Vertebrate mitochondrial DNA has two strands of different buoyant densities, i.e., the H-strand and the L-strand. The H-strand is the sense strand for 1 protein-coding gene (ND6) and 8 tRNA genes and the L-strand is the sense strand for 12 protein-coding genes, 2 rRNA genes and 14 tRNA genes. The two strands have different nucleotide frequencies, with the H-strand being GT-rich and the L-strand AC-rich 
. This asymmetrical distribution of nucleotides has been explained 
in terms of the strand-displacement model of mitochondrial DNA (mtDNA) replication 
. In short, the H-strand is left single-stranded for an extended period and subject to spontaneous deamination of A and C 
to G and U. In particular, the C→U mutation mediated by the spontaneous deamination is known to occur in single-stranded DNA about 100 times as frequently as in double-stranded DNA 
. Therefore, the H-strand tends to accumulate A→G and C→U mutations and become GT-rich while the L-strand tends to become AC-rich. This pattern is similar to the strand bias observed in eubacterial genomes 
The strand-biased mutation spectrum has profound consequences on codon usage in mitochondrial protein-coding sequences (CDSs) and the anticodon of tRNA genes 
. First, the codons of the 12 CDS sequences (that are collinear with the AC-rich L-strand) end mainly with A or C, and the codon bias in the ND6 gene collinear with the opposite strand is the opposite. Second, the 8 tRNA sequences collinear with the GT-rich H-strand is more GT-rich than the 14 tRNA sequences collinear with the AC-rich L-strand. Third, because the overall codon usage is mainly determined by the 12 CDSs collinear with the AC-rich L strand, the A-ending and C-ending codons are almost always the most frequently used codons. The anticodon of 21 tRNA genes (out of a total of 22), regardless of which strand they are located, have anticodons with their wobble site forming Watson-Crick base-paring with the most abundant codons in each codon family, i.e., the wobble site of the tRNA genes is either a U to pair with the abundant A-ending codons or a G to pair with the abundant C-ending codons 
The codon-anticodon adaptation is long known 
, and the pattern described above would have been nice but boring had there not been an interesting and singular exception to the general pattern of tRNA anticodon matching the most abundant codon. The tRNAMet
anticodon is 3′-UAC-5′ (or CAU for short), with the wobble site being C instead of U, and forms a Watson-Crick match with the AUG codon instead of the AUA codon, in spite of the fact that the latter is used much more frequently than the former. The ability of the CAU anticodon to pair with the AUA codon is achieved by modifying the C in the anticodon CAU to 5-formylcytidine 
. A similar case involves the methylation of guanine in starfish tRNASer
to translate all four AGN codons 
The use of the CAU anticodon instead of a UAU anticodon in vertebrate mitochondrial tRNAMet
is unexpected from two existing hypotheses of anticodon usage. The codon-anticodon adaptation hypothesis 
predicts that the anticodon should match the most abundant codon. Because AUA is much more frequent than AUG, the hypothesis predicts that the anticodon of the tRNAMet
gene should be UAU instead of the observed CAU. The hypothesis of selection on anticodon wobble versatility 
, which was implicitly proposed before 
and may be more appropriate for vertebrate mitochondrial genomes because each codon family is translated by a single tRNA species, states that the anticodon should maximize its wobble versatility in paring with synonymous codons. Because U in general is more versatile than C in wobble pairing with both A and G 
, the hypothesis of selection on anticodon versatility also predicts an UAU anticodon to maximize its paring versatility with the AUA and AUG codons. The fact that the observed tRNAMet
anticodon is CAU instead of the predicted UAU is intriguing.
