RNA-FrameFlow is a generative model for 3D RNA backbone design based on SE(3) flow matching.
Accompanying paper: 'RNA-FrameFlow: Flow Matching for de novo 3D RNA Backbone Design', by Rishabh Anand*, Chaitanya K. Joshi*, Alex Morehead, Arian Rokkum Jamasb, Charles Harris, Simon V Mathis, Kieran Didi, Bryan Hooi, Pietro Liò.
- Oral at ICML 2024 Structured Probabilistic Inference & Generative Modeling Workshop:
openreview- Spotlight at ICML 2024 AI4Science Workshop:
openreview
# clone repository
git clone https://github.com/rish-16/rna-backbone-design.git
cd rna-backbone-design/
# create python environment
conda create -n rna-bb-design python=3.10 -y
conda activate rna-bb-design
# install dependencies
chmod +x master_rna_install.sh
bash master_rna_install.shCaution
Do take note of the compatibility between PyTorch and CUDA versions. We used pytorch-cuda v11.8 in the installation script but you should change this depending on your system's NVCC version. You can find the target version using nvidia-smi and looking for CUDA Version.
Download RNASolo (3.48GB) containing the ~14K structures used to train our model at a resolution .pdb file format (we used a cutoff date 5 January 2024). In the project rna-backbone-design repository,
# create data directory
mkdir data/rnasolo; cd data/rnasolo
# download RNAsolo
wget https://rnasolo.cs.put.poznan.pl/media/files/zipped/bunches/pdb/all_member_pdb_4_0__3_326.zip
unzip all_member_pdb_4_0__3_326.zip # unzips all PDB files
mv all_member_pdb_4_0__3_326.zip ../ # moves ZIP archive out of new file directoryWe provide a preprocessing script process_rna_pdb_files.py that prepares the RNA samples used during training. Again, in the main project directory,
# from ./
python process_rna_pdb_files.py --pdb_dir data/rnasolo/ --write_dir data/rnasolo_proc/When you visit data/rnasolo_proc/, you should see a bunch of subdirectories representing the root of the PDB entry names. Each subdirectory contains pickled versions of the PDB files which capture some important structural descriptors extracted from the atomic records; this is for easier bookkeeping during training. You'll also notice a rna_metadata.csv file. Keep track of the filepath to this CSV file – it contains metadata about the pickled RNA files and the relative filepaths to access them during training.
Your directory should now look like this:
.
├── src
│ ├── analysis
│ ├── data
│ ├── experiments
│ ├── models
│ ├── openfold
│ └── tools
├── configs
│ ├── base.yaml
│ └── inference.yaml
├── data
│ ├── rnasolo
│ └── rnasolo_proc
│ └── rna_metadata.csv
└── camera_ready_ckpts/
Important
Our training relies on logging with wandb. Log in to WandB and make an account. Authorize WandB here.
We use 4 RTX3090 40GB GPUs via DDP to train our model for 120K steps, which took ~15 hours. We train on sequences of length between 40 and 150. For more specific experimental details, look at configs/config.yaml.
# run training
python train_se3_flows.pyAfter training, the final saved checkpoint can be found at ckpt/se3-fm/rna-frameflow/last.ckpt directory saved locally (not part of this repo); this ckpt directory is created automatically by wandb. We also store intermediate checkpoints for your reference. You can rename and shift this last.ckpt file where necessary to run inference.
Alternatively, you can use our camera-ready baseline checkpoint. The config files necessary can be found inside camera_ready_ckpts/.
We provide a brief description of the different schemes to represent all-atom RNA molecules used throughout the codebase in
src/DATA_REPR.md.
Our strongest baseline model's trained weights are hosted on Google Drive: link. Add it into the camera_ready_ckpts/ subdirectory. If you are on a remote server, there are libraries like gdown that help you retrieve this directly from GDrive. You need to pass in the file ID 1AnDMUa6ZnaRQonQje3Sfo1KBSQAsNKXe:
# download checkpoints
cd camera_ready_ckpts/
gdown 1AnDMUa6ZnaRQonQje3Sfo1KBSQAsNKXeYour subdirectory should look like this now:
.
