Designing RNA sequences that fold into target 3D structures is a fundamental challenge in therapeutics. Existing RNA inverse‑folding methods typically condition on backbone coordinates and generate only discrete 1D sequences, neglecting continuous side‑chain orientations that are crucial for structural stability.\ To address this gap, we explicitly incorporate side‑chain geometry into the RNA generative process. Unlike co‑design frameworks that diffuse in separate raw spaces, we propose Geometric Optimized Latent Diffusion (GOLD), which projects both sequence and structure onto a unified latent manifold. GOLD employs a novel GeoVAE that elegantly integrates discrete RNA sequences $x$ with continuous side‑chain angles $\theta$ via an encoder $z = f_{\text{enc}}(x, \theta)$, mapping them into a smooth latent space $\mathcal{Z}$ and enabling synergistic co‑diffusion.\ To improve structural validity, we introduce a training‑free, gradient‑guided sampling mechanism. During sampling, decoded sequence probabilities and side‑chain geometries are instantly converted into 3D structures, and analytical gradients $\nabla_{z}\mathcal{L}_{\text{phys}}$ derived from physical constraints are computed. This dynamic feedback actively steers the generated sequence toward a physically stable state.\ Extensive experiments demonstrate that GOLD achieves state‑of‑the‑art performance across sequence recovery, secondary‑structure validity, and 3D foldability metrics, outperforming existing baselines. The code and data are publicly available at https://github.com/darkflash03/GOLD.\
Review