RNA editing is the process whereby an RNA sequence is modified from the sequence corresponding to the DNA template. A particular form of RNA editing in plant mitochondria, by which some cytidines are converted to uridines before translation, occurs in many land plant lineages. Although cytidine to uridine conversion is most common, the reverse conversion is sometimes observed 1234. In plants, the phenomenon is best studied, albeit still poorly understood, in the mitochondria and plastids of angiosperms 5678.

The majority of plant mitochondrial RNA editing occurs in coding sequences, and editing frequently changes codons, resulting in changes of amino acids, or, in some cases, creation of entirely new open reading frames 1910. These changes often result in an increase in similarity with respect to homologous protein sequences among different organisms (such as in wheat 11), and Gray has postulated that the RNA editing process functions as a repair mechanism to correct otherwise-deleterious genomic mutations 12. RNA editing has also been detected in introns, where it is conjectured to improve splicing efficiency 13.

The precise biochemical basis for C-to-U editing in plant mitochondria is unknown, although experimental evidence suggests a deamination reaction 1415161718. Despite substantial study, the molecular biological mechanism by which C-to-U RNA editing proceeds remains relatively obscure, although several experimental studies have implicated a role for cis-recognition 192021. The mechanism by which edited sites are recognized is also still poorly understood, but the importance of surrounding nucleotides has been noted 22. A highly non-random distribution of nucleotides in the immediate vicinity of edited sites (within 10–20 nucleotides 5' and 3') is observed, but no precise consensus motif has been identified 916. Additionally, previous studies suggest that inferred secondary structure is not important in site recognition for C-to-U conversion 1619.

Identifying edited sites thus remains an open problem, one to which we have applied tree-based statistical models and an extension of such models. When applied to a similar problem (predicting peptide binding to major histocompatibility complex (MHC) class I molecules 23), tree-based statistical methods generated very accurate models, identifying specific important residues when no precise sequence motif had previously been identified. Therefore, we were motivated to apply tree-based statistical models and an extension, random forests, to the problem of C-to-U RNA editing in angiosperm mitochondria using complete mitochondrial genome data for three species: Arabidopsis thaliana, Brassica napus and Oryza sativa. The objective for the current research was to identify sequence features that may provide insights into C-to-U editing of plant mitochondrial RNA. We address the following specific questions. Is there evidence that sufficient information exists within sequence regions flanking edited sites to accurately predict editing? Is there an association between estimated free energy of folding for short sequence regions containing edited sites and C-to-U editing? We report tree-based statistical analysis of three complete mitochondrial genomes and show that relatively simple models provide moderately accurate prediction of C-to-U edited sites.

Despite their simplicity, the tree-based statistical models derived here performed moderately well, with mean accuracies across species generally ~0.71. Single trees were improved upon by constructing models based on ensembles of tree-based models (random forests) each of which was built using random subsamples of the data. This sub-sampling has the effect of reducing the variance through averaging and also reducing the correlation among models.

One of the advantages that random forests have over single classification trees is that they provide quantitative measures of variable importance, whereas with a simple classification tree, one is primarily limited to inferring variable importance from the frequency and location of the occurrence of variables in the model. One measure of variable importance is the decrease in the Gini index (a measure of impurity of observations at a particular node) induced by splitting on the variable, averaged over all trees 24.

In order to infer the relative importance of the predictor variables, we considered the measure of variable importance produced during the random forest run on the combined dataset, which is the most broadly representative dataset considered here. A plot of the variable importance measure for this dataset is shown in Figure 2; more important variables are shown as higher bars. The measure strongly indicates that the residue immediately 5' of the edited site (-1 position) is very important. These variable importance results are in agreement with previous work on C-to-U editing in mitochondria of Arabidopsis thaliana, which noted the -1, and -2 positions had highly non-random nucleotide distributions 9. However, the results here differ from the past study of Arabidopsis in that we find no indication that the -17 position has much importance in edited site recognition. Also previously noted was that for 93.1% of the time 9, the -1 position contained a pyrimidine, which is the data partition found by the classification trees.

The free energy results contrast with previous studies indicating that secondary structure was not important in edited site recognition 1619. Our results show free energy is a relatively important variable in the random forest analyses. These results therefore indicate that secondary structure, as measured by free energy of folding for the 41 nt region centered on an edited/non-edited site, does help in distinguishing edited from non-edited sites. Previous studies determined putative secondary structures for mRNA regions containing edited sites and looked for conserved structural motifs. In contrast, we used estimates of free energy of folding, which are much easier to compare quantitatively. It may be that secondary or tertiary structure is even more important in determining edited sites than shown here; however, secondary structure may not be effectively represented by the calculated estimates of free energy of folding analyzed.

Simple models based on nucleotides surrounding edited/non-edited sites and on estimated folding energies of those regions provide moderately accurate prediction of C-to-U RNA edited sites. More nuanced representation of secondary or higher-order structure in combination with variables based on the nucleotide positions found important here might improve models. Overall, the results strongly suggest that the C-to-U editing mechanism in plant mitochondria does not depend exclusively on the primary sequence immediately in the vicinity of the edited site.

MPC conceived, designed and coordinated the study. DSM carried out the programming and statistical analyses. Both authors wrote and approved the final manuscript.