The nucleotide content of rice and Arabidopsis coding sequences (expressed as percent GC) is summarized in Figure 1. The Figure shows that there is a distinctly bimodal distribution of GC content among the 14,005 rice genes, which is consistent with previous reports 151723. In contrast to this, the Arabidopsis genes are characterized by a unimodal distribution with a relatively low average value for GC content. In the Figure, the vertical line at 60% GC indicates the point at which we separated the rice genes into two classes: High GC genes and Low GC genes. The average GC content of these two classes, along with the average for all Arabidopsis genes is shown in Table 1. From the Table, we can see that the GC content of the Arabidopsis genes (44.5%) is comparable to that of the Low GC rice genes (50.1%).

Table 1 also presents the data for the third positions of codons only. In this case, we see the same trends as for all of the codon positions, but the differences are much greater. For instance, the GC content of the third codon positions of the High GC rice genes (80.4%) is almost twice the values for the Arabidopsis genes (42.8%). Given that variations in codon usage will affect the third codon position primarily, this result leads us to expect significant differences in codon usage between the two classes of rice genes. We investigated this using Correspondence Analysis (see below).

We also wished to investigate the possible clustering of GC-rich genes within the rice genome. To do this, we took a sample of two rice chromosomes and plotted the GC content at the third codon positions (GC3) against the position of the genes along the chromosome. For comparison, we did the same analysis for the GC3 content of Arabidopsis genes along the chromosome. The results (see 33) show that genes with varying levels of nucleotide composition are interspersed along the chromosome.

Correspondence analysis 21 was used to explore the variation in Relative Synonymous Codon Usage (RSCU). Since there are a total of 59 synonymous codons (61 sense codons, less the unique methionine and tryptophan codons), this analysis partitions the variation along 59 orthogonal axes, with 41 degrees of freedom. The first axis is the one that captures most of the variation in codon usage, with each subsequent axis explaining a diminishing amount of the variance. In contrast to other types of variance component analysis, such as Principal Component Analysis (PCA), correspondence analysis has the advantage of allowing one to not only show the distribution of genes in the multidimensional space, but also to show the corresponding distribution of synonymous codons (as shown in Figures 2A and 2B). Correspondence Analysis is primarily designed for use with data tables containing counts, e.g., numbers of synonymous codons, whereas PCA is a general method of data reduction that is more suitable for continuous measurement data. Perriere and Thioulouse 22 have provided a critical review of the use of Correspondence Analysis for studies of codon usage.

Figure 2 shows a correspondence analysis of the synonymous codon usage (RSCU) among the rice genes. The origin in Figure 2A represents the average RSCU for all genes, with respect to the first two axes. The distance between genes on this plot is a reflection of their dissimilarity in RSCU, with respect to the two axes. In this case, the two axes account for 36.9% and 6.9% of the variation in the data, respectively. The third axis accounts for approximately 3% of the variation and the remaining axes for even smaller amounts of the variance each. Thus the first axis reflects the primary factor that explains the differences in codon usage among the rice genes. From Figure 2A, we can see that the rice High GC genes (colored red in the Figure) and Low GC genes (colored blue) separate along this primary axis. The corresponding distribution of synonymous codons (see Figure 2B) shows the separation of C/G-ending codons and A/U-ending codons along this same axis. This indicates that the variations in synonymous codon usage among the rice genes are based on the nucleotide content of the genes. The separation of genes on the second axis appears to be largely due to frequency differences in C-ending and G-ending codons among the GC rich genes (see right side of Fig. 2B).

Although the color coding in Figure 2A suggests a general relationship between the nucleotide content of genes and their position on the first axis of the correspondence analysis, it does not give us any statistical measure of this relationship. To do this, we calculated the correlation between the GC content of individual rice genes and their location on the primary axis of the Correspondence Analysis. The results were highly significant (R = 0.96, p < 0.00001), indicating that the variations in codon usage are strongly correlated with the nucleotide content (i.e., GC content) of the genes.

We further investigated the relationship between nucleotide content and codon usage by calculating the effective number of codons for each of the rice genes. The effective number of codons 20 is a measure of the evenness of codon usage among the 61 sense codons. At one extreme is all codons are used equally frequently (given the observed frequencies of amino acids) the effective number of codons is 61. If, at the other extreme, a single codon only is used for each amino acid, then the effective number of codons is reduced to 20. In most cases, the observed number falls somewhere between these extremes. Figure 3 shows the relationship between the effective number of codons (Nc) and the GC content at the third position of each gene (GC3). This Figure also contains a reference line (GCref) showing the expected position of genes whose codon usage is constrained solely by the nucleotide composition at the third codon position 20. From the Figure, it can be seen that the observed value of Nc tracks the reference line quite closely. This indicates that the nucleotide composition at the third codon position is a major determinant of the effective number of codons. A polynomial line, to the power of 2, that regresses Nc on GC3s (not shown in the figure) fits the data very well (R² = 0.82, p < 0.00001). Essentially, the effective number of codons decreases as the GC content increases.

