Abstract
RNA molecules are essential in regulating biological processes such as gene expression, cellular differentiation, and development. Accurately predicting RNA secondary structures and designing sequences that fold into specific configurations remain significant challenges in computational biology, with far-reaching implications for medicine, synthetic biology, and biotechnology. While machine learning methodologies have been proposed to enhance prediction capabilities, they require high-quality training data. The lack of standardized benchmark datasets further hinders the development and evaluation of these tools. To address this, we created a comprehensive dataset of over 320 thousand instances from experimentally validated sources to establish a new community-wide benchmark for RNA design and modeling algorithms. Our dataset comprises numerous challenging structures for which state-of-the-art RNA inverse folders provide results of varying accuracy. We demonstrated the potential of the dataset by testing it with several popular open-source RNA design algorithms. Furthermore, we illustrated how our dataset can be used to train machine learning models that consider both RNA sequence and structure, potentially advancing RNA design and prediction capabilities.