Meta-Scoring Approach to Curate Synthetic RNA 3D Structures

The structural diversity of RNA is central to molecular medicine, RNA-targeted drug discovery, and synthetic biology, where engineered RNA molecules are increasingly used in therapeutic, regulatory, and nanotechnology applications. While the rapid development of deep learning methods has heightened the demand for large, diverse structural datasets, high-quality RNA data remain scarce compared to protein datasets. Synthetic RNA structure datasets can help address this gap by enabling large-scale exploration of conformational space; however, their practical utility depends on careful curation and validation to ensure reliability. Here, we present a pipeline, RNA-SyntHub, for systematically constructing curated synthetic RNA datasets. The workflow integrates data from multiple structure-generation approaches with a meta-scoring function to identify low-quality models and select high-confidence candidates. As a proof of concept, we applied this method to a dataset of 447,402 synthetic RNA structures predicted by RFDiffusion, obtaining a curated subset of approximately 16,000 structures, further complemented with predictions from Boltz-1 and RNAComposer. This study provides both a methodological framework for generating and curating large collections of synthetic RNA 3D models under explicitly stated assumptions and a concrete dataset that supports methodological development, exploratory analyses, decoy-based applications, and selected deep learning workflows in RNA design, synthetic biology, and nanotechnology.