Deep generalizable prediction of RNA secondary structure via base pair motif energy

Heqin Zhu · Fenghe Tang · Quan Quan · Ke Chen · Peng Xiong · S. Kevin Zhou

Submitted

bioRxiv | PDF | Code

• Introduction
• Installation
  • Requirements
  • Instructions
• Usage
  • BPfold motif library
  • BPfold Prediction
• Reproduction
• Acknowledgement
• LICENSE
• Citation

Introduction

Deep learning methods have demonstrated great performance for RNA secondary structure prediction. However, generalizability is a common unsolved issue on unseen out-of-distribution RNA families, which hinders further improvement of the accuracy and robustness of deep learning methods. Here we construct a base pair motif library that enumerates the complete space of locally adjacent three-neighbor base pair and records the thermodynamic energy of corresponding base pair motifs through de novo modeling of tertiary structures, and we further develop a deep learning approach for RNA secondary structure prediction, named BPfold, which learns relationship between RNA sequence and the energy map of base pair motif. Experiments on sequence-wise and family-wise datasets have demonstrated the great superiority of BPfold compared to other state-of-the-art approaches in accuracy and generalizability. We hope this work contributes to integrating physical priors and deep learning methods for the further discovery of RNA structures and functionalities.

Installation

Requirements

• Linux system
• python3.6+
• anaconda

Instructions

1. Clone this repo.

git clone git@github.com:heqin-zhu/BPfold.git
cd BPfold

2. Create and activate BPfold environment.

conda env create -f BPfold_environment.yaml
conda activate BPfold

3. Download model_predict.tar.gz in releases and decompress it.

wget https://github.com/heqin-zhu/BPfold/releases/download/v0.1/model_predict.tar.gz
tar -xzf model_predict.tar.gz -C paras

4. Download datasets BPfold_data.tar.gz in releases and decompress them.

wget https://github.com/heqin-zhu/BPfold/releases/download/v0.1/BPfold_data.tar.gz
mkdir BPfold_data && tar -xzf BPfold_data.tar.gz -C BPfold_data

Usage

BPfold motif library

The base pair motif library is publicly available in releases, which contains the motif:energy pairs. The motif is represented as sequence_pairIdx_pairIdx-chainBreak where pairIdx is 0-indexed, and the energy is a reference score of statistical and physical thermodynamic energy. For instance, CAAAAUG_0_6-3 -49.7835 represents motif CAAAAUG has a known pair C-G whose indexes are 0 and 6, with chainBreak lying at position 3.

Note

The base pair motif library can be used as thermodynamic priors in other models.

BPfold Prediction

Use BPfold to predict RNA secondary structures. The following are some examples.

python3 -m BPfold.predict --input examples/examples.fasta
python3 -m BPfold.predict --seq  UUAUCUCAUCAUGAGCGGUUUCUCUCACAAACCCGCCAACCGAGCCUAAAAGCCACGGUGGUCAGUUCCGCUAAAAGGAAUGAUGUGCCUUUUAUUAGGAAAAAGUGGAACCGCCUG   AGGCAGUGAUGAUGAAAAAAGAUUACCAUCAAACUUUGAGAGAUUCACAGCUCGUUGAUGCAUACUUCUUUAUAUUACCUGAGCCU
python3 -m BPfold.predict --input examples/URS0000D6831E_12908_1-117.bpseq --output results --save_type ct

Example of BPfold prediction

Here are the outputs after running python3 -m BPfold.predict --input examples/examples.fasta:

