Plant codon preferences varied considerably between monocots and eudicots. The monocots analyzed exclusively preferred C- and G-ending codons in their highly expressed genes; fifteen of their overrepresented codons were C-ending, while the remaining five were G-ending (Table 2). These patterns were consistent with the codon preferences of all monocot genes [22]. Conversely, eudicots preferred a combination of NNT (six) and NNG (four) codons in their most highly expressed genes, in addition to two NNC codons (Table 2), which also agreed with their overall codon preferences. In all cases where plants preferred an NNG codon, no pyrimidines were possible in the third position (they were two-fold redundant amino acids, or the two-fold redundant portion of six-fold redundant amino acid codons). The RSCU of the highly expressed genes in the monocots and eudicots [22] were strongly correlated for A or T-ending codons (r=0.93), and for G or C-ending codons (r=0.82, Figure 1). However, best-fit lines for the two groups of codons both differ from a line with a slope of 1 through the origin, indicative of divergent codon preferences. Consequently, monocots and eudicots represent distinct translational environments, and should exert dissimilar translational selection pressures on the CUB of their respective viruses.

The potyviruses showed a consistent pattern of elevated adenine in their genomes, regardless of host, and also contained correspondingly lower levels of cytosine and guanine. The third position nucleotide content in CP genes differed significantly from that of the overall genome in every potyvirus we examined (chi-square tests, p<0.05). Luteoviruses showed consistent genomic base composition, having slightly elevated genomic adenine content, and relatively equitable use of cytosine, guanine, and thymine. Third position base frequencies were also consistent regardless of host, but differed significantly from genomic nucleotide composition in most luteoviruses. Third position base usage in two of the four eudicot-infecting luteoviruses did not differ significantly from genomic base content (chi-square tests, p>0.1). In the two remaining eudicot-infecting, and all four monocot-infecting luteoviruses, third positions diverged significantly from the genomes (chi-square tests, p<0.05). These findings indicate that genomic base composition is a poor predictor of CUB in luteoviruses and potyviruses.

Average third position base composition of CP genes in begomoviruses (dicot-infecting geminiviruses) and mastreviruses (monocot-infecting geminiviruses) also varied greatly from their respective genomic nucleotide contents (18 begomoviruses and 4 mastreviruses, chi-square tests, p<0.001). Base composition of synonymous sites in the Rep genes of all begomoviruses (n=14) and two out of three mastreviruses also diverged substantially from overall genomic nucleotide content (chi-square tests, p<0.05). As is the case for the ssRNA viruses, these results strongly suggest that genomic base composition does not drive CUB in geminiviruses.

All potyviruses had somewhat similar codon preferences, independent of host: monocot- and eudicot-infecting potyviruses both generally preferred A- and T-ending codons (Table 3). Luteoviruses, both monocot- and eudicot-infecting, exhibited preferences for fewer codons overall, but tended to favor NNC codons. Despite this overall similarity they shared only two preferred codons (Table 3).

Eudicot-infecting potyvirus RSCU was moderately correlated with eudicot RSCU (r=0.55), and monocot-infecting RSCU was actually weakly anti-correlated with monocot RSCU (r=-0.29, Figure 2a). Surprisingly, monocot-infecting potyvirus RSCU correlated with eudicot RSCU nearly as well as eudicot-infecting potyviruses (r=0.46). Monocot- and eudicot-infecting luteovirus RSCU weakly correlated with that of their hosts (r=0.29 and 0.16, respectively, Figure 2b).

ssDNA dicot-infecting begomovirus CP genes exhibited a strong preference for NNT codons, while begomovirus Rep sequences strongly favored NNA codons (Table 4). In the monocot-infecting mastreviruses, CP genes preferentially used C- and G-ending codons, but the Rep sequences did not exhibit a specific preference; overrepresented codons ended in all four bases (Table 4). Begomovirus genomes are ambisense; genes are encoded in the coding and complimentary sense [31]. The coding sequence of the Rep gene is complimented on the virion strand. As a consequence, third positions in this gene are present as the first base of anti-codons in the single-stranded viral genome. Therefore, these findings indicate begomovirus genomes are enriched for thymine at synonymous sites in both the CP ORF (with T-ending codons) and Rep ORF (with T-beginning anticodons).

