NMRNet: toward a unified benchmark and framework for deep learning-based prediction of nuclear magnetic resonance chemical shifts

Fanjie Xu, Wentao Guo, Feng Wang, Lin Yao, Hongshuai Wang, Fujie Tang*, Zhifeng Gao*, Linfeng Zhang, Weinan E, Zhong-Qun Tian, Jun Cheng* (* indicates corresponding authors)

---

This is the official implementation of the code related to the paper "Toward a unified benchmark and framework for deep learning-based prediction of nuclear magnetic resonance chemical shifts".

📖 Overview

NMRNet is a unified deep learning framework for NMR chemical shift prediction. It consists of four synergistic modules:

Module	Description
Data Preparation	Provides structure and NMR data
Pre-training	Uses pure structural information for self-supervised tasks, including masked atom prediction and 3D position recovery
Fine-tuning	Supervised NMR chemical shift prediction
Inference	Fine-tuned NMRNet model parameters are frozen and applied to various tasks

Pre-training weights and datasets for all fine-tuning stages are available on Zenodo. An online web app is available for NMR chemical shift prediction.

---

🗞️ News

⚠️ Note Please note that the Zenodo records may be updated. Make sure to check the latest version.

Date	Update
🎞️ 2026.03.21	The fine-tuned weights have been updated and released on Zenodo
📄 2025.03.28	🎉🎉🎉 Paper published on Nature Computational Science
🔗 2025.03.13	The web application is available on the AI4EC platform
🎞️ 2024.12.05	The LMDB-format datasets have been updated and released on Zenodo
📄 2024.08.28	Paper published on arXiv
🎞️ 2024.08.14	Dataset and trained weights released on Zenodo
🏷️ 2024.08.14	Code has been released on GitHub

---

⚙️ Installation

The installation steps for Linux systems are as follows:

pip install torch==1.13.1+cu116 torchvision==0.14.1+cu116 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cu116
pip install scikit-learn==1.3.2
pip install ase==3.22.1
pip install ./unicore-0.0.1+cu116torch1.12.0-cp38-cp38-linux_x86_64.whl
pip install pandas==2.0.3

Detailed installation tutorials for other versions of the unicore package can be found at: Uni-Core.

---

🚀 Usage

1. Prepare your dataset

Prepare your dataset for pre-training or fine-tuning in lmdb format and place it in the data folder. You may refer to the demo as a reference.

2. Download pre-trained weights

Place the pre-trained weights into the weights folder (skip this step if re-training from scratch). Pre-trained weights are available on Zenodo.

3. Run training or inference

Pre-training (cutoff radius)

sh script/pretrain_rcut.sh

Fine-tuning (5-fold cross-validation)

sh script/finetune_cv.sh

Details of the original Uni-Mol architecture can be found in the paper.

Inference

A demo notebook is available in the notebook folder.

An online service is also available at ai4ec and bohrium.

---

📜 Citation

If you find NMRNet useful in your research, please cite:

@article{xu2025toward,
  title={Toward a unified benchmark and framework for deep learning-based prediction of nuclear magnetic resonance chemical shifts},
  author={Xu, Fanjie and Guo, Wentao and Wang, Feng and Yao, Lin and Wang, Hongshuai and Tang, Fujie and Gao, Zhifeng and Zhang, Linfeng and E, Weinan and Tian, Zhong-Qun and others},
  journal={Nature Computational Science},
  volume={5},
  number={4},
  pages={292--300},
  year={2025},
  publisher={Nature Publishing Group US New York}
}

---

⚠️ License

This project is licensed under the terms of the MIT License. See LICENSE for additional details.

---

📬 Contact

For questions and issues, please contact the author xufanjie@stu.xmu.edu.cn or open a GitHub issue.

---

_{State Key Laboratory of Physical Chemistry of Solid Surfaces · Xiamen University}