LLM Agent–Driven ncRNA Design via Intrinsic Features and Structure-Guided Feedback

Develop an LLM agent–driven pipeline that generates and analyzes non coding RNA sequences, under guidance from human-designed prompts. Each pipeline automatically integrates RNA structural features, conditional diffusion-based generative model and internet search

• Tool Inputs for non-coding RNA sequence design and analysis:
  • 3D structural reconstructions
  • Conditional diffusion model outputs
  • Web search data
  • PubMed abstract retrieval, followed by targeted web searches
• Tool-Calling Agent implementation
• ReAct Agent implementation
• Click here for quick presentation of preliminary results

• Large Language Models (Claude Sonnet 4 (20250514) Anthropic.
• LangGraph (2024): Low-level orchestration framework for building, managing, and deploying long-running, stateful agents
• RiboDiffusion model (paper, GitHub)
• DRfold2 model (paper, GitHub)
• DSSR computational tool paper
• RNA-FM: Interpretable RNA Foundation Model from Unannotated Data for Highly Accurate RNA Structure and Function Predictions paper
• RNAcentral DATABASE of non-coding RNA (ncRNA) sequences
• NIST Chemistry WebBook to get some physicochemical properties
• ViennaRNA predicting and comparing RNA secondary structures
• IUPAC code for nucleotides and amino acids
• Docker
• Ubuntu 22.04, Windows 11
• Visual Studio Code
• Torch

I adapted the same conda environment for both LLM agents and RiboDiffusion model. However, I installed DRfold2 in a Docker container running Ubuntu 22.04 because ARENA package requires Linux for compilation (see RNA_agents/Dockerfile). The Large Language Model (LLM) requires the key, please get it here. I use an NVIDIA GeForce RTX 4060 with 8 GB VRAM and 32 GB of RAM to run DRfold2 and Ribodiffusion models. A single simulation step takes about 5–10 minutes.

• Clone the repository::

git clone https://github.com/PavelPll/RNA_agents.git cd RNA_agents 

• Install DRfold2 inside a Docker container:

cd RNA_agents git clone https://github.com/leeyang/DRfold2.git drfold2 git clone https://github.com/pylelab/Arena.git drfold2/Arena cd drfold2 mkdir file_exchange\fasta_input && mkdir file_exchange\pdb_output docker build -t drfold_image ../ docker run --gpus all -it --name drfold_container -v .:/opt/drfold2 drfold_image bash Run inside container: wget --header="User-Agent: Mozilla/5.0" https://zhanglab.comp.nus.edu.sg/DRfold2/res/model_hub.tar.gz tar -xzvf model_hub.tar.gz rm -rf model_hub.tar.gz cd Arena make Arena exit Go back to RNA_RAG folder: cd .. 

• Install RiboDiffusion:

cd RNA_agents git clone https://github.com/ml4bio/RiboDiffusion cd RiboDiffusion Model checkpoint can be downloaded from here. https://drive.google.com/drive/folders/10BNyCNjxGDJ4rEze9yPGPDXa73iu1skx Another checkpoint trained on the full dataset (with extra 0.1 Gaussian noise for coordinates) can be downloaded from here. https://drive.google.com/file/d/1-IfWkLa5asu4SeeZAQ09oWm4KlpBMPmq/view Download and put the checkpoint files in the RiboDiffusion/ckpts folder. 

• Set up a conda environment:

conda create -n rna_agents python=3.11.11 conda activate rna_agents pip install -q torch==2.5.1 torchvision --index-url https://download.pytorch.org/whl/cu121 pip install requests matplotlib pip install transformers sentence-transformers langchain langchain-community langchain_anthropic pip install torch_geometric==2.3.1 torch_scatter==2.1.1 torch_cluster==1.6.1 pip install fair_esm==2.0.0 ml_collections==0.1.1 conda install -c conda-forge dm-tree=0.1.7 pip install biopython==1.80 pip install -U ddgs pip install wikipedia pip install easy-entrez pip install langgraph 

• Install ViennaRNA from here

• Large Language Model (LLM) ReAct Agent–Driven RNA Design via Structural Features and conditional ncRNA Diffusion Model

  notebooks/RNA_reactAgent.ipynb 

  Possible prompts/queries are summarized in the PDF presentation

• A speculative attempt to model RNA evolution using an LLM-based tool agent

  notebooks/RNA_toolAgent.ipynb 

  • Model input: FASTA file with initial RNA sequence (e.g. trnaGlycine_Asgard_group_archaeon.fasta from data/processed/rna_evolution_seed folder;
  • Model output: ancestral_sequence in FASTA and PDB formats with detailed description of each evolution step (evolution_steps.txt) in data/processed/rna_evolution folder.

This project is licensed under the [NAME HERE] License - see the LICENSE.md file for details