└── camera_ready_ckpts/
├── inference.yaml # for inference/sampling
├── rna_frameflow_public_weights.ckpt # for inference / sampling / finetuning
└── config.yaml # for training
If you've retrained RNA-FrameFlow from scratch, as mentioned above, your checkpoint can be found at ckpts/se3-fm/rna-frameflow/last.ckpt. After renaming and shifting this last.ckpt file where necessary, visit configs/inference.yaml and change the path in the inference YAML file:
inference:
ckpt_path: configs/<insert_ckpt_name>.ckpt # path to model checkpoint of interest
This ensures the correct model checkpoints are used to generate new samples.
By default we sample 50 sequences per length between 40 and 150. Generated all-atom RNA backbones are stored as PDB files in the inference.output_dir directory listed in the inference YAML file.
# run inference
python inference_se3_flows.pyRunning inference also performs evaluation on the generated samples to compute local and global structural metrics. Inference together with evaluation takes around 2 hours on our hardware. See the subsequent section for setting up and running evaluation separately from inference.
We provide an evaluation pipeline called EvalSuite that computes local and global structural metrics when pointed at a directory of our model's RNA backbone samples. We use gRNAde (Joshi et al., 2023) as our inverse folder and RhoFold (Shen et al., 2022) as the structure predictor. First, download the RhoFold checkpoints (we didn't include this because of its size):
cd src/tools/rhofold_api/
mkdir checkpoints
cd checkpoints/
wget https://proj.cse.cuhk.edu.hk/aihlab/RhoFold/api/download?filename=RhoFold_pretrained.pt -O RhoFold_pretrained.ptIf the above URL does not work, we have stored the RhoFold checkpoints in a Google Drive for easier access:
https://drive.google.com/drive/folders/1umg0hgkBl7zsF_2GdCIKkfsRWbJNEOvp?usp=drive_link
Go back to the project's root directory. Here is a minimal example of EvalSuite in action. The API takes care of the computation, storage, and management of local structural measurements as well as global metrics (desigability, diversity, and novelty). Set-up instructions can be found in inference_se3_flows.py.
from src.analysis.evalsuite import EvalSuite
rna_bb_samples_dir = "generated_rna_bb_samples/" # generated samples for each sequence length
saving_dir = "rna_eval_metrics" # save temp files and metrics
evalsuite = EvalSuite(
save_dir=saving_dir,
paths=cfg.inference.evalsuite.paths,
constants=cfg.inference.evalsuite.constants,
gpu_id1=0, # cuda:0 -> for inverse-folding model
gpu_id2=1, # cuda:1 -> for forward-folding model
)
# compute local structural measurements and
metric_dict = evalsuite.perform_eval(
rna_bb_samples_dir,
flatten_dir=True
)
evalsuite.print_metrics(metric_dict) # print eval metrics
"""
Diversity (#clusters / #designable): 0.55
Novelty (pdbTM): 0.63
Designability (% scTM >= 0.45) 0.457
Designability (% scRMSD <= 4 ang): 0.433
"""This work is presented by Rishabh Anand to fulfill the Bachelor's Dissertation requirements at the Department of Computer Science, School of Computing, National University of Singapore (NUS). It is done in collaboration with Pietro Liò's group at the University of Cambridge, UK.
Our codebase builds on the open-source contributions from the following projects:
@article{anand2024rnaframeflow,
title={RNA-FrameFlow: Flow Matching for de novo 3D RNA Backbone Design},
author={Anand, Rishabh and Joshi, Chaitanya K. and Morehead, Alex and Jamasb, Arian R. and Harris, Charles and Mathis, Simon and Didi, Kieran and Hooi, Bryan and Li{\`o}, Pietro},
journal={arXiv preprint arXiv:2406.13839},
year={2024},
}