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*

Paper | PDF | poster | GitHub | PyPI

• Introduction
• Installation
  • Requirements
  • Use base pair motif library
  • Predict RNA secondary structure
• Usage
  • Base pair motif library
  • BPfold for secondary structure prediction
    • Run command line
    • Import python code
  • Evaluation
• 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

• python3.8+
• anaconda

Use base pair motif library

pip3 install BPfold

Predict RNA secondary structure

0. Clone this repo

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

1. Create and activate BPfold environment.

conda env create -f BPfold_environment.yaml
conda activate BPfold

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

wget https://github.com/heqin-zhu/BPfold/releases/latest/download/model_predict.tar.gz
tar -xzf model_predict.tar.gz

3. Optional (for training and evaluation): Download datasets BPfold_data.tar.gz in releases and decompress them.

wget https://github.com/heqin-zhu/BPfold/releases/latest/download/BPfold_data.tar.gz
tar -xzf BPfold_data.tar.gz 

Usage

Base pair 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.

For an input RNA sequence seq, the base pair motif energy matrix mat can be directly obatined as follows:

from BPfold.util.base_pair_motif import BPM_energy

BPM = BPM_energy()

seq = 'AUGCGUAGTa'
# default, recommended, normed according to the major energy of each motif type, BPfold used, shape 2xLxL
mat = BPM.get_energy(seq)

# origin energy, value may be -50.3, 49.7, ..., shape 1xLxL
mat2 = BPM.get_energy(seq, normalize_energy=False, dispart_outer_inner=False)

To get a folding score of a sequence for reference, we accumulate all energy items from canonical base pairs assigned by greedy search matching algorithm (issue #2).

from src.BPfold.util.base_pair_motif import BPM_energy()

seq = 'GCGCAGGACUCGGCUUCUUCGGAAGGGACGAGGGGCGC'
score = BPM_energy().get_score(seq)
print(score)

BPfold for secondary structure prediction

Run command line

Args:

• --checkpoint_dir: required, specify checkpoint dir path.
• --seq: specify one or more input RNA sequences.
• --input: specify input file of RNA seqs in format of .fasta(multiple seqs are supported), .bpseq, .ct, or .dbn.
• --output: output dir (will be created automatically), default BPfold_results.
• --out_type: out format of RNA secondary structures, can be .csv, .bpseq, .ct, or .dbn, default .csv Here are some examples:

BPfold --checkpoint_dir PATH_TO_CHECKPOINT_DIR --seq GGUAAAACAGCCUGU AGUAGGAUGUAUAUG --output BPfold_results
BPfold --checkpoint_dir PATH_TO_CHECKPOINT_DIR --input examples/examples.fasta --out_type csv # (multiple sequences are supported)
BPfold --checkpoint_dir PATH_TO_CHECKPOINT_DIR --input examples/URS0000D6831E_12908_1-117.bpseq

>> Welcome to use "BPfold" for predicting RNA secondary structure!
Loading model_predict/BPfold_1-6.pth
Loading model_predict/BPfold_2-6.pth
Loading model_predict/BPfold_3-6.pth
Loading model_predict/BPfold_4-6.pth
Loading model_predict/BPfold_5-6.pth
Loading model_predict/BPfold_6-6.pth
[      1] saved in "BPfold_results/1M5L.bpseq", CI=0.913
GCGCAGGACUCGGCUUCUUCGGAAGGGACGAGGGGCGC
((((....((((.(((((..)))))...))))..))))
............(..............).......... NC
((((....((((((((((..)))))..)))))..)))) MIX
[      2] saved in "BPfold_results/URS0000D6831E_12908_1-117.bpseq", CI=0.892
UUAUCUCAUCAUGAGCGGUUUCUCUCACAAACCCGCCAACCGAGCCUAAAAGCCACGGUGGUCAGUUCCGCUAAAAGGAAUGAUGUGCCUUUUAUUAGGAAAAAGUGGAACCGCCUG
......((((((.....((((.......))))..(((.((((.((......))..))))))).................))))))..(((......)))..................
..................................................................................................................... NC
......((((((.....((((.......))))..(((.((((.((......))..))))))).................))))))..(((......))).................. MIX
Confidence indexes are saved in "BPfold_results_confidence_20250915_03h19m33s.yaml"
Program Finished!

[!NOTE] Results (dbn, connects, bpseq...) with no tag are predicted canonical pairs, tagged with _nc are predicted non-canonical pairs, and tagged with _mix are mixed canonical and non-canonical pairs (i.e., all base pairs). If you want to ignore non-canonical pairs, pass argument --ignore_nc to BPfold.

Run command BPfold -h for more help information.

