Loading…
NMR foundation model that turns 1D 1H spectra into reusable embeddings, with a shared encoder serving denoising, peak detection, and retrieval.
No providers recorded yet. Browse all providers
A lab that wants to denoise a proton spectrum, pick its peaks, and then search a structure library against it has historically needed three separate models, each trained end-to-end on its own labelled corpus. Nothing learned about lineshape, solvent effects, or field dependence in the first task carries over to the second. ROSE inverts that arrangement: it learns one representation of a 1D proton spectrum and hangs lightweight heads off it, so the expensive part is paid for once and reused.
The model was developed by Rombo AI, a spectral-analysis startup in Cagliari, with the University of Cagliari, and released in August 2026 as a ChemRxiv preprint. Its full name, Reusable One-dimensional Spectrum Embeddings, states the design goal directly. A spectrum is resampled onto a fixed linear grid of 4,096 points spanning 0–14 ppm, optionally annotated with the acquisition field in MHz and a solvent identifier, and encoded into a 256-dimensional vector. That vector is the product; the five task heads are demonstrations that it carries enough information to be useful.
ROSE is deliberately narrower than spectra-to-structure systems such as UltraNMR. The authors state that it is not a structure-elucidation solver: its retrieval and forward-prediction heads are auxiliary, not a replacement for assignment workflows. What it offers instead is a pretrained spectral front end for analytical tasks where labelled data is scarce.
The spectral encoder is a 1D vision transformer: a three-layer dilated convolutional stem feeds 32-point patches into eight transformer layers of width 256 with eight attention heads, plus four metadata tokens and a 32-entry solvent vocabulary. A separate four-layer graph transformer encodes molecular structure for the contrastive retrieval and forward-shift heads. The whole model is 7.8 million parameters, roughly an order of magnitude below UltraNMR's 120 million. The released checkpoint, ROSE-Pretrain-L, was trained on roughly 3.2 million experimental and simulated spectra assembled from fifteen public sources including HMDB, BMRB, NMRShiftDB2, MetaboLights, and GISSMO.
On identity-disjoint internal holdouts the native heads reach 0.95 cosine similarity on denoising, 87.8% Top-1 on spectrum pairing, 79.4% Top-1 on retrieval, and 0.80 F1 on peak detection at 0.05 ppm tolerance. Externally, a frozen encoder with a linear head reaches 98.8% balanced accuracy classifying 60-MHz edible-oil spectra from the QIB dataset, cuts NMRNet structure-to-spectrum Chamfer loss to 0.79 against 1.29 for a Morgan-fingerprint ridge baseline under the same protocol, and reaches 69% F1 on NMRformer peak detection. After five epochs of domain adaptation it reaches 37.1% Top-1 and 67.7% Top-10 retrieval on NMR-Solver against a roughly 30,000-molecule gallery. The authors are explicit that these are same-protocol comparisons, not head-to-head results against published stacks, which use different retrieval machinery and far larger galleries. Performance is uneven across field strengths: on the low-field slice at 100 MHz and below, retrieval Top-1 falls to 45.2% and peak F1 to 0.20, while denoising cosine similarity holds at 0.99.
The natural use is as a feature extractor for benchtop NMR in industrial and analytical settings, where spectra are plentiful but labels are expensive. Classifying edible oils from 60-MHz spectra with a linear probe is the worked example, and the same pattern applies to raw-material identity checks and batch-to-batch quality control. In metabolomics and natural-product work, the denoising and peak-detection heads offer preprocessing that transfers across instruments, and the embedding supports similarity search over spectral collections.
ROSE's contribution is less a benchmark result than a claim about factoring: that the spectrum side of NMR machine learning can be pretrained once and reused, in the way sequence encoders are reused across protein tasks. Releasing code and weights under Apache 2.0, together with the InChIKey-14 holdout and split manifests, makes that claim testable — though the training spectra themselves are not redistributed, since several upstream sources carry share-alike or non-commercial terms. The gap between high-field and low-field performance on structure-linked tasks is the honest limit of the current checkpoint, for a model targeting benchtop industrial analysis.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.