Dataset for DiffSpectra

Model Description

DiffSpectra is a generative framework for molecular structure elucidation from multi-modal spectral data. Unlike retrieval-based approaches that rely on finite molecular libraries or SMILES-based autoregressive models that often ignore 3D geometry, DiffSpectra formulates structure elucidation as a conditional diffusion process.
The framework integrates two core components:

• Diffusion Molecule Transformer (DMT): An SE(3)-equivariant denoising network that models both 2D topology and 3D geometry of molecules.
• SpecFormer: A transformer-based spectral encoder that captures intra- and inter-spectrum dependencies across diverse spectral modalities (e.g., IR, Raman, UV-Vis).

Through spectrum-conditioned diffusion modeling, DiffSpectra unifies multi-modal reasoning with 2D/3D generative modeling. Extensive experiments demonstrate that DiffSpectra achieves 40.76% top-1 accuracy and 99.49% top-10 accuracy in recovering exact molecular structures.
To our knowledge, DiffSpectra is the first framework that unifies multi-modal spectral reasoning and joint 2D/3D generative modeling for de novo molecular structure elucidation.

• Paper: https://arxiv.org/abs/2507.06853
• Code: https://github.com/AzureLeon1/DiffSpectra
• Model: https://huggingface.co/AzureLeon1/DiffSpectra

---

Usage

1. Place the dataset in your configured path/to/dataset directory.
2. Modify the configuration file in configs/ to set data.root to your configured path.
3. Use the provided scripts in scripts/ for training or evaluation.

Note: Please ensure required dependencies and environment settings match those described in the repository instructions.

Contact

For questions, feedback, or collaborations, please reach out to: 📧 [email protected]

Cite Us

If you use DiffSpectra in your research or applications, please cite:

@article{wang2025diffspectra,
  title={DiffSpectra: Molecular Structure Elucidation from Spectra using Diffusion Models},
  author={Liang Wang and Yu Rong and Tingyang Xu and Zhenyi Zhong and Zhiyuan Liu and Pengju Wang and Deli Zhao and Qiang Liu and Shu Wu and Liang Wang and Yang Zhang},
  journal={arXiv},
  volume={abs/2502.09511},
  year={2025}
}