Codon usage bias in prokaryotic pyrimidine-ending codons is associated with the degeneracy of the encoded amino acids

Abstract

Synonymous codons are unevenly distributed among genes, a phenomenon termed codon usage bias. Understanding the patterns of codon bias and the forces shaping them is a major step towards elucidating the adaptive advantage codon choice can confer at the level of individual genes and organisms. Here, we perform a large-scale analysis to assess codon usage bias pattern of pyrimidine-ending codons in highly expressed genes in prokaryotes. We find a bias pattern linked to the degeneracy of the encoded amino acid. Specifically, we show that codon-pairs that encode two- and three-fold degenerate amino acids are biased towards the C-ending codon while codons encoding four-fold degenerate amino acids are biased towards the U-ending codon. This codon usage pattern is widespread in prokaryotes, and its strength is correlated with translational selection both within and between organisms. We show that this bias is associated with an improved correspondence with the tRNA pool, avoidance of mis-incorporation errors during translation and moderate stability of codon–anticodon interaction, all consistent with more efficient translation.

INTRODUCTION

The degeneracy of the genetic code implies that most amino acids are encoded by a family of synonymous codons, usually with the first two nucleotides fixed and the third one varied. While synonymous codons encode the same amino acid, their relative usage varies dramatically between different species and between genes of the same species. The deviation from uniform codon distribution is referred to as codon usage bias (1,2). Many theories have been proposed to explain the forces shaping codon usage bias (3–7). For example, mutational biases influence the genome base composition, which in turn affects codon usage considerably (8). Asymmetry in DNA replication and repair of the leading and lagging DNA strands (9) creates further codon usage bias. Furthermore, it has been shown that certain types of codon usage bias are associated with gene expression levels (10–17). This association is often explained through selection towards translational efficiency (18–22), which regards optimization of both translation rate (23) and fidelity (24,25). Selection towards codon usage that promotes efficient translation has both local effects on specific genes (22,26) as well as global effects on the organism’s fitness. The latter is achieved by promoting rapid recycling of ribosomes, reducing costs wasted on correcting translation errors (27) and lowering the production of inactive and at times toxic proteins (28).

While the concept of selecting codons to enhance translation efficiency is easily comprehended, the underlying principles that determine the preferred codons are still under debate. Previous studies identified several codon properties as associated with expression-dependent selection: (i) robustness—preference of certain codons that reduce the impact of mutations and errors in translation (29,30), (ii) compatibility with the cellular tRNA pool—preference of codons matching the more abundant tRNA iso-acceptor, which will be translated faster and with fewer errors than alternative synonymous codons (10,15–17,19,21,23,31–37). Furthermore, it has been suggested that the tRNA pool itself is regulated to optimize gene expression under different growth conditions (38,39). However, differences in tRNA abundance cannot explain codon usage bias for synonymous codons that are translated by the same tRNA. This is often the case in prokaryotes, where due to the striking absence of tRNAs containing adenine (A) at the wobble position (40), synonymous codons ending with a pyrimidine (Y) residue [uracil (U) or cytosine (C)] are translated by a single tRNA containing guanine (G) at the wobble position. We took advantage of this phenomenon to explore the pattern of codon usage bias that seems independent of tRNA abundance, focusing on pyrimidine-ending codons (hereinafter N₁N₂Y₃). Intriguingly, we have observed a clear association between the number of synonymous codons encoding an amino acid (degeneracy level) and the preference for U or C at the third codon position, which is emphasized in organisms where translational selection is operational in highly expressed genes. Specifically, we find that U is preferred over C at position Y₃ in amino acids encoded by four codons, while C is preferred over U in amino acids encoded by two and three codons. We analyze and discuss these observations in light of existing and new theories regarding properties of codon–anticodon interaction.

MATERIALS AND METHODS

Genome data

We retrieved the genomic sequence and genome-related data of 1346 prokaryotes from the NCBI FTP site (ftp://ftp.ncbi.nih.gov/genomes/Bacteria, April 2011) including 1245 bacteria and 101 archaea. To reduce bias caused by inclusion of closely related organisms we randomly selected one organism per genus as representative. This resulted in a data set of 481 prokaryotes including 424 bacteria and 57 archaea (Supplementary Data set S1).

Codon-pair bias calculation

In each organism, for each of the codon-pairs of the N₁N₂Y₃ type, we compared the counts of the two synonymous codons (ending with C and U) in the ribosomal genes to their expected counts based on the nucleotide distribution in the third codon position of the same genes (Supplementary Data set S1). The comparison was done using χ² test with Yates correction. False discovery rate (FDR) correction was applied to all the tests of a specific organism.