>> Welcome to use "BPfold" for predicting RNA secondary structure!
Loading paras/model_predict/BPfold_1-6.pth
Loading paras/model_predict/BPfold_2-6.pth
Loading paras/model_predict/BPfold_3-6.pth
Loading paras/model_predict/BPfold_4-6.pth
Loading paras/model_predict/BPfold_5-6.pth
Loading paras/model_predict/BPfold_6-6.pth
[    1] saved in "BPfold_results/SS/5s_Shigella-flexneri-3.bpseq", CI=0.980
CUGGCGGCAGUUGCGCGGUGGUCCCACCUGACCCCAUGCCGAACUCAGAAGUGAAACGCCGUAGCGCCGAUGGUAGUGUGGGGUCUCCCCAUGCGAGAGUAGGGAACUGCCAG
(((((((.....((((((((.....((((((.............))))..))....)))))).)).((.((....((((((((...))))))))....)).))...)))))))
[    2] saved in "BPfold_results/SS/URS0000D6831E_12908_1-117.bpseq", CI=0.931
UUAUCUCAUCAUGAGCGGUUUCUCUCACAAACCCGCCAACCGAGCCUAAAAGCCACGGUGGUCAGUUCCGCUAAAAGGAAUGAUGUGCCUUUUAUUAGGAAAAAGUGGAACCGCCUG
......((((((..(.(((((.......))))))(((.((((.((......))..))))))).................))))))..(((......)))..................
Finished!

For more help information, please run command python3 -m BPfold.predict -h to see.

Reproduction

For reproduction of all the quantitative results, we provide the predicted secondary structures and model parameters of BPfold in experiments. You can directly downalod the predicted secondary structures by BPfold or use BPfold with trained parameters to predict these secondary structures, and then evaluate the predicted results.

Directly download

wget https://github.com/heqin-zhu/BPfold/releases/download/v0.1/BPfold_test_results.tar.gz
tar -xzf BPfold_test_results.tar.gz

Use BPfold

1. Download BPfold_reproduce.tar.gz in releases.

wget https://github.com/heqin-zhu/BPfold/releases/download/v0.1/model_reproduce.tar.gz
tar -xzf model_reproduce.tar.gz -C paras

2. Use BPfold to predict test sequences.

Evaluate

python3 -m BPfold.evaluate --data_dir BPfold_data --pred_dir BPfold_reproduce/pred_results

After running above commands for evaluation, you will see the following outputs:

Outputs of evaluating BPfold

Time used: 23s
[Summary] eval_BPfold_test_results.yaml
 Pred/Total num: [('PDB_test', 60, 60), ('Rfam14.5-14.10', 2048, 2048), ('archiveII', 3966, 3966), ('bpRNA', 1305, 1305)]
---------------------len>600---------------------
dataset         & num   & f1    & p     & r    \\
archiveII       & 55    & 0.377 & 0.547 & 0.313\\
--------------------len<=600---------------------
dataset         & num   & f1    & p     & r    \\
PDB_test        & 60    & 0.727 & 0.779 & 0.695\\
Rfam14.5-14.10  & 2048  & 0.791 & 0.777 & 0.824\\
archiveII       & 3911  & 0.926 & 0.927 & 0.929\\
bpRNA           & 1305  & 0.701 & 0.670 & 0.757\\
-----------------------all-----------------------
dataset         & num   & f1    & p     & r    \\
PDB_test        & 60    & 0.727 & 0.779 & 0.695\\
Rfam14.5-14.10  & 2048  & 0.791 & 0.777 & 0.824\\
archiveII       & 3966  & 0.918 & 0.922 & 0.921\\
bpRNA           & 1305  & 0.701 & 0.670 & 0.757\\

Acknowledgement

We appreciate the following open source projects:

• UFold
• vigg_ribonanza
• e2efold

LICENSE

MIT LICENSE

Citation

If you use our code, please kindly consider to cite our paper:

@article {Zhu2024.10.22.619430,
    author = {Zhu, Heqin and Tang, Fenghe and Quan, Quan and Chen, Ke and Xiong, Peng and Zhou, S. Kevin},
    title = {Deep generalizable prediction of RNA secondary structure via base pair motif energy},
    elocation-id = {2024.10.22.619430},
    year = {2024},
    doi = {10.1101/2024.10.22.619430},
    publisher = {Cold Spring Harbor Laboratory},
    URL = {https://www.biorxiv.org/content/early/2024/10/25/2024.10.22.619430},
    eprint = {https://www.biorxiv.org/content/early/2024/10/25/2024.10.22.619430.full.pdf},
    journal = {bioRxiv}
}