Deep learning unlocks de novo design of complex RNA pseudoknots

Complex RNA structures govern fundamental processes like translation and viral replication, but designing entirely new RNA folds from scratch has remained out of reach—until now. Scientists report today in Science that deep learning methods can now design intricate RNA pseudoknots as reliably as expert human players, marking a turning point for RNA engineering.

The work, led by Rhiju Das at Stanford University School of Medicine, with first author Jill Townley, harnessed the citizen science platform Eterna to pit AI algorithms against experienced human designers. In a series of challenges covering 57 different pseudoknot targets, AI methods solved over 95% of the problems—tested by single-nucleotide resolution chemical mapping, compensatory mutagenesis across ~50,000 sequences, and cryo-electron microscopy.

The AI-generated molecules not only folded correctly but formed well-ordered three-dimensional architectures, sometimes incorporating noncanonical tertiary interactions that the models had never been instructed to build. Cryo-EM revealed that the designs adopted entirely new 3D folds, distinct from any known natural RNA structure. This success was guided by an RNet foundation model trained on chemical mapping data, suggesting that accurate RNA design may be achievable without first solving 3D structure prediction—a long-standing bottleneck.

Why this matters: The ability to reliably create new RNA pseudoknots on demand opens the door to custom-designed RNA tools for therapeutics, diagnostics, and synthetic biology, such as new ribozymes, aptamers, and even redesigned molecular machines. With AI now matching human expertise in less than a year, the field shifts from trial-and-error to a design-build-test cycle that can be dramatically accelerated.

Caveats: The methods excel at secondary structure design but do not yet predict atomic-level 3D interactions, which remain important for more sophisticated RNA functions. Extending the approach to other complex functional RNA architectures will require further validation. Additionally, one target was not designable by any method, indicating room for improvement.

Reference: Townley J, Kladwang W, Baker D, et al. De novo design of RNA pseudoknots with deep learning. Science. 2026;393(6814):931-937. DOI: 10.1126/science.aeg6829