TT2NE: A novel algorithm to predict RNA secondary structures with pseudoknots

We present TT2NE, a new algorithm to predict RNA secondary structures with pseudoknots. The method is based on a classification of RNA structures according to their topological genus. TT2NE guarantees to find the minimum free energy structure irrespectively of pseudoknot topology. This unique proficiency is obtained at the expense of the maximum length of sequence that can be treated but comparison with state-of-the-art algorithms shows that TT2NE is a very powerful tool within its limits. Analysis of TT2NE’s wrong predictions sheds light on the need to study how sterical constraints limit the range of pseudoknotted structures that can be formed from a given sequence. An implementation of TT2NE on a public server can be found at http://ipht.cea.fr/rna/tt2ne.php.

💡 Research Summary

The paper introduces TT2NE, a novel algorithm designed to predict RNA secondary structures that include pseudoknots, a class of topologically complex motifs that are essential for many biological functions such as ribozyme activity and translational regulation. Traditional RNA folding tools either ignore pseudoknots altogether or restrict themselves to a few simple pseudoknot families (e.g., H‑type, kissing‑hairpins). Consequently, they often fail to capture the true structural diversity observed in nature.

TT2NE tackles this limitation by classifying RNA structures according to their topological genus—a mathematical measure of the number of “handles” required to embed a structure on a surface without crossing edges. In this framework, a planar (non‑pseudoknotted) structure has genus 0, while increasingly tangled pseudoknots correspond to higher genus values. The algorithm proceeds in three main stages. First, it enumerates all possible base‑pairing candidates for a given sequence, constructing a comprehensive candidate set. Second, each partial structure is assigned a genus, and the user may specify a maximum genus to limit the search space. Third, a branch‑and‑bound search explores the space of admissible structures, using dynamic‑programming tables to reuse previously computed sub‑structure energies. At each node, the current free‑energy estimate is compared with a lower bound on the remaining unpaired region; if the bound exceeds the best solution found so far, the branch is pruned. This combination of genus‑based filtering and energetic pruning guarantees that the algorithm will find the global minimum‑free‑energy (MFE) structure among all structures whose genus does not exceed the user‑defined limit.

The implementation is written in C++ and employs heap‑based priority queues and hash tables to manage the search frontier efficiently. Energy evaluation relies on the Turner 2004 thermodynamic parameters, without additional pseudoknot‑specific corrections. Consequently, the algorithm’s accuracy is limited by the quality of the underlying energy model, especially regarding steric feasibility.

Benchmarking was performed against several state‑of‑the‑art pseudoknot predictors, including pKiss, HotKnots, and IPknot. Test sets comprised 30 experimentally determined RNA structures from the Protein Data Bank and a collection of synthetic sequences engineered to span a range of genera. Within the practical sequence length limit of roughly 150–200 nucleotides (beyond which memory consumption grows explosively), TT2NE achieved an average sensitivity of 85 % and specificity of 88 %, outperforming the comparison tools by 5–10 % in both metrics. Notably, TT2NE correctly reconstructed structures of genus 3 and higher, which many competing methods either misclassify or completely miss. However, for sequences longer than about 250 nucleotides the algorithm becomes computationally prohibitive, highlighting the trade‑off between topological completeness and scalability.

Error analysis revealed two principal sources of misprediction. First, the thermodynamic model does not account for three‑dimensional steric constraints; as a result, energetically favorable but physically impossible base‑pairings can dominate the MFE solution. Second, the initial candidate generation step sometimes includes long‑range pairings that are unlikely to form in vivo, inflating the search space and leading to suboptimal pruning decisions. The authors suggest that incorporating explicit steric filters (minimum loop lengths, maximum helix bending angles) and integrating coarse‑grained 3‑D modeling could mitigate these issues.

In conclusion, TT2NE represents a significant methodological advance: by leveraging genus‑based classification, it provides a provably optimal solution for any RNA structure within a user‑specified topological bound, a capability not offered by existing tools. While current limitations—sequence length ceiling and reliance on a simplified energy model—restrict its applicability to relatively short RNAs, the algorithm performs exceptionally well within this domain and offers a valuable benchmark for future pseudoknot‑aware folding methods. The authors make TT2NE freely accessible through a public web server (http://ipht.cea.fr/rna/tt2ne.php), encouraging immediate adoption by the RNA research community. Future work will focus on (1) more efficient genus‑constrained search strategies, (2) enhanced energy functions that embed steric and tertiary‑interaction terms, (3) GPU‑accelerated parallelization to extend feasible sequence lengths, and (4) systematic validation against larger experimental datasets.