Understanding patterns of rRNA evolution is critical for a number of fields, including structure prediction and phylogeny. The standard model of RNA evolution is that compensatory mutations in stems make up the bulk of the changes between homologous sequences, while unpaired regions are relatively homogeneous. We show that considerable heterogeneity exists in the relative rates of evolution of different secondary structure categories (stems, loops, bulges, etc.) within the rRNA, and that in eukaryotes, loops actually evolve much faster than stems. Both rates of evolution and abundance of different structural categories vary with distance from functionally important parts of the ribosome such as the tRNA path and the peptidyl transferase center. For example, fast-evolving residues are mainly found at the surface; stems are enriched at the subunit interface, and junctions near the peptidyl transferase center. However, different secondary structure categories evolve at different rates even when these effects are accounted for. The results demonstrate that relative rates and patterns of evolution are lineage specific, suggesting that phylogenetically and structurally specific models will improve evolutionary and structural predictions.

Structural elements in RNA molecules have a distinct nucleotide composition, which changes gradually over evolutionary time. We discovered certain features of these compositional patterns that are shared between all RNA families. Based on this information, we developed a structure prediction method that evaluates candidate structures for a set of homologous RNAs on their ability to reproduce the patterns exhibited by biological structures. The method is named SPuNC for 'Structure Prediction using Nucleotide Composition'. In a performance test on a diverse set of RNA families we demonstrate that the SPuNC algorithm succeeds in selecting the most realistic structures in an ensemble. The average accuracy of top-scoring structures is significantly higher than the average accuracy of all ensemble members (improvements of more than 20% observed). In addition, a consensus structure that includes the most reliable base pairs gleaned from a set of top-scoring structures is generally more accurate than a consensus derived from the full structural ensemble. Our method achieves better accuracy than existing methods on several RNA families, including novel riboswitches and ribozymes. The results clearly show that nucleotide composition can be used to reveal the quality of RNA structures and thus the presented technique should be added to the set of prediction tools.