A new artificial intelligence method developed at Virginia Tech has matched one of the world's most advanced AI systems at predicting RNA structures, achieving the feat with significantly less training data than competitors require. The breakthrough, called RNAbpFlow, outperformed Google DeepMind's AlphaFold 3 in a blind test, correctly predicting the three-dimensional shape of 12 out of 14 RNA targets compared to AlphaFold 3's eight out of 14.

Why Does RNA Structure Prediction Matter So Much?

RNA molecules fold into specific three-dimensional shapes that serve as targets for drug design. Understanding these shapes is critical because drugs work by attaching to specific pockets in the folded structure. Without accurate shape predictions, researchers cannot design effective treatments. The real-world impact is already clear: risdiplam, one of the first small-molecule drugs designed to target RNA directly, works by latching onto a specific folded shape in an RNA molecule to treat spinal muscular atrophy, a leading genetic cause of infant death.

"How can you target an RNA if you don't have its shape? 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," said Sumit Tarafder, lead author of the study and a doctoral student in the Department of Computer Science at Virginia Tech.

Sumit Tarafder, Doctoral Student, Department of Computer Science, Virginia Tech

The challenge has always been that RNA is structurally flexible and severely underrepresented in biological databases, making it far harder to model than proteins. Most leading AI tools depend on large evolutionary sequence databases to infer structure, but these databases are notoriously difficult to assemble for RNA molecules.

How Does RNAbpFlow Work Differently?

Rather than searching through thousands of related sequences across different species to infer structure, RNAbpFlow uses a technique called flow matching. This approach belongs to the same broad class of generative AI that powers image generators like DALL-E and Midjourney. The method generates complete, all-atom three-dimensional structures in a single end-to-end process using only the RNA sequence and base pair information.

The elegance of the approach lies in its simplicity and efficiency. The model starts from complete noise and, guided by base pair information, folds into the correct three-dimensional shape. This process can generate multiple possible structures, allowing researchers to capture how the molecule actually moves and behaves in different conditions.

Steps to Understanding RNAbpFlow's Advantages

• Data Efficiency: RNAbpFlow requires no large evolutionary sequence databases, making it especially useful for RNA molecules with few known relatives in biological databases.
• Speed and Simplicity: The flow matching technique generates structures in a single process rather than requiring multiple inference steps, reducing computational complexity.
• Flexibility in Output: The method can generate as many different structural predictions as needed, helping researchers understand the range of shapes a molecule can adopt.
• Performance on Difficult Cases: RNAbpFlow excels when evolutionary data is thin or unavailable, including challenging cases like conserved structural elements from the SARS-CoV-2 genome and laboratory-built ribozymes tested in the study.

The research was published on June 30, 2026, in Nature Methods, one of the most selective journals in computational life sciences. Debswapna Bhattacharya, associate professor in the Department of Computer Science and senior author of the study, explained the team's philosophy behind the work.

"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.

Debswapna Bhattacharya, Associate Professor, Department of Computer Science, Virginia Tech

What Are the Limitations and Next Steps?

The researchers acknowledge that on larger, more complex RNA molecules, established tools that draw on evolutionary data still maintain an advantage. RNAbpFlow performs best in cases where that evolutionary data is limited or unavailable. However, this is precisely where the method offers the most value, since those are the cases where traditional approaches struggle most.

Tarafder is now leading development of an improved version of RNAbpFlow that will be submitted to CASP, the community-wide prediction competition where Google DeepMind's protein-folding breakthrough AlphaFold first drew global attention. This competition will provide another opportunity to benchmark the method against other state-of-the-art approaches.

In keeping with a growing push for reproducible science, the Virginia Tech team has released the full implementation, training data, and code publicly. The work was supported by the National Institutes of Health and the National Science Foundation.

"We owe a debt to taxpayers, and everything we're doing is open source and public. It's for the public good," said Bhattacharya.

Debswapna Bhattacharya, Associate Professor, Department of Computer Science, Virginia Tech

The implications for drug discovery are significant. Tools that can predict RNA shapes quickly and accurately could accelerate the search for breakthrough therapies for diseases including Huntington's disease, ALS (amyotrophic lateral sclerosis), certain cancers, and viral infections. By reducing the data requirements and computational burden of RNA structure prediction, RNAbpFlow could democratize access to these tools for researchers worldwide who lack access to massive computational resources or extensive sequence databases.