Scientific Computing: Artificial Intelligence Learns the Language of RNA

The versatile biomolecule ribonucleic acid (RNA) performs numerous tasks in cells and serves as an important tool in modern medicine. RNA-based vaccines and therapeutics use RNA to specifically intervene in biological processes. However, researchers still know very little about the structures and functions of many RNAs within the cell. Researchers at the Karlsruhe Institute of Technology (KIT) and Forschungszentrum Jülich have now developed a new AI foundation model. The NucleicBERT model achieves high accuracy in various prediction tasks and provides insights into the biological relationships it has learned. Pre-trained on supercomputers using approximately 30 million RNA sequences, it can be adapted to a wide range of scientific applications. The researchers report their findings in Nature Machine Intelligence.

A Foundation Model for Diverse Research Applications

For example, NucleicBERT can predict which regions of an RNA molecule interact with one another, how its secondary structure forms, and how mutations affect its function. All the model requires is a single sequence. “Our model can do more than just accelerate RNA research,” says Dr. Alexander Schug, department head at the Scientific Computing Center at KIT. “We can also investigate which biological relationships NucleicBERT has learned, thereby gaining a better understanding of how it arrives at its predictions.” The new model bridges two worlds of data: the rapidly growing collection of known RNA sequences and the limited experimental knowledge of their structure and function.

or, September 10, 2026