The same family of artificial intelligence that powers today's image generators is now being aimed at one of biology's hardest puzzles: the ever-changing, three-dimensional shapes of RNA. These are the molecules behind mRNA vaccines like the ones that prevent serious COVID-19 cases.
A new method built by two Virginia Tech computer scientists is matching one of the world's most advanced AI systems at the task of mapping RNA and is doing it with far less data.
The method, called RNAbpFlow, was described in a study published on June 30 in Nature Methods , among the most selective journals in computational life sciences. In a blind test against a widely used community benchmark, RNAbpFlow produced a correct overall structure for 12 of 14 RNA targets, compared with eight out of 14 for AlphaFold 3, the system from Google DeepMind. And RNAbpFlow reached that result without the large evolutionary sequence databases that most leading tools require.
"We asked whether we could leverage what data we have, and use additional knowledge from experiments to fill the data-gap and give RNA-based drug discovery a fair shot," said Debswapna Bhattacharya , associate professor in the Department of Computer Science and the study's senior author.
The impact of this work is not hypothetical. In 2020, the U.S. Food and Drug Administration approved risdiplam, a daily oral medication that treats spinal muscular atrophy, a leading genetic cause of death in infants. The treatment works by latching onto a specific folded shape in an RNA molecule and correcting how a gene is read. It was among the first small-molecule drugs designed to target RNA directly, and it helped turn a disease that once killed children before their second birthday into a manageable condition.
But finding drugs like risdiplam can be a slow process, largely because scientists cannot easily see the 3D shapes RNA folds into. RNA is structurally flexible and badly underrepresented in databases, which has made it far harder to model than proteins. Tools that can predict those shapes quickly and accurately could speed the search for the next breakthrough therapy for diseases from Huntington's and ALS to certain cancers and viral infections.
"How can you target an RNA if you don't have its shape?" said Sumit Tarafder, the study's lead author and a doctoral student in the department. "In the shape, there are pockets where a drug can attach. If you can't predict the shape, your pockets are wrong — and the drug won't work."
RNAbpFlow takes a different route than its better-known competitors. Rather than searching for thousands of related sequences to infer structure as AlphaFold-style systems do, it uses a technique called flow matching. This approach uses the same broad class of generative AI used to create images and generates complete, all-atom 3D structures in a single end-to-end process.
"We wanted to keep it simple and predict the structure from scratch, using just the sequence and the base pairs," Tarafder said. "The model starts from complete noise and, guided by those base pairs, folds into the right 3D shape. That's the beauty of flow matching, and we can generate as many structures as you want, which lets us capture how the molecule actually moves."
The method’s low data-dependence is the heart of the work. Most top AI approaches depend on deep data, including collections of related sequences from across species. These are notoriously difficult to assemble for RNA. RNAbpFlow needs none of them. That makes it especially useful for the many RNAs with few known relatives. Among the cases the team modeled were a conserved structural element from the SARS-CoV-2 genome and a laboratory-built ribozyme.
The researchers also highlight the limits of their work so far. On larger, more complex RNAs, established servers that draw on evolutionary data still hold an edge. But RNAbpFlow excels in challenging cases where that data is thin.
The work was led by Tarafder, a doctoral student and first author on the paper, and Bhattacharya, with support from the National Institutes of Health and the National Science Foundation. Tarafder is now leading work on an improved version of the method that will go to this summer's CASP, the community-wide prediction competition where Google DeepMind's protein-folding breakthrough first drew global attention.
In keeping with a growing push for reproducible science, the team has released the full implementation, training data, and code publicly.
"We owe a debt to taxpayers, and everything we're doing is open source and public," Bhattacharya said. "It's for the public good."
Original study : DOI 10.1038/s41592-026-03128-4
Nature Methods
Data/statistical analysis
Cells
RNAbpFlow: base pair-augmented SE(3) flow matching for conditional RNA 3D structure generation
30-Jun-2026