the clustering of codons, which are separated into two distinct clusters in which either a codon preference for A/T (lower half) or G/C (upper half) in the third position for dCAI is evident (GC3/AT3 skew dominates over translational bias). However, although the AT content appears to be a significant factor in the clustering, merely ordering by AT content does not yield the same highly distinguishable clusters. Consequently, the correlation between the level of translational codon adaptation (measured by the correlation between tCAI and dCAI) and the genomic AT content was indeed very low but still significant (rho about -0.14, P
value about 0.015), supporting the minor although unmistakable correlation between AT content and clustering order visible in Figure . Furthermore, from the color bar in Figure , indicating the phylogeny of each microbe, we observe that the clustering is not related to known phylogenetic relationships based on sequence homology. Although smaller clusters of microbes of the same bacterial species are indeed observed, this is perhaps not surprising because genomes of the same species would be expected to have essentially the same codon usage preferences. However, microbes from the same phylum are not clustered but rather are scattered throughout the figure, while many clusters contain organisms that are quite far apart phylogenetically.
Figure 2 Two-dimensional cluster analysis of differential codon preferences for tCAI and dCAI. The differences in relative adaptiveness of each codon (wij for tCAI minus wij for dCAI) for each Genbank entry were clustered into two dimensions, one clustering codons (more ...)
The middle area of Figure appears most diverse and can be divided into three distinct regions (ignoring a few smaller clusters on its left side). This division results in a total of five distinct regions, as illustrated in Figure . Figure provides a zoom of the third and fourth region from the left. The third region consists mainly of 'enterics' (intestinal bacteria) living in the human intestine (for example, Escherichia, Shigella, Salmonella, Bacteroides), the fly intestine (Yersinia pestis), and the animal intestine (Yersinia pseudotuberculosis). The yeast genome, S. cerevisiae, clusters with the enterics. Although fungi are clearly quite distant from bacteria phylogenetically, both can be relatively fast replicating and hence would face the same selective pressure on codon usage. Moreover, Kluyveromyces lactis also groups with the enterics, including E. coli K-12, with whom it is often grown together in fermentors to produce chymosin (rennet) on a commercial scale, reflecting similar preferences on growth environment.
The fourth region mostly consists of bacteria living in aquatic environments such as marine waters (Thermotoga maritima
, Prochlorococcus marinus
, Desulfotalea psychrophila
, Synechococcus species
), groundwater (Dehalococcoides
), freshwater (Synechococcus elongatus
), and hot springs (Thermosynechococcus elongatus
). Although other P. marinus
strains cluster in the first region, strain MIT9313 is low-light-adapted and has almost as many strain-specific genes as it has genes in common with its high-light-adapted relative, strain MED4 [16
], which reflects the differing environmental preferences of the two strains.
Looking at the remaining regions in Figure , we observe that the first (left-most) region consists of slow-growing intracellular pathogens (Mycoplasma
, and Chlamydia
, among others) and other small pathogens (Bartonella
, and Campylobacter
), mostly with genome sizes less than or close to 1 megabase (Mbp). The content of this region reflects the observation that many organisms with reduced genomes have very low GC content and supports the speculations that there is a selective pressure in this group of bacteria to lower the nitrogen requirement for DNA synthesis [17
] by adapting the codon usage to favor codons with more As and Us. The second region mainly consists of spore formers, including Gram-positive bacteria. Many of the bacteria in this region can replicate quite rapidly, and exhibit other evidence of selective pressure for optimization of the genome for quick replication on demand. For example, the Vibrio
(a Gram-negative, non-spore-former) and Bacillus
(a Gram-positive spore-former) cluster close together; and they have the largest number of rRNAs and tRNAs out of several hundred bacterial genomes sequenced so far. Finally, the fifth (right-most) region mainly consists of soil bacteria, soil symbionts and plant pathogens, as well as a few mammalian pathogens. Among additional bacteria in this region, we found an intercellular pathogen, Brucella melitensis
, that may have evolved from soil and plant associated bacteria [18
] and a pathogen, Wolinella succinogenes
, in which several soil-related genes have been identified [19
]. Thus, we find that, upon closer inspection, apparently misplaced genomes in a region may reflect similar shared ecologic niches in the past.