Clustering procedure

χ² test results were transformed into a numeric representation. Statistically significant preferences for C at Y₃ (C-Bias) or U at Y₃ (U-Bias) were assigned the values 1 and −1, respectively. Otherwise, for insignificant results or insufficiency of data to employ the test (N-Bias) a value of 0 was assigned. Results were arranged in a matrix where columns and rows indicated the codon-pairs and organisms, respectively. Hierarchical clustering was performed using a hamming distance metric.

Computation of organism genome bias—ENC′_diff

ENC′ is a variant of the effective number of codons (ENC) index (41), accounting for background nucleotide composition (42). It takes the value of 61 when all codons are used at the frequency expected given the nucleotide composition, and its value decreases as codon usage deviates from the expected. The nucleotide composition of a gene was used as background. To obtain an estimate of translational selection at an organism scale, we computed ENC′_diff for each organism according to Equation (1).

(1)

ENC′_diff is the difference between the averages of ENC′ of all genes and the genes encoding ribosomal proteins, normalized by the average ENC′ of all genes.

Fold rule scores

We defined two scores of codon bias; one at an organism scale and one at a gene scale. The scores take into account the type of bias found in N₁N₂Y₃ codon-pairs and the degeneracy level of the amino acid the pair encodes (2-/3-fold in cases of amino acids encoded by pairs or a triplet of codons and 4-fold in cases of amino acids encoded by quartets of codons; codons for amino acids encoded by six codons were divided to a pair and a quartet of codons).

Organism fold rule score (OFRS)

A score describing the conformity of an organism to the fold rule in genes encoding ribosomal proteins.

(2)

where CP_i is the i-th out of 16 N₁N₂Y₃ codon-pairs, 4-fold and 2-/3-fold indicate the level of degeneracy, and χ²(CP_i) indicates the type of bias observed in a χ² test.

Gene fold rule score (GFRS)

A score describing the conformity of a gene to the fold rule.

(3)

where N_codons is the number of codons in the gene, C_i is the i-th codon, 4-fold and 2-/3-fold indicate the level of degeneracy, and N₃ is the nucleotide found at position N₃ of the codon.

Gene expression data

Normalized mRNA expression profiles of nine organisms were extracted from the GEO database as detailed in the ‘Supplementary Methods’.

tRNA repertoire

We scanned the sequences of all non-coding RNAs identified in each organism using the program tRNAscan-SE (43). This comprised the repertoire of identified tRNAs and their respective anticodons.

RESULTS AND DISCUSSION

Association between codon usage bias and amino acid degeneracy

We compiled genome sequence data of 481 prokaryotes, one representative per genus (Supplementary Data set S1), to be included in the analysis. For each organism, we used genes encoding ribosomal proteins as proxy for highly expressed genes. We employed χ² test to ask whether the 16 codon-pairs of the type N₁N₂Y₃ exhibit codon preferences beyond those expected from the nucleotide content in the third codon positions of these genes (Supplementary Data set S1). A statistically insignificant result was designated N-Bias whereas a statistically significant result was designated C-Bias or U-Bias depending on bias direction.

To systematically analyze the patterns of bias, we transformed the results into a numeric representation (see ‘Materials and Methods’ section) and clustered the resulting matrix in both dimensions (Figure 1). Remarkably, the clustering clearly distinguishes between two groups of codon-pairs based on the level of degeneracy of the amino acid they encode. The codon-pair group that shows distinct C-Bias (marked by the cyan bar) contains all N₁N₂Y₃ codon-pairs that encode two-fold degenerate amino acids, the codon-pair AUY that encodes the three-fold degenerate Ile and the two-fold component of the six-fold Ser. We refer to this group as the 2-/3-fold group. The second group, which shows distinct U-Bias (marked by the yellow bar), includes all the N₁N₂Y₃ codon-pairs that encode the four-fold degenerate amino acids, both those that are strictly four-fold and the four-fold component of six-fold degenerate amino acids (Ser, Leu, Arg). We refer to this group as the 4-fold group. These results are consistent with and lend statistical support to similar observations by Ran and Higgs (44,45) in a smaller set of organisms. We define a fold rule, by which the preferred codon in codon-pairs belonging to the 2-/3-fold group is N₁N₂C₃, whereas the preferred codon in codon-pairs belonging to the 4-fold group is N₁N₂U₃.

Figure 1.