In both begomoviruses and mastreviruses, the RSCU of their CP genes was moderately correlated to the RSCU of the highly expressed genes of their respective hosts (r=0.69 for begomoviruses, 0.53 for mastreviruses, Figure 3a). However, the Rep gene was not well correlated to their host RSCU in either begomoviruses (r=0.38) or mastreviruses (r=0.06, Figure 3b). Mastrevirus Rep RSCU actually matched that of eudicots better than the begomovirus Rep RSCU (r=0.71).

The conservation of codon usage between monocot- and eudicot- infecting viruses within each group varied significantly. Potyvirus RSCUs were strongly correlated to each other (r=0.90), despite their hosts having divergent preferences (Figure 4a). Similar to the potyviruses, the two luteovirus groups exhibited a moderate correlation to each other (r=0.66, Figure 4a). Begomovirus and mastrevirus RSCUs were uncorrelated in the CP ORF (r=0.16), but moderately correlated in the Rep ORF (r=0.51, Figure 4b).

The null hypothesis, that synonymous codon usage is purely a function of the nucleotide base frequencies in the viral genome, was insufficient to explain codon preferences in all of the groups we examined. The discord between genomic frequencies and third position frequencies did not often manifest as more equitable nucleotide use in the third position; instead, codon use was more biased than genomic nucleotide frequencies in several cases (eudicot-infecting potyviruses, begomovirus CP). Similarly, the alternative hypothesis of translational selection failed to explain much of the plant virus CUB. Dicot-infecting potyviruses and geminivirus CP CUB were moderately correlated with their host CUB, but we did not find as high a correlation as in phage [21] or human RNA viruses [20].

The most common methods of assessing the relationship between genomic base composition and CUB involve using GC₃, the GC content of the third codon position, as a measure of codon bias. GC₃ is then compared to GC_1,2, the GC content of the first and second codon positions, or ENC, the effective number of codons [32,33]. Both measures vary along predictable lines or curves when base composition drives CUB. By these measures, CUB in many viruses is strongly affected by overall base composition [33,34,35]. However, metrics that group AT and GC are unable to account for the over- or underrepresentation of a specific nucleotide in the third position at the expense of its complement, which is of special concern when analyzing CUB in single-stranded viruses. In the begomoviruses we analyzed, the average GC₃ is approximately equal in the CP, Rep, and whole genome, but the CP (and genome) favor guanine while cytosine is overrepresented in the Rep. Consequently, our methods require a higher degree of agreement between the overall genomic base content and gene third position for base composition to be considered a plausible explanation for CUB.

Furthermore, many studies of viral CUB do not explicitly evaluate other factors that can drive codon bias. Rather, the null hypothesis is tested—does CUB follow the predicted relationship between GC₃ and GC_1,2 or ENC—and if not rejected, the effects of translational selection and other possible factors are not subsequently analyzed [15,16,34,36]. Other studies attribute the rejection to translational selection, but fail to consider additional factors [37,38,39].

Genes of propagative arthropod-vectored viruses (those that replicate within their vectors) should be under dual selective pressures to maximize replication speed within their plant hosts and their vectors. Consequently, vector codon preferences could influence codon bias in these viruses. However, the potyviruses and luteoviruses are nonpropagative [40], and while the evidence is more ambiguous in geminiviruses, they are generally considered nonpropagative as well [41,42], so translational selection is not acting on these viral genomes in their respective vectors. There is evidence of heightened purifying selection on capsid structure in vectored RNA viruses due to specific interactions between CP and vector [26], but in nonpropagative viruses, this pressure is independent of translation kinetics, instead acting solely on amino acid sequence. Therefore, translational selection within arthropod vectors cannot explain the observed CUB.

As potyviruses have high mutation rates [43], have diverged over at least thousands of years [44], and the different species analyzed were at least 25% divergent by nucleotide [23], it is impossible that the common synonymous codon usage we observed is an accident of recent fixation. It is similarly unlikely that a recent host-shift from eudicots (to which potyvirus CUB is better matched) brought recently diverged potyviruses into monocots [44]. Luteoviruses [45] and geminiviruses [46,47,48] evolve at similarly high speeds, so it is unlikely that these correlations are due to accidental historical contingency. It would further be expected that third positions would be saturated after thousands, if not millions of years of divergence [49]. Given these factors, it is most likely that the correlation among monocot- and eudicot-infecting members of each group is due to a similar set of pressures affecting CUB of each group as a whole.

