RNA molecules do much more than carry genetic messages. They fold into specific shapes that help control how genes are regulated, how proteins are made, and how enzymes function. One of the most challenging features to predict in RNA folding is the pseudoknot, a structure where loops and base pairs intertwine in complex ways. These structures are biologically important, but they are notoriously difficult and slow to predict using standard computational approaches.

Researchers at the University of Tokyo describe a new strategy that makes pseudoknot prediction far more efficient while preserving accuracy. Traditional methods attempt to model the entire RNA molecule at once, which leads to extremely high computational costs as sequence length increases. This makes large scale or routine analysis impractical.

Graphical representations of RNA structures and section-based pseudoknot classifications

Yellow lines: secondary structure base pairs; red lines: pseudoknot base pairs; blue regions: sections (unpaired nucleotides); black regions: non-sections. a) RNA structure in stem-loop and linear representations. b) Stem-loop (left) and linear (right) representations of the same structure. Blue arrows indicate sections (contiguous unpaired regions). The black arrow indicates a non-section where nucleotides participate in secondary structure pairing. c) Topologically distinct pseudoknot structures of both 2-clusters and 3-clusters.

The new approach takes a different path. Instead of treating the RNA as a single continuous structure, the algorithm first divides it into smaller, manageable regions called sections. These sections are stretches of unpaired bases identified from known secondary structures. The method then focuses on how pairs of sections interact to form pseudoknots, using a nearest-neighbor energy model combined with dynamic programming.

By narrowing the search space to the most informative section pairs, the algorithm dramatically reduces computational demands. The researchers show that biologically meaningful pseudoknots tend to appear in section pairs with large gains in minimum free energy. In fact, more than 90 percent of relevant interactions are found within just the top few percent of section pairs ranked by energy gain.

When tested on hundreds of transfer messenger RNA and RNase P RNA sequences, the method performed well. For simpler pseudoknot groupings, prediction sensitivity exceeded 0.9 and precision remained high. For more complex cases, the results revealed an interesting pattern. Early forming section interactions were predicted accurately, while later ones were not, suggesting that RNA folds sequentially as it is transcribed rather than settling into a single global energy minimum.

These findings provide both a practical tool and a deeper biological insight. The algorithm offers a scalable way to analyze RNA structure and also supports the idea that RNA folding is a stepwise, time-dependent process. Together, these advances help bridge computational efficiency with biological realism in RNA structure prediction.

Availability – All source code used to perform the analyses in this study, along with the complete datasets, is publicly available on Github: https://github.com/donn-liew/rna-pseudoknot-hierarchical