RNArefine

RNArefine is an AI-guided framework for atomic-level refinement of RNA three-dimensional structures, developed by Yang Zhang's group at the National University of Singapore together with collaborators at the Institute of Science Tokyo and the Kyushu Institute of Technology. It was posted to bioRxiv in June 2026 and addresses a persistent gap in RNA structural modeling: while AI-driven predictors have advanced rapidly, their output models frequently lack complete atomic detail or contain severe stereochemical distortions and incorrect local interactions that limit downstream use.

Rather than predicting folds from sequence, RNArefine takes an existing RNA model and repairs it. The framework couples a learned interaction predictor with a physics-based optimization engine in a hierarchical pipeline. A geometric attention network first predicts base-pairing and base-stacking interactions directly from the input 3D coordinates; these predicted interactions are then combined with physics-based force fields to guide a two-step refinement consisting of Monte Carlo conformational sampling followed by L-BFGS energy optimization. The pipeline also completes coarse-grained or reduced models into full-atom structures via structural fragment superimposition.

This hybrid strategy places RNArefine alongside other learning-plus-physics RNA modeling tools, but focused specifically on the refinement stage rather than de novo folding. By improving the physical realism of predicted and experimentally derived models, it complements rather than competes with sequence-based structure predictors.

#Key Features

• Geometric attention interaction predictor: A geometric attention network predicts base-pairing and base-stacking interactions from 3D coordinate inputs, supplying the learned restraints that drive refinement.
• Two-step physics-based refinement: Predicted interactions are integrated with physics-based force fields to guide Monte Carlo conformational sampling followed by L-BFGS energy optimization, improving structures while preserving global topology.
• Coarse-grained to full-atom completion: Reduced or coarse-grained models are converted into complete atomic structures through structural fragment superimposition.
• Broad applicability: Refinement improves both sequence-based prediction models and cryo-EM-derived structures across stereochemical quality, interaction fidelity, and physically penalized accuracy.
• Open-source release: Source code, pretrained model weights, and a hosted web server are all publicly available.

#Technical Details

RNArefine's core learned component is a geometric attention network trained to predict RNA base-pairing and base-stacking interactions from 3D coordinates. The predicted interactions act as restraints within a force-field-guided refinement protocol: an initial Monte Carlo conformational sampling stage explores nearby conformations, and a subsequent L-BFGS energy minimization stage drives the structure toward a physically realistic, lower-energy state while maintaining the input's overall fold. The implementation is released primarily in C++ with a Python component for interaction prediction; pretrained models, statistical parameters, and structure fragments are distributed via Zenodo. In large-scale benchmarks spanning sequence-based prediction models and cryo-EM-derived structures, RNArefine consistently improved stereochemical quality, interaction fidelity, and physically penalized structural accuracy. Applied to blind CASP16 RNA prediction models, it improved ranking scores for 28 of the top 30 groups.

#Applications

RNArefine is intended as a post-processing step for RNA structural modeling workflows. Researchers who generate RNA models from sequence-based predictors can pass those models through RNArefine to remove stereochemical distortions and correct local base interactions before downstream analysis. The same refinement improves cryo-EM-derived models, making it useful in experimental structure determination pipelines. By producing physically realistic full-atom coordinates, it supports structural and therapeutic applications such as RNA function analysis and structure-guided design, and the hosted web server lowers the barrier for users without local computational infrastructure.

#Impact

RNArefine targets a recognized bottleneck: the structures produced by modern RNA predictors are often topologically plausible but physically imperfect, limiting their utility for mechanistic and therapeutic work. By demonstrating consistent gains across diverse model sources and improving ranking scores for 28 of the top 30 groups on blind CASP16 RNA targets, it establishes refinement as a robust, broadly applicable complement to RNA structure prediction. As an open-source framework with released code, weights, and a web server, it is positioned for adoption across the RNA modeling community. The work is a preprint awaiting peer review, and its benchmarks are computational evaluations of structural quality.