Import python code

Specify arguments:

• checkpiont_dir
• at least one of input_seqs (list of seqs) and input_path (fasta_path)

from BPfold.predict import BPfold_predict
from BPfold.util.RNA_kit import connects2dbn


## arguments
checkpoint_dir = '' # to be specified
input_seqs = ['GCGCAGGACUCGGCUUCUUCGGAAGGGACGAGGGGCGC', 'AUGUAUGUCCUGUCGUA']
input_path = 'examples/examples.fasta'

## init model
BPfold_predictor = BPfold_predict(checkpoint_dir)

## BPfold predict  # specify at least one of input_seqs and input_path
pred_results = BPfold_predictor.predict(input_seqs=input_seqs, input_path=input_path, ignore_nc=False)

for dic in pred_results:
    print(f'>{dic["seq_name"]}')
    print(dic["seq"])
    print(connects2dbn(dic["connects"]), f'CI={dic["CI"]:.3f}')

Loading /public2/home/heqinzhu/gitrepo/RNA/SS_pred/BPfold/src/BPfold/paras/model_predict/BPfold_1-6.pth
Loading /public2/home/heqinzhu/gitrepo/RNA/SS_pred/BPfold/src/BPfold/paras/model_predict/BPfold_2-6.pth
Loading /public2/home/heqinzhu/gitrepo/RNA/SS_pred/BPfold/src/BPfold/paras/model_predict/BPfold_3-6.pth
Loading /public2/home/heqinzhu/gitrepo/RNA/SS_pred/BPfold/src/BPfold/paras/model_predict/BPfold_4-6.pth
Loading /public2/home/heqinzhu/gitrepo/RNA/SS_pred/BPfold/src/BPfold/paras/model_predict/BPfold_5-6.pth
Loading /public2/home/heqinzhu/gitrepo/RNA/SS_pred/BPfold/src/BPfold/paras/model_predict/BPfold_6-6.pth
>seq_20250929_14h23m28s_1
GCGCAGGACUCGGCUUCUUCGGAAGGGACGAGGGGCGC
((((....((((.(((((..)))))...))))..)))) CI=0.913
>seq_20250929_14h23m28s_2
AUGUAUGUCCUGUCGUA
.....((......)).. CI=0.807
>1M5L
GCGCAGGACUCGGCUUCUUCGGAAGGGACGAGGGGCGC
((((....((((.(((((..)))))...))))..)))) CI=0.913
>URS0000D6831E_12908_1-117
UUAUCUCAUCAUGAGCGGUUUCUCUCACAAACCCGCCAACCGAGCCUAAAAGCCACGGUGGUCAGUUCCGCUAAAAGGAAUGAUGUGCCUUUUAUUAGGAAAAAGUGGAACCGCCUG
......((((((.....((((.......))))..(((.((((.((......))..))))))).................))))))..(((......))).................. CI=0.892

Evaluation

Specify pred_dir and gt_dir. In each directory, there are secondary structures in format of bpseq, ct, or dbn.

BPfold_eval --pred_dir BPfold_results  --gt_dir PATH_TO_NATIVE_STRUCTURES

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 v0.2.0 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/v0.2/BPfold_test_results.tar.gz
tar -xzf BPfold_test_results.tar.gz

Use BPfold

1. Download the checkpoints of BPfold: BPfold_reproduce.tar.gz.

wget https://github.com/heqin-zhu/BPfold/releases/download/v0.2/model_reproduce.tar.gz
tar -xzf model_reproduce.tar.gz

2. Install BPfold version 0.2.4.

pip install BPfold==0.2.4

3. Use BPfold to predict RNA sequences in test datasets.

Evaluate

BPfold_eval --gt_dir BPfold_data --pred_dir BPfold_test_results

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

Time used: 29s
[Summary] eval_BPfold_test_results.yaml
 Pred/Total num: [('PDB_test', 116, 116), ('Rfam12.3-14.10', 10791, 10791), ('archiveII', 3966, 3966), ('bpRNA', 1305, 1305), ('bpRNAnew', 5401, 5401)]
-------------------------len>600-------------------------
dataset         & num   & INF   & F1    & P     & R    \\
Rfam12.3-14.10  & 64    & 0.395 & 0.387 & 0.471 & 0.333\\
archiveII       & 55    & 0.352 & 0.311 & 0.580 & 0.242\\
------------------------len<=600-------------------------
dataset         & num   & INF   & F1    & P     & R    \\
PDB_test        & 116   & 0.817 & 0.814 & 0.840 & 0.801\\
Rfam12.3-14.10  & 10727 & 0.696 & 0.690 & 0.662 & 0.743\\
archiveII       & 3911  & 0.829 & 0.827 & 0.821 & 0.843\\
bpRNA           & 1305  & 0.670 & 0.658 & 0.599 & 0.770\\
bpRNAnew        & 5401  & 0.655 & 0.647 & 0.604 & 0.723\\
---------------------------all---------------------------
dataset         & num   & INF   & F1    & P     & R    \\
PDB_test        & 116   & 0.817 & 0.814 & 0.840 & 0.801\\
Rfam12.3-14.10  & 10791 & 0.694 & 0.689 & 0.660 & 0.741\\
archiveII       & 3966  & 0.823 & 0.820 & 0.818 & 0.834\\
bpRNA           & 1305  & 0.670 & 0.658 & 0.599 & 0.770\\
bpRNAnew        & 5401  & 0.655 & 0.647 & 0.604 & 0.723\\

Acknowledgement

We appreciate the following open source projects:

• UFold
• vigg_ribonanza
• e2efold

LICENSE

MIT LICENSE

Citation

If you find our work helpful, please cite our paper:

@article{BPfold,
    title = {Deep generalizable prediction of {RNA} secondary structure via base pair motif energy},
    author = {Zhu, Heqin and Tang, Fenghe and Quan, Quan and Chen, Ke and Xiong, Peng and Zhou, S. Kevin},
    volume = {16},
    issn = {2041-1723},
    url = {https://doi.org/10.1038/s41467-025-60048-1},
    doi = {10.1038/s41467-025-60048-1},
    number = {1},
    journal = {Nature Communications},
    month = jul,
    year = {2025},
    pages = {5856},
}