Programmable RNA design with diffusion models and structure-aware conditioning
RDiff is an RNA design framework for generating functional RNA candidates through programmable structural and interaction constraints. The repository provides two complementary inference modules:
RDiff_seq— inverse-folding sequence design conditioned on RNA 3D structures, with optional secondary-structure, noncanonical-interaction, and fixed-motif constraints.RDiff_str— RNA 3D backbone generation conditioned on secondary structure, noncanonical interactions, or user-provided scaffolding motifs.
Repository scope. This public release is focused on inference. Training scripts, training datasets, and training dataloaders are not included.
- 🧬 Structure-conditioned RNA sequence design
- 🧱 Conditional RNA 3D backbone generation
- 🔗 Watson–Crick and noncanonical base-pair conditioning
- 🧩 Fixed-sequence constraints and motif scaffolding
- ⚙️ Batch inference with configurable diffusion sampling
RDiff/
├── RDiff_seq/ # RNA inverse folding and sequence design
├── RDiff_str/ # RNA 3D backbone generation and scaffolding
├── env_RDiff.yaml # Conda environment specification
└── LICENSE # MIT License
conda env create -f env_RDiff.yaml
conda activate RDiffPretrained checkpoints are available in the following directory:
https://drive.google.com/drive/folders/11o--a0hqlBrGhnD7BaJ1Wpfw3VinjDur?usp=sharing
After downloading, place the checkpoints in the corresponding directories:
RDiff_seq/RDiff_seq_ckpts/ # Sequence-design checkpoints
RDiff_str/RDiff_str_ckpts/ # Structure-generation checkpoints
RDiff_str/RDiff_scaffolding_ckpts/ # Scaffolding checkpoints
You may also provide absolute checkpoint paths in the commands below. Checkpoint names with an nc suffix denote subtype-level non-Watson–Crick-conditioned models; use these models when supplying detailed noncanonical interaction labels through NC_sparse or input_nc_npz.
Design RNA sequences conditioned on one or more RNA 3D structures:
cd RDiff_seq
CUDA_VISIBLE_DEVICES=0 python main.py \
--config configs/rdiff_seq_ss.py \
--no_wandb \
--mode inf_dir \
--inf_folder /path/to/pdb_or_directory \
--config.eval.output_fasta_file output.fasta \
--config.eval.n_samples=2500 \
--config.eval.model_path RDiff_seq_ckpts/ribo_woss_cluster/step_500000.pth--inf_folder accepts either a single .pdb file or a directory containing .pdb files. For fixed-sequence masks, secondary-structure conditioning, and detailed noncanonical interaction conditioning, see RDiff_seq/README.md.
Generate RNA backbones conditioned on a dot-bracket secondary structure:
cd RDiff_str
python inference_rdiff_str.py --config-name=rna_cww_ss_training \
inference.name=hhr_generation \
inference.weights_path=RDiff_str_ckpts/cww_ss_1250k.pth \
inference.num_samples=10 \
inference.target_length=49 \
"inference.ss_db_constrain='(((((....(..(((((..))))).).(((((((..))))))).)))))'" \
inference.diffusion.num_t=500 \
inference.seed=42For secondary-structure plus noncanonical interaction conditioning, motif scaffolding, and contig sweeps, see RDiff_str/README.md and RDiff_str/scaffolding_pipeline.md.
RDiff_seqwrites designed sequences to FASTA files.RDiff_strwrites generated RNA backbone structures as PDB files underinference.output_dir/inference.name/.
If you use RDiff in your work, please cite:
Programmable functional RNA design with a virtual scientist
Han Huang, Tianqing Zhang, Liang Hong, Ziqian Lin, Dongchen He, Jiuming Wang, Tian-Ying Wu, Huimin Li, Da Han, Le Yu, Jiyue Jiang, Yimin Fan, Yanruisheng Shao, Irwin King, Pengan Chen, Chun Kit Kwok, James J. Collins, and Yu Li. Nature Biotechnology (2026).
@article{rhoagent,
title = {Programmable functional RNA design with a virtual scientist},
author = {Han Huang, Tianqing Zhang, Liang Hong, Ziqian Lin, Dongchen He, Jiuming Wang, Tian-Ying Wu, Huimin Li, Da Han, Le Yu, Jiyue Jiang, Yimin Fan, Yanruisheng Shao, Irwin King, Pengan Chen, Chun Kit Kwok, James J. Collins, and Yu Li},
journal = {Nature Biotechnology},
year = {2026},
doi = {10.1038/s41587-026-03362-3}
}This project is released under the MIT License.