The redundancy within the genetic code accommodates a variable number of codons to encode for the same amino acid. Codon usage biases have been found to exist, ranging from relatively neutral to extremely strong [^1-6]. Debate within the scientific community continues as to exactly why codon biases exist. Theories based on translational selection, mutational biases, and drift have all been found to contribute to codon biases [^7-11]. Sequence-based analysis revealed that within organisms having a biased genome, the most frequently occurring codons often reflect the most abundant transfer RNA (tRNA) available [^12-15]. Furthermore, through direct molecular manipulation protein-throughput was increased by designing coding sequences to utilize the most abundant tRNAs [^16-18]. These bioinformatic and experimental studies suggest that translational selection may be the primary factor shaping codon usage.

Correspondences between codon usage and tRNA availability may signify selection for translational accuracy, selection for translational efficiency, or both [^19-30]. Theories for selection for translational accuracy assume that the codon with the highest tRNA abundance has a lower missense error rate than its synonymous codons, considering tRNA gene copy numbers of both cognate and near-cognate tRNA abundances (e.g. [¹¹,³¹]). If codon bias exists as a result of selection to maximize protein-throughput, codon usage often reflects available tRNAs. Because codon bias is often strongest within highly expressed genes and there exists a correspondence between generation time and the strength of codon bias within these highly expressed genes [⁶], it is believed that selection for translational efficiency plays some role in determining a species’ codon usage. Moreover, recent studies have shown that the usage of particular synonymous codons can also impact protein folding or misfolding [^32-34].

As the abundances of individual tRNAs vary from one species to the next, so too do the preferences in codons. Translational selection seems most prevalent in viral species. Correspondence in virus-host codon usage has been observed in DNA-based, RNA-based, and retro-transcribing viruses (e.g. [^35-39]). This observation is not isolated to eukaryote-infecting viruses; bacteriophages also exhibit codon biases similar to their host species (e.g. [^40-42]). Given that viruses are heavily dependent upon their host for biosynthesis, utilizing the most prevalent tRNAs within the host would likely give the virus a translational advantage.

In order to quantify the bias present within a species, a number of metrics have been developed. Calculating the codon usage frequencies within the genome can reveal biases, in particular when comparing these usage profiles to expected usage patterns and/or the usage profiles of synonymous codons, e.g. the frequency of preferred codons (FOP) metric [¹³], the synonymous codon usage order (SCUO) metric [⁴³], and the %MinMax algorithm [⁴⁴]. In order to compare the strength of biases within and between species, the relative synonymous codon usage (RSCU) and the geometric mean of the RSCU values, the codon adaptation index (CAI), were developed in which codon usage frequencies within highly expressed genes (HEGs) are specifically examined [⁶,²³,⁴⁵]. Numerous extensions of the CAI metric have been proposed, e.g. the self-consistent codon index (SCCI) [⁴⁶] and relative codon adaption index (rCAI) [⁴⁷]. Rather than looking at the codons themselves to ascertain biases, a second approach exists in which biases are assessed relative to individual tRNA abundances. The tRNA adaptation index (tAI), also inspired by the CAI metric, considers not only the gene copy number of the tRNA with the perfectly matched anticodon but also those tRNAs which can bind imperfectly [²⁴]. The “local tAI” metric takes a similar approach, however the tAI measure is averaged for sliding windows across the gene rather than for the whole gene sequence [²⁶]. In many species the codon usage in HEGs matches tRNA abundance [¹²,⁴⁸].

Given the information encoded within the usage of codons, a number of tools and databases have been developed for analyzing the codon and tRNA content of genic and genomic sequences. CAIcal [⁴⁹], CodonExplorer [⁵⁰], and CodonW [⁵¹] calculate CAI values for user input sequences while E-CAI [⁵²] calculates the expected CAI values by generating random sequences with a G+C content and amino acid composition similar to the user input sequence. The application CodonO [⁵³] can analyze individual genomes or compare genomes using the SCUO metric. JCat [⁵⁴] and GCUA [⁵⁵] both calculate the RSCU values of user input sequences relative to a reference organism’s usage profile. The RSCU values for many species are contained in the Codon Usage Database [⁵⁶]; this collection, however, does not appear to have been updated since June 15, 2007. The Microbial Genome Codon Usage Database [⁵⁷] and Prokaryotes Codon Usage Database [⁵⁸] have lookup tables of codon counts and frequencies for over 500 and 800 species, respectively (as of June 2011). The most comprehensive database of codon statistics for individual microbial strains can be found in the Codon Usage Bias Database (CUB-DB) [⁵⁹]; individual links guide the user to individual strain values for many of the aforementioned metrics as well as two metrics developed by the database’s author [⁶⁰].

We have developed a new web resource called the Codon Bias Database (CBDB) which lists RSCU, normalized RSCU, and frequency biases (FB) values for 300+ and counting bacterial strains. The genera selected for this first release of CBDB all have sequenced, annotated bacteriophages, thus providing a reference for those studying adaptation in phage. Following the metric proposed by Paul Sharp and collaborators [⁶,²³,⁴⁵], analysis is performed for the set of highly expressed genes (HEGs) and these gene sets have been manually curated. CBDB is organized in such a way to easily accommodate codon usage comparisons between strains and species.

Previous analysis of the variation in codon usage between bacterial species selected just a few representatives of a genus and species [⁶]. We were interested to see how codon usage varied between species in the same genus as well as between different strains of the same species. For each of the three metrics included in CBDB, we calculated the correlation in codon usage for each pair of strains. A distance matrix was then computed as (1-r)/2 such that the instance of a pair being anticorrelated has a distance of 1 and a perfectly correlated pair has a distance of 0. The FITCH application of the PHYLIP package [⁶²] was used to derive a tree to visualize similarities/dissimilarities in codon usage. Figure2A shows the tree derived when the NRSCU values were compared for all strains. (Clades were collapsed when the subtree contained only one genus. Shigella was found to group both with Escherichia and Salmonella. Chlamydia and Chlamydophila were also collapsed into single nodes as some strains were closer to strains of the other species than of their own species.) From Figure2A, we also identified two branches including Pseudomonas (maroon circle) and Salmonella & Shigella (orange square) species. The leftmost Salmonella &Shigella branch, separating the pseudomonads is a single species – Salmonella enterica subsp. arizonae serovar 62:z4,z23:-- str. RSK2980 (GenBank: NC_010067). Also from this visualization of codon usage bias we noticed that the 34 Bacillus strains are interspersed amongst other species in the tree. The two B. lincheniformis and single B. clausii strain show a more similar codon usage pattern to the Vibrio fisheri sequences than they do to the other bacilli, as is shown in Figure2B. Looking at the Bacillus genus data file available from CBDB which includes statistics for all of the strains, one can see variation in usage between species.

The genera included in this first release vary in the strength of their bias. For instance, as has been previously documented [⁶], the Chlamydiae phylum does not exhibit a significant bias. The eight Chlamydia and Chlamydophila species examined here do not show a strong bias. ANOVAs were performed for all three statistics revealing that the variation between the species is not statistically significant. The Vibrio species, however, are strongly biased [⁶] and exhibit differences in codon usage between species. For instance, Figure3 shows the biases for the 11 Vibrio species examined for the six leucine codons. Referring to the Genomic tRNA database [⁶³], one can find that the non-cholerae species have more TAG-anticodon tRNAs (blue) than any other Leucine tRNAs. This thus can explain the preference within these species for CTA codon. The V. cholerae species also have more TAG-anticodon tRNAs than any other Leucine tRNAs, but in contrast to the non-cholerae species they also have CAG-anticodon tRNAs (green) [⁶³].