An association between the degeneracy level of amino acids and codon bias pattern in N₁N₂Y₃ codon-pairs. The matrix of codon-pair biases in multiple organisms (Green: U-Bias, magenta: C-Bias, black: N-Bias) was clustered using a hamming distance metric. The clustering reveals a clear distinction between codon-pairs belonging to codon families with four-fold degeneracy (yellow bar) and those belonging to two- and three-fold degenerate families (cyan bar), where the former are statistically significantly biased towards the U-ending codon and the latter are statistically significantly biased towards the C-ending codon. This distinction is most pronounced in a subset of organisms (blue bar).

The fold rule is associated with translational selection

Approximately 25% of the organisms included in the analysis (marked by the blue bar in Figure 1) show a very clear demarcation between the two groups of codons. Of note, these organisms represent multiple clades throughout the taxonomic tree (Supplementary Figure S1). If this widespread phenomenon represents variance in translation efficiency of synonymous codons, we would expect to see an association between the level of translational selection and the extent of codon bias following the fold rule. To this end, we computed for each organism a measure of agreement with the fold rule and a measure of translational selection. The organism fold rule score (OFRS) quantifies the adherence of an organism to the fold rule (see ‘Materials and Methods’ section). This measure is the difference between the numbers of codon-pairs demonstrating statistically significant bias conforming to the fold rule and opposing it. To quantify the level of translational selection of an organism, we employed a measure similar to the one proposed by Rocha (5). We calculated ENC′_diff, the normalized difference between the average ENC′ of all genes and the average ENC′ of ribosomal genes, where ENC′ is a measure of codon bias of a gene (see ‘Materials and Methods’ section). The stronger the selection towards efficient translation in highly expressed genes, the larger the difference between codon usage in ribosomal genes and all other genes, resulting in higher ENC′_diff. The values of ENC′_diff in our data set range from −0.084 in the archaeon Thermofilum pendens to 0.274 in the bacterium Leuconostoc mesenteroides. In 460 out of 481 organisms, the value of ENC′_diff is positive, indicating that genes encoding ribosomal proteins use a less random set of codons as suggested by Rocha (5). We find a correlation between OFRS and ENC′_diff (r = 0.582, P = 6.4 × 10⁻⁴⁵; Figure 2A), suggesting that this fold-dependent codon bias pattern is associated with translational selection. To further establish this conclusion, for each codon-pair, we divided the organisms into three groups according to the type of the χ² result (C-Bias, U-Bias and N-Bias). We compared the ENC′_diff distribution between the C-Bias and U-Bias organism groups using the Mann–Whitney test (Supplementary Figure S2). In 2-/3-fold codon-pairs the C-Bias group has higher ENC′_diff values than the U-Bias group. The only exception is the Cys-UGY codon-pair in which the U-Bias group is too small to produce a statistically significant result. In this case, however, it is possible to see that the C-Bias group has higher ENC′_diff values than the N-Bias group (P = 9.8 × 10⁻⁴). For most 4-fold codon-pairs we see the opposite phenomenon, where the U-Bias group has higher ENC′_diff values compared to the C-Bias group. One exception is the Val codon-pair GUY, which shows the same trend although statistically insignificantly. Still, comparison between the U- and N-Bias groups shows higher ENC′_diff values in the U-Bias group (P = 4.1 × 10⁻⁶). The other exception is the Leu codon-pair CUY in which the U-Bias and C-Bias groups have similar ENC′_diff distributions. These results suggest that codon bias in most N₁N₂Y₃ codon-pairs is associated with selection towards translational efficiency.

Figure 2.

Organisms conforming to the fold rule demonstrate strong translational selection and fast growth. OFRS and ENC′_diff were calculated as described in ‘Materials and Methods’ section and minimal generation times were taken from (46). (A) Association between OFRS and ENC′_diff was evaluated by computing Pearson correlation coefficient. (B) Association between OFRS and minimal generation time was evaluated by computing Spearman correlation coefficient (since the generation time variable is not normally distributed). Organisms most compatible with the fold rule (marked by the blue bar in Figure 1) are colored dark gray and have relatively high ENC′_diff and low minimal generation time values.

It has been suggested that selection for fast growth affects codon bias by promoting the use of translationally efficient codons (46,47). We evaluated the association between minimal generation-times taken from Vieira-Silva and Rocha (46) and OFRS values. Figure 2B demonstrates a strong correlation (r_s = −0.654, P = 4.3 × 10⁻²⁷) as would be expected if the preferred codons according to the fold rule indeed promote translational efficiency. Notably, for 17 of the organisms marked by the blue bar in Figure 1, there are generation-time values, and 14 of them show minimal generation-time of 2.5 h or less.