One possible factor that may influence codon bias in ssRNA viruses is selective constraints on secondary structure. ssRNA viral genomes often contain complex secondary structures that are important for replication or gene expression [50]. Disruption of these structures can inhibit one or both processes, reducing viral fitness. Substitutions are often observed in pairs: an initial mutation and a compensatory mutation that restores base pairing across stems in stem-loop structures, for instance [51]. These factors should manifest as more constrained codon usage at specific sites, though the effects on overall codon usage are ambiguous.

The begomovirus CP, which strongly preferred NNT codons, aligned well with the preference of their eudicot hosts for T-ending codons and correlated strongly with host RSCU. Despite the significant differences between third position and genomic base content, these results also indicate the potential importance of the thymine enrichment of high-AT begomovirus genomes. Therefore, it is tempting to explain these data as the result of the combination of compositional constraints and strong translational selection, even if third position base use significantly differed from that of the entire genome. Conversely, begomovirus Rep sequences have different preferences and demand a different explanation. CUB is not explained by base composition, but the prevalence of A-ending codons and the weak correlation (r=0.37, Figure 4b) between host and virus RSCU suggests weak translational selection. It is unlikely that these two genes are subject to such divergent pressures that they would exhibit such inverse biases, as CUB tends to be similar within species [52]. However, neither base composition nor translational selection is sufficient to explain begomovirus CUB.

Similarly, CUB in the mastreviruses is not easily explained. In particular, the Rep has no well-defined codon preferences, which is not predicted by the genomic nucleotide composition, translational selection for their hosts’ CUB, nor an antagonistic relationship between the two. Consequently, these two factors alone are not sufficient to explain mastrevirus CUB.

The ssDNA architecture of geminiviruses provides a possible explanation for their CUB. ssDNA is prone to rapid cytosine deamination to uracil [53], and this process may explain the preference forT-ending codons in ssDNA phages, even in hosts with low AT% [21]. If this process also affects eukaryotic ssDNA viruses, we would expect a very different CUB profile compared to that which is determined by only base content and translational selection. Specifically, strong translational selection predicts uniform codon usage in both CP and Rep, but a strong, biased mutational pressure predicts the begomovirus preference for A-ending codons in the Rep sequences, given that they are encoded in the negative sense. C→T transitions may be tolerated only at synonymous sites, resulting in an overabundance of thymine in the genomic sequence, and a corresponding preference for adenine in the Rep coding sequence. When viewed as they are encoded in the genome, begomovirus CP and Rep nucleotide preferences at synonymous sites are remarkably consistent: both strongly prefer T-ending codons/T-beginning anticodons, suggesting that this biased mutational pressure may contribute to geminivirus CUB. A recent study of the ssDNA porcine circovirus also shows a preference for T-ending codons that differs from the codon preferences of their swine hosts, and is not due to genomic composition [54].

We believe it is very likely that a biased C→T mutational pressure affects eukaryotic ssDNA viruses. Eudicot-infecting begomoviruses are known to exhibit a long-term C→T substitution bias [46,47]. Additionally, ssDNA phages typically exhibit little secondary structure [55,56,57,58,59,60], and while very limited degrees have been documented in eukaryotic ssDNA viruses [61], it has not been observed in the ORFs we examined here. Consequently, these genomes are unconstrained by structural constraints at synonymous sites, and, because unpaired DNA is 100 times more susceptible to oxidative cytosine deamination to uracil than dsDNA [53], highly vulnerable to C→T transition. These oxidative deaminations are common in cellular genomes, but are efficiently repaired [62], while such changes in ssDNA viruses might simply go unrepaired [63]. Alternatively, host cytidine deaminases could be increasing viral thymine content by enzymatically deaminating cytosines. These enzymes are an innate mammalian anti-viral defense, and are active against both viral RNA and ssDNA [64]. Regardless of the exact mechanism, the evidence points to a biased mutational pressure at cytosines contributing to begomovirus evolution.

