Atomic Diffusion Models for Small Molecule Structure Elucidation from NMR Spectra

AI summary

Overview

• Research area: Machine learning for chemistry, specifically generative deep learning applied to NMR-based small-molecule structure elucidation.
• Technical level: Intermediate. The paper is readable for someone with general deep learning familiarity; NMR-specific terminology (chemical shifts, J-coupling, 1H/13C spectra) is defined in the background section, but some chemistry context helps.
• Scope: This paper introduces ChefNMR, an end-to-end conditional diffusion model that predicts a molecule's 3D atomic structure directly from its 1D 1H and 13C NMR spectra plus chemical formula, along with a new large-scale simulated dataset (SpectraNP) of 111,181 natural products.

What This Paper Is About

Determining the structure of an unknown small molecule from its NMR spectra is normally a slow, manual process that requires extensive expert interpretation of peak positions, intensities, and splitting patterns. This paper asks whether a deep generative model can solve that inverse problem directly, taking only raw 1D NMR spectra and a chemical formula and outputting the molecule's atoms and their 3D coordinates. The goal is to handle chemically complex natural products, not just the small, simple molecules that prior machine learning methods have targeted.

Key Contributions

1. ChefNMR, an end-to-end diffusion framework for NMR structure elucidation. The model generates 3D atomic coordinates conditioned on 1H and 13C NMR spectra and the chemical formula, using a conditional atomic diffusion model built on a non-equivariant Diffusion Transformer (DiT) architecture.

2. NMR-ConvFormer spectral encoder. A hybrid convolutional transformer that processes 1H and 13C spectra with a convolutional tokenizer for local peak features, a transformer encoder for intra- and inter-spectral dependencies, and multihead attention pooling (MAP) to produce a fixed-size conditioning embedding.

3. SpectraNP, a large-scale simulated NMR dataset of natural products. The authors curated simulated 1D NMR spectra for 111,181 natural products with up to 274 atoms, which the paper states significantly expands the chemical complexity of prior datasets (≤101 atoms).

4. A new graph-based baseline (NMR-DiGress) and systematic comparison of molecular representations. NMR-DiGress combines the discrete graph diffusion model DiGress with NMR-ConvFormer to test whether graph representations can match 3D atomic diffusion for this task.

Main Findings

• State-of-the-art accuracy on synthetic benchmarks: On SpectraNP, ChefNMR-L reaches 40.15% top-1 and 65.74% top-10 exact-match accuracy, versus 19.26% top-1 and 39.87% top-10 for Hu et al. and 2.12% top-1 and 9.17% top-10 for NMR-DiGress. On SpectraBase, ChefNMR-L reaches 72.04% top-1 and 88.20% top-10. On USPTO, ChefNMR-L reaches 81.57% top-1 and 93.01% top-10, with Tanimoto similarity of 0.973 at top-10.

• Reported headline claim: The abstract states ChefNMR predicts structures of challenging natural product compounds with an unsurpassed accuracy of over 65%, corresponding to the 65.74% top-10 accuracy on SpectraNP.

• Performance scales with model and dataset size: ChefNMR-S (134M parameters) outperforms baselines across all datasets, and ChefNMR-L (462M parameters) improves further. The paper reports the highest performance on the largest dataset, USPTO (745k molecules), followed by SpectraBase (141k) and SpectraNP (111k), and suggests expanding SpectraNP could further improve results on complex natural products.

• Zero-shot generalization to experimental spectra: Trained on USPTO, ChefNMR achieves 56% top-1 accuracy on SpecTeach and 21% on NMRShiftDB2, outperforming Hu et al. and NMR-DiGress, which the paper states generalize poorly to both experimental benchmarks. Notably, the model can find correct top-1 structures despite solvent effects and impurities.

• Graph representations underperform 3D atomic diffusion: NMR-DiGress reaches 43.56% top-1 on SpectraBase but collapses to 2.12% top-1 on SpectraNP, while ChefNMR-S reaches 40.37% top-1 on SpectraNP.

• Coordinate augmentation matters most in ablations: On SpectraBase, removing coordinate augmentation drops ChefNMR-S from 69.15% to 49.75% top-1 (described in the paper as coordinate augmentation improving accuracy by 20%). Removing smooth LDDT loss gives 68.31%, removing the convolutional tokenizer gives 61.78%, removing token count reduction gives 66.28%, removing the transformer block gives 68.12%, removing MAP pooling gives 62.97%, and removing dropout gives 65.48%.

• Incorrect predictions are often chemically plausible: Qualitative examples on SpectraNP show that many erroneous top-k predictions remain chemically valid and structurally similar to the ground truth.

Methodology in Plain English

The problem is framed as conditional generation: given a molecule's chemical formula (atom types) and its 1H and 13C NMR spectra, produce the 3D coordinates of its atoms.

The spectra are handled by NMR-ConvFormer. Each spectrum is first passed through two 1D convolutional layers with ReLU and max-pooling, which compress the signal and summarize local patterns like peak intensity and splitting. The resulting token sequence, augmented with positional and type embeddings, goes through a standard transformer encoder with multi-head self-attention, letting the model relate peaks within a spectrum and across the 1H and 13C spectra. A learnable [CLS] token aggregated by a multihead attention pooling layer produces a single fixed-size embedding that conditions the generator.

