ChironRNA

ChironRNA is an all-atom diffusion model for refining RNA three-dimensional structures, developed by Jingyi Li, Jian Wang, and Nikolay V. Dokholyan at the University of Virginia School of Medicine and posted to bioRxiv in March 2026. Experimentally derived and computationally predicted RNA structures frequently contain geometrically invalid features—severe steric clashes where atoms overlap, and missing atoms that leave the model incomplete. ChironRNA targets this refinement step, cleaning up such defects so that downstream analysis and modeling can proceed on physically plausible structures.

The model applies E(3)-equivariant graph neural networks within a denoising diffusion framework, a combination that has become influential in molecular structure generation because it respects the rotational and translational symmetries of 3D coordinates. Rather than predicting structure from sequence, ChironRNA occupies the increasingly important niche of structure refinement, complementing RNA folding predictors whose raw outputs often require cleanup before they are usable.

#Key Features

• Steric clash resolution: ChironRNA regenerates the atoms of clashing nucleotides step by step, removing overlaps that violate basic physical constraints; it achieves roughly 80% clash reduction on more than 80% of the test set.
• Missing-atom completion: The model reconstructs absent atoms, completing partial structures that conventional refinement tools struggle to repair.
• E(3)-equivariant denoising: An equivariant graph neural network (EGNN) predicts the noise removed at each diffusion step, ensuring outputs transform correctly under rotation and translation.
• Hierarchical coarse-to-fine design: A coarse-grained diffusion stage, where each nucleotide is represented by a five-point summary, is paired with an all-atom stage to capture both global geometry and atomic detail.

#Technical Details

ChironRNA combines two diffusion stages: a coarse-grained model that represents each nucleotide with a five-point coarse representation and an all-atom model that operates on full atomic coordinates. Both stages use E(3)-equivariant graph neural networks to predict and remove noise across the diffusion trajectory, iteratively regenerating the atoms of clashing or incomplete nucleotides. The method operates in distinct training and generation phases and is evaluated on a held-out test set of RNA structures, where it reports about 80% clash reduction on over 80% of cases. Performance is strongest on structures with fewer than 200 nucleotides, with the model successfully reconstructing missing atoms in cases where conventional refinement approaches prove insufficient.

#Applications

ChironRNA is intended for structural biologists and RNA modelers who need to clean up RNA structures before downstream use—whether those structures come from experimental determination with residual geometric errors or from computational folding predictors whose outputs contain clashes and gaps. By serving as a refinement step, it can be slotted into existing RNA structure-prediction and analysis pipelines to produce physically valid models suitable for visualization, docking, or further simulation.

#Impact

ChironRNA brings equivariant diffusion, already prominent in protein and small-molecule generation, to the comparatively underserved problem of all-atom RNA structure refinement. By directly tackling steric clashes and missing atoms, it addresses a practical bottleneck that limits the usability of both experimental and predicted RNA models. The approach is narrow by design—it refines rather than predicts structure—and its strongest results are on smaller RNAs under 200 nucleotides. As a recent preprint without a confirmed public code or weights release, independent validation and broader adoption remain to be established.