RNA molecules are essential players in our cells, involved in many crucial processes such as carrying genetic information, regulating genes, and catalyzing chemical reactions. To perform these functions, RNA folds into specific shapes, known as secondary structures, formed by the way its building blocks, nucleotides, pair up and interact. Understanding these folded shapes helps scientists learn how RNA works and can even guide the development of new therapies.

However, predicting RNA secondary structures accurately has long been a challenge. Traditional laboratory methods that determine RNA structures are often slow, expensive, and not suited to analyze large datasets. Computational tools exist but face limitations in both speed and accuracy, especially when dealing with complex structures like pseudoknots, where parts of the RNA fold back and pair in intricate ways.

A team of researchers at Northwestern Polytechnical University has developed a new computational approach that uses deep learning to improve the prediction of RNA secondary structures. Their model, called DSRNAFold, combines information from RNA sequences with the structural context, meaning it considers both the linear sequence of nucleotides and how different parts of the molecule may interact over short or long distances.

One key innovation in their approach is the use of a “phased learning strategy.” This means the model is trained step-by-step, reducing the risk of overfitting, where a model learns to perform well on specific training data but fails to generalize to new RNA sequences. Additionally, the model incorporates biological constraints on how nucleotides pair, guiding it to focus on realistic folding patterns instead of random or biologically impossible structures.

When tested on multiple datasets, DSRNAFold showed superior performance compared to existing methods. It was particularly strong in recognizing pseudoknots and in predicting outcomes of chemical mapping experiments, which help scientists verify RNA structures in the lab. These improvements mean researchers can now rely on more accurate computational predictions to better understand RNA’s role in health and disease.

Availability – The DSRNAFold source code used in this study and all analyzed datasets are available at Zenodo (https://doi.org/10.5281/zenodo.15545891). The DSRNAFold web server is available for use at http://www.nextbioai.com/.