Research map / Vision-language and multimodal / Cross-modal learning
Unraveling Molecular Structure: A Multimodal Spectroscopic Dataset for Chemistry
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 a simulated multimodal spectroscopy dataset (IR, 1H-NMR, 13C-NMR, HSQC-NMR, positive/negative MS/MS) for ~790k USPTO molecules, with baseline benchmarks for structure elucidation, spectrum prediction, and functional group prediction.
Problem
Machine learning for molecular structure elucidation has been limited to single spectroscopic modalities because no large, open, multimodal spectral dataset exists.
Approach
Simulate IR (molecular dynamics dipole-based), NMR (MestReNova), and MS/MS (CFM-ID 4.0, SCARF, ICEBERG) spectra for 794k patent-derived molecules from USPTO, and benchmark transformer, XGBoost, and 1D-CNN baselines on single-modality tasks.
Key results
- Structure elucidation from combined 1H+13C NMR reaches 73.38% Top-1 accuracy, versus 64.99% for 1H-NMR alone and 9.97% for IR.
- Transformers achieve the best F1 on most functional group tasks, e.g. 0.935 on 1H-NMR and 0.913 on 13C-NMR, while 1D-CNN is marginally better on IR (0.895 vs 0.881).
- Simulated spectra show higher similarity to matched experimental spectra than to other experimental spectra across modalities (e.g. 1H-NMR cosine 21.9 vs 6.8 in the initial comparison).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| combine 1H-NMR and 13C-NMR spectra instead of using 1H-NMR alone | Transformer for structure elucidation · Structure elucidation | Top--1% accuracy: 64.99 $\pm$ 0.31 -> 73.38 $\pm$ 0.08 | Helped |
| use 1H-NMR rather than 13C-NMR spectra | Transformer for structure elucidation · Structure elucidation | Top--1% accuracy: 51.95 $\pm$ 0.29 -> 64.99 $\pm$ 0.31 | Helped |
| use combined 1H-NMR and 13C-NMR rather than 1H-NMR alone | Transformer for structure elucidation · Structure elucidation | Top--1% accuracy: 64.99 $\pm$ 0.31 -> 73.38 $\pm$ 0.08 | Helped |
| use 1D-CNN rather than Transformer | Functional group prediction from IR · Functional group prediction from IR | F1 score: 0.881 $\pm$ 0.021 -> 0.895 $\pm$ 0.002 | Helped |
| use Transformer rather than 1D-CNN | Functional group prediction · Functional group prediction from MS/MS (CFM-ID, Positive) | F1 score: 0.645 $\pm$ 0.001 -> 0.897 $\pm$ 0.012 | Helped |
| increase MS/MS ionisation energy from 10 eV to 40 eV | Transformer for spectrum prediction · Spectrum prediction from molecule: MS/MS (Positive) | Cosine Similarity: 83.94 $\pm$ 0.10 -> 66.35 $\pm$ 0.15 | Hurt |
Surprising findings
- SCARF-simulated MS/MS yields very low structure elucidation accuracy (1.92% Top-1) compared with CFM-ID (23.53%)
- CFM-ID models trained on simulated spectra score 0% zero-shot on the experimental van Bramer dataset
Limitations
- All spectra are simulated, causing a sim-to-real distribution shift that likely requires finetuning on experimental spectra
- Chemical space is limited to patent-derived, synthesizable molecules, so natural products are underrepresented
- Only single-modality benchmarks are provided; multimodal models are left to future work
Experimental setup
Datasets: USPTO reaction dataset van Bramer et al. experimental spectra NIST Gas-Phase IR Database nmrshiftdb2 Pistachio GNPS
Benchmarks: Multimodal Spectroscopic Dataset (structure elucidation) van Bramer et al. experimental spectra (zero-shot)
Metrics: Top-1/5/10 accuracy F1 cosine similarity token accuracy
Baselines: XGBoost 1D-CNN CFM-ID 4.0 SCARF ICEBERG
Builds on: USPTO reaction dataset CFM-ID 4.0 SCARF ICEBERG OpenNMT-py
Closest papers by meaning
- MS-BART: Unified Modeling of Mass Spectra and Molecules for Structure Elucidation (NeurIPS 2025 · similarity 0.87)
- Atomic Diffusion Models for Small Molecule Structure Elucidation from NMR Spectra (NeurIPS 2025 · similarity 0.86)
- ChemPile: A 250 GB Diverse and Curated Dataset for Chemical Foundation Models (NeurIPS 2025 · similarity 0.86)
- Can LLMs Solve Molecule Puzzles? A Multimodal Benchmark for Molecular Structure Elucidation (NeurIPS 2024 · similarity 0.86)
- DiffMS: Diffusion Generation of Molecules Conditioned on Mass Spectra (ICML 2025 · similarity 0.85)
- RamanBench: A Large-Scale Benchmark for Machine Learning on Raman Spectroscopy (NeurIPS 2026 · 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.