BPS2026 – NuFold and NuFold DB: Deep learning approach for RNA structure prediction and database of predicted structures

RNA plays various biological functions as non-coding RNAs (ncRNAs). Obtaining tertiary structure information of those RNAs is crucial to reveal their function in detail. However, only a fraction of the vast number of known RNA sequences has been experimentally resolved, mainly due to the time and cost of experiments. To address this challenge, we developed NuFold, a novel computational method that accurately predicts RNA 3D structures by leveraging a state-of-the-art deep learning architecture. Our NuFold is an end-to-end deep neural network that outputs all-atom 3D structures of RNA directly from its sequence. Notably, NuFold introduces a unique nucleobase-centric representation, which enables the modeling of flexible conformations, including details such as ribose puckering, contributing to achieving high accuracy in local structures. In benchmarks, NuFold outperforms existing energy-based methods and shows comparable performance to other recent deep learning-based approaches. While NuFold is publicly available, to further enhance accessibility for a broader community, we have performed large-scale structure predictions on known RNA sequences and established the NuFold RNA structure database (NuFold DB). NuFold DB stores more than 100,000 predicted structures for various model organisms known ncRNAs. Through our user-friendly interface, NuFold DB provides the predicted 3D structure, its confidence score (pLDDT), and an additional confidence measure from a comparison our predicted 3D structure with known secondary structures. This database serves as a central resource for researchers to quickly retrieve structural information for their RNAs of interest. NuFold and NuFold DB provide a powerful and accessible solution to accelerate RNA structure studies and facilitate RNA drug discovery. Our NuFold DB is available at https://nufold.kiharalab.org.