From Human Labels to Literature: Semi-Supervised Learning of NMR Chemical Shifts at Scale

Accurate prediction of nuclear magnetic resonance (NMR) chemical shifts is fundamental to spectral analysis and molecular structure elucidation, yet existing machine learning methods rely on limited, labor-intensive atom-assigned datasets. We propose a semi-supervised framework that learns NMR chemical shifts from millions of literature-extracted spectra without explicit atom-level assignments, integrating a small amount of labeled data with large-scale unassigned spectra. We formulate chemical shift prediction from literature spectra as a permutation-invariant set supervision problem, and show that under commonly satisfied conditions on the loss function, optimal bipartite matching reduces to a sorting-based loss, enabling stable large-scale semi-supervised training beyond traditional curated datasets. Our models achieve substantially improved accuracy and robustness over state-of-the-art methods and exhibit stronger generalization on significantly larger and more diverse molecular datasets. Moreover, by incorporating solvent information at scale, our approach captures systematic solvent effects across common NMR solvents for the first time. Overall, our results demonstrate that large-scale unlabeled spectra mined from the literature can serve as a practical and effective data source for training NMR shift models, suggesting a broader role of literature-derived, weakly structured data in data-centric AI for science.

Nuclear magnetic resonance (NMR) spectroscopy is a key tool chemists use to determine the structure of molecules, but predicting NMR chemical shifts with machine learning has been limited by the small size of carefully labeled datasets. We developed a semi-supervised learning framework that learns from millions of NMR spectra extracted from the scientific literature, which lack explicit atom-level assignments, by treating the prediction problem as matching sets of numbers rather than individual labels. A key theoretical insight shows that under mild conditions, this matching simplifies to a sorting-based loss, making large-scale training stable and practical. Our models outperform existing methods in accuracy and generalization across larger and more chemically diverse molecules. For the first time, we also capture how different solvents systematically shift NMR peaks by training on solvent-annotated literature data at scale. This work demonstrates that weakly structured data mined from the scientific literature can serve as a powerful resource for building more accurate and generalizable models in AI for science.