Highly expressed genes demonstrate fold-dependent codon bias

We showed an association between fold-related codon bias in genes encoding ribosomal proteins and translational selection strength at the organism level. Since the strength of translational selection is likely to vary between genes of the same organism depending on their expression level, we examined the relationship between gene codon preference and expression level. To this end, we computed a gene fold rule score (GFRS) that represents the general tendency of a gene to use N₁N₂C₃ over N₁N₂U₃ codons for 2-/3-fold codon-pairs and vice versa for 4-fold codon-pairs (see ‘Materials and Methods’ section). We computed the correlation between GFRS values and gene expression levels obtained from microarray results in nine prokaryotes, representing several major taxonomic clades and a range of ENC′_diff values (Supplementary Table S1). Since GFRS is normalized by the total number of codons in the open reading frame (ORF), there might be bias towards high scores in short ORFs. This may bias the correlation, especially given the known inverse correlation between expression and ORF length (48). We therefore performed a partial Pearson correlation while controlling for ORF length. Four of the five organisms that have high ENC′_diff (Shewanella oneidensis MR-1, Streptococcus mutans UA159, Escherichia coli K12 and Haemophilus influenzae Rd KW20) show a marked correlation between GFRS and gene expression in the 50% of genes with the highest expression levels but not in the 50% of genes with low expression levels, suggesting that translational selection is most prominent in genes expressed above a certain expression level threshold (Supplementary Figure S3). Bacillus subtilis 168, despite its high ENC′_diff value, shows only a weak correlation between the GFRS and gene expression. It is, however, evident that B. subtilis genes with high GFRS other than those encoding ribosomal proteins (marked red), tend to be highly expressed. Organisms with low ENC′_diff values (Rhodobacter sphaeroides 2.4.1, Pseudomonas aeruginosa PAO1, Thermus thermophilus HB8 and Nitrosomonas europaea ATCC19718) exhibited weaker correlations between GFRS and expression and less pronounced differences between weakly and highly expressed genes. These latter results are consistent with our hypothesis, as in the absence of translational selection, we would not expect a correlation between translation-related codon property and gene expression.

Possible mechanisms underlying fold-dependent codon bias

We have observed that highly expressed genes show a clear bias in N₁N₂Y₃ codons that is associated with the degeneracy level of the encoded amino acid. Remarkably, these tendencies are mostly observed in organisms with high translational selection and seem to be associated with more efficient translation. In the next sections, we analyze several possible mechanisms in an attempt to understand the forces shaping this pattern of codon bias.

tRNA availability

We stated above that N₁N₂Y₃ codon-pairs are translated by a single tRNA with G at the wobble position. However, over the years the wobble rules have been modified and expanded (49). It is now acknowledged that an unmodified U in the first position of the tRNA anticodon (U₃₄) is capable of reading N₁N₂U₃ codons and to a lesser extent N₁N₂C₃ codons, in addition to reading the N₁N₂A₃ (Watson–Crick base pairing) and N₁N₂G₃ (conventional wobble) codons. Certain U₃₄ modifications abolish the ability to read the N₁N₂C₃ codon whereas others limit the wobble even further to only purine-ending codons (N₁N₂R₃). While it is suggested that the price for such ‘superwobbling’ is reduced translational efficiency (50), in 4-fold codon-pairs it might tilt the balance towards U-Bias due to an increase in the available tRNA pool. We would expect C-Bias in codon-pairs of 4-fold amino acids only when the pool of G₃₄N₃₅N₃₆ tRNA fitting both codons is larger than the pool of U₃₄N₃₅N₃₆ tRNA, thereby minimizing the effect of tRNA availability. To this end, we analyzed the effect of the two tRNA species on codon bias across the organisms in our study. Since the actual tRNA pool is not available for most organisms, we used the number of tRNA genes as a proxy for tRNA levels. Figure 3 demonstrates that as the U₃₄/G₃₄ tRNA ratio increases, the bias leans in the direction of the N₁N₂U₃ codon. Only in relatively low ratios, do we see C-Bias in 4-fold codon-pairs. This is most pronounced in Pro-CCY, Gly-GGY and Ser-UCY, where all organisms with U₃₄/G₃₄ tRNA ratio above 1 are U-biased. A similar trend is observed in Leu-CUY, Val-GUY, Thr-ACY and Ala-GCY. Since for Arg-CGY the major tRNA is A₃₄C₃₅G₃₆ while U₃₄C₃₅G₃₆ is very rarely used (mostly in archaea), it is impossible to determine how the U₃₄/A(I)₃₄ tRNA ratio will affect its codon bias. Ran and Higgs (45) also showed U-preference but claimed that it stems from a surprisingly well-translated non-standard U₃:U₃₄ base pairing. Since U₃₄ tRNA simultaneously increases the tRNA pool of N₁N₂U₃ codons and hypothetically improves their translation, we were unable to evaluate the selective significance of each factor.

