gRNAde is a generative AI framework for RNA inverse design. gRNAde learns RNA design rules from 3D RNA structures, allowing it to capture complex tertiary interactions (like pseudoknots and non-canonical base pairs) with expert-level accuracy and design functional RNAs, including aptamers and ribozymes.

gRNAde stands for Geometric RNA Design; pronounced as ‘grenade’

⚙️ Need help with new RNA design problem? Check out the Example design notebook and please contact Chaitanya for collaborations!

🧬 New to 3D RNA modeling? Check out this curated reading and watch list for beginners

🤗 Data and Checkpoints: HuggingFace repository

🧪 Experimental Validation: Generative inverse design of RNA structure and function with gRNAde (bioRxiv preprint)

📄 Methodology: gRNAde: Geometric Deep Learning for 3D RNA inverse design (ICLR 2025 Spotlight)

• 3D RNA Design: Conditions on 3D coordinates to capture non-canonical pairings and tertiary motifs that 2D-only methods miss. Supports multi-modal design specifications including pseudoknotted secondary structures, 3D backbones, and partial sequence constraints (e.g., from fitness landscapes)

• Expert-Level Accuracy on Pseudoknots: Matches human expert performance on the Eterna OpenKnot Benchmark while being fully automated.

• Functional Design Campaigns: Capable of generating active ribozyme variants with high sequence divergence from the wild type.

• High-Throughput Pipeline with Screening: Includes built-in wrappers for RibonanzaNet and RhoFold to generate and filter millions of candidates automatically.

Set up your Python environment, install dependencies, and download checkpoints from HuggingFace following the instructions in docs/installation.md.

Optionally, download pre-processed datasets and/or raw 3D structures from HuggingFace, following the instructions in docs/data_setup.md. This data can be used for training new gRNAde models.

The main.py script can be used to train new gRNAde models. Simply run:

python main.py --config configs/default.yaml

The src directory houses all the code for the gRNAde design pipeline. The codebase is integrated with Weights&Biases for logging as well as hyperparameter tuning/running sweeps.

The fastest way to get started is with the design.py command-line script, which provides a production-ready interface for the complete design workflow:

python design.py --config configs/design.yaml

This automated pipeline will:

• Load model checkpoints and target structures
• Generate diverse candidate sequences with gRNAde (conditioned on 3D coordinates, secondary structure, and optional sequence constraints)
• Screen candidates with RibonanzaNet for structural accuracy
• Filter and rank designs based on configurable metrics
• Save top designs for experimental validation

See docs/creating_designs.md for detailed documentation, configuration options, and best practices.

For interactive exploration and customization, see the example design notebook.

All design campaigns from the paper are available in the projects/ directory with complete code and step-by-step READMEs to reproduce results and generate figures:

• projects/openknot_benchmark/: Eterna OpenKnot designs for riboswitches, ribozymes, viral elements, and other complex RNA pseudoknot structures.
• projects/rna_polymerase_ribozyme/: Functional RNA polymerase ribozyme engineering campaign.

The codebase corresponds to the first stable, experimentally validated version of gRNAde (v1.0.0) used in "Generative inverse design of RNA structure and function with gRNAde". All the code and data used for the previous ICLR 2025 Spotlight paper introducing the gRNAde methodology is available in release v0.3.2, with detailed instructions for reproducing the computational benchmarks: https://github.com/chaitjo/geometric-rna-design/releases/tag/v0.3.2

.
├── README.md
├── LICENSE
|
├── main.py                         # Main script for training and evaluating models
├── design.py                       # Command-line script for RNA sequence design
|
├── .env.example                    # Example environment file
├── .env                            # Your environment file
|
├── checkpoints                     # Model checkpoints
├── configs                         # Configuration files directory
├── data                            # Dataset and data files directory
├── wandb                           # W&B output directory
|
├── docs                            # Documentation directory
|   ├── creating_designs.md         # Guide to using design.py for RNA design
|   ├── installation.md             # Installation instructions
|   └── data_setup.md               # Data preparation guide
|
├── tools                           # Directory for external tools
|   ├── rhofold                     # RNA sequence to 3D structure prediction tool
|   ├── ribonanzanet                # RNA sequence to chemical mapping prediction tool
|   └── ribonanzanet_sec_struct     # RNA sequence to pseudoknotted secondary structure prediction tool
|
├── src                             # Source code directory for gRNAde
|   ├── constants.py                # Constant values for data, paths, etc.
|   ├── evaluator.py                # Evaluation metrics and computational filtering pipeline
|   ├── layers.py                   # PyTorch modules for building gRNAde models
|   ├── models.py                   # Main gRNAde model code
|   ├── trainer.py                  # Training loop for gRNAde
|   |
|   └── data                        # Data-related code
|       ├── clustering_utils.py     # Methods for clustering by sequence and structural similarity
|       ├── data_utils.py           # Methods for loading PDB files and handling coordinates
|       ├── dataset.py              # Dataset and batch sampler class
|       ├── featurizer.py           # Featurizer class
|       └── sec_struct_utils.py     # Methods for secondary structure prediction and determination
|
└── projects                        # Code for reproducing design campaigns in the paper
    ├── openknot_benchmark          # Eterna OpenKnot Benchmark for psuedoknotted RNA design
    └── rna_polymerase_ribozyme     # Generative Design of RNA polymerase ribozymes

@article{joshi2025generative,
  title={Generative inverse design of RNA structure and function with g{RNA}de},
  author={Joshi, Chaitanya K and Gianni, Edoardo and Kwok, Samantha LY and Mathis, Simon V and Lio, Pietro and Holliger, Philipp},
  journal={bioRxiv},
  year={2025},
  publisher={Cold Spring Harbor Laboratory}
}

@inproceedings{joshi2025grnade,
  title={g{RNA}de: Geometric Deep Learning for 3D RNA inverse design},
  author={Joshi, Chaitanya K and Jamasb, Arian R and Vi{\~n}as, Ramon and Harris, Charles and Mathis, Simon V and Morehead, Alex and Anand, Rishabh and Li{\`o}, Pietro},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2025},
}

@incollection{joshi2024grnade,
  title={g{RNA}de: A Geometric Deep Learning pipeline for 3D RNA inverse design},
  author={Joshi, Chaitanya K and Li{\`o}, Pietro},
  booktitle={RNA Design: Methods and Protocols},
  pages={121--135},
  year={2024},
  publisher={Springer}
}