RNADyn: A Unified Benchmark and Model for Predicting RNA Molecular Dynamics
RNADyn：預測與理解 RNA 動態變化的統一基準與 AI 模型
Traditional approaches separate RNA trajectory generation from dynamics understanding, hindered by a lack of standardized datasets. To address this, researchers developed RNADynBench, containing 2585 high-quality, 100-ns all-atom molecular dynamics (MD) trajectories. Built on this, RNADynNet utilizes a shared backbone to perform both trajectory generation and dynamics fingerprint extraction from a single conformer. The model achieves high RMSF correlations of up to 0.875, demonstrating excellent prediction accuracy.
Key points
Introducing RNADynBench
Provides 2585 quality-controlled, 100-ns all-atom molecular dynamics trajectories with leakage-controlled splits.
Unified RNADynNet Architecture
Employs a shared backbone to integrate both all-atom trajectory generation and dynamics fingerprint extraction.
Physical Grounding & Alignment
Combines coordinate denoising, single-frame-to-trajectory alignment, and physics-based constraints to enhance accuracy.
High-Precision Dynamics Prediction
Achieves RMSF correlations between 0.766 and 0.875, closely matching traditional, computationally expensive MD simulations.
How it works
Why it matters
RNA function is dictated by dynamic conformational changes rather than static structures, but traditional molecular dynamics (MD) simulations are computationally prohibitive. By delivering a standardized benchmark and a unified AI model, RNADyn enables researchers to accurately predict and analyze RNA dynamics from a single static conformer at a fraction of the computational cost, accelerating RNA drug discovery and structural biology.
Who it affects
- AI Researcher
- AI Developer
How to use it
- 1RNA Drug Discovery and Target Identification
- 2Accelerating MD Simulations as a Surrogate Model
Limitations & caveats
- Trajectories are limited to 100-ns, which might not capture longer-scale RNA folding or massive conformational transitions.
- The model's accuracy heavily relies on simulation data, which may still have gaps compared to physical wet-lab RNA observations.
Related
Less Decoder is More Encoder: Extracting Robust 3D Geometric Representations via Novel View Synthesis
減少解碼器反而增強編碼器：從新視角合成中提煉強大三維幾何表徵
This paper reveals how expressive decoders dilute geometric learning in Novel View Synthesis, and proposes SNAP—a self-supervised framework that restricts decoders to force encoders to learn robust 3D representations.
4DCodeBench: Benchmarking AI Agents on 4D Inverse Graphics and Dynamic Scene Code Generation
4DCodeBench：評估 AI Agent 動態場景 4D 反向圖形學與程式碼生成能力的全新基準
4DCodeBench is a new benchmark designed to evaluate AI agents' ability to reconstruct 4D dynamic scenes from videos by generating executable graphics and physics code.
What Should World Models Forget? Stratified Retention for Continual Adaptation
世界模型該遺忘什麼？以「分層保留」實現持續適應環境的能力
This paper argues that world models must not avoid all forgetting, proposing 'stratified retention' to distinguish permanent physical laws from dynamic, environment-specific facts that require timely revision.