Although the preceding results clearly establish a strong correlation between nucleotide content and codon usage within the rice genome, they do not tell us which of the two is the causal factor. In an effort to understand the biological basis of these differences in codon usage and nucleotide content within the rice genome, we compared these rice genes with their homologs in Arabidopsis. We used a BLAST search to identify 7,160 pairs of homologous genes in rice and Arabidopsis (see Methods).

We first calculated the GC content of the homologous gene pairs. Among these gene pairs, we found that the GC content of the Arabidopsis homologs remains unimodal, as is shown for the full set Arabidopsis genes in Figure 1, whereas the content of the rice homologs remains bimodal, again as shown in Figure 1 for the complete rice data set. Thus, the overall differences between the genomes that are seen in Figure 1 cannot be due to differences in gene content between the two species because they are still present in the homologous gene set.

We then computed relative synonymous codon usage values for the homologous genes and performed a new correspondence analysis. The results are shown in Figure 4. Here we see that the High GC and Low GC rice genes (now defined within the homologous set) again separate along the first axis of the analysis, whereas all of the Arabidopsis genes are clustered on the left side of the plot. In other words, the use of homologous genes does not alter the result that we observed in Figure 2. Furthermore, all of the Arabidopsis genes have generally similar patterns of codon usage, regardless of whether they are homologs of High GC or Low GC rice genes. This suggests that the divergence in codon usage patterns among rice genes has occurred since the evolutionary divergence of the dicots and monocots approximately 200 million years (My) ago, i.e., over a relatively short evolutionary time.

Although our results suggest that the GC content of the High GC rice genes has increased significantly since the divergence of the monocots and dicots, there remains the formal possibility that, instead, the Arabidopsis genes have recently converged toward a common, lower GC content. To distinguish between these possibilities, we extracted 92 homologous sequences from the genome of Pinus taeda, and we used these as an out-group to infer the direction of the change. Whereas the High GC rice genes have an average GC content of 66.6 (SE 0.08) and their Arabidopsis homologs have an average GC content of 46.0 (SE 0.07), the average GC content of the P. taeda homologs is 45.2 (SE 0.04). Thus we can infer that the ancestral condition was similar to that currently seen in Arabidopsis.

Gene length has previously been shown to be negatively correlated with codon usage in C. elegans, Drosophila and Arabidopsis 25. We tested to see whether the same relationship holds true for rice genes. We compared the average gene lengths of the two groups of rice genes (High GC and Low GC), as defined in Figure 1. Not only did we find that the High GC genes were shorter, as suggested by previous studies in other species, but the magnitude of this length difference was surprisingly large and highly significant (p < 0.0001). Specifically, the average length of the Low GC coding sequences (1417 +/- 13 bp) is approximately 500 bp larger than the average for the High GC genes (921 +/- 9 bp). Although there is a wide range of individual gene lengths within each class, this highly significant average difference suggests that the length of the rice genes was a significant factor in the evolutionary increase in GC content.

14005 rice coding sequences that are longer than 75 codons were obtained from Gramene database 16 and the EMBL as previously described 17. For the A. thaliana coding sequences we used the file containing 25,625 Arabidopsis coding sequences (all greater than 75 codons) that we obtained previously 17. The homologous sequences from Pinus taeda (Loblolly pine) were extracted, using BLASTN searches with a cutoff Expect value of 1e-20, from the dataset of 14,198 P. taeda sequences in the NCBI Unigene database 34.

Homologous pairs between O. sativa and A. thaliana were identified by performing BLASTP searches 18 of the rice protein sequences against Arabidopsis sequences with a cutoff Expect value of 1e-20. When a rice protein has more than one Arabidopsis protein hit, the pair having the lowest Expect value was retained. Using this method, we identified 7,160 homologous gene pairs between the two species, of which 895 gene pairs are rice genes homologous to Arabidopsis chromosome 4 genes. In order to see the relationship between codon bias and the evolutionary rate, we used the method of Yang and Nielsen 19 to calculate the synonymous rates for the 895 gene pairs.

Relative synonymous codon usage (RSCU) is the observed frequency of a codon divided by the frequency expected if all synonyms for that amino acid were used equally. An RSCU value close to 1.0 indicates a lack of codon bias. The RSCU was computed for each gene. The RSCU values were then analyzed using correspondence analysis (see below).

The effective number of codons (Nc) is a commonly used measurement to quantify codon usage bias of a gene 20. The Nc takes a value between 20, when only one synonymous codon is used for each amino acid, and 61, when all codons are uniformly used. Lower Nc values indicate stronger bias. Since Nc is constrained by G+C content of the gene, it is often plotted against GC3s (the frequency of G+C at the third synonymous codon positions) of the gene to investigate patterns of codon usage 20.

Correspondence analysis 21, was used to explore the variation in the 59 RSCU values for each of the 61 sense codons, other than the unique methionine and tryptophan codons. This multivariate statistical method creates a series of orthogonal axes to identify trends that explain the data variation, with each subsequent axis explaining a decreasing amount of the variation. The method, as implemented in CodonW version 1.4 2, was used in this study. Correspondence analysis assigns ordination for each gene and codon on these axes, and the ordination of the genes and codons can be superimposed. Since the first two axes capture a larger fraction of the variance of the data than any of the other axes, genes and codons were plotted on these two axes only.