We have discovered dramatic variation in rates of mitochondrial synonymous substitutions within the angiosperm family Geraniaceae. Our principal findings are three-fold: 1) The synonymous rate increased hugely during the evolution of the Geraniaceae, by at least a factor of 100. 2) This overall rate increase occurred via multiple (at least two, possibly more) stepwise and temporally separated rate increases. 3) Following these rate increases, at least three, and quite possibly more, rate decreases occurred, with many lineages reverting to the very low synonymous rates typical of most angiosperms.

Our estimates for the highest and lowest rates of synonymous substitutions, and of the number of rate changes, are inherently conservative ones, being constrained by the limited taxonomic sampling of this study. The R_S values by definition average the substitution rate across an entire branch, the length/time of that branch being a function of the number of extant Geraniaceae taxa and our sampling of those taxa. Any one R_S value may in fact average across periods of substantially higher and lower absolute substitution rate than would be indicated by the R_S value itself. For example, the highest branch-wise (R_S) rate estimated in this study, of 38 SSB, may turn out to be a significant underestimate of the fastest period of mitochondrial evolution within Pelargonium. We sampled only 7 of ~280 species in the genus, and thus there is ample opportunity to more finely dissect the pattern of rate variation and thereby discover greater peaks (and valleys) of substitution rates, as well as detect a larger number of rate changes.

If synonymous rates are more or less equal across genes within a mitochondrial genome (as seems to be the case for P. hortorum; Fig. 1), then sampling multiple protein genes (and among many taxa) should provide another opportunity to better elucidate the tempo and pattern of rate changes. In particular, this should significantly reduce the standard errors on the estimates of R_S, many of which are quite high for cox1 alone (Fig. 4 and Additional File 1). Sequencing additional non-mitochondrial genes, to reduce the standard errors on the divergence time estimates (see Additional File 1), should also be helpful in this regard.

The mitochondrial rate variation in Geraniaceae uncovered in this study is similar in several respects to that recently described in Plantago 17. In both cases, 1) there is dramatic variation in synonymous substitution rates; 2) taxa fall into multiple (three or four) groups with respect to the amount of accumulated sequence divergence (Fig. 2); 3) the most divergent groups have accumulated comparable levels of synonymous site divergence (Fig. 2); 4) at least two successive major increases in the synonymous rate occurred relatively early on; 5) multiple rate decreases occurred subsequent to the initial increases; and 6) rates of chloroplast and nuclear evolution are not elevated, such that in the fastest lineages mitochondrial rates significantly exceed those of the chloroplast and nuclear genomes (Fig. 2; ref. 17).

The Geraniaceae and Plantago situations also differ in several ways: 1) The Geraniaceae situation encompasses a broader group taxonomically. 2) The initial speed-up occurred much earlier in the Geraniaceae (ca. 80 Myr) than in Plantago (ca. 30 Myr). 3) The peak rates estimated in Plantago (ca. 200 SSB and possibly even 700 SSB) far exceed those in Geraniaceae (38 SSB), although this difference could in large part be a sampling artifact (i.e., if the fastest period of mitochondrial evolution extended throughout a single sampled branch in Plantago, whereas that in Pelargonium was averaged, on the fastest measured branch, with periods of significantly slower evolution). 4) There seems to be less of a tendancy to revert to a typically low substitution rate (of ca. 0.5 SSB) in Plantago compared to Geraniaceae. Overall, then, the Plantago speed-up is more recent, apparently more intense, and apparently less reverted. Like Pelargonium, Plantago is a large genus (in fact the two are almost exactly the same size), and it will be interesting to see how the two situations compare once more taxa and genes (and even some complete genomes) are sampled from both groups.

Geraniaceae and Plantago are, to our knowledge, the only two cases of such dramatic variation in synonymous substitution rates – both in magnitude and rollercoaster-like behavior – documented in any group of plant mitochondrial genomes, or, for that matter, in any genome/group whatsoever. However, several published, mostly phylogenetic studies have identified other plants for which mitochondrial divergence is disproportionately high compared to chloroplast divergence. These long-branched mitochondrial clades include Ephedra in the Gnetales 212223, the monocots Acorus and Alismatales 24, the lycophyte Selaginella 25, the asterids Goodenia and Polemonium 1726, and the simple thalloid and leafy liverworts 27. With the exception of Goodenia 17, none of these cases has been quantitatively treated, and the published trees include all sites rather than just synonymous sites. Nonetheless, in each case, the accumulated divergence appears to be lower than in Pelargonium or Plantago. However, in most of these cases, only one or two species were examined, and thus there is plenty of opportunity to find even more divergent lineages in these groups that may approach or even exceed Pelargonium and Plantago in amount of divergence and peak substitution rate.

Because the extensive variation in substitution rates in the Geraniaceae is restricted to the mitochondrial genome and because nonsynonymous rates are proportionately less elevated than synonymous rates (Fig. 2, Additional File 1), the root cause of this rate variation is almost certainly changes in the mitochondrial mutation rate. The same conclusion was reached for Plantago mitochondrial rate variation 17, whereas most other cases of (much more modest) synonymous site variation in plant mitochondrial DNA have been ascribed to factors that would affect all three plant genomes, such as generation time, population size, and/or speciation rate 678.

