Research map / Image generation / Autoregressive and latent generation
Atomic Diffusion Models for Small Molecule Structure Elucidation from NMR Spectra
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces ChefNMR, a conditional 3D atomic diffusion model (NMR-ConvFormer encoder + DiT) that generates small-molecule 3D structures from 1D 1H/13C NMR spectra and formula, trained on a new 111k natural-product simulated-spectra dataset SpectraNP.
Problem
Automating de novo small-molecule structure elucidation directly from raw 1D NMR spectra and chemical formula, especially for complex natural products with many atoms and large rings, where prior methods handle only simple molecules.
Approach
Frame structure elucidation as conditional 3D atomic coordinate generation with an EDM-style diffusion transformer conditioned on a hybrid convolutional-transformer spectral encoder, using random rigid-transform augmentation, smooth LDDT loss, and classifier-free guidance over the two spectra.
Key results
- On SpectraNP, ChefNMR-L reaches 40.15% top-1 and 65.74% top-10 accuracy versus 19.26% and 39.87% for Hu et al.
- On USPTO top-1 accuracy is 81.57% for ChefNMR-L versus 38.02% for Hu et al. (Alberts et al. reported 73.38%)
- Zero-shot on experimental data: 56% top-1 on SpecTeach and 21% on NMRShiftDB2
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ChefNMR-L instead of Hu et al. | NMR structure elucidation · SpectraBase | Top-1 accuracy: 45.24{ $\pm$.18} -> \textbf{72.04{ $\pm$.02}} | Helped |
| Use ChefNMR-L instead of Alberts et al. | NMR structure elucidation · USPTO | Top-1 accuracy: 73.38{ $\pm$.08} -> \textbf{81.57{ $\pm$.09}} | Helped |
| Use ChefNMR-S instead of Hu et al. | NMR structure elucidation · SpectraNP | Top-1 accuracy: 19.26{ $\pm$.10} -> \textbf{40.37{ $\pm$.33}} | Helped |
| Increase model size from ChefNMR-S to ChefNMR-L | ChefNMR · SpectraBase | Top-1 accuracy: 69.15{ $\pm$.08} -> \textbf{72.04{ $\pm$.02}} | Helped |
| Increase model size from ChefNMR-S to ChefNMR-L | ChefNMR · USPTO | Top-1 accuracy: 81.16{ $\pm$.08} -> \textbf{81.57{ $\pm$.09}} | Helped |
| Increase model size from ChefNMR-S to ChefNMR-L | ChefNMR · SpectraNP | Top-1 accuracy: \textbf{40.37{ $\pm$.33} -> 40.15{ $\pm$.29} | Mixed |
| Add coordinate augmentation (ablation) | ChefNMR-S · SpectraBase | Top-1 accuracy: 49.75 -> 69.15 | Helped |
| Add smooth LDDT loss (ablation) | ChefNMR-S · SpectraBase | Top-1 accuracy: 68.31 -> 69.15 | Helped |
| Add convolutional tokenizer (ablation) | ChefNMR-S · SpectraBase | Top-1 accuracy: 61.78 -> 69.15 | Helped |
| Add token count reduction (ablation) | ChefNMR-S · SpectraBase | Top-1 accuracy: 66.28 -> 69.15 | Helped |
| Add transformer block (ablation) | ChefNMR-S · SpectraBase | Top-1 accuracy: 68.12 -> 69.15 | Helped |
| Add MAP pooling (ablation) | ChefNMR-S · SpectraBase | Top-1 accuracy: 62.97 -> 69.15 | Helped |
Surprising findings
- NMR-DiGress, a graph-diffusion baseline, performs very poorly on SpectraNP (2.12% top-1), far below the chemical language model baseline
- Removing coordinate augmentation drops top-1 accuracy from 69.15% to 49.75% on SpectraBase
Limitations
- Trained only on simulated spectra; experimental performance is zero-shot and lower
- Does not resolve stereochemistry; requires 2D NMR for stereo information
- No confidence module to assess reliability of predictions
Experimental setup
Datasets: SpectraBase USPTO SpectraNP SpecTeach NMRShiftDB2 NPAtlas NP-MRD
Benchmarks: SpectraBase USPTO SpectraNP SpecTeach NMRShiftDB2
Metrics: top-k matching accuracy top-k maximum Tanimoto similarity
Baselines: Hu et al. Alberts et al. NMR-DiGress
Builds on: EDM Diffusion Transformer DiGress AlphaFold3 smooth LDDT loss classifier-free guidance
Closest papers by meaning
- DiffMS: Diffusion Generation of Molecules Conditioned on Mass Spectra (ICML 2025 · similarity 0.89)
- Can LLMs Solve Molecule Puzzles? A Multimodal Benchmark for Molecular Structure Elucidation (NeurIPS 2024 · similarity 0.87)
- Unraveling Molecular Structure: A Multimodal Spectroscopic Dataset for Chemistry (NeurIPS 2024 · similarity 0.86)
- Learning Over Molecular Conformer Ensembles: Datasets and Benchmarks (ICLR 2024 · similarity 0.86)
- Swallowing the Bitter Pill: Simplified Scalable Conformer Generation (ICML 2024 · similarity 0.85)
- Geometric-Facilitated Denoising Diffusion Model for 3D Molecule Generation (AAAI 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.