Unifying evolutionary and thermodynamic information for RNA folding of multiple alignments

Seemann, Stefan E.; Gorodkin, Jan; Backofen, Rolf

We presented a new method, named PETfold, which unifies probabilistic evolutionary and thermodynamic models to predict global RNA structures from multiple alignments. Highly conserved RNA structures are probably naturally selected due to their functional relevance, while less conserved sequences tend to fold in a state of minimum free energy. Combining both types of information increases the predictive power: we rely on phylogenetic information if the structure elements are highly conserved and extend the model by including folding energies to build the surrounding structure. The 46 Rfam seed alignments used for benchmarking are selected for their consistency. They might include manually optimized structure information, that are not achieved by sequence alignment algorithms. However, we are interested in the best possible structure prediction, which is often only found using conserved structure information in the alignment. The analyses of PETfold revealed that the selection of a reliably conserved substructure rel is less important than expected. This has been confirmed by estimating the statistical significance of the reliability scores, which shows that only high-reliability scores are significant. On the other hand, the combination of an evolutionary model with a thermodynamic model outperforms the widely used RNA structure predictors Pfold and RNAalifold.