TransUFold: Unlocking the structural complexity of short and long RNA with pseudoknots

The RNA secondary structure is like a blueprint that holds the key to unlocking the mysteries of RNA function and 3D structure. It serves as a crucial foundation for investigating the complex world of RNA, making it an indispensable component of research in this exciting field. However, pseudoknots cannot be accurately predicted by conventional prediction methods based on free energy minimization, which results in a performance bottleneck. To this end, we propose a deep learning-based method called TransUFold to train directly on RNA data annotated with structure information. It employs an encoder-decoder network architecture, named Vision Transformer, to extract long-range interactions in RNA sequences and utilizes convolutions with lateral connections to supplement short-range interactions. Then, a post-processing program is designed to constrain the model's output to produce realistic and effective RNA secondary structures, including pseudoknots. After training TransUFold on benchmark datasets, we outperform other methods in test data on the same family. Additionally, we achieve better results on longer sequences up to 1600 nt, demonstrating the outstanding performance of Vision Transformer in extracting long-range interactions in RNA sequences. Finally, our analysis indicates that TransUFold produces effective pseudoknot structures in long sequences. As more high-quality RNA structures become available, deep learning-based prediction methods like Vision Transformer can exhibit better performance.