Translational selection on synonymous codon use is indicated if frequencies of preferentially used, 'optimal', codons increase with expression level and correspond to the most abundant tRNA or to the tRNA with which they form the strongest binding - for several organisms, this seems to be the case (see for reviews [1–4]). Additional support for a beneficial role of certain 'optimal' codons in translation comes from laboratory studies [5–8]. Translational selection may act to maximise the speed of elongation, minimize the costs of proofreading or maximise the accuracy of translation , and depending on the selective target, one can test further distinct predictions. Under selection for translational accuracy we expect for example: (i) selection against translational errors to relate to the error's costs. As costs of an erroneous protein should accrue with each added amino acid during protein synthesis, one may expect long genes to experience higher optimal codon use than short genes . Supporting selection for translational accuracy, in E. coli and yeast, relative optimal codon use indeed increases with gene length [9–12]. (ii) We also expect selection against translational errors to relate to the error's functional effect: translational errors for some amino acids may have no functional effects, while for other amino acids, translational errors render a protein non-functional. The latter should be under stronger selection for translational accuracy. As Akashi  points out, the functional importance of amino acid site may be approximated by its evolutionary conservation. Under translational selection for accuracy one may hence expect higher optimal codon frequencies at conserved than at non-conserved amino acid sites. This is indeed the case in D. melanogaster, C. elegans, E. coli [12–14]; a recent study  indicates this pattern may also apply to mouse and human using a modified measure of optimal codons.
However, surprisingly, in D. melanogaster, C. elegans, A. thaliana, and humans, optimal codon use decreases with gene length, thereby opposing the prediction under selection for translational accuracy [16–18]. This decrease is particularly surprising for species, in which selection for translational accuracy is indicated by the aforementioned higher optimal codon use at conserved amino acid sites. The explanation may be that the negative correlation between optimal codon use and gene length simply is a side effect: highly expressed genes with high optimal codon use tend to be short, possibly to be more economic . Yet, while control of expression level affects the correlation of optimal codon use with gene length in yeast, causing a change from negative to positive [10, 11], in D. melanogaster, C. elegans, A. thaliana or humans, the negative correlation of optimal codon use with gene length does not seem to be due to a correlation of gene length with expression level only [11, 17, 18, 20, 21].
Two explanations for the negative correlation have been proposed, both of which are based on translational selection. First, under selection for translational efficiency, selection for optimal codons may decrease with gene length due to the decrease in the relative fitness effect per optimal codon . The second hypothesis invokes Hill-Robertson interference , which considers the reduction in selective efficacy due to linkage among sites: weakly or strongly selected sites that evolve adaptively or under constraints may affect evolutionary dynamics of linked sites. As Comeron et al.  suggest selection efficacy on optimal codons may decrease with gene length, as long genes with higher numbers of potentially interfering sites may experience a stronger Hill-Robertson effect. The Hill-Robertson effect has been considered for various effects on synonymous codons [e.g. [23–29]]. As recombination breaks down linkage, the observed positive correlation of optimal codon use with recombination rate was taken as support for the Hill-Robertson effect reducing the efficacy of translational selection on optimal codons [30–32].
Yet, optimal codons in several metazoans, such as the ones for which the negative correlation was first reported for, i.e. D. melanogaster, C. elegans and A. thaliana, but also for humans are mostly ending with G or C (-GC) [see codon tables in [17, 18]], and compositionally biased mutation or repair processes may indirectly affect optimal codon use. Recombination-dependent repair (gene conversion) is indeed biased towards -GC in many organisms including yeast, mice, humans and Drosophila [20, 33–36], and hence may be the potential force. Effects of GC-biased gene conversion will be most obvious at sites that evolve neutrally or under weak selection, and the substitution patterns it leaves resembles that of directional selection [see for review ]. GC-biased gene conversion has been indicated to affect optimal codon use before: optimal codon frequencies increase with recombination rate, a patterns consistent with population genetic predictions under translational selection on optimal codons [30, 31]. However, in D. melanogaster and C. elegans not only optimal codon frequencies increase with recombination, but also non-optimal ones, as long as they end with -GC [20, 37]. The positive correlation of non-optimal GC-ending codon frequencies with recombination indicates the observed positive correlation optimal (GC-ending) codons - that was taken as evidence for reduced efficacy of translational selection due to Hill-Robertson interference - is likely affected by compositionally biased processes such GC-biased gene conversion [20, 37]. Whether or not GC-biased gene conversion or Hill-Robertson effects the positive correlation between optimal codons and recombination attracted controversy [see for example [20, 37] versus [31, 32]], but with respect to the observed negative correlation of optimal codon use with gene length, GC-biased gene conversion has never been considered.
The negative correlation of optimal codon use with gene length is found in organisms whose optimal codons are biased towards GC-ending ones, and may hence be caused by forces acting on optimal codons or on base composition. As translational selection affects optimal codons, while a compositional bias like gene conversion affects GC-ending codons, one may disentangle the effects by looking at optimal and non-optimal GC- and AT-ending codons separately. Saccharomyces cerevisiae is a good organism to disentangle the two forces because translationally optimal codons are not biased towards GC-ending ones as in the above mentionned organisms. Furthermore, translational selection and GC-biased gene conversion are comparably well-studied and supported in S. cerevisiae [e.g. [34, 38, 39]].
Results of this study demonstrate in S. cerevisiae the frequency of GC-ending (optimal AND non-optimal) codons decreases with gene length and increases with recombination. Also a decrease of GC-ending codons along genes is indicated. This distinction between AT- and GC-ending codons cannot be explained by variation in strength and efficiency of translational selection, while GC-biased gene conversion may explain the observation. Substitutions at four-fold degenerated sites differ between AT->GC and GC->AT changes, further supporting an effect of GC-biased gene conversion. Initiation of gene conversion events in promoter regions and the presence of a gene conversion gradient most likely explain the observed decrease of GC-ending codons with gene length and gene position.