Vibeleaderboard
← All Intel
Intel / article

Programmable functional RNA design with a virtual scientist

Source
nature.com
Date

Context

RNA molecules do useful jobs when they fold into the right shape. Some fold into ribozymes, which cut RNA at a chosen site; others fold into aptamers, which grip a specific target, including dyes that only glow once held. Designing new ones is hard because the function follows from the folded three-dimensional structure, and the published methods and human designers mostly work on the sequence and its flat base-pairing pattern instead. A paper in Nature Biotechnology on 19 November 2026, Programmable functional RNA design with a virtual scientist, describes a system built to attack that gap directly. The authors, a group at the Chinese University of Hong Kong with collaborators including James J. Collins at MIT, call the system RhoAgent, and its abstract states both the problem and what it is for: human expertise and existing computational methods struggle to program the structures that govern function.

RhoAgent is described as an agentic system, which here means a program that generates candidate RNAs, takes experimental measurements of how they performed, and refines the next round from that feedback instead of producing one batch and stopping. The generator underneath it is RDiff, a diffusion model, the same family of generative method used for images, conditioned on which bases should interact with which and on the functional motifs that have to be preserved. The authors' code release splits it into two halves that match the task: one generates an RNA three-dimensional backbone from the required structure and motifs, and the other designs a sequence that would fold into a given structure. Holding a functional core fixed while the model builds new structure around it is what the authors' code release calls motif scaffolding, and it is what lets the system keep a part that is known to work and change everything else.

The reported results are four, in rising order of difficulty. The system generated ribozymes that were active while sharing only 40% to 70% of their sequence with the natural versions they were derived from, which is the basic check that it is designing rather than copying. It then raised the cutting efficiency of the hovlinc ribozyme, whose structure has not been characterized, by up to sixfold against the construct a human expert had trimmed by hand. It produced a fluorogenic aptamer scaffold that improved fluorescence by 230% over the wild-type aptamer. Finally it designed scaffolds holding two different functional cores at a set distance from each other, producing aptamer-based FRET systems, where energy passes from one component to the other only when the two are correctly positioned; the authors report that both components kept their binding activity, which they name as a key barrier in combining RNA parts. These are the authors' own measurements as stated in the abstract, and only the abstract is readable without a subscription, so the experimental conditions behind each number, and how the comparisons were run, are not checkable from outside the paywall.

Read the source www.nature.com
Recommended reads
Comments

Checking sign-in…

Loading comments…