Limits of deep-learning-based RNA prediction methods

Abstract In recent years, tremendous advances have been made in predicting protein structures and protein-protein interactions. However, progress in predicting RNA structure, either alone or in complex with other macromolecules, has been less prominent, though some recent developments have been reported. It remains unclear whether the improved prediction accuracy is sustained for novel RNA structures. Here, we use an independent benchmark to evaluate the performance of the latest methods. First, we show that state-of-the-art methods can sometimes predict the structure of single-chain RNA strands, with accurate models observed for RNAs with well-defined or regular secondary structures. Next, our evaluation was extended to RNA complexes, where prediction accuracy was notably higher for those involving extensive canonical base pairing. Additionally, a structural similarity analysis revealed that prediction success strongly correlates with resemblance to known structures, indicating that current methods recognise recurring motifs rather than generalising to novel folds. Finally, we also noted that the accuracy estimates for RNA models are far from accurate. Therefore, it is not possible to reliably identify the correctly predicted models with today’s methods. Competing Interest Statement The authors have declared no competing interest. Footnotes Update version with many changed/improvements. In short, we do find that the best methods accurately predict the structure of one third to slightly more than half of the structures/complexes. The performance is similar between the best methods (AlphaFold3 and Boltz-1) and only slighlty worse for other methods. However, we do note that basically all structures that can be predicted accurately share structural similarity (not necessarily homology though) with entries in PDB. Thie means that curren tstate of the art methods can not be used to predict the structure of "novel" RNA-molecules. Further, the ability to estimate the accuracy of the predictions (the pTM and ipTM values) is significantly worse than for proteins. This limits the ability to do large scale structure prediction studies and we do believe this is an important area for improvements