Figure 3.

U-Bias in 4-fold codon-pairs is associated with the tRNA repertoire. Circles represent organisms with the indicated U₃₄N₃₅N₃₆ (X-axis) and G₃₄N₃₅N₃₆ (Y-axis) tRNA gene copy numbers. The only exception is the Arg codon-pair CGY for which the common tRNA is A₃₄C₃₅G₃₆ (where A is modified to I). Circle size represents the number of organisms with the specific tRNA gene counts. Circle color represents the direction of bias in the codon-pair (Green: U-Bias; magenta: C-Bias). Circles along the diagonal indicate equal number of gene copies of the two tRNAs. Most organisms with C-Bias have a low U₃₄/G₃₄ ratio.

Ran and Higgs (44,45) further suggested that the C-preference observed in the 2-fold group results from a more rapid processing of the strongly interacting Watson–Crick C:G base pair by the ribosome than of the weakly interacting wobble G:U base pair. We hypothesized that the U₃₄N₃₅N₃₆ tRNA may have an indirect contribution to this bias. Translation of N₁N₂Y₃ codons by U₃₄N₃₅N₃₆ tRNA would cause a major disruption of the genetic code in the 2-fold group. However, it is conceivable that the mechanisms restricting the specificity of the U₃₄N₃₅N₃₆ tRNA to N₁N₂R₃ codons are not absolute, leading to a certain rate of amino acid mis-incorporation. If such mistakes occur, we would expect selection to favor the use of N₁N₂C₃ that is less likely to be misread by the U₃₄N₃₅N₃₆ tRNA, as is indeed the case. Another expectation arising from such a mechanism is stronger selection when substitution of the correct amino acid with its N₁N₂R₃ encoded neighbor is not well tolerated. Although it is known that two amino acids that share the first two nucleotides in their codons tend to have similar properties (51,52), some neighboring amino acids are more likely to be replaced during evolution than others, indicating that putative mistranslations between them are better tolerated. Table 1 shows the substitution scores given by a range of blocks substitution matrices (BLOSUM) (53), which can be used as an indication of how conservative a substitution is. We used BLOSUM70–BLOSUM90 matrices in order to focus on substitutions likely to happen in relatively similar proteins without changing or impairing their function. The most extreme cases are the Ser/Arg pair, which is substituted less than expected in all the BLOSUM matrices checked, and the Asp/Glu pair, which is substituted more than expected. As anticipated, the Ser 2-fold codon-pair shows C-Bias in 199 organisms and U-Bias in only one organism (C/U ratio 199), whereas the Asp codon-pair has almost the same number of organisms demonstrating C- and U-Bias (69 and 62, respectively, C/U ratio 1.1). The His/Gln pair, which is also substituted more than expected, has a relatively low C/U ratio for His (4.2), whereas the Phe/Leu and Asn/Lys pairs which are relatively neutral have a high C/U ratio for Phe and Asn (52.5 and 27, respectively), although not as high as Ser. These results lend further support to the possible role of error minimization in shaping codon bias in 2-/3-fold codon-pairs. Note that Tyr, Cys and Ile were not included in this analysis since there are no U₃₄N₃₅N₃₆ tRNAs that can misread them.

Table 1.

Propensity towards C-Bias depends on the similarity between amino acids sharing the codon quartet

	Ser(AGY)–Arg(AGR)	Phe(UUY)–Leu(UUR)	Asn(AAY)–Lys(AAR)	His(CAY)–Gln(CAR)	Asp(GAY)–Glu(GAR)
BLOSUM substitution valuesa	−1:−2	0	0	1	1:2
C-Biasb	199	210	189	140	69
U-Biasb	1	4	7	33	62
N-Biasb	41	27	45	68	110
C/U Ratio	199	52.5	27	4.2	1.1

^aRange is the minimal and maximal substitution values observed in BLOSUM substitution matrices 70, 75, 80, 85 and 90

^bCalculated from organisms with high ENC′_diff values (upper 50%) for Ser, Phe, Asn, His and Asp.

