Sprecher
Beschreibung
The growing interest in RNA therapeutics and dynamic nucleic acid systems has drawn increasing attention to ribozymes as post- or cotranscriptional editors of RNA oligonucleotides. However, both the ribozyme and the remaining transcript consist of ribonucleotides that can potentially interact in undesirable ways, disrupting the ribozyme structure and reducing catalytic activity. While the prediction of ribozyme functionality from sequence using machine learning or physics-based models is a very active field of research, predicting the influence of the surrounding up- and downstream ribonucleotides on ribozyme functionality has received less attention.
In this work, we investigate whether 2D and 3D structure prediction methods, such as RibonanzaNet and Boltz-2, can be exploited to predict ribozyme functionality within the context of a larger oligonucleotide. By combining structural predictions from multiple models with a machine learning classifier, we develop an approach to assess context-dependent ribozyme activity based on deviations from a reference structure predicted for the isolated ribozyme. We evaluated the performance of several prediction algorithms and classifiers. Our final predictor requires only sequence input and does not rely on ribozyme-specific annotations such as catalytic residue positions or relevant secondary interactions. Instead, structural similarity metrics, such as root mean square deviation (RMSD) and a normalised two-dimensional base-pairing similarity score, serve as features to estimate ribozyme activity.
We validate our approach using experimental cleavage data for four ribozyme classes (HDV-like, Twister, Pistol, and Hatchet) across a diverse set of sequence contexts, achieving approximately 80% prediction accuracy. These results show that structure prediction models can serve as reliable indicators of ribozyme activity in complex synthetic contexts. This work supports the rational design of functional RNA molecules embedded in nucleic acid nanotechnology platforms and invites further exploration of predictive models for context-dependent RNA activity.
