RNAFlow is a flow matching model for protein-conditioned RNA sequence-structure design. Its denoising network integrates an RNA inverse folding model and a pre-trained RosettaFold2NA network for generation of RNA sequences and structures.
Paper: RNAFlow: RNA Structure & Sequence Design via Inverse Folding-Based Flow Matching, ICML 2024
External dependencies are contained in environment.yml
conda env create -f environment.yml
Additionally, please install torch-scatter, torch-cluster, and openmm (compatible with your CUDA version).
You also need to install NVIDIA's SE(3)-Transformer (please use SE3Transformer in this repo to install).
conda activate rnaflow
cd RoseTTAFold2NA/SE3Transformer
pip install --no-cache-dir -r requirements.txt
python setup.py install
cd ../../
Please see tutorial.ipynb for information on how to design your own RNA sequences and backbone structures, given a target backbone protein pocket and protein sequence.
To generate the necessary files to run RF2NA during the inference process, run the following commands from the main directory (need to run both commands regardless of what split you are using):
python rnaflow/data/process_data.py --dataset rnaflow/data/rf2na_dataset.pickle --full_process False
python rnaflow/data/process_data.py --dataset rnaflow/data/seq_sim_dataset.pickle --full_process False
This will create a folder called rnaflow/data/rf_data containing the necessary files to fold sequences with RF2NA. Then run the following command from the main directory:
python scripts/inference_rnaflow.py
The script will print RMSD and sequence recovery for each generated sample. PDBs for the final complex and each structure in the trajectory will also be saved.
Pre-processed datasets for both splits are given at rnaflow/data/rf2na_dataset.pickle and rnaflow/data/seq_sim_dataset.pickle
To preprocess the dataset on your own, run the following command from the main directory:
python rnaflow/data/process_data.py --pdbbind_csv rnaflow/data/pdbbind_na.csv \
--pdb_out_dir rnaflow/data/pdbs \
--rf2na_test_csv rnaflow/data/rf2na_test_val_set.csv \
--dataset rnaflow/data/rf2na_dataset.pickle \
--full_process True
To retrain RNAFlow, get RF2NA weights from https://files.ipd.uw.edu/dimaio/RF2NA_apr23.tgz and save to RoseTTAFold2NA/network/weights/RF2NA_apr23.pt. To generate the necessary files to run RF2NA during the training process, run the following commands from the main directory:
python rnaflow/data/process_data.py --dataset rnaflow/data/rf2na_dataset.pickle --full_process False
python rnaflow/data/process_data.py --dataset rnaflow/data/seq_sim_dataset.pickle --full_process False
This will create a folder called rnaflow/data/rf_data containing the necessary files to fold sequences with RF2NA. Then run the following command from the main directory:
python scripts/train_rnaflow.py