Nucleotide bias

In a previous study of codon bias in the prokaryotic world, Hershberg and Petrov (54) showed that codons are biased beyond, but in the same direction of the general nucleotide content of the organism. Although there is no correlation between organism GC content and ENC′_diff (r = 1.3 × 10⁻⁴, P = 1), we wished to explore the possibility that nucleotide content has an effect on the observed codon bias. When examining ENC′_diff versus GC content, no consistent effect of GC content on the bias pattern could be observed (Supplementary Figure S4). However, when concentrating on organisms with high ENC′_diff values (above 0.1), where translational selection is strong, it is possible to see a complicated pattern of GC content influence (Supplementary Figure S5 and discussion therein). We hypothesized that since GC content is mostly reflected in the N₃ codon position (55), it might also be associated with the number of tRNAs that read the resulting codons. While GC content is not statistically significantly associated with the number of G₃₄N₃₅N₃₆ tRNA genes (r = 0.07, P = 0.1), it is associated with U₃₄N₃₅N₃₆ and C₃₄N₃₅N₃₆ tRNA genes (r = −0.26, P = 5.8 × 10⁻⁹ and r = 0.63, P = 2.4 × 10⁻⁵⁵, respectively). When evaluating how the GC content affects the ratio of U₃₄/G₃₄ (Supplementary Figure S6), we see that this ratio tends to be high (above 1) in organisms with low GC content and low (below 1) in organisms with high GC content. We have shown before that the U₃₄/G₃₄ ratio affects bias direction in 4-fold codon-pairs. We can therefore surmise that in organisms with high GC content, the tRNA pool available to the N₁N₂U₃ codon is not significantly larger than that of the N₁N₂C₃ codon and therefore, we see less cases of U-Bias. In 2-fold codon-pairs, the same reasoning may apply since a low U₃₄/G₃₄ ratio indicates a low probability of mistranslation and therefore reduced selection towards C-Bias. Interestingly, of the three codon-pairs in which there is no risk of mistranslation by the tRNA of the neighboring codon-pair (Cys-UGY, Ile-AUY and Tyr-UAY), only Cys demonstrates a small association with GC content. It therefore appears that the effect of GC content on the tRNA pool available for each codon is responsible for at least some of the observed bias.

Optimal stability of codon–anticodon interaction

Grosjean et al. (56–58) noticed that N₁N₂Y₃ codon-pairs beginning with G or C in the first two positions were biased towards the N₁N₂U₃ codon, while those beginning with A or U in the first two positions were biased towards the N₁N₂C₃ codon. They noted that since A and U are weak (W) binders that form only two hydrogen bonds each, while C and G are strong (S) binders that can form three hydrogen bonds each, the bias yields codon–anticodon binding of intermediate strength, suggesting that too strong and too weak codon–anticodon binding is selected against, probably because it hampers translation efficiency. Grosjean’s optimal stability model was further supported by Gouy and Gautier (16), Sharp et al. (10) and Percudani and Ottonello (59). Our results regarding codon-pairs of the type WWY and SSY are in complete accordance with this theory, as all of the WWY pairs are of the 2-/3-fold C-Bias type and all the SSY pairs are of the 4-fold U-Bias type. Since the original optimal stability model does not address codon bias in SWY and WSY codon-pairs and yet they appear to be biased in the same direction as their fold counterparts, we hypothesized that there might be a more general stability model linking fold and bias for all N₁N₂Y₃ codons.

In 1978, Lagerkvist noticed that the genetic code is arranged such that splitting of a codon quartet into more than one amino acid (or stop signal) happens only when the first two nucleotides common to the quartet (N₁ and N₂) are both weak or if one is weak and the other is strong and N₂ is a purine (60). He suggested that in these cases the codon–anticodon interaction is weak and requires stabilization by base pairing in the third position through Watson–Crick or wobble interaction. As a result, at least one tRNA is needed to efficiently read each codon-pair (the G₃₄N₃₅N₃₆ tRNA for the N₁N₂Y₃ codons and the U₃₄N₃₅N₃₆ tRNA for the N₁N₂R₃ codons), and therefore two amino acids can share the quartet with little risk of codon mistranslation. In the other codon quartets, the first two nucleotides bind the anticodon with enough strength to compensate for a mismatch in the third position. In this case, since a single tRNA can read all four codons, the quartet must encode a single amino acid to avoid mistranslation. Lehmann and Libchaber (61) provided further support to this theory by showing, based on structural considerations, the effect of N₂ on the stability of the interaction. A Y₂ in the codon implies an R₃₅ in the anticodon, which forms an anticodon-loop-stabilizing hydrogen-bond with the U₃₃ tRNA nucleotide to allow efficient binding (62). They assigned each codon quartet a score based on the interaction of N₁–N₃₆ and N₂–N₃₅ where an A:U base pair contributes two hydrogen bonds, a C:G base pair contributes three hydrogen bonds, and a Y₂ (reflecting an R₃₅) contributes one hydrogen bond (Figure 4). We speculated that by combining their score with a measure of the stability of the N₃–N₃₄ base pair, we might arrive at a more accurate representation of codon–anticodon stability, which should provide a more general insight into codon usage bias. Hence, to the original score, we added the number of hydrogen bonds formed in the wobble position between the codon and G₃₄N₃₅N₃₆ tRNA (the major tRNA responsible for translating N₁N₂Y₃ codons). In the case of the Arg C₁G₂Y₃ codon-pair, which is read in most organisms by the I₃₄C₃₅G₃₆ tRNA, we added two hydrogen bonds to both codons. However, since the two hydrogen bonds stabilizing the I:C base pair are stronger than those stabilizing the I:U base pair (63), we considered the total score of the CGC codon as higher than the total score of CGU. The extended stability scores (ESS) are summarized in Table 2. Remarkably, in organisms under translational selection, codon bias follows the optimal stability rule. In the 2-/3-fold group where ESS is low (6 or 7 for N₁N₂U₃ and 7 or 8 for N₁N₂C₃) the ‘not too weak’ N₁N₂C₃ is preferred, while in the 4-fold group where ESS is high (8 or 9 for N₁N₂U₃ and 9 or 10 for N₁N₂C₃), the ‘not too strong’ N₁N₂U₃ is preferred (except for Leu). This notable choice of moderate stability appears to support the superiority of such an interaction over both weaker and stronger interactions.

