rnaglib

RNA Geometric Library (rnaglib)

RNAglib is a Python package for studying RNA 2.5D and 3D structures. Functionality includes automated data loading, analysis, visualization, ML model building and benchmarking.

We host RNAs annotated with molecule, base pair, and nucleotide level attributes. These include, but are not limited to:

• Secondary structure
• 3D coordinates
• Protein binding
• Small molecule binding
• Chemical modifications
• Leontis-westhof base pair geometry classification

Installation

The package can be cloned and the source code used directly. We also deploy it as a pip package and recommend using this install in conda environments.

If one wants to use GPU support, one should install Pytorch and DGL with the appropriate options. Otherwise, you can just skip this step and the pip installs of Pytorch and DGL will be used.

Then, one just needs to run :

pip install rnaglib

Then one can start using the packages functionalities by importing them in one's python script.

Optional Dependencies

We use fr3d-python to build 2.5D annotations from PDBs and mmCIFs. This has to be installed manually with:

pip install git+https://github.com/cgoliver/fr3d-python.git

To load graphs and train with dgl:

pip install dgl

To load graphs and train with pytorch geometric:

pip install torch_geometric

Advanced data splitting of datasets (i.e. by sequence or structure-based similarity) depends on executables cd-hit and RNAalign. You may install these yourself or use the convenience script provided in this repo as follows, though we do not guarantee it will work on any system and has only bee tested on linux:

chmod u+x install_dependencies.sh
./install_dependencids /my/path/to/executables

Setup: updating RNA structure database

Run the rnaglib_download script to fetch the database of annotated structures. By default it will download to ~/.rnaglib. Use the --help flag for more options.

$ rnaglib_download

In addition to analysing RNA data, RNAglib also distributes available parsed RNA structures. Databases of annotated structures can be downloaded directly from Zenodo, either the non-redundant subset link or all rna structures link. They can also be obtained through the provided command line utility $ rnaglib_download -r all|nr.

Version	Date	Total RNAs	Total Non-Redundant	Non-redundant version	rnaglib commit
1.0.0	15-02-23	5759	1176	3.269	5446ae2c
0.0.0	20-07-21	3739	899	3.186	eb25dabd

To speed up some functions we also build an efficient index of the data. This is needed for the rnaglib.tasks functionality.

$ rnaglib_index

What can you do with rnaglib?

Benchmark ML models on RNA 3D structures (new)

Datasets of RNA 3D structures ready-to-use for machine learning model benchmarking in various biologically relevant tasks.

• One line to fetch annotated dataset and splits loading
• Encode structures as graphs, voxels and point clouds

All tasks take as input an RNA 3D structure and have as output either residue or whole RNA-level properties.

We currently support the following 6 prediction tasks and the ability to create new tasks:

• Residue-Level RNA-Protein Binding
• Chemical Modification
• Inverse Folding
• Small Molecule Ligand Classification
• Small Molecule Binding Site Detection
• RNA Virtual Screening

See docs for more info on usage.

Each link for the tasks above takes you to a demo.py file with example usage for each task.

Here is a snippet for a full data loading and model training of the RNA Binding Site Detection task. All tasks follow a similar pattern.

from torch.nn import BCELoss
from rnaglib.tasks import BindingSiteDetection
from rnaglib.transforms import GraphRepresentation

# Load the task data and annotations
ta = BindingSiteDetection("my_root")

# Select a data representation and framework (see docs for support of other data modalities and deep learning frameworks)

ta.dataset.add_representation(GraphRepresentation(framework="pyg"))

train_loader, val_loader, test_loader = ta.get_split_loaders()

# most tasks ship with a dummy model for debugging, gives random outputs of the right shape
model = ta.dummy_model

# Access the predefined splits
for batch in train_loader:
    pred = ta.dummy_model(batch["graph"])
    y = batch["graph"].y
    loss = BCELoss()(y, pred)

Have a look at a full implemenation in one of the demos for a full example.

Create your own ML tasks (new)

The rnaglib.tasks subpackage defines an abstract class to quickly implement machine learning tasks. You can create a task from the databases available through rnaglib as well as custom annotations to open up your challenge to the rest of the community. If you implement a task we encourage you to submit a pull request.

To implement a task you must subclass rnaglib.tasks.Task and define the following methods:

• default_splitter(): a method that takes as input a dataset and returns train, validation, and test indices.
• build_dataset(): a method that returns a rnaglib.RNADataset object containing the RNAs needed for the task.

