RNAgrail: graph neural network and diffusion model for RNA 3D structure prediction

The function of RNA is intrinsically tied to its 3D structure traditionally explored by X-ray crystallography, NMR, and Cryo-EM. However, these experiments often lack atomic-level resolution, creating the need for accurate in silico RNA structure prediction tools. This need has driven advances in artificial intelligence (AI), which has already revolutionized protein structure prediction. Unfortunately, similar breakthroughs in the RNA field remain limited due to sparse and unbalanced structural data. Here, we introduce RNAgrail, a novel RNA 3D structure prediction method that focuses on RNA substructures using a denoising diffusion probabilistic model (DDPM). Unlike AlphaFold 3 (AF3), considered by many to be an oracle, RNAgrail allows expert users to define base pair constraints, offering superior flexibility and precision. Our method outperformed AF3 by 12% in terms of mean RMSD and by 24% in terms of mean eRMSD. Additionally, it perfectly reproduced the canonical secondary structure outperforming Af3 by 40% in terms of interaction network fidelity (INF). RNAgrail demonstrated robustness across diverse RNA motifs and families. Despite being trained exclusively on rRNA and tRNA, it effectively generalizes to new RNA families, thus, addressing one of the major challenges in RNA 3D structure prediction. These results underscore the potential of focusing on small RNA components and integrating user-defined constraints to significantly enhance RNA 3D structure prediction, setting a new standard in RNA modeling.