Figure 4.

Hydrogen bonds stabilizing codon–anticodon interactions [an extension of the model suggested in (61)]. In black are the standard Watson–Crick interactions formed between N₁:N₃₆ and N₂:N₃₅, representing two or three hydrogen bonds each. In light gray is the loop-stabilizing bond between U₃₃ and N₃₅, which occurs only when N₃₅ is a purine. In dark gray is the interaction formed between N₃:N₃₄, which represents two hydrogen bonds for U₃:G₃₄ and three hydrogen bonds for C₃:G₃₄.

Table 2.

Extended stability score of N₁N₂Y₃ codon-pairs

Amino acid	Codon pair	Fold	N1:N36 H-bonds	N2:N35 H-bonds	N2 Y/R	C3:G34 H-bonds	U3:G34 H-bonds	N1N2C3 Extended stability score	N1N2U3 Extended stability score
Asn	AAY	2	2	2	0	3	2	7a	6
Tyr	UAY	2	2	2	0	3	2	7a	6
Ser	AGY	2	2	3	0	3	2	8a	7
Ile	AUY	3	2	2	1	3	2	8a	7
His	CAY	2	3	2	0	3	2	8a	7
Asp	GAY	2	3	2	0	3	2	8a	7
Cys	UGY	2	2	3	0	3	2	8a	7
Phe	UUY	2	2	2	1	3	2	8a	7
Arg	CGYb	4	3	3	0	2(+)	2(−)	8(+)	8(−)a
Thr	ACY	4	2	3	1	3	2	9	8a
Leu	CUY	4	3	2	1	3	2	9	8
Gly	GGY	4	3	3	0	3	2	9	8a
Val	GUY	4	3	2	1	3	2	9	8a
Ser	UCY	4	2	3	1	3	2	9	8a
Pro	CCY	4	3	3	1	3	2	10	9a
Ala	GCY	4	3	3	1	3	2	10	9a

^aMarks the extended stability score of the preferred codon according to the observed fold rule.

^bThe Arg-CGY codon-pair is read by the ACG (modified to ICG) tRNA. Since bonds formed within an I:C base pair are stronger than within an I:U pair, the relevant hydrogen bonds and the extended stability scores are marked with (+) and (−), respectively.

