Theoretical Biology and Medical Modelling Open Access Does Codon Bias Have an Evolutionary Origin?

Background: There is a 3-fold redundancy in the Genetic Code; most amino acids are encoded by more than one codon. These synonymous codons are not used equally; there is a Codon Usage Bias (CUB). This article will provide novel information about the origin and evolution of this bias.


Background
The genetic code is redundant: 20 amino acids plus start and stop signals are coded by 64 codons. This redundancy increases the resistance of genes to mutation: the third codon letters (wobble bases) can often be interchanged without affecting the primary sequence of the protein product. Nevertheless, wobble base usage is highly conserved in mRNA sequences (there is no or very little individual or intra-species variation) and, interestingly, some wobble mutations (though they are called silent mutations) are known to cause genetic disease with no change in the amino acid sequences [1].
However, the wobble bases are not randomly selected, as they might be if interchangeability were unrestricted.
There is codon bias, i.e. codon usage is not equally distributed between the possible synonyms; some redundant codons are preferentially used. This bias is described in Codon Usage Frequency (CUF) Tables [2].
These observations may not be universally valid because some statistically significant observations in one species are not reproduced in another. However, there is a strong expectation that codon bias, which is obviously well conserved in different species, reflects a general biological function because of the universal nature of the Genetic Code and the structure and function of nucleic acids and proteins.
The aim of this study is to investigate the possible origin of so-called "codon bias", measure it quantitatively and compare it among many species.

Materials and methods
Codon Usage Frequency (CUF) Tables were obtained for 113 different organisms from the Codon Usage Database (NCBI-GenBank, update: November 16, 2006 [24]). The organisms were selected from KEGG (Kyoto Encyclopedia of Genes and Genomes, [25]) and represented a wide variety of species from different evolutionary lines [Additional file 1].
To calculate Codon Usage Bias (CUB) numerically, I assumed that statistically equal usage of all available synonymous codons is the neutral "starting point" for the development of species-specific codon usages, and the CUB is the sum of the deviations from such random, equal usage.
The codons (i, 64) were divided into 21 subgroups (j, corresponding to the 20 amino acids and 1 stop signal). The number of occurrences of a codon was normalized and the frequencies of the codons (CUF ij ) in each fraction were calculated. The sum of CUF if in a fraction was always treated as 100% so the sum of all fractions was 2100%. n i is the number of synonymous codons in the j th fraction and n j = 64 CUF ij is the frequency (%) of the i th codon in the j th fraction encoded by n i synonymous codons.
These fractional frequencies were compared to the random fractional frequencies (rCUF ij ), defined as the fractional frequency that a codon would have if all alternative codons were used randomly and equally. rCUF (1j) = rCUF (2j) = rCUF (n)j = rCUF (ij) = 100/n i (%) The sum of rCUF in a fraction is also 100% and in each fraction altogether is 2100%.
CUB is defined as the absolute difference between CUF and rCUF:-More simply, CUB is the absolute number of fractional frequencies minus the number expected if usage of synonymous codons was uniform.
CUB may be used in some cases with its +/-orientation indicated. In these cases, positive values indicate over-utilization of a codon (e.g. dominant codons) while negative values indicate under-utilization (suppression). CUBmin = 0 if CUF ij = rCUF ij and the Calculated Maximal Possible CUBmax is 2416.7%. This is the value when only one of all the possible synonymous codons is used (100% frequency) for every amino acid and for the stop signal.
Further explanation of the CUB calculation is given in [Additional file 2], together with an example. CUF ij (%) is not to be confused with a "regular" codon frequency (CUF i ), which indicates the frequency of a codon in the entire genome (all 21 fractions) and is usually given in the CUF Tables in #/1000 units.
The definition of CUB in this article is not directly comparable to other widely used definitions such as CUI.

