• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Sunday, November 22, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Biology

Virtual Scientist RhoAgent Designs Functional Programmable RNAs and Enhances Molecular Activity

by
November 21, 2026
in Biology
Reading Time: 6 mins read
0
Virtual Scientist RhoAgent Designs Functional Programmable RNAs and Enhances Molecular Activity

Virtual Scientist RhoAgent Designs Functional Programmable RNAs and Enhances Molecular Activity

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

A research team led by Han Huang, Tianqing Zhang, Liang Hong, Ziqian Lin, Dongchen He, Jiuming Wang, and colleagues, alongside senior investigators James J. Collins and Yu Li, has introduced RhoAgent, an artificial intelligence system that autonomously designs functional ribonucleic acid molecules and refines them using iterative experimental feedback. Reported in Nature Biotechnology, the platform addresses a longstanding obstacle in synthetic biology and therapeutics: the difficulty human experts and conventional computational methods face when attempting to directly program the complex three-dimensional structures that govern RNA functions. By coupling an autonomous reasoning framework with RDiff, a specialized generative diffusion tool conditioned on base interactions and functional structural motifs, the system successfully engineered active ribozymes exhibiting only 40 to 70 percent sequence similarity to their wild-type counterparts, increased the cleavage efficiency of the structurally uncharacterized hovlinc ribozyme by up to sixfold relative to human-expert-truncated baselines, and produced fluorogenic aptamer scaffolds that improved fluorescence output by 230 percent.

Functional RNA design has historically relied on optimizing primary nucleotide sequences to fold into predefined, rigid secondary structures or static backbones, an approach that frequently fails to achieve target biological activities or accommodate tertiary interactions. RhoAgent overcomes this limitation by integrating hierarchical planning, fine-grained geometric control, and closed-loop experimental feedback into a coordinated virtual scientist architecture. At the center of this framework operates an open-source large language model, Qwen3-235B-A22B, functioning as the core cognitive engine. The system bridges the divide between high-level natural language instructions and low-level geometric and biophysical constraints through a specialized planner module, which loads past dialogue history from a memory module and incorporates workflow templates to establish comprehensive context for each biomolecular design task.

When presented with an abstract biological objective, such as generating a stable scaffold for a twister ribozyme, the planner module employs chain-of-thought reasoning to deconstruct the instruction into explicit, executable specification sets for the downstream generative model. This decomposition requires the planner to systematically extract essential geometric parameters, sequence constraints, and relevant biological prior knowledge, select the most effective conditioning modalities among secondary structure representations, base-interaction tensors, or fixed motif definitions, and confirm constraint feasibility using an integrated computational tool library. To generalize across diverse biochemical design challenges without necessitating parameter fine-tuning, the planner relies on in-context learning, dynamically retrieving proven tool-usage routines from a bank of few-shot examples embedded directly within its system prompt.

Execution of the planned design specifications is directed by an executor module, which orchestrates the structural and sequence sampling routines of RDiff while managing a multidimensional prioritization pipeline. The executor translates the high-level directives into discrete tool calls and oversees a cascaded generative process that first produces three-dimensional backbone structures and subsequently designs corresponding nucleotide sequences. To ensure candidate viability, the executor evaluates generated structures through a hierarchical verification pipeline of increasing complexity, starting with geometric sanity checks that identify steric clashes and abnormal adjacent nucleotide distances, progressing to secondary structure consistency evaluations, and culminating in tertiary structure verification using predictive models such as AlphaFold3.

This verification process yields a composite numerical score vector that measures steric geometry validity, adherence to specified base interactions, overall structural consistency, and the retention of critical functional motifs. To enable the central language model to interpret and reason over these technical outputs, the numerical metrics are translated into structured textual descriptions. Generated RNA candidates are then ranked and prioritized using a weighted combination of these scores, with broad weight ranges established by the planner according to the specific functional goals of the task and granular weights dynamically fine-tuned by the executor during generation rounds.

The platform maintains design quality through a dual-loop automated optimization architecture operating across distinct operational timeframes. Within an individual generation round, an inner-loop critic agent monitors intermediate outputs, performing immediate quality assurance by assessing whether the volume of candidates meeting all constraints satisfies predefined thresholds; if strict failure criteria are met, the critic triggers the immediate re-execution of specific generation substeps. Across broader design cycles, an outer-loop planner acts as a meta-optimizer by analyzing candidate performance, including empirical data returned from laboratory experiments. By examining these validation results, the planner updates evaluation metric weights and tightens constraint parameters, such as structural consistency thresholds if structural collapse is detected, effectively learning from the posterior distribution of prior iterations to progressively narrow the structural search space.