By analogy to the extensive literature on mutators (and antimutators) in bacteria and yeast, mutations affecting a broad range of processes that impinge on the mutation rate could be responsible for the extensive variation in mitochondrial mutation rates in Geraniaceae. The only mitochondrial mutators characterized thus far, in lab mutants of the yeast Saccharomyces cerevisiae, affect DNA replication and mismatch repair 28. These are also the most common mutators in microorganisms 2930. In plants and virtually all other mitochondrial systems, these processes are controlled entirely by nuclear genes. Thus, the major increases in the mitochondrial mutation rate that took place early in the evolution of the Geraniaceae and then Pelargonium could be the consequence of nuclear mutations resulting in error-prone replication or defective mismatch repair, as postulated for other cases of synonymous rate variation in plant organelle genomes 71731. Subsequent rate decreases could reflect direct reversal or compensatory suppression of these mutator mutations. Mitochondrial-specific forces could also be at work, e.g., major increases in the amount of oxygen free-radical damage to mitochondrial DNA 3233.

That Geranicaeae mitochondrial genes seem, based on limited samping (Table 1), to be relatively bereft of RNA editing, raises the possibility of a novel molecular mechanism, which we term "mutagenic retroprocessing" (see also 34). Reverse transcriptases are extremely error-prone polymerases, and so exceptionally high levels of reverse transcription in Geraniaceae mitochondria, coupled with high levels of homologous recombination, could in principle be responsible for at least part of the elevation in mitochondrial mutation rate seen in the family. If this hypothesis is correct, we would expect nontranscribed regions to be much less divergent than genes. Unfortunately, identifying and amplifying homologous intergenic spacer DNA is not straightforward given the high rate of rearrangement in plant mitochondrial genomes 2, and our attempts to do so in Pelargonium have been unsuccessful. According to this hypothesis, the existence of limited RNA editing in extant Geraniaceae (Table 1) would reflect mutational drift to unedited C residues following the evident reversals in mutation rate, i.e, cessation of the hypothetical mutagenic retroprocessing activity.

In all 15 examined Geraniaceae, the mitochondrial mutation rate has dropped from its peak levels, with at least three (and probably more) separate rate decreases implied by the R_S patterns of Fig. 4. In many taxa, the mutation rate appears to have dropped more or less to the low rates (ca. 0.3 SSB) typical of most rosid lineages (Fig. 4). And even those taxa (four species of Pelargonium) whose terminal branches possess relatively high R_S values may nonetheless have more fully reverted mutation rates, the present-day rates being averaged with potentially higher rates earlier on these branches.

Such a pervasive local pattern of rate decreases, combined with pervasively slow mitochondrial evolution across land plants in general, suggests that high mutation rates are selected against in the evolution of plant mitochondrial genomes. Chance mutations that directly reverse or otherwise suppress the earlier mutator-like mutations in Geraniaceae may be favored and fixed by natural selection. From this perspective, plant mitochondrial mutators may be viewed as ephemeral, like bacterial mutators but on a much longer time scale.

If most or all Geraniacease have reverted to low mitochondrial mutation rates, then it may be difficult if not impossible to figure out the nature of the mutator mutations that caused the rapid periods of mitochondrial evolution. Mutagenic retroprocessing, admittedly a long-shot hypothesis, is nonetheless attractive in this regard because it would leave obvious signatures in the genome (see above).

Methods for plant DNA isolation, Southern and Northern blot hybridization, PCR isolation, DNA cloning, cDNA preparation, and DNA sequencing are as in 121335. Additional sequences used in this study were taken from GenBank and are listed in Additional File 2, as are GenBank numbers for the sequences generated in this study. PCR primer sequences and aligned data sets are available upon request.

Phylogenetic relationships within the rosids were determined from a concatenated data set consisting of nuclear SSU rDNA and the chloroplast genes atpB, matK, and rbcL. Total aligned length was 6,226 NT. Poorly alignable regions and regions with gaps in most taxa were excluded from the analyses. A maximum likelihood (ML) tree was constructed with PAUP*, version 4.0b10 36, by using the general time-reversible model, a gamma distribution with four rate categories, and an estimate of the proportion of invariant sites. The rate matrix, base frequencies, shape of the gamma distribution, and proportion of invariant sites were estimated before the ML analysis from a parsimony tree constructed from the data. Support for the ML tree was evaluated by the bootstrap procedure with 500 replicates using parameters estimated from the ML tree.

Divergence times for all nodes within rosids were calculated using a penalized likelihood approach as implemented in the r8s program 37. A fixed time constraint of 110 million years 3839 was used for the crown group age of rosids. The ML tree from the phylogenetic analysis of rosids (Fig. 3) was used as the starting tree. A smoothing factor of 18 was determined by using the r8s cross-validation procedure 37. Different starting points of initial age estimates and reanalysis after perturbation of the final age estimates had no effect on the results. Standard errors for the divergence time of each node were calculated by rerunning the divergence time analyses on 500 bootstrapped data sets.