Results
Quantitative evaluation of codon bias CUB = 0% when all available synonymous codons are equally used. The maximal calculated bias, CUB max = 100%, indicates that only one codon is used for each amino acid (and for the stop signal), while the remaining 43 codons are not used at all. I calculated CUB in 113 species and found that the average value is 29.3 +/-1.1% (S.E.M, n = 113). There seems to be a modest but significant decrease in the bias during evolution: bacteria and archeoata have the highest bias while vertebrates have the lowest. Eukaryotes have significantly lower CUB than prokaryotes. Humans have the lowest value (18.9%) (Figure 1).
There is a slight negative correlation between the size of the codon-and gene-pool of an organism and its CUB (p < 0.01, n = 113, not shown). The size and complexity of both genome and proteome increase with evolution, while the CUB decreases. A larger codon pool seems to utilize more codon variation, which leads to lower differences between the usage frequencies of synonymous codons.

Qualitative evaluation of CUB
Detailed analysis of different species reveals wide variations in CUB ( Figure 2). There is a seemingly random variation in CUB between amino acids and different groups of organisms. However, a comparison of closely-related species with large codon pools shows very similar patterns. For example, all mammals have very similar CUB patterns.

Pan-genomic codon usage
I accumulated the CUF data from the 113 species into a single CUF Table (Table 1). This Table is intended to give a virtual representation of all organisms (Pan-Genome) and a numerical representation of the "universal" translation machinery. As many as 288 × E10 codons are represented in this collection. The distribution of CUB values in the Pan-Genomic CUF Table is illustrated in Figure 3. The transition from maximum-positive to maximum-negative values is smooth and there is no obvious or unambiguous border between the so-called dominant and prohibited codons. All possible codons are used.
There is a significant positive correlation between the number of synonymous codons (n i , #/amino acid) and the propensity of amino acids in the proteome (#/1000 amino acid residues). A similar correlation exists between synonymous codon frequency and CUB ( Figure 4). These important correlations were discovered by analyzing the Pan-Genomic CUF Table (64 values) and were confirmed using individual data from all species (113 × 21 values).
Another possible way to evaluate the possible phylogenetic relationships among CUBs in different species is to use the Pan-Genomic CUB I found that the CUB of vertebrates is most similar (least distant) to the average CUB, while bacteria and viruses are most distant from it. This correlation analysis involves all codons and gives no information about the development of individual CUBs. I therefore compared the codon-specific CUB values in the 113 species to obtain a rough estimate of the stability of (commitment to) a CUB through evolution. The mean/SD of the 113 amino acid-specific CUB values gives a good estimate how this stability ( Figure 5).

Internal dynamics of codons Correlations between individual CUB frequencies
When one of the synonymous codons is used more frequently than expected (positive CUB), another will be less frequently used (negative CUB). More generally, this means that codon usage changes in a subgroup of the 64 codons will be accompanied by changes in the opposite direction in the remaining codons.
I sorted the CUB values (64 × 113 = 7,232 listed in total) in the Pan-Genomic CUB Table according to their sizes and +/-directions [Additional file 4]. This sorting divided the 64 codons (c) into two subgroups (Ac and Bc) and the 113 species (s) into two additional groups (As and Bs). The Ac-As and Bc-Bs subgroups contained predominantly over-represented (positive CUB) codons and are located in the opposite diagonal corners of the Table. The Ac-Bs and Bc-As fields contained predominantly under-represented (negative CUB) codons and are located in the other opposite diagonal corners of the Table. There is an internal inverse relationship between codons, which is valid and the same for all species. This inverse relationship is shown in a compressed and simplified form in Figure 6a, b.  Negative correlations were expected between some subgroups of CUBs and others in the same species. Surprisingly, however, all codons and all species belong to only 2 clusters with highly correlated, opposite dynamics.

