HomeAIAI Science & HealthCUHK’s RhoAgent Virtual Scientist Designs Working RNA

CUHK’s RhoAgent Virtual Scientist Designs Working RNA

A team at the Chinese University of Hong Kong has published RhoAgent, an AI “virtual scientist” that designs RNA candidates and then refines them with experimental feedback, producing working ribozymes and aptamers in laboratory tests. The paper, “Programmable functional RNA design with a virtual scientist,” appeared in Nature Biotechnology on 19 November 2026. A CUHK Department of Computer Science and Engineering release dated 20 November says the work was led by Professor Yu Li. The results are laboratory findings on specific ribozymes and aptamers. They are not a therapy, and neither the abstract nor the university release presents the work as a clinical finding.

What laboratory tests of RhoAgent showed

The Nature Biotechnology abstract describes the system as an agentic one that “designs RNA candidates and iteratively refines them using experimental feedback.” At its core is RDiff, “a diffusion-based generative tool conditioned on base interactions and functional motifs.” The CUHK release adds that RDiff is a two-stage cascaded diffusion model: it first generates RNA backbones from those constraints, then designs nucleotide sequences compatible with the generated structures, while allowing specified sequence segments to be preserved. RhoAgent, the university says, translates natural-language design instructions into structural and sequence constraints, scores candidates on geometric plausibility, structural consistency and motif preservation, and can send designs back for another computational round or take in wet-lab feedback.

The published tests, as summarised in the abstract, come in four parts. First, the system generated “diverse and active ribozymes with only 40–70% sequence similarity to their wild-type counterparts.” The CUHK release says selected designs from several ribozyme families showed cleavage activities “comparable to or greater than those of the corresponding wild-type ribozymes” despite that sequence distance. Second, the system was applied to the hovlinc ribozyme, which the abstract calls “structurally uncharacterized.” Cleavage was improved “up to sixfold relative to the human-expert-truncated construct.” CUHK describes that figure as the best-performing design, at “approximately six times” the expert-truncated construct, and says a second round of diffusion-based resampling left the mean cleaved fraction of the top five variants 12% higher than in the first round.

Third, it generated a fluorogenic aptamer scaffold that, the abstract says, “improves fluorescence by 230% relative to the wild-type aptamer.” The university names that scaffold as a Broccoli aptamer. Fourth, the system designed three-dimensional scaffolds that place two distinct functional cores in space, creating aptamer-based Förster resonance energy transfer (FRET) systems and, the abstract says, “preserving the binding activities of both components.”

The paper’s code-availability note, visible on the Nature page, says RhoAgent and RDiff are available on GitHub under the ml4bio organisation. The authors list is headed by Han Huang and several co-first authors, with Yu Li as a senior author at CUHK, matching the university’s account of who led the work.

That combination of a language-driven design loop and a laboratory check sits close to other recent TSN coverage of AI for molecular design. Boltz’s $110 million Series A was raised for molecular design tools, while a Westlake team has described Pinal as designing working proteins from text. The work also follows TSN’s account today of Mass General Brigham’s THREADS pathology foundation model, another peer-reviewed research result rather than a clinical product.

A sixfold lab gain is not a therapy

It is easy, with phrases such as “virtual scientist” and a 230% fluorescence gain, to read this paper as a step into RNA medicine. That reading overreaches the published record. The Nature abstract places programmable RNA design in the setting of synthetic biology and therapeutics as a research problem; the results it then reports are cleavage of ribozymes, fluorescence of an aptamer scaffold, and a FRET construction that keeps both binding activities. Those are laboratory assays. CUHK’s closing remark that linking RhoAgent to automated experimental platforms “could further accelerate the design and validation of functional RNA for applications in RNA biology, synthetic biology, and therapeutics” is a forward look, not a finding from this study.

The comparisons themselves are also tightly specified. The 40–70% sequence-similarity result is about activity against wild-type ribozymes, not about a drug-like molecule. The “up to sixfold” hovlinc figure is relative to a human-expert-truncated construct, and the 12% second-round lift is a mean over the top five variants in the university’s account, not a clinical endpoint. The 230% fluorescence improvement is against the wild-type aptamer. None of those baselines is a patient, a trial arm or a licensed product.

Until the authors publish something else, RhoAgent is a research system that designed working RNA in the lab, with an iterative loop between a diffusion model and experimental feedback. That is already a substantial result. It is not a therapeutic RNA, and it should not be written up as one.

Sources

Related stories

Follow @tsnmediahub on X for breaking AI news as it happens.

Follow on X

Share this story

Latest stories

  1. ECMWF Starts Forecast-in-a-Box Weather Pilot in Malawi
    AI Science & Health

More in this category

Latest stories

  1. ECMWF Starts Forecast-in-a-Box Weather Pilot in Malawi
    AI Science & Health

Explore the TSN Atlas: Every topic we cover, mapped. Click a bubble to dive in.