UC San Diego details $15M effort to teach AI the shapes of RNA

AI models need more examples of RNA’s three-dimensional shapes before they can reliably predict them. UC San Diego described a plan on November 2, 2026, to supply that missing data: a three-year, $15 million Department of Energy project aiming to determine 1,000 protein-free RNA structures and train models to predict shape from genetic sequence alone.

The award itself was announced by DOE on October 8, as one of 12 Phase II Genesis Mission awards totaling $159 million. The university’s new account lays out the experimental approach and research targets for the project, led by biochemistry and molecular biophysics professor Navtej Toor.

The missing examples are physical structures

RNA is a single-stranded molecule present in all living cells. It folds into complex shapes that determine how it carries genetic information, helps make proteins and drives chemical reactions. Most RNA structures remain unknown, leaving researchers without a clear view of how these molecules perform their jobs.

The training-data gap is stark. According to the university, fewer than 200 protein-free RNA structures have been determined over the past 50 years. By comparison, AlphaFold, the protein-structure prediction tool, trained on 200,000 experimentally determined proteins. Toor’s team focuses on RNA without proteins attached so its shape can be studied without proteins obscuring it.

Obtaining those examples has been slow. X-ray crystallography can take years to resolve a protein-free RNA structure; one record-setting structure from Toor’s lab took five years. Another method, cryo-electron microscopy, sends electrons through frozen samples. But protein-free RNA can be too small and flexible to produce clear images.

A data shortage, not computing power, blocks accurate AI prediction of RNA shapes and limits the design of organisms for energy and biomanufacturing.

Navtej Toor, UC San Diego project lead

Stabilize the RNA, then let uncertainty guide experiments

Toor’s lab developed a scaffold technology to stabilize RNA for imaging. The university says this has reduced the time needed to determine a structure to two weeks, compared with the years crystallography can require. The lab determined 60 protein-free RNA structures in the past year, providing the experimental foundation for the larger project.

The proposed workflow connects model development directly to laboratory work. Models will identify RNA molecules whose shapes they predict least confidently. Researchers will image those molecules next, then retrain the models on the measured structures. Rather than collecting a fixed dataset before training begins, the team intends to use each round of predictions to guide further measurements.

Energy organisms first, human RNA later

The project will examine selected DOE model organisms, including two photosynthetic microbes that make fuels and chemicals from carbon dioxide, plus the bioenergy crops sorghum and poplar. The university identifies a practical problem behind that choice: current models do not reliably predict how genetic changes will affect an organism’s traits, partly because they omit RNA’s three-dimensional shape.

Toor’s intended output is an online program that accepts any RNA sequence and returns an accurate three-dimensional structure. He says it could eventually extend to human RNA and help drug developers obtain structures of their targets. Those are research ambitions, not capabilities delivered by the award.

UC San Diego leads with co-investigators Rhiju Das of Stanford, Wah Chiu of SLAC National Accelerator Laboratory and Daniel Haack of A-Form Solutions. Additional partners include Google and NVIDIA. The university disclosed that Toor co-founded A-Form Solutions, which could benefit financially from the research results.