Limits of deep-learning-based RNA prediction methods

In recent years, tremendous advances have been made in predicting protein structures and protein-protein interactions, but progress on RNA structure, alone or in complex with other macromolecules, has lagged behind despite recent methodological developments. It remains unclear whether reported gains in accuracy generalize to novel RNA structures. Here, we benchmark 10 deep learning methods on an independent, non-redundant dataset of monomeric RNAs and RNA complexes. For monomeric RNA, AlphaFold3 and Boltz-1 achieve the best overall performance, but accuracy depends greatly on similarity to the training set: predictions are reliable mainly for RNAs with well-defined or recurring folds, such as L-shaped and simple loop structures, while structurally divergent RNAs remain poorly modelled. Extending the evaluation to RNA complexes, AlphaFold3 and Boltz-1 substantially outperform HelixFold3 and RosettaFoldNA, with no statistically significant difference between the two leading methods on most measures. Together, these results show that current methods are constrained by limited, structurally biased training data and recognize recurring motifs rather than generalize to novel RNA folds.