A fully open structure-guided RNA foundation model for robust structural and functional inference

Heqin Zhu · Ruifeng Li · Feng Zhang · Fenghe Tang
Tong Ye · Xin Li · Yunjie Gu · Peng Xiong* · S. Kevin Zhou*

Submitted

bioRxiv | PDF | GitHub | PyPI

• Overview
  • Abstract
  • Key Achievements
• Installation
  • Requirements
  • Instructions
• Usage
  • Extract RNA sequence features
  • Build structRFM for finetuning
• Pretraining
  • Download sequence-structure dataset
  • Run pretraining
• Downstream Tasks
• Acknowledgement
• LICENSE
• Citation

Overview

Abstract

RNA language models have achieved strong performance across diverse downstream tasks by leveraging large-scale sequence data. However, RNA function is fundamentally shaped by its hierarchical structure, making the integration of structural information into pretraining essential. Existing methods often depend on noisy structural annotations or introduce task-specific biases, limiting model generalizability. Here, we introduce structRFM, a structure-guided RNA foundation model that is pretrained by implicitly incorporating large-scale base pairing interactions and sequence data via a dynamic masking ratio to balance nucleotide-level and structure-level masking. structRFM learns joint knowledge of sequential and structural data, producing versatile representations-including classification-level, sequence-level, and pairwise matrix features-that support broad downstream adaptations. structRFM ranks top models in zero-shot homology classification across fifteen biological language models, and sets new benchmarks for secondary structure prediction, achieving F1 scores of 0.873 on ArchiveII and 0.641 on bpRNA-TS0 dataset. structRFM further enables robust and reliable tertiary structure prediction, with consistent improvements in both 3D accuracy and extracted 2D structures. In functional tasks such as internal ribosome entry site identification, structRFM achieves a 49% performance gain. These results demonstrate the effectiveness of structure-guided pretraining and highlight a promising direction for developing multi-modal RNA language models in computational biology.

Key Achievements

• Zero-shot homology classification: Top-ranked among 15 biological language models.
• Secondary structure prediction: Sets new state-of-the-art performances.
• Tertiary structure prediction: Derived method Zfold improves RNA Puzzles accuracy by 19% over AlphaFold3.
• Functional inference: Boosts F1 score by 49% on IRES identification.

Installation

Requirements

• python3.8+
• anaconda

Instructions

0. Clone this repo.

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

1. Create and activate conda environment.

conda env create -f environment.yaml
conda activate structRFM

2. Install structRFM.

pip3 install structRFM

3. Download and decompress pretrained structRFM (305 MB).

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

4. Set environment varible structRFM_checkpoint.

export structRFM_checkpoint=PATH_TO_CHECKPOINT # modify ~/.bashrc for permanent setting

Usage

Extract RNA sequence features

import os

from structRFM.infer import structRFM_infer

from_pretrained = os.getenv('structRFM_checkpoint')
model = structRFM_infer(from_pretrained=from_pretrained, max_length=514)

seq = 'AGUACGUAGUA'

print('seq len:', len(seq))
feat_dic = model.extract_feature(seq)
for k, v in feat_dic.items():
    print(k, v.shape)

'''
seq len: 11
cls_feat torch.Size([768])
seq_feat torch.Size([11, 768])
mat_feat torch.Size([11, 11])
'''

Build structRFM for finetuning

import os

from structRFM.model import get_structRFM
from structRFM.data import preprocess_and_load_dataset, get_mlm_tokenizer

from_pretrained = os.getenv('structRFM_checkpoint')

tokenizer = get_mlm_tokenizer(max_length=514)
model = get_structRFM(dim=768, layer=12, from_pretrained=from_pretrained, pretrained_length=None, max_length=514, tokenizer=tokenizer)

Pretraining

Download sequence-structure dataset

The pretrianing sequence-structure dataset is constructed using RNAcentral and BPfold. We filter sequences with a length limited to 512, resulting about 21 millions sequence-structure paired data. It can be downloaded at Zenodo (4.5 GB).

Run pretraining

Modify variables USER_DIR, PROGRAM_DIR, DATA_DIR, and OUT_DIR in run.sh, then run:

bash ./run.sh --print --batch_size 128 --epoch 100 --lr 0.0001 --tag mlm --mlm_structure

Downstream Tasks

Download all data (3.7 GB) and checkpoints (2.2 GB) from Zenodo, and then place them into corresponding folder of each task.

• Zero-shot inference
  • Zero-shot homology classfication
  • Zero-shot secondary structure prediction
• Structure prediction
  • Secondary structure prediction
  • Tertiary structure prediction
• Function prediction
  • ncRNA classification
  • Splice site prediction
  • IRES identification

Acknowledgement

We appreciate the following open-source projects for their valuable contributions:

• RNAcentral
• BPfold
• RNAErnie
• trRosettaRNA
• BEACON
• MXfold2

LICENSE

MIT LICENSE

Citation

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

@article {structRFM,
    author = {Zhu, Heqin and Li, Ruifeng and Zhang, Feng and Tang, Fenghe and Ye, Tong and Li, Xin and Gu, Yujie and Xiong, Peng and Zhou, S Kevin},
    title = {A fully-open structure-guided RNA foundation model for robust structural and functional inference},
    elocation-id = {2025.08.06.668731},
    year = {2025},
    doi = {10.1101/2025.08.06.668731},
    publisher = {Cold Spring Harbor Laboratory},
    URL = {https://www.biorxiv.org/content/early/2025/08/07/2025.08.06.668731},
    journal = {bioRxiv}
}