By the above described approach, we were able to divide the organisms into three overall groups reflective of the genomic AT/GC content as previously demonstrated, based on distances between binarized codon weights from dCAI [7
]. However, rather than merely discriminating between classes of lifestyle in terms of mesophily, thermophily and hyperthermophily - as previously shown based on either amino acid composition [20
] or by codon usage [7
] - we obtained an environmental signature based on differences in codon weights between evolutionary more dominant codons and codons preferred by the translational machinery. Consequently, we demonstrate that differences in codon usage bias by tCAI and dCAI provide an environmental signature by which it is possible to group bacteria into environmental groups, such as soil bacteria, enterics, sporeformer, and intracellular pathogens. Moreover, this environmental signature does not reflect already known phylogenetic relationships, and as such the approach described above is not intended to replace or extending the existing methods in phylogeny that are based on sequence homology. These results build on a previous finding that GC content of microbial communities is influenced by the environment [22
Prediction of highly expressed genes
tCAI is a 'forced' measure of translational bias, whereas dCAI is a measure of the most dominating bias for an organism independently of the type of bias (GC skew bias, strand bias, and so on). For this reason, the correlation between these two measures is a simple and intuitive yet strong indication of whether the most dominating bias is translational, and consequently of how well the dCAI values explain gene expression. In this sense, it is not surprising that the correlation between the two CAI measures also gives an indication of how well tCAI explains gene expression levels. This trend holds true at least for the six organisms for which we compared CAI values with microarray data, where the correlations between the two CAI measures are significantly correlated with the degree of how well tCAI correlates with gene expression (rho = 0.6).
To further analyze and compare genes predicted as being highly expressed by tCAI versus genes having extreme codon bias according to dCAI values and versus the highly expressed genes estimated by microarray analysis, the overlap between the top 10% genes was found and visualized in Venn diagrams (Figure ). For both S. cerevisiae
and E. coli
there is good overlap of all three circles; that is, many of the same genes with high tCAI also have high dCAI values, and furthermore these genes are also found to be highly expressed in microarray experiments. For Bacillus subtilis
, a smaller but similar trend is evident. For the remaining bacteria, a significantly higher number of genes with high expression values (microarray data) overlap with genes with high tCAI values than with genes having high dCAI values. An investigation of the functional categories to which the dCAI reference genes (top 1% of genes) belonged revealed that for S. cerevisiae
, E. coli
and B. subtilis
, a significant fraction of ribosomal proteins were included, whereas for Pseudomonas aeruginosa
, Campylobacter jejuni
and Geobacter sulfurreducens
, no ribosomal proteins where found among dCAI reference genes. This is in agreement with the ribosomal criterion defined by Carbone and coworkers [7
], which states that that ribosomal proteins have significantly higher dCAI values than other protein encoding genes in translationally biased organisms. Thus, organisms having few or no ribosomal proteins among dCAI reference genes exhibit little translational codon usage bias as compared with organisms having many ribosomal proteins among dCAI reference genes.
Figure 3 Venn diagram evaluating the prediction of highly expressed genes (tCAI and dCAI) by comparison with microarray gene expression data (Expr). The overlap between genes with top 10% tCAI, dCAI, and Expr values are pictured as overlapping circles, in which (more ...)
The above comparison of microarray data with tCAI values demonstrates that even for organisms that are evolutionarily far from E. coli
(for which the bacterial reference set of highly expressed genes was derived), it is possible to predict highly expressed genes by their tCAI values even when the most dominating bias in an organism is not translational, by comparing codon usage for each gene to that of genes in the reference set of highly expressed genes using tCAI. This demonstrates that although the assumption that the same 27 genes are highly expressed in all bacteria may not be entirely true, the codon usage pattern for these genes do provide a useful signature for predicting highly expressed genes. However, the level of confidence decreases with decreasing levels of translational codon adaptation in the dominating codon usage biases (as estimated from the correlation between tCAI and dCAI). Thereby, better performance was obtained than by employing merely the most dominating codon usage bias identified by dCAI, especially for organisms for which translational bias is not dominant (as also observed by Carbone and coworkers [7
]); in the latter case, dCAI would not be useful for predicting gene expression at all.