Generation uses an EDM-style diffusion model. During training, clean coordinates are perturbed with Gaussian noise at a randomly sampled noise level, and the network is trained to predict the clean coordinates back. The training objective combines a mean-squared-error term for global structure with a smooth Local Distance Difference Test (LDDT) loss adapted from AlphaFold3, which penalizes pairwise distance deviations at thresholds of 0.5, 1.0, 2.0, and 4.0 Å; unlike the protein version, it is computed over all atom pairs because small molecules are more compact. Three ground-truth conformers are generated per molecule with RDKit's ETKDGv3, and one is randomly selected per training step and given a random rigid translation and rotation to encourage SE(3)-invariant representations.

The denoising network is a Diffusion Transformer that takes atom tokens formed by concatenating noisy coordinates and atom types, with the noise level embedded via frequency encoding and added to the spectral embedding, injected into the DiT blocks through adaptive layer normalization. At inference, classifier-free guidance combines conditional and unconditional predictions; the 1H spectrum, the 13C spectrum, and both are each dropped with probability 0.1 during training, with guidance scale ω set to 1, 1.5, or 2 depending on the dataset.

Data sources: synthetic spectra are simulated with MestReNova, conformers come from RDKit ETKDGv3, spectra are interpolated to 10,000-dimensional vectors, 13C spectra for SpectraBase and the experimental datasets are binned into 80 binary vectors, and molecules whose SMILES cannot be reconstructed from atom types and 3D coordinates via RDKit's DetermineBonds are discarded. Evaluation uses top-k exact SMILES matching and top-k maximum Tanimoto similarity of Morgan fingerprints (length 2048, radius 2).

Why This Matters

• Impact on research: The work argues that automating structure elucidation from raw 1D NMR data could accelerate chemistry, biomedicine, and natural product drug discovery, and positions 3D atomic diffusion as a viable representation for chemical inverse problems. It also provides a new large, chemically complex benchmark (SpectraNP) for the community.

• Natural product drug discovery: The paper notes that natural products have delivered more than half of FDA-approved small-molecule agents, including the majority of antibiotics and antitumor drugs in current clinical use, and that finding new ones currently requires de novo experimental work on every molecule.

• Antibiotic and therapeutic development: Because natural products underpin antibiotics and antitumor drugs such as penicillin and taxol, faster elucidation could shorten discovery cycles for new therapeutics.

• Metabolomics and cellular metabolite discovery: Identifying unknown metabolites from biological samples is a bottleneck for understanding metabolic pathways, which spectral interpretation currently slows down.

• Chemistry education and routine lab analysis: Datasets like SpecTeach, built for spectroscopy education, indicate a role for automated interpretation tools in training and in routine analytical workflows.

• Industry relevance: Pharmaceutical, agrochemical, and biotech organizations that screen natural products or synthesize novel compounds rely on NMR characterization; a model that returns ranked candidate structures could reduce expert time and prioritize which candidates to pursue. The acknowledgements list support from Janssen Pharmaceuticals, Generate Biomedicines, the Chan Zuckerberg Imaging Institute, and the Princeton Catalysis Initiative.

Future Directions

1. Expanding training data: The authors state that adding experimental spectra and more natural products could further improve performance, and their scaling observation suggests SpectraNP itself could be enlarged.

2. Incorporating 2D NMR: The paper identifies 2D NMR spectra as additional information that could be used to resolve stereochemistry, which the current model does not address (SMILES are canonicalized with stereochemistry removed).

3. Confidence estimation: Adding a confidence module is proposed so chemists can better assess the reliability of predicted structures before acting on them.

4. Closing the synthetic-to-experimental gap: Zero-shot accuracy drops to 56% on SpecTeach and 21% on NMRShiftDB2 when training only on USPTO, leaving open how best to handle solvent peaks, impurities, and baseline noise.

5. Validation and responsible deployment: The conclusion emphasizes that careful validation and responsible deployment will be essential for safe and impactful real-world use.

Target Audience

• Machine learning researchers working on generative models, diffusion models, and scientific applications of deep learning.
• Computational chemists and cheminformatics practitioners interested in automating spectral interpretation and structure elucidation.
• Natural products chemists, metabolomics researchers, and drug discovery scientists who handle NMR data and complex molecules.
• Benchmark and dataset builders interested in SpectraNP as a resource for chemically complex structure elucidation, and researchers comparing molecular representations such as chemical language models, graph models, and 3D atomic diffusion.

Authors’ abstract

Nuclear Magnetic Resonance (NMR) spectroscopy is a cornerstone technique for determining the structures of small molecules and is especially critical in the discovery of novel natural products and clinical therapeutics. Yet, interpreting NMR spectra remains a time-consuming, manual process requiring extensive domain expertise. We introduce ChefNMR (CHemical Elucidation From NMR), an end-to-end framework that directly predicts an unknown molecule's structure solely from its 1D NMR spectra and chemical formula. We frame structure elucidation as conditional generation from an atomic diffusion model built on a non-equivariant transformer architecture. To model the complex chemical groups found in natural products, we generated a dataset of simulated 1D NMR spectra for over 111,000 natural products. ChefNMR predicts the structures of challenging natural product compounds with an unsurpassed accuracy of over 65%. This work takes a significant step toward solving the grand challenge of automating small-molecule structure elucidation and highlights the potential of deep learning in accelerating molecular discovery. Code is available at https://github.com/ml-struct-bio/chefnmr.