Distribution of Pan-Genomic CUB
The above figures indicate that there is a close internal and inverse correlation between the CUBs of different codons. The magnitude and orientation of a CUB shows wide variation between species. Our collection of 113 species is too limited for any conclusion about the phylogenetic rules of development of CUB to be drawn, but the first impression is an absence of phylogenetic rules: -about half the species under-utilize about half the codons, while the other half show the opposite behavior in respect of the remaining codons.
-It is difficult to find a correlation between CUB and taxon boundaries. All mammals (in the table) show a homogenous CUB pattern, while other taxa are much more diverse.
-Most codons show a wide pangenomic variation in CUB, but some vary much less than others ( Figure 5). Some codons (TAG, GGG, CGA, CTA) are under-utilized by more than 80% of the 113 species listed, i.e. these synonymous codons have become committed to a given CUB orientation while others have not. There is a significant negative correlation between the proportion of codons committed to a given CUB orientation and the extent to which CUB varies (also apparent in Figure 5). There are some highly significant correlations among codon bases. The fractional frequency of each nucleotide base in every codon position correlates positively with its complementary codon ( Table 2).

Internal relationship among codon bases in codon usage tables
The sum of both complementary codon pairs (A+T and G+C) in every codon position is positively correlated to the sum of the same codon pair in the other two codon positions (Table 3). These correlations are valid for every species.
This strong positional correlation between codon bases suggests that it is possible to predict the frequency of usage of a nucleotide in the codon usage table from the frequencies of other nucleotides. Predictions regarding the third nucleotides in codons are especially interesting, because these are wobble bases for most amino acid codons.  Figure 7). This is of course a prediction of the frequencies of the four wobble bases in all 64 possible codons and has no predictive value for individual wobble bases belonging to individual amino acids. All these correlation were of course carefully compared to corresponding random controls. Care was taken to ensure that the randomized control samples had the same size and distribution as the test samples. The sum of randomized fractions was kept equal to 1, as in the test samples. There were no correlations between the corresponding nucleotides in the control samples. This simple but highly significant and species-independent positional relationship between NUFs provides fur-ther strong support for the view that the genetic code is the result of development and not at all a "frozen accident".

Correlation between individual codons
The detection of a strong internal pangenomic relationship among codons in the CUF Tables and the positional correlation among the base residues of these codons led to an even deeper correlation analysis. The correlations between every single codon frequency and every other codon frequency (64 × 64/2 = 2,048) were calculated using linear regression analysis [Additional file 5]. Accuracy of codon predictions I used the strongest correlations [Supplementary File 6] to predict codon frequencies, and the mean of several predictions was used as the averaged predicted value (p). Four different approaches were used to evaluate the predictions quantitatively.
The correlation between real (r) and predicted (p) values belonging to the same codons was significant (p < 0.05) in 54 cases but not the other 10 ( Figure 9a).
The correlation between real (r) and predicted (p) values belonging to the same species was significant (p < 0.05) in all 113 cases and The p value was below 10E-07 in all but 2 species (Figure 9b).
The average accuracy of individual CUF predictions in 113 species and 87 individual proteins was estimated by com- paring the average real and predicted frequencies. The significance of the correlation between real and predicted CUF was 1.3E-64 when data from 113 species were averaged and compared (n = 64) and 1.9E-28 when data derived from 87 individual proteins (n = 64) were used ( Figure 10).

