Synonymous codons, which encode the same amino acid, are not used randomly. Such codon usage biases are explained as the balance between mutational drift and natural selection 1. In unicellular organisms 23456 and invertebrate metazoans 7891011, the levels of gene expression can be used to interpret their codon biases. Specifically, highly expressed genes, compared with weakly expressed ones, selectively use 'optimal codons' that correspond to abundant tRNAs so as to improve their translational efficiency 1112131415.

Nevertheless, in vertebrates, whose genes display more dramatic codon usage biases than those of simple organisms 14, the correlations between codon usage and patterns of gene expression (that is, the levels and breadth of gene expression) remain a subject of controversy 1116. In a number of rodent and human tissues, recent studies have indicated positive correlations between levels of gene expression, as estimated by SAGE and/or microarray analysis, and GC3 16171819. However, these results are in contradiction with observations made by analyzing expressed sequence tags (ESTs) 1116. Among extremely highly expressed genes, the H3 histone gene family is biased to use GC-ending codons 20. However, there is no difference in codon usage between ribosomal protein genes, which are also expressed at very high levels, and other genes 14. As to correlations between breadth of gene expression and codon usage, some studies suggest that housekeeping genes, with a wider breadth of expression, are biased to use GC-ending codons 1821222324 (also see the debate between 25 and 16); however, other papers have described different observations 1126272829. Although codon usage has been found to exhibit variations in human genes specifically expressed in six tissues 30, the effect is very weak 31 and cannot be generalized to interpret the global variation (the preference of AT-ending or GC-ending codons) of synonymous codons in the thousands of mammalian genes.

Moreover, in vertebrates, the reasons why there are correlations between codon usage and patterns of gene expression remain to be elucidated. By using multivariance analyses (MVA), highly expressed genes have been observed to have excessive usage of T-ending codons in Xenopus 32 and the Cyprinidae family 33. However, both natural selection and 'transcriptional associated mutation bias' (TAMB) 343536 would account for these observations. In the tissues with no evidence of TAMB, a set of GC-ending codons favored in highly expressed genes has been suggested to be optimal codons 19. Moreover, GC-ending codons are more abundant in highly expressed genes 18 and constitutively spliced exons 37. However, if GC-ending codons are optimal due to selective advantages, it is difficult to see why the synonymous substitution rate (Ks) would be increased with GC-ending codon usage 38394041 or why the Ks of alternatively spliced exons would be lower than that of constitutively spliced exons 42. It has been reported that highly expressed genes have higher recombination rates 434445. Moreover, according to the model of biased gene conversion (BGC), recombination rates are positively correlated with GC3 464748495051, indicating that both natural selection and BGC may be responsible for the correlations between the levels of gene expression and GC3. The variations of synonymous codon usage among tissue-specific genes have been suggested to be the consequence of translational selection 30; a recent study, however, has indicated that these observations were due to regional variations of substitutional patterns rather than translational selection 31. Taken together, further research is obviously still needed to explore the mechanisms of vertebrate codon usage bias.

In this paper, to investigate the regularity and mechanisms of mammalian codon usage, we have taken developmental stage-related patterns of gene expression into account in models of stem cell differentiation (Figure 1 and Table 1). Stem cells, progenitor cells and their derivates, defined by their distinct differentiation potential (Figure 1a), play critical roles in the early stages of metazoan ontogenesis and thus provide ideal models of the mammalian developmental hierarchy. Moreover, developmental processes are believed to be of critical importance to the investigation of evolutionary mechanisms 52, even at the genomic level 53. In the current study, therefore, we have investigated the correlations between developmental stage-related patterns of gene expression and codon usage in developmental hierarchies of stem cell differentiation. Specifically, we have taken advantage of two independent models of stem cell differentiation 5455 to identify developmental stage-related patterns of gene expression, as well as the correlations between these patterns of gene expression and codon usage.

To define the developmental stage-related patterns of gene expression in models of stem cell differentiation, we have introduced two parameters. First, the 'level of gene expression' has been defined as the intensity of gene transcription in a particular cell type. Second, the 'fold change of gene expression' has been defined as the ratio of the expression levels of the same gene in two cell types of two neighboring stages in the developmental hierarchy (Figure 1b). We have further defined one of these two cell types, in the upper developmental hierarchy, as the earlier cell type, and the other, in the lower developmental hierarchy, as the later cell type. These two cell types together constitute a 'differentiation pair'. Thus, the 'fold change of gene expression' is a descriptive index of the levels of gene enrichment in a given differentiation pair.

In the present work, we investigate the correlations between developmental stage-related patterns of gene expression (that is, the 'levels of gene expression' in each cell type in the models of stem cell differentiation and the 'fold changes of gene expression' in each differentiation pair) and the molecular features (GC3 and genomic GC (GCg) content) of these genes. We also explore possible mechanisms for these developmental stage-related patterns of codon usage. This study reveals that developmental stage-related patterns of gene expression are correlated with GC3 and GCg in models of stem cell differentiation. Moreover, these analyses suggest that the model of translational selection, rather than other known hypotheses that have been put forward, might be the most likely to account for the developmental stage-related patterns of codon usage, especially for the negative correlations between the levels of gene expression and GC3.