Skip to content

README.md

Latest commit

 

History

History
155 lines (107 loc) · 6.68 KB

README.md

File metadata and controls

155 lines (107 loc) · 6.68 KB

DiffNMR

DiffNMR: Diffusion Models for Nuclear Magnetic Resonance Spectra Elucidation

Qingsong Yang#, Binglan Wu#, **Xuwei Liu, Bo Chen, Wei Li, Gen Long†, Xin Chen†, Mingjun Xiao†

1 Department of Computer Science, University of Science and Technology of China (USTC)

2 Suzhou Laboratory

3 Baidu Inc.

# Equal contribution · † Corresponding authors

Abstract

DiffNMR is an end-to-end framework that infers molecular structures directly from 1H/13C NMR spectra using a conditional discrete diffusion model. It holistically refines molecular graphs through denoising steps (instead of autoregressive token-by-token generation), improving global consistency and reducing error accumulation. The system couples a two-stage pretraining pipeline—(i) Diffusion Autoencoder (Diff-AE) for molecular representation + graph decoder pretraining and (ii) contrastive alignment between spectra and molecular representations—with a domain-tailored NMR encoder that leverages RBF encodings for chemical shifts and coupling constants. Inference further benefits from similarity-based filtering and optional retrieval-initialized sampling, yielding strong Top‑k accuracy and high Tanimoto similarity across molecule sizes.

DiffNMR Overview

Highlights

  • End-to-end spectrum→structure on 1H & 13C NMR.
  • Discrete graph diffusion (denoise molecular graphs instead of autoregressive SMILES).
  • NMR encoder with RBF embeddings for shifts (and J-coupling) + learnable embeddings for multiplicity & integrals; bi-directional cross‑attention to fuse 1H/13C information.
  • Two-stage pretraining: Diff‑AE (molecular encoder + graph diffusion decoder) → contrastive learning to align NMR & molecular spaces.
  • Inference enhancements: similarity filtering (cosine scoring in latent space) and retrieval‑initialized sampling.

Dataset & Metrics

  • MSD multimodal spectroscopic dataset: ~7.9e5 molecules (from USPTO), each with simulated 1H NMR, 13C NMR, HSQC, IR, MS; molecules span 5–35 heavy atoms.
  • Evaluation: Top‑k accuracy (exact‑match structure among top candidates) and Tanimoto similarity (Morgan fp, radius 2, 2048 bits).

Observations

  • 1H+13C outperforms single‑modality inputs.
  • Providing molecular formula improves accuracy across sizes.
  • Similarity filtering and retrieval initialization substantially boost Top‑1 accuracy and average Tanimoto, especially for larger molecules.

Method at a Glance

Architecture

  • Molecular encoder (graph transformer) → compact molecular representation.
  • Graph diffusion decoder (discrete denoising on nodes/edges).
  • NMR encoder: RBF(δ) for shifts; embeddings for multiplicity/integrals; RBF(J) for couplings; transformers per modality; bi‑directional cross‑attention (1H↔13C); pooled & fused to a conditioning vector.

Training

  1. Stage‑1: Diff‑AE — pretrain molecular encoder + diffusion decoder via reconstruction.
  2. Stage‑2: Contrastive — freeze molecular encoder; train NMR encoder to align to the molecular space with InfoNCE.
  3. Fine‑tuning — end‑to‑end spectra→graph with diffusion decoder conditioned on NMR embedding.

Inference

  • Similarity filtering: rank sampled candidates by cosine similarity between predicted molecule & input NMR embeddings.
  • Retrieval initialization (start from nearest neighbor in latent space) improves Top‑1 accuracy.
  • Use formula outputs atoms enforced to reset according to formula.

Results (summary)

  • 1H+13C achieves the best Top‑1/Top‑k; adding molecular formula improves all sizes (≤15/≤20/≤25 HAC).
  • Similarity filtering and retrieval initialization notably improve Top‑1 and avg. Tanimoto, with the largest gains on higher HAC.

Results Table1 Results Table2

Repository Layout

PaddleMaterials/
  spectrum_elucidation/
    configs/DiffNMR/
      DiffNMR_DiffGraphFormer.yaml
      DiffNMR_NMRNet.yaml
      DiffNMR.yaml
    train.py
    sample.py

How to Use

Refer to the install doc to install PaddleMaterials.

1. Prepare Vocabulary Table & Retrieval Database:

cd spectrum_elucidation
wget https://paddle-org.bj.bcebos.com/paddlematerials/checkpoints/spectrum_elucidation/diffnmr/vocab.tar.gz
tar -xvf vocab.tar.gz
wget https://paddle-org.bj.bcebos.com/paddlematerials/checkpoints/spectrum_elucidation/diffnmr/retrival_database.zip
unzip retrival_database.zip

2. Generate molecular structures from NMR spectra with DiffNMR:

cd pretrained
wget https://paddle-org.bj.bcebos.com/paddlematerials/checkpoints/spectrum_elucidation/diffnmr/DiffNMR_nless15_best.pdparams
cd ..
python spectrum_elucidation/sample.py --config_path='spectrum_elucidation/configs/diffnmr/DiffNMR.yaml' --weights_name='DiffNMR_nless15_best.pdparams' --save_path='result_diffnmr_nless15/' --checkpoint_path="pretrained"

3. Generate molecular structures with Retrieval-initializion/Similarity Filtering/Formula:

revise the spectrum_elucidation/configs/diffnmr/DiffNMR.py and replace the following arguments according to your needs:

flag_retrival_sampling: True
flag_use_formula: True
flag_retrival_initilization: True
num_candidates: 1 # recommend set to 20 for retrieval-sampling similarity filtering

run the command reference to step 2.

Citation

If you use DiffNMR, please cite:

@misc{yang2025diffnmr,
  title         = {DiffNMR: Diffusion Models for Nuclear Magnetic Resonance Spectra Elucidation},
  author        = {Yang, Qingsong and Wu, Binglan and Liu, Xuwei and Chen, Bo and Li, Wei and Long, Gen and Chen, Xin and Xiao, Mingjun},
  year          = {2025},
  eprint        = {2507.08854},
  archivePrefix = {arXiv},
  primaryClass  = {physics.chem-ph},
  doi           = {10.48550/arXiv.2507.08854},
  url           = {https://arxiv.org/abs/2507.08854}
}

License

This repository is released under the Apache-2.0 license (unless otherwise stated). See LICENSE for details.

Acknowledgements

Supported by the National Science and Technology Major Project (2023ZD0120702) and Basic Research Program of Jiangsu (BK20231215). We thank contributors of PaddlePaddle & PaddleMaterials