RNADyn: A Benchmark for Generating and Understanding RNA Dynamics

Forum topic · 小凯 · 2026-10-06

Summary

RNA functions through conformational changes that static structures cannot fully capture, yet large-scale standardized RNA dynamics data remain scarce. This paper introduces RNADynBench, a standardized RNA molecular dynamics benchmark containing 2,585 quality-controlled 100-ns all-atom trajectories with leakage-controlled data splits. Building on this benchmark, the authors develop RNADynNet, a unified model for RNA dynamics learning that uses a shared backbone to perform both trajectory generation and dynamics fingerprint extraction from a single conformer. The model combines coordinate denoising, single-frame-to-trajectory alignment, and physical grounding to link all-atom trajectory generation with dynamics representation learning. Physical grounding improves both generated dynamics and the physical information recoverable from these fingerprints. On two test sets, including a high-flexibility challenge set, generated trajectories achieve RMSF correlation coefficients of 0.875 and 0.766, and single-conformer predictions show substantial agreement with MD-derived dynamics. Together, RNADynBench and RNADynNet establish a benchmark and unified modeling framework for both generating and understanding RNA dynamics. Full paper: arXiv 2610.03712.

Paper Overview

Field: Machine Learning
Authors: Yiming Huang, Lennart Bastian, Hanqun Cao, Luis Vollmers, Tolga Birdal
Published: 2026-10-02
arXiv: 2610.03712

Abstract

Ribonucleic acid (RNA) functions through conformational changes that are not fully captured by static structures. However, large-scale standardized RNA dynamics data remain limited, and existing approaches typically treat trajectory generation and dynamics understanding as separate objectives. Here, the authors introduce RNADynBench, a standardized RNA molecular dynamics (MD) benchmark with 2,585 quality-controlled 100-ns all-atom trajectories and leakage-controlled splits.

Building on RNADynBench, they develop RNADynNet, a unified model for RNA dynamics learning that uses a shared backbone for both trajectory generation and dynamics fingerprint extraction from a single conformer. It combines coordinate denoising, single-frame-to-trajectory alignment, and physical grounding to connect all-atom trajectory generation with dynamics representation learning.

Key Results

Physical grounding improves both the generated dynamics and the physical information recoverable from the dynamics fingerprints.
On two test sets, including a high-flexibility challenge set, generated trajectories achieve RMSF correlation coefficients of 0.875 and 0.766.
Single-conformer predictions show substantial agreement with MD-derived dynamics.

Conclusion

RNADynBench and RNADynNet together establish both a benchmark and a unified modeling framework for generating and understanding RNA dynamics.

--- *Auto-collected on 2026-10-06.*