RNA design stops rewriting sequences alone: RhoAgent constrains 3D geometry and lifts HOV ribozyme cleavage up to sixfold
Synopsis
RhoAgent moves RNA design from sequence and secondary-structure optimization to closed-loop design under tertiary geometric constraints. It pairs the diffusion generator RDiff with planner, executor, critic and memory modules, and iterates candidates on wet-lab feedback. On the HOV ribozyme, which lacks an experimental 3D structure, the best variant cleaved more than fivefold faster than the expert construct, with a reported maximum of sixfold; the Bro-Scaf6 aptamer reached 230% higher fluorescence than wild type.
Interpretation
RNA design tools can now generate candidates conditioned directly on base interactions and 3D functional motifs, rather than optimizing sequence or secondary structure alone. Earlier computational design mostly optimized sequence, built on secondary structure, or combined the two, thereby bypassing the noncanonical pairs, base stacking and long-range contacts that stabilize the global fold; fragment assembly was bounded by library availability, and deep learning models mostly designed sequences on fixed backbones. On a time-censored test set, RDiff designs recovered the input base interactions, with mean distance differences between designed and native structures of 0.12 Å, 0.53 Å and 0.10 Å for Watson–Crick, non-Watson–Crick and stacking constraints, respectively.
RhoAgent couples that generator to an agentic workflow that plans tasks, evaluates its own intermediates and adjusts strategy from experimental feedback. Existing generative paradigms were largely single-shot and lacked the iteration needed for complex engineering objectives, while general-purpose LLMs and biomedical agents were limited by missing domain tools or overly general workflows. Across 33 representative Rfam clans covering most annotated RNA space, RhoAgent reached a mean Watson–Crick base-pair recall of 0.93 for its top five designs, compared with 0.76 for the biomedical agent Biomni, while general-purpose LLMs performed poorly.
On the HOV ribozyme, which has no experimental 3D structure, the agent started from an 83-nucleotide expert truncation and produced more active truncated variants over two rounds. The expert construct came from systematic deletion assays that identified the minimal functional core, and the best first-round variant, HOV-08, already cleaved more than fivefold faster than that construct at pH 7.5. Round 2 tested ten candidates and five achieved more than fivefold improvement; the mean cleaved fraction of the top five variants was 12% higher in round 2 than in round 1, and pairwise sequence identity to the expert construct ranged from 39.8% to 62.7%.
The same pipeline scaffolds a fixed functional core and places two distinct aptamer cores within one RNA chain. Inverse-folding methods generate sequences for a fixed, complete tertiary structure and struggle to position a compact functional motif within geometrically diverse contexts. In the first round, Bro-Scaf6 reached 230% higher fluorescence than the wild-type Broccoli aptamer; 14 of 18 FRET candidates showed fluorescence in both donor and acceptor emission bands, with FRET ratios spanning 0.2 to 7.0.
Perspective
The demonstrated scope is ribozyme cleavage, fluorogenic aptamer brightness and two-aptamer FRET constructs characterized by in vitro transcription and gel electrophoresis, with designs generated by RDiff and prioritized by RhoAgent. The authors release RhoAgent and RDiff as open source as a foundation for a more autonomous design ecosystem.
The authors describe RhoAgent as currently a powerful executor rather than a full collaborator: it depends on explicit constraints, responds unpredictably to vague prompts, cannot automatically retrieve operable design principles from the literature, and its planner does not yet reliably turn underspecified high-level biological goals into testable strategies. RDiff optimizes a single conformation and does not capture ion-dependent folding, conformational switching or allosteric regulation; its optimal working scope of about 200 nucleotides limits large RNA devices. The scarcity of long, high-resolution RNA structures is a field-wide constraint.
