Research map / Probabilistic and scientific modeling / Molecules and proteins
DiffMS: Diffusion Generation of Molecules Conditioned on Mass 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.
Proposes DiffMS, a formula-constrained encoder-decoder that uses a MIST transformer encoder and a discrete graph diffusion (DiGress-style) decoder, pretrained on 2.8M fingerprint-structure pairs, to generate molecules from mass spectra.
Problem
De novo generation of molecular structures from mass spectra is very hard: prior methods reach near-zero structural accuracy and autoregressive SMILES models ignore permutation invariance and formula constraints.
Approach
Condition a discrete graph diffusion decoder on heavy-atom composition from a known formula and on spectrum embeddings from a MIST formula transformer, and pretrain the decoder on fingerprint-to-structure pairs, which are far more plentiful than spectrum-structure pairs.
Key results
- On MassSpecGym, DiffMS reaches 2.30% top-1 accuracy versus 1.31% for MADGEN, with MCES 18.45 vs 27.47.
- On NPLIB1, DiffMS reaches 8.34% top-1 and 15.44% top-10 accuracy, versus 5.40% and 11.04% for the next best baseline (MIST + MSNovelist).
- Decoder pretraining on 0 to 2.8M structures raises NPLIB1 top-1 accuracy from 2.22% to 8.34%, with performance increasing with dataset size.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DiffMS instead of MIST + MSNovelist | molecule generation from spectra · NPLIB1 Top-1 | accuracy: 5.40% -> 8.34% | Helped |
| use DiffMS instead of MIST + Neuraldecipher | molecule generation from spectra · NPLIB1 Top-1 | MCES: 12.11 -> 11.95 | Helped |
| use DiffMS instead of MADGEN | molecule generation from spectra · MassSpecGym Top-1 | accuracy: 1.31% -> 2.30% | Helped |
| add spectrum-to-fingerprint encoder pretraining (ablation) | DiffMS · NPLIB1 Top-1 | accuracy: 4.36% -> 8.34% | Helped |
| add spectrum-to-fingerprint encoder pretraining (ablation) | DiffMS · NPLIB1 Top-1 | MCES: 12.34 -> 11.95 | Helped |
| add spectrum-to-fingerprint encoder pretraining (ablation) | DiffMS · NPLIB1 Top-1 | Tanimoto: 0.31 -> 0.35 | Helped |
| add spectrum-to-fingerprint encoder pretraining (ablation) | DiffMS · NPLIB1 Top-10 | accuracy: 11.46% -> 15.44% | Helped |
| add spectrum-to-fingerprint encoder pretraining (ablation) | DiffMS · NPLIB1 Top-10 | MCES: 9.31 -> 9.23 | Helped |
| add spectrum-to-fingerprint encoder pretraining (ablation) | DiffMS · NPLIB1 Top-10 | Tanimoto: 0.44 -> 0.47 | Helped |
| increase fingerprint-molecule decoder pretraining data from 0 to 2.8M structures (ablation) | DiffMS · NPLIB1 Top-1 | accuracy: 2.22% -> 8.34% | Helped |
| increase fingerprint-molecule decoder pretraining data from 0 to 2.8M structures (ablation) | DiffMS · NPLIB1 Top-1 | Tanimoto: 0.22 -> 0.35 | Helped |
| replace fully connected edge prior with marginal edge prior (ablation) | DiffMS · NPLIB1 Top-1 | accuracy: 3.36% -> 8.34% | Helped |
Surprising findings
- Using MIST-CF predicted formulae instead of true formulae yields comparable MCES and Tanimoto scores in some cases, since sampling across multiple formulae increases diversity.
Limitations
- Exact-structure accuracy remains low (2.30% top-1 on MassSpecGym).
- Hydrogen atoms are not modeled explicitly, so generated molecules may differ from the true formula in H count.
- Baselines were re-implemented by the authors, and MS2Mol could not be compared because its implementation was unavailable.
Experimental setup
Datasets: NPLIB1 MassSpecGym DSSTox HMDB COCONUT MOSES
Benchmarks: NPLIB1 MassSpecGym
Metrics: top-k accuracy top-k MCES top-k Tanimoto similarity % valid meaningful match rate close match rate
Baselines: Spec2Mol MADGEN MIST + Neuraldecipher MIST + MSNovelist SMILES Transformer SELFIES Transformer Random Generation
Builds on: DiGress MIST MSNovelist MADGEN Spec2Mol
Closest papers by meaning
- Geometric-Facilitated Denoising Diffusion Model for 3D Molecule Generation (AAAI 2024 · similarity 0.89)
- Navigating the Design Space of Equivariant Diffusion-Based Generative Models for De Novo 3D Molecule Generation (ICLR 2024 · similarity 0.89)
- Atomic Diffusion Models for Small Molecule Structure Elucidation from NMR Spectra (NeurIPS 2025 · similarity 0.89)
- All-atom Diffusion Transformers: Unified generative modelling of molecules and materials (ICML 2025 · similarity 0.89)
- Swallowing the Bitter Pill: Simplified Scalable Conformer Generation (ICML 2024 · similarity 0.89)
- Equivariant Blurring Diffusion for Hierarchical Molecular Conformer Generation (NeurIPS 2024 · similarity 0.88)
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.