Tsukiyama, S.; Li, Y.; Sato, K.; Kurata, H.; Zhang, Y.

Considerable progress has been made in AI-driven RNA structure prediction, but the resulting models often lack complete atomic details or suffer from severe stereochemical distortions and incorrect local interactions. We present RNArefine, an AI-guided hierarchical framework for atomic-level RNA structure refinement. RNArefine first predicts base-pairing and base-stacking interactions using geometric attention networks and then integrates the interactions with physics-based force fields to guide a two-step refinement strategy consisting of Monte Carlo conformational sampling followed by L-BFGS energy optimization. Large-scale benchmark experiments on both sequence-based prediction models and cryo-EM-derived structures demonstrated that RNArefine consistently improves stereochemical quality, interaction fidelity and physically penalized structural accuracy while preserving global topology. When applied to blind CASP16 RNA prediction models, RNArefine improved ranking scores for 28 of the top 30 groups. These results establish RNArefine as a robust open-source framework for transforming raw RNA folds into physically realistic atomic models for downstream structural and therapeutic applications.