ROSE: a Foundation Model for Reusable One-dimensional Spectrum Embeddings in 1H NMR
- 1Italy
- 2Università di Cagliari
Abstract
Most deep learning models for 1H NMR are trained for a single endpoint and do not provide representations that can be reused across analytical tasks. We introduce ROSE (Reusable One-dimensional Spectrum Embeddings), a 7.8-million-parameter spectrum-side foundation model pretrained on 3.2 million experimental and simulated spectra using an InChIKey-14 identity holdout that excludes all molecules appearing in the external benchmark test sets. Its shared spectral encoder supports denoising, peak detection, classification, cross-field alignment, molecular retrieval and structure-tospectrum prediction. Beyond internal evaluation on identity-disjoint holdouts, ROSE was assessed on external benchmarks excluded from pretraining, matching the published QIB PLS-DA result for 60-MHz edible-oil classification (98.8% balanced accuracy vs 98.9% accuracy), reducing NMRNet structure-to-spectrum Chamfer loss by 39% over Morgan+Ridge (0.79 vs 1.29), achieving 69% peak-detection F1 on NMRformer, and, after short domain adaptation, reaching 37.1% Top-1 and 67.7% Top-10 retrieval on NMR-Solver. ROSE represents a first step toward a general-purpose NMR foundation model for scalable industrial analytical applications.
Related
- Covalently Functionalized Two-Dimensional Red Phosphorus Nanosheets: Efficient Platform for Broad-Spectrum Biofilm Eradication
- Impact of silica nanoparticles architectures on the photosensitization of O2 by immobilized Rose Bengal
- ROSE: Deep Learning exchange-correlation functional from reduced density-matrix representations
- Ultrasound-assisted cloud point extraction of phenolamide from rose residues: optimization, characterization and molecular interactions
- Aqueous-rich flow intensification for enzymatic transesterification of rose geranium oil in a coiled flow inverter microreactor
BibTeX
@misc{iorio2026rose,
title = {{ROSE: a Foundation Model for Reusable One-dimensional Spectrum Embeddings in 1H NMR}},
author = {Mattia Di Iorio and Carmine Mattia and Andrea Zanda and Maurizio Atzori},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15007823/v2},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15007823/v2},
note = {CC BY},
}
RIS
TY - UNPB
TI - ROSE: a Foundation Model for Reusable One-dimensional Spectrum Embeddings in 1H NMR
AU - Mattia Di Iorio
AU - Carmine Mattia
AU - Andrea Zanda
AU - Maurizio Atzori
PY - 2026
DA - 2026/09/30
PB - ChemRxiv
DO - 10.26434/chemrxiv.15007823/v2
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15007823/v2
AB - Most deep learning models for 1H NMR are trained for a single endpoint and do not provide representations that can be reused across analytical tasks. We introduce ROSE (Reusable One-dimensional Spectrum Embeddings), a 7.8-million-parameter spectrum-side foundation model pretrained on 3.2 million experimental and simulated spectra using an InChIKey-14 identity holdout that excludes all molecules appearing in the external benchmark test sets. Its shared spectral encoder supports denoising, peak detection, classification, cross-field alignment, molecular retrieval and structure-tospectrum prediction. Beyond internal evaluation on identity-disjoint holdouts, ROSE was assessed on external benchmarks excluded from pretraining, matching the published QIB PLS-DA result for 60-MHz edible-oil classification (98.8% balanced accuracy vs 98.9% accuracy), reducing NMRNet structure-to-spectrum Chamfer loss by 39% over Morgan+Ridge (0.79 vs 1.29), achieving 69% peak-detection F1 on NMRformer, and, after short domain adaptation, reaching 37.1% Top-1 and 67.7% Top-10 retrieval on NMR-Solver. ROSE represents a first step toward a general-purpose NMR foundation model for scalable industrial analytical applications.
ER -