See here for an example of a full custom task implementation.

Fetch and browse annotated RNA 3D structures

Current release contains annotations generated by x3dna-dssr as well as some additional ones that we added for all available PDBs at the time of release.

Each RNA is stored as a networkx graph where nodes are residues and edges are backbone and base pairing edges. The networkx graph object has graph-level, node-level and edge-level attributes. Here is a reference for all the annotations currently available.

from rnaglib.utils import available_pdbids
from rnaglib.utils import graph_from_pdbid

# load a graph containing annotations from a PDBID
rna = graph_from_pdbid('4v9i')

# you can get list of pdbids currently available in RNAglib
pdbids = available_pdbids()

# print(rna.graph)
{'dbn': {'all_chains': {'num_nts': 143, 'num_chars': 144,
                        'bseq': 'GCCCGGAUAGCUCAGUCGGUAGAGCAGGGGAUUGAAAAUCCCCGUGUCCUUGGUUCGAUUCCGAGUCUGGGCAC&CGGAUAGCUCAGUCGGUAGAGCAGGGGAUUGAAAAUCCCCGUGUCCUUGGUUCGAUUCCGAGUCCGGGC',
                        'sstr': '(((((((..((((.....[..)))).(((((.......))))).....(((((..]....))))))))))))..&((((..((((.....[..)))).(((((.......))))).....(.(((..]....))).)))))...',
                        'form': 'AAAAAA...AA...A.......AAA.AAAA.......A.AAA......AAAAA..A....AAAAAAAAAAAA.-&.AA...AA...A.......AAA.AAAA.......A.AAA......AAAAA..A....A...AAAA.A.-'}...,

Annotate your own structures

You can extract Leontis-Westhof interactions and convert 3D structures to 2.5D graphs. We wrap a fork of fr3d-python to support this functionality.

from rnaglib.prepare_data import fr3d_to_graph

G = fr3d_to_graph("../data/structures/1fmn.cif")

Warning: this method currently does not support non-standard residues. Support coming soon. Up to version 1.0.0 of the RNA database were created using x3dna-dssr which do contain non-standard residues.

Quick visualization of 2.5D graphs

We customize networkx graph drawing functionalities to give some convenient visualization of 2.5D base pairing networks. This is not a dedicated visualization tool, it is only intended for quick debugging. We point you to VARNAhttps://varna.lisn.upsaclay.fr/ or RNAscape for a full-featured visualizer.

from rnaglib.drawing import rna_draw

rna_draw(G, show=True, layout="spring")

2.5D graph comparison and alignment

When dealing with 3D structures as 2.5D graphs we support graph-level comparison through the graph edit distance.

from rnaglib.ged import graph_edit_distance
from rnaglib.utils import graph_from_pdbid

G = graph_from_pdbid("4nlf")
print(graph_edit_distance(G, G))  # 0.0

Testing library functionality

Go to root of the rnaglib directory from a git clone and run pytest.

pip install pytest
pytest tests/

Cite

@article{mallet2022rnaglib,
  title={RNAglib: a python package for RNA 2.5 D graphs},
  author={Mallet, Vincent and Oliver, Carlos and Broadbent, Jonathan and Hamilton, William L and Waldisp{\"u}hl, J{\'e}r{\^o}me},
  journal={Bioinformatics},
  volume={38},
  number={5},
  pages={1458--1459},
  year={2022},
  publisher={Oxford University Press}
}

Projects using rnaglib

If you use rnaglib in one of your projects, please cite and feel free to make a pull request so we can list your project here.

• RNAMigos2
• Structure-and Function-Aware Substitution Matrices
• MultiModRLBP: A Deep Learning Approach for RNA-Small Molecule Ligand Binding Site Prediction using Multi-modal features
• VeRNAl
• RNAMigos

Resources

• Documentation
• Twitter
• Contact: [e-mail]

References

1. Leontis, N. B., & Zirbel, C. L. (2012). Nonredundant 3D Structure Datasets for RNA Knowledge Extraction and Benchmarking. In RNA 3D Structure Analysis and Prediction N. Leontis & E. Westhof (Eds.), (Vol. 27, pp. 281–298). Springer Berlin Heidelberg. doi:10.1007/978-3-642-25740-7_13