The elapsed time along each branch, T_ij, was calculated as the difference in divergence times of the starting node, T_i, and the ending node, T_j, that define the branch. Standard errors for T_ij were calculated according to the formula

σ T i j = ( σ T i ) 2 + ( σ T j ) 2 MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciGacaGaaeqabaqabeGadaaakeaacqaHdpWCdaWgaaWcbaGaemivaq1aaSbaaWqaaiabdMgaPjabdQgaQbqabaaaleqaaOGaeyypa0ZaaOaaaeaacqGGOaakcqaHdpWCdaWgaaWcbaGaemivaq1aaSbaaWqaaiabdMgaPbqabaaaleqaaOGaeiykaKYaaWbaaSqabeaacqaIYaGmaaGccqGHRaWkcqGGOaakcqaHdpWCdaWgaaWcbaGaemivaq1aaSbaaWqaaiabdQgaQbqabaaaleqaaOGaeiykaKYaaWbaaSqabeaacqaIYaGmaaaabeaaaaa@43E8@

Branch lengths, representing the number of substitutions per synonymous site (d_S) or number of substitutions per nonsynonymous site (d_N), were determined for protein genes using codeml in the PAML package, version 3.14 40. The Muse-Gaut (MG94) codon model was used with separate d_N/d_S ratios for each branch. Codon frequencies were computed by using the F3 × 4 method. The transition/transversion rate ratio and d_N/d_S ratios were estimated during the analysis with initial values of 2 and 0.4, respectively. Standard errors for total branch lengths (t) were reported by PAML, and these values were propagated to calculate standard errors for their corresponding d_S and d_N branch lengths according to the formulas

σ d N = ( N + S ) σ t N + S ω MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciGacaGaaeqabaqabeGadaaakeaacqaHdpWCdaWgaaWcbaGaemizaq2aaSbaaWqaaiabd6eaobqabaaaleqaaOGaeyypa0ZaaSaaaeaacqGGOaakcqWGobGtcqGHRaWkcqWGtbWucqGGPaqkcqaHdpWCdaWgaaWcbaGaemiDaqhabeaaaOqaaiabd6eaojabgUcaRmaalaaabaGaem4uamfabaGaeqyYdChaaaaaaaa@3FD0@

σ d S = ( N + S ) σ t N ω + S MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciGacaGaaeqabaqabeGadaaakeaacqaHdpWCdaWgaaWcbaGaemizaq2aaSbaaWqaaiabdofatbqabaaaleqaaOGaeyypa0ZaaSaaaeaacqGGOaakcqWGobGtcqGHRaWkcqWGtbWucqGGPaqkcqaHdpWCdaWgaaWcbaGaemiDaqhabeaaaOqaaiabd6eaojabeM8a3jabgUcaRiabdofatbaaaaa@3FCA@

where ω is the d_N/d_S ratio and N and S are the number of nonsynonymous and synonymous sites, respectively, as reported by PAML. Branch lengths representing total substitutions per site for rRNA genes were estimated with baseml in PAML. The general time-reversible nucleotide model was used with a gamma distribution for rate variation with four categories. The rate matrix, nucleotide frequencies, and shape of the gamma parameter (initial value of 0.5) were estimated during the analysis.

All branch length analyses employed user-defined topologies. Within rosids, topologies were constrained according to the phylogenetic analysis carried out in this study (Fig. 3). Relationships in the Plantaginaceae were constrained according to 17. All other relationships were constrained according to the Angiosperm Phylogeny Website 41.

Absolute rates of synonymous substitution per branch (R_S) were calculated by dividing the synonymous branch length, d_S, by the length of time, T, for that branch. Standard errors for R_S were determined by propagating the errors associated with branch length and time according to the formula

σ R S = R S ( σ d S d S ) 2 + ( σ T T ) 2 MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciGacaGaaeqabaqabeGadaaakeaacqaHdpWCdaWgaaWcbaGaemOuai1aaSbaaWqaaiabdofatbqabaaaleqaaOGaeyypa0JaemOuai1aaSbaaSqaaiabdofatbqabaGcdaGcaaqaamaabmaabaWaaSaaaeaacqaHdpWCdaWgaaWcbaGaemizaq2aaSbaaWqaaiabdofatbqabaaaleqaaaGcbaGaemizaq2aaSbaaSqaaiabdofatbqabaaaaaGccaGLOaGaayzkaaWaaWbaaSqabeaacqaIYaGmaaGccqGHRaWkdaqadaqaamaalaaabaGaeq4Wdm3aaSbaaSqaaiabdsfaubqabaaakeaacqWGubavaaaacaGLOaGaayzkaaWaaWbaaSqabeaacqaIYaGmaaaabeaaaaa@4701@

R_N values and their standard errors were calculated similarly.