Discussion
There are basically two approaches to measuring CUB. First, relative synonymous codon usage (RSCU) values can be calculated [5]. RSCU is the observed number of codon occurrences divided by the number expected if synonymous codons were used uniformly. Second, the relative merits of different codons can be assessed from the viewpoint of translational efficiency. This second approach led to the development of the Codon Adaptation Index (CAI, [6]). The CAI model assigns a parameter, termed 'relative adaptiveness', to each of the 61 codons (stop codons excluded). The relative adaptiveness of a codon is defined as its frequency relative to the most often-used synonymous codons and is computed from a set of highly expressed genes. The CAI is widely used even though the subjectivity involved in selecting the reference codons is well recognized [26,27].
My way of calculating CUB is very close to the original suggestion [5] and regards uniform codon usage as the "null hypothesis"; any deviation from this is the bias. This approach made it possible to avoid subjectivity and species limitations in choosing the reference set of codons, and I can build the concept of CUB on the massive foundation of statistical laws and the large collection of sequence data collected in Codon Usage Frequency Tables.
The origin and biological significance of CUB is not well understood, therefore I tried to find the rules (if any) of its evolutionary development and gain new insights about its possible function. I sort my findings into two main categories: I found a.) some (few) signs of the evolutionary origin and development of CUB; b.) unexpectedly large number of highly significant intern correlations between different codon residues (bases) at different codon positions (first, central, wobble) as well as between individual codons.
Inter-species variation in CUB is about 10%, but it is obvious that prokaryotes have significantly larger CUBs than eukaryotes. Bacteria may show the greatest bias because these primitive organisms are rich in highly-expressed genes and often use only one dominant codon. CUB decreases progressively with evolution and humans have the lowest bias (only about 20%). Evolutionary increase in codon number and genome complexity seems to reduce the CUB. It is noticeable that the average CUB (29.3 ± 1.1% (S.E.M.) n = 113) means that synonymous codon usage frequencies are 29.3% distant from the "all codons are equally good" hypothesis, and 70.7% distant from the "one codon is the best 'codon" alternative.
A more detailed qualitative analyzes of CUB is possible using a pan-genomic CUF Table. The original purpose of this virtual table was to create a reference for comparison of CUBs, but it turned out to reveal other codon-related connections too. The pan-genomic CUF Table is based on only 113 species, so it might be the first but not the last of its kind. It makes it possible to detect major, universal trends in codon usage behind small individual (or even species-wide) variations.
CUB is often correlated to the intensity of translation and has even been used to predict highly-expressed genes [6]. It is also known to be related to tRNA copy number, and co-evolution of tRNA gene composition and codon usage bias in genomes has been suggested [28]. I found a very strong correlation between the number of synonymous  codons and the frequency of the amino acids they encoded, as well as the CUB. More synonymous codons encode more amino acids of the same kind and cause greater bias. This (rather logical) connection is not described in the literature, probably because the definition of CUB is very different from mine.
I tried to define a kind of "phylogenetic tree" of CUBs using the pan-genomic CUF table as reference. The significance of correlations between species-specific CUF and pan-genomic CUF gave a qualitative, theoretical measure of distances between codon usages. However this correlation-based approach did not successfully detect any recognizable, species-related evolutionary pattern.
Estimation of codon commitments through evolution showed that some codons are clearly over-utilized while other are avoided in most species. This finding is compatible with the concept of dominant and suppressed codons, but without stating that this difference is the result of evolution [29].  The non-randomness of synonymous codon usage is widely accepted today, and it has been suggested that independent forces (such as tRNA pool size [30]) have a role in the reading frame and there are contextual constraints on synonymous codon choice [31][32][33].
Other lines of evidence suggest that the Genetic Code itself (the 64 codons in toto as a system) has an inherited, internal structure [34,35]. Statistical studies on the nucleotide compositions of codons and of different codon positions support this concept [36][37][38][39][40][41].
I searched for the origin end development of codon bias and I found an extensive network of internal correlations between codons of a species and the nucleotides that define them. The correlations described in this article are: -Correlation between the frequency of any single codon residue (base) at any codon position (first, central, wobble) and the frequency of any other single codon residue (base) at any other codon position (also first, central, wobble); -Correlation between the sum of frequencies of any two codon residues (bases) at any two codon positions ((first, central, wobble) and the sum of any two other codon res-idues (bases) at any two other codon positions (also first, central, wobble); -Correlation between A+T, G+C frequencies at the 1 st , 2 nd codon positions and A+T, G+C frequencies at the 3 rd codon position; -Correlations between any two codons.
There seems to be a simple rule behind all these statistically significant correlations: the correlation between any two nucleotides at any two codon positions is positive if the two nucleotides are complementary to each other and negative if they are not (illustrated in Figure 8).
The large number of statistically highly significant correlations made it possible to predict the frequencies of synonymous codons (in 113 species and 87 individual proteins) from the general overall frequencies of codons. The reliability of predictions was tested.

Conclusion
The cumulative Codon Usage Frequency of any codon is strongly dependent on the cumulative Codon Usage Frequency of other codons belonging to the same species. The rules of this codon dependency are the same for all species and reflect WC base pair complementarity. This internal connectivity of codons indicates that all synonymous codons are integrated parts of the Genetic Code