Decoupled peak property learning for efficient and interpretable ECD spectra prediction

Accepted by Nature Computational Science 2025

If you like our project, please give us a star ⭐ on GitHub for the latest update.

📰 News

• [2025.1.12] 🤝 We have released the ECD and IR spectra data at https://huggingface.co/datasets/OzymandisLi/ECDFormer_Datasets.
• [2024.12.31] 🤝 We have released the IR prediction codes.
• [2024.12.25] 🔥 Our ECDFormer has been accepted by Nature Computational Science!
• [2024.12.21] 🤝 We have released the ECD prediction codes for natural product and multi-chiralty molecules.

😮 Highlights

ECDFormer exhibits rearkable spectra modeling capability for molecule structure and IR, ECD, Mass spectra.

💡 Well performance and Generalization for spectra modeling and prediction

• With the decoupled peak property learning algorithm, we enable the ECDFormer to model spectra from discrete perspective, which is suitable for multiple spectra types including IR, ECD, and Mass spectra.

🔥 Well Explanability for spectrum peaks

• The decoupled peak property learning module facilitates the modeling and visualization of the relative relationships between spectral peaks and molecular functional groups, addressing the issue of interpretability in spectral prediction.

🛠️ Data Preparation

For training and inference, please download the IR&ECD spectra and put them into the folder datasets/ please download and put the descriptor_all_column.npy into the folder utils/

utils/descriptor_all_column.npy

🛠️ Requirements and Installation

• Python == 3.8
• Pytorch == 1.13.1
• CUDA Version == 11.7
• torch_geometric, torch-scatter, torch-sparse, torch-cluster, torch-spline-conv
• Install required packages:

git clone [email protected]:HowardLi1984/ECDFormer.git
cd ECDFormer
pip install -r requirements.txt

PS: you can follow this link for faster torch_geometric install

## First install these related packages
pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-${TORCH}+${CUDA}.html
pip install torch-sparse -f https://pytorch-geometric.com/whl/torch-${TORCH}+${CUDA}.html
pip install torch-cluster -f https://pytorch-geometric.com/whl/torch-${TORCH}+${CUDA}.html
pip install torch-spline-conv -f https://pytorch-geometric.com/whl/torch-${TORCH}+${CUDA}.html

## Then install the torch-geometric package
pip install torch-geometric

🗝️ Train and Inference

The training and inferencing instruction for ECD spectra is in main_func_pos.py.

For training the ECDFormer:

CUDA_VISIBLE_DEVICES=${0/1/2/3} python main_func_pos.py --model_name gnn_allthree --batch_size 256 --emb_dim 128 --epochs 1000 --lr 1e-3 --mode Train

For inference using the model's 400 epoch checkpoint:

CUDA_VISIBLE_DEVICES=${0/1/2/3} python main_func_pos.py --model_name gnn_allthree --batch_size 256 --emb_dim 128 --epochs 1000 --lr 1e-3 --mode Test --visual_epoch 400

To visualize the excellent spectra predictions or generate Atom-Attention-Weight for molecules:

CUDA_VISIBLE_DEVICES=${0/1/2/3} python main_func_pos.py --model_name gnn_allthree --batch_size 256 --emb_dim 128 --epochs 1000 --lr 1e-3 --mode Visual --visual_epoch 500

For further experiments, to predict the molecules from natural products or multi-chiralty-centric molecules, set the mode to Real or multi_carbon:

CUDA_VISIBLE_DEVICES=${0/1/2/3} python main_func_pos.py --model_name gnn_allthree --batch_size 256 --emb_dim 128 --epochs 1000 --lr 1e-3 --mode ${your mode} --visual_epoch ${your ckpt epoch}

The training and inferencing instruction for IR spectra is in main_ir.py.

For training and testing on the IR spectra dataset, set mode to Train or Test:

CUDA_VISIBLE_DEVICES=7 python main_ir.py --model_name gnn_ir --batch_size 128 --emb_dim 256 --epochs 360 --lr 1e-3 --mode ${your mode}

To visualize the excellent spectra predictions or generate Atom-Attention-Weight for molecules:

CUDA_VISIBLE_DEVICES=6 python main_ir.py --model_name gnn_ir --batch_size 128 --emb_dim 256 --epochs 360 --lr 1e-3 --visual_epoch 320 --mode Visual

🚀 Experimental Results

Quantitively, we propose the experimental results on our ECDFormer framework and the corresponding baselines including machine learning models and deep learning models. Focusing on peak property prediction, our ECDFormer model surpasses baselines under all evaluation metrics.

🚀 Generalization and Interpretation

ECD spectra predictions on natural products with pharmaceutical effects from recent journals demonstrate the effectiveness and generalization ability of our ECDFormer.

The ECDFormer framework has excellent generalization ability on the tandem mass spectra prediction and infrared spectra prediction tasks.

ECDFormer exhibits commendable interpretability due to its attribute prediction for spectral peak disentanglement. To elucidate the process of how ECDFormer generates spectra, we employed the Integrated Gradients to delineate regions within the entire molecule that contribute most significantly to spectrum generation. We observe that chromophores within the molecules generally contribute significantly to the peaks, which is consistent with the chemical principles of ECD spectroscopy. In the computational results, by examining the characteristic peaks of the ECD spectra for these molecules, we discovered a correlation with specific electronic transitions within the molecules.

✏️ Citation

@article{li2025decoupled,
  title={Decoupled peak property learning for efficient and interpretable electronic circular dichroism spectrum prediction},
  author={Li, Hao and Long, Da and Yuan, Li and Wang, Yu and Tian, Yonghong and Wang, Xinchang and Mo, Fanyang},
  journal={Nature Computational Science},
  pages={1--11},
  year={2025},
  publisher={Nature Publishing Group US New York}
}