Spectra to structure elucidation

spectra-to-structure-elucidation

Summary

Apply spectrum-conditioned diffusion modeling to generate candidate molecular structures from multi-modal spectroscopic input (IR, Raman, UV-Vis, mass spectra, NMR). This skill unifies multi-modal spectral reasoning with joint 2D topology and 3D geometry generative modeling to perform de novo molecular structure elucidation.

When to use

You have one or more spectroscopic datasets (IR, Raman, UV-Vis, mass spectra, NMR) from an unknown compound and need to generate candidate molecular structures ranked by likelihood. Use this when retrieval-based approaches are infeasible (e.g., compound is not in a library) and you require both 2D connectivity and 3D geometry in the output.

When NOT to use

• Spectroscopic input has not been preprocessed or normalized according to DiffSpectra pipeline requirements — preprocessing is mandatory.
• The unknown compound is known to be in an existing chemical library; use retrieval-based methods instead.
• Only 2D topology is needed and 3D geometry is not relevant to the application.

Inputs

• Multi-modal spectroscopic data (IR, Raman, UV-Vis, mass spectra, or combinations)
• Preprocessed and tokenized spectra following DiffSpectra normalization requirements
• Pretrained model checkpoint (DMT + SpecFormer weights)
• Configuration file specifying spectra_version, model.name, eval.num_samples, eval.ckpts

Outputs

• Ranked list of candidate molecular structures in SMILES, InChI, or molecular graph format
• Confidence scores or reconstruction likelihood metrics for each candidate
• Top-k accuracy metrics (e.g., top-1 and top-10 exact match rates)
• 3D structural metrics (e.g., RMSD) if evaluated against reference structures

How to apply

Load preprocessed spectroscopic input data and initialize a pretrained diffusion model (DMT + SpecFormer) conditioned on the spectra. The SpecFormer encoder captures intra- and inter-spectrum dependencies across modalities; the SE(3)-equivariant DMT denoises iteratively to generate 2D/3D molecular candidates. Run the diffusion process (typically 40 checkpoint steps) to generate a sample pool (e.g., 10,000 molecules). Decode latent representations into SMILES, InChI, or molecular graph format. Rank candidates by model confidence or reconstruction likelihood and filter by chemical validity. Evaluate top-k accuracy (e.g., top-1 and top-10 exact match rates) against ground truth if available.

Related tools

• DiffSpectra (Diffusion Molecule Transformer + SpecFormer) (Pretrained generative model that performs spectrum-conditioned diffusion to jointly model 2D topology and 3D geometry; SE(3)-equivariant denoising network with multi-modal spectral encoder) — https://github.com/AzureLeon1/DiffSpectra
• SpecFormer (Transformer-based spectral encoder within DiffSpectra that captures intra- and inter-spectrum dependencies across IR, Raman, UV-Vis, and other modalities) — https://github.com/AzureLeon1/DiffSpectra
• Diffusion Molecule Transformer (DMT) (SE(3)-equivariant denoising network that models both 2D topology and 3D geometry of molecules during iterative generation) — https://github.com/AzureLeon1/DiffSpectra

Examples

CUDA_VISIBLE_DEVICES=0,1 python main.py --config configs/diffspectra_qm9s.py --config_original_qm9 configs/base_qm9.py --mode eval --workdir exp/allspectra --config.eval.ckpts '40' --config.eval.num_samples 10000 --config.eval.save_mols true --config.data.spectra_version allspectra

Evaluation signals

• Top-1 and top-10 exact match accuracy: percentage of test cases where ground-truth structure ranks in top-k predictions (expected: ≥40.76% top-1, ≥99.49% top-10 based on paper results)
• Chemical validity: all generated molecules pass RDKit validity checks and have correct valence/aromaticity
• Reconstruction likelihood: confidence scores for top-ranked predictions are higher than those for lower-ranked candidates
• 3D geometry fidelity: RMSD between predicted 3D coordinates and reference structures falls within acceptable range for the application domain
• Spectra consistency: generated structures, when re-embedded and spectrum-predicted, should reconstruct input spectra with low reconstruction error

Limitations

• Performance is spectrum-dependent: single-modality predictions (IR or Raman only) show lower accuracy than multi-modal fusion; UV-Vis alone is less informative than combined modalities
• Requires substantial training data and pretrained model checkpoints; performance degrades on out-of-distribution spectroscopic inputs or rare molecular scaffolds not well-represented in training set
• Computational cost scales with number of diffusion steps (40 steps per model) and sample pool size; generating 10,000 candidates requires multi-GPU resources
• Does not recover stereochemistry information that is not encoded in spectroscopic modalities; 2D exact match may not imply correct stereoisomer assignment

Evidence

• [readme] multi-modal spectral data: "generative framework for molecular structure elucidation from multi-modal spectral data"
• [readme] SE(3)-equivariant denoising network: "Diffusion Molecule Transformer (DMT): An SE(3)-equivariant denoising network that models both 2D topology and 3D geometry of molecules"
• [readme] SpecFormer transformer encoder: "SpecFormer: A transformer-based spectral encoder that captures intra- and inter-spectrum dependencies across diverse spectral modalities (e.g., IR, Raman, UV-Vis)"
• [readme] 40.76% top-1 accuracy and 99.49% top-10 accuracy: "Extensive experiments demonstrate that DiffSpectra achieves 40.76% top-1 accuracy and 99.49% top-10 accuracy in recovering exact molecular structures"
• [readme] conditional diffusion process: "DiffSpectra formulates structure elucidation as a conditional diffusion process"
• [readme] spectrum-conditioned diffusion modeling: "Through spectrum-conditioned diffusion modeling, DiffSpectra unifies multi-modal reasoning with 2D/3D generative modeling"
• [readme] first framework unifying multi-modal spectral reasoning: "DiffSpectra is the first framework that unifies multi-modal spectral reasoning and joint 2D/3D generative modeling for de novo molecular structure elucidation"
• [other] Preprocess spectra and initialize diffusion model: "Preprocess spectra according to DiffSpectra pipeline normalization and tokenization requirements. 3. Initialize the pretrained diffusion model"
• [other] Decode into SMILES, InChI, or molecular graph: "Decode the generated latent representations into molecular structure format (SMILES, InChI, or molecular graph)"