The structural backbone and sequence generation itself is driven by RDiff, a deep generative model that factors RNA design into a two-stage cascaded conditional diffusion process modeling the joint distribution of structure and sequence. First, the structure module samples three-dimensional backbone coordinates from a conditional distribution via diffusion over the rigid-body group SE(3). Second, the sequence module performs sequence assignment conditioned on both the sampled backbone geometry and target constraints. Both stages are governed by conditional reverse-time stochastic differential equations where drift and diffusion terms define the noise schedule, a standard Wiener process supplies stochastic noise, and the score function guides trajectories toward the learned data manifold. To incorporate experimental feedback directly into generation, RDiff features a resampling mechanism capable of corrupting viable designs back to intermediate timesteps and regenerating them under revised guidance.

RDiff handles biological requirements using complementary soft and hard conditioning strategies. Secondary structure patterns and base-interaction tensors are encoded as continuous latent representations that softly condition the denoising neural network rather than imposing rigid, unyielding boundaries. In contrast, essential functional elements, such as catalytic ribozyme cores or specific aptamer binding pockets, are implemented as hard constraints through motif scaffolding, operating analogously to conditional in-painting. Target sequences are divided into fixed functional motif segments and designable scaffold regions; during reverse diffusion, coordinates and nucleotides at motif positions remain strictly fixed while the surrounding non-motif scaffold is progressively denoised, compelling the generative network to construct supporting frameworks that are sterically and geometrically compatible with the active site.

To train RDiff, the researchers assembled a composite structural dataset combining high-resolution experimental records from the Protein Data Bank with synthetic structural models. Experimental RNA structures from the Protein Data Bank were filtered to retain only models with resolutions of 4 angstroms or better, complete chains, and at least 25 percent base-pairing across their nucleotides. To broaden structural diversity beyond the limited collection of experimentally resolved RNAs, the team augmented the training corpus with sequences sourced from the bpRNA and Ribonanza databases, using Boltz-2 to predict their three-dimensional conformations. Only synthetic structures demonstrating high agreement with secondary structure annotations, defined by an F1-score of at least 0.8, were retained, producing an augmented dataset of 28,768 structures clustered using MMseqs2 at 80 percent sequence identity into 11,849 clusters. Evaluation of RDiff was conducted on an independent test set of structures deposited after October 1, 2022, alongside 33 representative RNA clans retrieved from the Rfam database.

Laboratory validation confirmed the practical efficacy of RhoAgent across multiple distinct functional RNA classes. In addition to generating active synthetic ribozymes with significant sequence divergence from wild-type forms and engineering fluorogenic aptamer scaffolds that enhanced target fluorescence by 230 percent, the platform successfully designed complex three-dimensional architectures that spatially positioned two separate functional cores. By maintaining the independent binding activities of both constituent domains, RhoAgent produced dual-core constructs that functioned as aptamer-based Förster resonance energy transfer systems, resolving a persistent barrier in compositional RNA design where combining separate functional domains typically disrupts individual binding conformations.

Empirical characterization of the designed RNAs involved systematic in vitro biochemical protocols. Short templates were prepared by appending reverse-complement T7 promoter sequences to ribozyme sequences and transcribing them using T7 RNA polymerase in reactions containing ribonucleotide triphosphates, ribonuclease inhibitor, dithiothreitol, and inorganic pyrophosphatase at 37 degrees Celsius for 14 hours, followed by DNase I digestion. Long templates were constructed with flanking XhoI and BamHI restriction sites, amplified via polymerase chain reaction using PrimeSTAR Max DNA polymerase, restriction digested, gel purified, and transcribed under identical enzymatic conditions. Products including hammerhead, twister, and CPEB3 ribozymes were quantified on a NanoDrop One spectrophotometer and resolved via denaturing polyacrylamide gel electrophoresis using 10 percent polyacrylamide gels containing 8 molar urea under 300 volts, followed by SYBR Gold nucleic acid staining and gel imaging.

Cleavage activity of the generated HOV ribozymes was evaluated across varying buffer conditions by denaturing and renaturing transcribed RNAs in thermal cyclers across a temperature gradient from 75 down to 12 degrees Celsius in Tris-HCl buffers at pH 7.5 and pH 8.5 containing 150 millimolar potassium chloride. Cleavage reactions were initiated by introducing magnesium chloride at concentrations ranging from 0 to 10 millimolar, incubated at 37 degrees Celsius for 36 hours, quenched with ethylenediaminetetraacetic acid, and separated by denaturing electrophoresis. Aptamer fluorescence characterization utilized spin-column purified, Qubit-quantified RNAs, incubating Broccoli variants with 10 micromolar DFHBI-1T fluorophore at 472 nanometer excitation and 507 nanometer emission, and Pepper variants with 2 micromolar HBC-620 at 577 nanometer excitation and 620 nanometer emission on a Tecan Spark plate reader, with Förster resonance energy transfer monitored at 472 nanometer excitation and 620 nanometer emission.