CONCLUDING REMARKS

It is widely accepted that in organisms undergoing translational selection the preferred codons used in highly expressed genes are those that are associated with more efficient translation and hence promote organism fitness. What determines the identity of preferred codons of an organism? Our study identifies a distinctive codon preference pattern in N₁N₂Y₃ codon-pairs: preference of C at position N₃ of 2-/3-fold codon-pairs and preference of U at position N₃ of 4-fold codon-pairs. Of the factors examined in our study as possibly playing a role in this type of codon bias, the stability of codon–anticodon interaction was found to be associated with both C-Bias and U-Bias, while other codon properties were associated with only one type of bias (Table 3). Thus, the stability of the interaction in combination with other codon–anticodon properties seems to underlie the observed codon bias. To summarize, 2-/3-fold codon-pairs are biased towards the N₁N₂C₃ codons that form a perfect Watson–Crick pairing with the available G₃₄N₃₅N₃₆ tRNAs, create a moderately stable interaction with the tRNA instead of a weak one and reduce the risk of mistranslation. The preference of N₁N₂U₃ codons in 4-fold codon-pairs is also supported by more than one explanation. First, since N₁N₂U₃ codons exhibit moderate interaction with their G₃₄N₃₅N₃₆ anticodons compared to N₁N₂C₃ codons, they assure a not too sticky and presumably more efficient codon-anticodon interaction. Second, these codons can be read, although less efficiently, by U₃₄N₃₅N₃₆ anticodons and therefore the ratio of U₃₄/G₃₄ tRNAs determines which codon has a larger tRNA pool and consequently affects the direction of bias. Since for most organisms and most codon-pairs this ratio is high, U-Bias is usually promoted. Interestingly, we have observed an influence of the GC content in organisms demonstrating translational selection in certain codon-pairs. We propose that the explanation may be linked in a non-trivial way to the tRNA repertoire that affects the tRNA pool in 4-fold codon-pairs and the chance of mistranslation in 2-fold codon-pairs. However, since this does not explain the C-Bias observed in the 2-/3-fold codon-pairs Cys-UGY, Ile-AUY and Tyr-UAY that are not at risk of mistranslation, it appears that essentially there may be two major forces affecting codon bias. One regards the frequency of the various possibilities of codon–anticodon interaction and the consequences of a correct or incorrect translation stemming from it, and the other regards the stability of this interaction and the way it influences translation efficiency. In order to better understand the very fundamental concept of translation selection that has implications on the evolution of all organisms, it would be necessary to perform carefully designed experiments that will differentiate between the contributions of the various forces to codon bias.

Table 3.

Summary of potential forces affecting codon bias in N₁N₂Y₃ codon-pairs

Amino acid	Codon pair	Preferred codona	Watson–Crickb	GCc	tRNA poold	Error minimizatione	Stabilityf	Extended stabilityg
Asn	AAY	C	+	− −	NR	+	+	+
Tyr	UAY	C	+	− −	NR	NR	+	+
Ser	AGY	C	+	−	NR	+	NR	+
Ile	AUY	C	+	−	NR	NR	+	+
His	CAY	C	+	− −	NR	+	NR	+
Asp	GAY	C	+	− −	NR	+	NR	+
Cys	UGY	C	+	+	NR	NR	NR	+
Phe	UUY	C	+	−	NR	+	+	+
Arg	CGY	U	NR	−	−	NR	+	+
Thr	ACY	U	−	++	+	NR	NR	+
Leu	CUY	C/U	NR	++	+	NR	NR	NR
Gly	GGY	U	−	−	++	NR	+	+
Val	GUY	U	−	+	+	NR	NR	+
Ser	UCY	U	−	++	++	NR	NR	+
Pro	CCY	U	−	+	++	NR	+	+
Ala	GCY	U	−	+	+	NR	+	+

^aCodon preference is determined by bias direction observed in organisms with high ENC’_diff as seen in Supplementary Figure S2.

^bAgreement with a prefect Watson–Crick base pairing. (+) The preferred codon forms a perfect Watson–Crick pairing with the available tRNA, (−) the preferred codon forms a wobble pairing with the tRNA, (NR) both codons form wobble pairing (due to A to I modification) or there is no preferred codon.

^cGC effect on bias direction as seen in Supplementary Figure S5: (−) no effect, (− −) opposite effect, (+) weak effect, (++) strong effect.

^dtRNA pool size effect on bias direction: (−) no effect, (+) weak effect, (++) strong effect, (NR) not relevant.

^eIndication of error minimization: (+) observed, (NR) not relevant.

^fAgreement with Grosjean’s stability model (56): (+) the preferred codon forms moderate interaction with the tRNA compared to its synonym, (NR) not relevant to SWY and WSY codon-pairs.

^gAgreement with the Extended Stability model: (+) the preferred codon forms a moderate interaction with the tRNA compared to its synonym, (NR) there is no distinctly preferred codon.

SUPPLEMENTARY DATA

Supplementary Data are available at NAR Online: Supplementary Table 1, Supplementary Figures 1–6, Supplementary Methods, Supplementary Dataset 1, and Supplementary References [64,65].

FUNDING

The Israel Science Foundation administered by the Israeli Academy for Sciences and Humanities [1171/09 awarded to H.M.] Ministry of Science and Technology Eshkol scholarship awarded (to N.W). Funding for open access charge: The Israel Science Foundation administered by the Israeli Academy for Sciences and Humanities [1171/09 awarded to H.M.].

Conflict of interest statement. None declared.