Graph neural network and diffusion model for modeling RNA interatomic interactions.

Motivation: Ribonucleic acid (RNA) function is inherently linked to its 3D structure, traditionally determined by X-ray crystallography, Nuclear Magnetic Resonance, and Cryo-EM. However, these techniques often lack atomic-level resolution, highlighting the need for accurate in silico RNA structure prediction tools. Current state-of-the-art methods, such as AlphaFold3, Boltz1, RhoFold, or trRosettaRNA, rely on deep learning models that represent residues as frames and use transformers to learn relative positions. While effective for known RNA families, their performance drops for synthetic or novel families.

Results: In this work, we explore the potential of graph neural networks and denoising diffusion probabilistic models for learning interatomic interactions. We model RNA as a graph in a coarse-grained, five-atom representation and evaluate our approach on a dataset of small RNA substructures, known as local RNA descriptors, which recur even in non-homologous structures. Generalization is assessed using a dataset partitioned by RNA family: the training set consists of rRNA and tRNA structures, while the test set includes descriptors from all other families. Our results demonstrate that the proposed method reliably predicts the structures of unseen descriptors and effectively adheres to user-defined constraints, such as Watson-Crick-Franklin interactions.

Availability and implementation: The GraphaRNA source code is available on GitHub (github.com/mjustynaPhD/GraphaRNA); training/test datasets and pre-trained model weights are provided on Zenodo (zenodo.org/records/13750967).