Predicting the 3D structure of RNA has long remained a bottleneck in structural biology. Modeling RNA is significantly more complex than proteins: the molecule possesses extreme flexibility, and experimental data is catastrophically scarce. According to a publication in Nature Machine Intelligence, the Yang-Server team presented trRosettaRNA2—a deep learning system that attempts to break these limitations. The main architectural solution here is an auxiliary secondary structure module (SS-prior). It was trained on extensive secondary structure data without waiting for rare 3D templates to appear. This allows the model to turn the chaos of sequences into rigid spatial frameworks, relying on information about base pairing. The result is evident: in the CASP16 blind testing, the Yang-Server group using this model became the top automated server, leaving even AlphaFold 3 behind.
Technical superiority of trRosettaRNA2 is provided by the structure-aware attention mechanism. Unlike standard models that only look for statistical correlations, the system accounts for the physical geometry of the molecule directly during calculations. As follows from the developers' report, this allows for the generation of not just one frozen image, but a set of different conformers—spatial variations of the same molecule. At the same time, trRosettaRNA2 makes do with significantly fewer parameters and computational resources than competitors. During a demonstration using ribonuclease P RNA, the model successfully restored the conformational ensemble without any experimental inputs, identifying structural diversity that traditional algorithms simply ignore.
For business, the success of trRosettaRNA2 means a transition from piece-meal modeling to the serial design of synthetic RNA sensors and targeted drugs. In our view, this is a direct path to a radical reduction in R&D cycles: expensive "wet" laboratories are partially replaced by high-performance digital forecasting. However, questions remain regarding the stability of predictions in dynamic environments. The accuracy boundary of the current approach hits architectural limits of neural networks when working with extremely mobile molecules. Pharma giants should take note: for now, these are high-precision digital approximations, and final validation in the cell still remains a mandatory stage, albeit on a much narrower selection of candidates.