This unexpected tRNAMet
anticodon has been attributed to a compromise between translation initiation and elongation 
as follows. AUG is not only the most frequently used initiation codon, but also the most efficient initiation codon in Escherichia coli 
and Saccharomyces cerevisiae 
. In E. coli
, the most efficient non-AUG initiation codon is AUA and its rate of initiation is only 7.5% of AUG 
. In yeast mitochondria, a mutation of the initiation AUG to AUA in the COX2 gene caused at least a five-fold decrease in translation 
, and similar finding was also duplicated in another yeast mitochondrial gene COX3 
. Assuming the generality of these findings, an anticodon matching AUG will increase the initiation rate and would be favored by natural selection because translation initiation is often the limiting step in protein production 
. This presents a conflict between translation initiation and translation elongation. An AUG-matching anticodon would increase the translation initiation rate but decrease the translation elongation rate because an overwhelming majority of methionine codons are AUA in vertebrate mitochondrial genomes. The fact that all known vertebrate tRNAMet
genes feature an AUG-matching anticodon implies that nature has chosen to maximize the translation initiation rate 
. This hypothesis that invokes a conflict between translation initiation and translation elongation to explain the usage of the CAU anticodon in tRNAMet
will be referred hereafter as the translation conflict hypothesis.
Two consequences can be derived from the translation conflict hypothesis. First, we should expect a relative reduction of AUA usage because the AUG-matching anticodon imposes selection against the use of AUA codons as AUA would need to be wobble-translated. To fix ideas, let us focus only on AUR (methionine) and UUR (leucine) codon families. The reason for choosing UUR instead of any other R-ending codon families is because other R-ending codon families do not have a middle U and the middle nucleotide in a codon is known to affect the nucleotide at the third codon position (P. Higgs, pers. comm.).
For the 12 CDSs that are collinear with the AC-rich L-strand, the mutation favors A-ending codon 
. For UUR codons, because the anticodon wobble site is U and form Watson-Crick base pair with A, we also expect UUA codon to be preferred against UUG codons. Thus, both mutation and the tRNA-mediated selection favor the use of UUA against UUG codons. However, for the methionine codons, the AUG-matching tRNAMet
anticodon would favor the AUG codon against the AUA codon. Thus, the tRNA-mediated selection and the mutation bias are in opposite directions. If we define
for each of these two codon families, where X is either A or U, and NXUA
are the number of XUA and XUG codons, respectively, we should find PAUA
to be smaller in the AUR codon family than PUUA
in the UUR codon family.
An argument against using Eq. (1) is that the result would be biased in favor of supporting the prediction of PAUA<PUUA because the initiation codon, which is AUG in most cases, was not excluded. A more convincing comparison should compute PAUA after excluding initiation codons entirely. This is what we used in this study.
For the ND6 gene collinear with the GT-rich H-strand, the strand-biased mutation spectrum favors G-ending codons in the two XUR codon families. For the methionine codon family, the AUG-matching anticodon also favors the AUG codon against the AUA codon. So the AUA codon will be depressed by both the strand-biased mutation and the tRNA-imposed selection. The tRNA-imposed selection is absent against UUA codon in the UUR codon families because their respective tRNA anticodons all match the A-ending codons 
. Thus, for the ND6 gene, we also expect PAUA
to be smaller in the AUR codon family than PUUA
in the UUR codon family.
The expected PAUA<PUUA, if confirmed, can have two possibilities. If the total number of methionine remains constant across mitochondrial genomes, then a deficiency of the AUA codons in one genome implies an equal amount of surplus of AUG codons. In contrast, if there is no selection maintaining a constant number of methionine codons but there is selection against AUA codons because it requires the inefficient wobble translation, then a genome with a deficiency of AUA codons would also exhibit a deficiency of methionine codons.
In this paper, we use mitochondrial genomes from representative vertebrates, urochordates and bivalves to test these predictions (with the relevance of the bivalve mitochondrial genomes pointed out to us by an anonymous reviewer). While vertebrate mitochondrial genomes all have just one tRNA-Met gene, urochordates have two tRNA-Met genes, with one having a CAU anticodon and the other having an UAU anticodon. The presence of the UAU anticodon in the tRNA-Met gene in urochordate mitochondrial genomes implies that the selection against AUA codon should be weaker in urochordates than in vertebrates. The bivalve mitochondrial genomes are particularly interesting because some of them have anticodon CAU in both of their tRNAMet genes whereas others have an anticodon CAU in one of their tRNAMet genes and an anticodon UAU in the other tRNAMet gene. We expect the selection against the AUA codon to be weaker in the latter than in the former.