A new generative model called La-Ribo designs an RNA molecule's sequence and its three-dimensional shape at the same time, instead of treating them as separate steps.
Researchers behind La-Ribo use a technique called flow matching - a generative method related to the diffusion models that power image generators - to jointly produce an RNA's nucleotide sequence and its folded structure. The model keeps a simplified backbone scaffold and encodes each nucleotide's identity and local shape as a compact latent representation, then a separate decoder fills in the full atomic detail. To train it, the team built a dataset of 168,561 RNA structures, combining real experimental data with predictions from three existing folding models, including 10,631 structures generated specifically for this work. In tests, La-Ribo scored higher than prior design methods on designability - whether a generated sequence actually folds into its intended shape - across different computational budgets and two separate folding-verification models.
RNA-based drugs and lab tools depend on getting sequence and fold right together, and most existing design software still treats those as two separate problems to solve in sequence. La-Ribo's trained model can also be reused to redesign a sequence for an already-fixed shape without retraining, which points toward cheaper, faster iteration for RNA engineering.
It's a computational result, not a lab-validated drug - the real test is whether any of these designable RNAs fold the way the model predicts once someone actually synthesizes them.