Mastreviruses present a contrary case: whether or not they experience this potential thymine-enriching factor, their CUB remains unlikely in the absence of additional drivers. Mastreviruses may not experience the same mutational pressure from deamination, may have developed ways to compensate for it, or monocot hosts may interact with ssDNA genomes differently than eudicots. Neither their CP nor Rep sequences carry the signature of rapid deamination, and their CP CUB strongly adheres to host preferences, indicating the primacy of translational selection over base composition and other potential factors. Furthermore, an examination of maize streak virus revealed no evidence of the long-term C→T substitution bias evident in begomoviruses [48]. Finally, unlike most organisms that have been studied, MSV exhibits high degrees of variance in CUB between different genes, and the reasons for this variation are unclear [33]. A single recently discovered eudicot-infecting mastrevirus sequence [65] exhibited codon usage preferences that differ from monocots, eudicots, and the other viruses we examined. Additional analysis is required to more precisely determine the forces affecting CUB in mastreviruses.

Codon preferences in highly expressed genes for five monocot and six eudicot plants were determined using the RSCU data from Wang and Roossinck [22]. Average monocot and eudicot RSCU was calculated for each codon, and preferred codons were defined by Bonferroni-corrected two-tailed t-tests (Microsoft Excel) of average RSCU for synonymous codons. Each set of redundant codons was analyzed individually; no comparisons were made between non-redundant codons. For these analyses, codons for the six-fold degenerate amino acids (L, R, S) were divided into two-fold and four-fold redundant groups, for which RSCU values were calculated independently. This was done because the two groups of codons for these amino acids differ at non-synonymous sites and are consequently recognized by different groups of tRNA species, making it inappropriate to treat them as a single set of redundant codons.

All available complete CP sequences of luteoviruses and potyviruses were collected from GenBank between March and May of 2012. Only species with 15 or more full CP gene sequences were analyzed, and sequences with ambiguous nucleotides were excluded. Complete CP and complete Rep gene sequences of Geminiviruses (monocot-infecting mastreviruses and dicot-infecting begomoviruses) were downloaded from GenBank between January and April of 2012. As with the ssRNA viruses, only species with at least 15 full gene sequences were analyzed, and sequences with ambiguous nucleotides were excluded. Consequently, some geminivirus species could be included in one of our analyses (CP analysis) but not the other (Rep analysis). In total, we analyzed 1285 geminivirus Rep gene sequences, 1481 geminivirus, 1210 potyvirus, and 315 luteovirus CP gene sequences (Supplemental File 1).

All sequences were formatted for analysis using ReadSeq (http://www-bimas.cit.nih.gov/molbio/readseq). CAICal [66] was used to calculate the viral base composition. Reference sequences for each viral species were collected from GenBank on June 12, 2012. Sequences were formatted with ReadSeq, and CAICal was used to determine overall and site-specific base composition for each sequence. Observed third position nucleotide counts were averaged for each species. Expected third position nucleotide counts were computed for each gene/species combination we analyzed based on the genomic nucleotide frequencies of the species’ reference genome and the length of the ORF in the reference genome. Chi-square tests were used to evaluate the differences between these observed and expected counts, with three degrees of freedom (MS Excel). In total, sixty-seven chi-square tests were carried out: one on the CP gene of each potyvirus, luteovirus, and geminivirus we examined, and one on the Rep gene of each geminivirus we analyzed.

Viral RSCU calculations were as for the plant hosts. RSCU values were calculated for each sequence in each viral genus (in the case of the potyviruses, monocot-infecting and dicot-infecting). Then, mean RSCU values were calculated for each viral genus/group. Preferred codons were again defined by Bonferroni-corrected two-tailed t-tests of average RSCU for synonymous codons, and determined separately for monocot- and dicot-infecting members of each viral group.

Average RSCU values for the monocot- and eudicot-infecting viruses of each viral group were compared to those of their respective hosts to determine the correlation between host and viral CUB. Average RSCU of monocot-infecting and dicot-infecting viruses within each group were also compared to each other to determine if the correlation among related viruses was stronger than the correlation to their respective hosts. RSCU between two groups was classified as uncorrelated (r<0.50), moderately correlated (0.50≤r<0.70), or strongly correlated (r≥0.70). Translational selection was rejected when host and virus RSCU were uncorrelated, but considered in cases of moderate or strong correlation.