Computational and experimental analyses also delineated key operational boundaries of the current system. RhoAgent presently functions primarily as an autonomous specification executor rather than an open-ended scientific collaborator, exhibiting unpredictable outputs when supplied with underspecified prompts such as requests to design a stable RNA without precise constraints. Furthermore, because the planner cannot yet independently extract operable molecular design principles directly from scientific literature, human-supplied domain knowledge remains necessary for each new task. Generative sampling via RDiff is optimized for static three-dimensional conformations and does not explicitly co-optimize multistate ensembles, leaving ion-dependent dynamic folding, allosteric regulation, and conformational switching unaddressed. The optimal generative window is bounded at approximately 200 nucleotides, limiting immediate application to very large assemblies like long non-coding RNAs, complete ribosomal elements, and complex allosteric aptazymes, constraints reinforced by the field-wide scarcity of long, high-resolution experimental RNA structures.

Subject of Research: Agentic artificial intelligence for programmable three-dimensional functional RNA design

Article Title: Programmable functional RNA design with a virtual scientist

Article References: Huang, H., Zhang, T., Hong, L., Lin, Z., He, D., Wang, J., Wu, T.-Y., Li, H., Han, D., Yu, L., Jiang, J., Fan, Y., Shao, Y., King, I., Chen, P., Kwok, C. K., Collins, J. J., & Li, Y. (2026). Programmable functional RNA design with a virtual scientist. Nature Biotechnology. https://doi.org/10.1038/s41587-026-03362-3

Image Credits: AI Generated

DOI: 10.1038/s41587-026-03362-3

Keywords: RhoAgent, RDiff, RNA design, synthetic biology, ribozymes, hovlinc ribozyme, fluorogenic aptamers, FRET systems, Qwen3-235B-A22B, diffusion models, motif scaffolding, in vitro transcription, denaturing PAGE, biomolecular engineering, Nature Biotechnology, AlphaFold3, Boltz-2

News Source: Gregory Coleman. (November 21, 2026). Virtual Scientist RhoAgent Designs Functional Programmable RNAs and Enhances Molecular Activity. Scienmag.

Tags: AlphaFold3biomolecular engineeringBoltz-2denaturing PAGEDiffusion modelsfluorogenic aptamersFRET systemshovlinc ribozymein vitro transcriptionmotif scaffoldingNature BiotechnologyQwen3-235B-A22BRDiffRhoAgentribozymesRNA designsynthetic biology
Share12Tweet7Share2ShareShareShare1

Related Posts

Review Maps Multistep Host Networks Directing Human Papillomavirus Entry and Trafficking

Review Maps Multistep Host Networks Directing Human Papillomavirus Entry and Trafficking

November 22, 2026
Improved Double-Buffer Extraction Yields High-Purity Genomic DNA from Rare Himalayan Tree

Improved Double-Buffer Extraction Yields High-Purity Genomic DNA from Rare Himalayan Tree

November 22, 2026

Wastewater Predator Outperforms Soil Microbe in Tackling Mixed Waterborne Bacterial Pathogens

November 22, 2026

Emirates Internal Medicine Conference Highlights Advances Across Critical Care and Pharmacotherapy

November 22, 2026

POPULAR NEWS

  • LLM-Enhanced BERTopic Improves Topic Coherence and Diversity in Specialized Tourism Corpus

    LLM-Enhanced BERTopic Improves Topic Coherence and Diversity in Specialized Tourism Corpus

    29 shares
    Share 12 Tweet 7
  • Hemodialysis Frailty Profiles Show No Significant Link to Hospitalization After Adjustment

    29 shares
    Share 12 Tweet 7
  • Rare Neural Autoantibodies Identified in Cerebrospinal Fluid in Schizophrenia Spectrum Disorders

    29 shares
    Share 12 Tweet 7
  • Pyrogallic Acid Restores Doxycycline Susceptibility in Multidrug-Resistant Enterotoxigenic Escherichia Coli

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

LLM-Enhanced BERTopic Improves Topic Coherence and Diversity in Specialized Tourism Corpus

Hemodialysis Frailty Profiles Show No Significant Link to Hospitalization After Adjustment

Rare Neural Autoantibodies Identified in Cerebrospinal Fluid in Schizophrenia Spectrum Disorders

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 89 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.