RhoAgent Designed Active RNA, Including a Sixfold Ribozyme Result
Artificial Intelligence

RhoAgent Designed Active RNA, Including a Sixfold Ribozyme Result

RhoAgent designed RNA molecules that were synthesized and tested, including a Hovlinc ribozyme with about six times the cleaved fraction of one expert-truncated construct and an aptamer scaffold with 230% higher fluorescence. The tasks remain limited laboratory benchmarks, not therapeutic results.

NewTqnia Artificial Intelligence Desk Updated 3 min read
RhoAgent Designed Active RNA, Including a Sixfold Ribozyme Result

Quick summary

  • RhoAgent combines an AI agent with a diffusion model to design RNA structures and sequences.
  • Laboratory tests found active ribozymes with only 40% to 70% sequence similarity to natural references.
  • Its best Hovlinc ribozyme design showed about six times the cleaved fraction of a human-designed truncated construct.
  • The system designed research molecules, not approved therapies, and the benchmarks were chosen by the developers.

A research team led by the Chinese University of Hong Kong has built an agentic system that designs functional RNA and revises candidates using laboratory feedback. RhoAgent combines natural-language task planning with RDiff, a two-stage diffusion model that generates three-dimensional RNA backbones and compatible nucleotide sequences.

The central evidence is experimental: selected molecules were synthesized and tested, rather than judged only by a structure-prediction score.

From instructions to molecules

RhoAgent translates a requested function into structural constraints, base interactions and motifs that must be preserved. RDiff first proposes RNA backbones conditioned on those requirements, then produces sequences predicted to fold into the proposed structures. The agent evaluates geometric plausibility and motif preservation before choosing candidates for laboratory work.

Experimental results can then guide another design round. In the Hovlinc ribozyme task, the average cleaved fraction of the top five second-round variants was 12% higher than in the first round, according to the paper.

What worked in the laboratory

Designed ribozymes remained active despite sharing only 40% to 70% of their sequence with wild-type references. For a Hovlinc ribozyme whose experimental three-dimensional structure was unavailable, the best design produced about six times the cleaved fraction of a human-expert truncated construct.

RhoAgent also produced a Broccoli aptamer scaffold with 230% higher fluorescence than the wild-type aptamer and combined two functional cores into a single scaffold for a fluorescence resonance energy transfer system. Those tests support the claim that the platform can preserve more than one functional component.

Why this is more than a language-model demo

The system coordinates specialized structural and sequence models instead of asking a general chatbot to invent a molecule directly. More importantly, wet-lab measurements close the loop between computational proposals and biochemical activity. That distinguishes this work from AI-generated scientific claims that have not yet been tested. NewTqnia previously examined that verification gap in a report on three AI-generated scientific claims awaiting external checking.

Reality check

The experiments involve a limited set of ribozymes and aptamers selected by the developers. A sixfold improvement compares with one truncated human-designed construct, not every expert method. The study does not demonstrate a medicine, performance inside a patient or a fully autonomous laboratory. Human researchers still define goals, run experiments and decide how feedback enters the next round.

What comes next

Independent teams will need to reproduce the designed molecules and test the approach on less familiar RNA functions. A stronger assessment would compare RhoAgent with expert and computational baselines under identical budgets, then measure design success rates rather than only the best candidates. Integration with automated laboratories could shorten iteration cycles, but it would also require controls for experimental quality and model error.

Verified topics and entities

Sources and citations3 sources

Published by

N

NewTqnia Artificial Intelligence Desk

An institutional editorial team within NewTqnia

A new version of NewTqnia is ready.