Metabolite annotation model that scores MS/MS spectra against candidate molecules by matching individual peaks to atoms, not pooled embeddings.
No providers recorded yet. Browse all providers
A tandem mass spectrum is produced when a molecule breaks apart, and each peak records one fragment. Which atoms end up in which fragment is decided locally, by the bonds that break. Contrastive annotation models such as JESTR nonetheless pool every peak into one spectrum vector and every atom into one molecule vector, then compare the two vectors. The pooling discards the peak-to-substructure correspondences that fragmentation actually produces, and it leaves the similarity score with nothing a chemist can inspect.
FLARE (Fine-grained Learning for Alignment of spectra–molecule REpresentations) keeps both modalities as sets of tokens. Every peak is compared with every atom of a candidate molecule. For each peak the model keeps its best-matching atom, and for each atom its best-matching peak. The final score is the average of the two directions, so a candidate ranks highly only when the spectrum and the structure support each other. The approach adapts FILIP's token-level matching from vision–language models to a setting where the training data never states which peak comes from which substructure.
FLARE was developed by Yan Zhou Chen and Soha Hassoun at Tufts University, with Blake Rushing of the Nutrition Research Institute at the University of North Carolina at Chapel Hill, who supplied the tumor metabolomics data. The bioRxiv preprint was posted in January 2026. FLARE follows JESTR and MVP in the Hassoun lab's line of joint-embedding annotation models.
The spectrum encoder takes the 60 most intense peaks. Each peak is given a subformula within ±20 ppm using MIST's enumeration algorithm and encoded as a 14-element count vector plus intensity. The precursor formula is added as an extra token. Peaks pass through an MLP projection and a two-layer, four-head transformer encoder. Molecules are atom graphs with ten atom-level features, encoded by a graph convolutional network. Peak and node embeddings are 512-dimensional. Training uses a symmetric InfoNCE loss in which the fine-grained score replaces the usual global cosine similarity. Optuna tuning set the temperature to 0.0227. The model assumes that the precursor formula and adduct are known.
FLARE was trained and evaluated on MassSpecGym (194,119 training spectra), whose test molecules are separated from all training molecules by an MCES distance of at least 10. With candidates retrieved by mass, it reaches 43.15% rank@1, 75.59% rank@5 and 92.89% rank@20. The best previous model, MVP, reaches 26.37% rank@1 and JESTR 15.13%. With candidates retrieved by formula, FLARE reaches 22.66% rank@1, against 11.85% for JESTR. Spectral embeddings correlate with Tanimoto similarity at Pearson 0.664 over more than 50,000 pairs. That is higher than MS2DeepScore and DreaMS in the same comparison. FLARE's gains over JESTR grow for test molecules least similar to the training set, and the two models' rankings agree only weakly (mean Spearman 0.34).
FLARE is built for untargeted LC–MS/MS metabolomics, where most spectra have no library match. In patient-derived xenograft tumors from metastatic and primary breast cancers, candidates were retrieved from HMDB, COCONUT and spectral-library molecules. Against in-house standards, FLARE ranked the correct structure first for 39.19% of the 74 spectra whose true structure was among the candidates. Differential analysis on its confident annotations found 54 features separating metastatic from primary tumors and two metastatic subgroups. For 23 differential features lacking confident annotations, nearest-neighbour search in the embedding space predicted a chemical superclass and peak–node maps suggested substructures. The data from this case study are deposited at Metabolomics Workbench as ST003752, under embargo until February 2027.
FLARE brings token-level cross-modal alignment into metabolite annotation. Its benchmark gain over global contrastive models such as JESTR, CSU-MS² and MS²-SMILES AlignNet matters, but the larger contribution may be that a ranking now comes with evidence a mass spectrometrist can check against known fragmentation. The code, training and evaluation scripts and the pretrained checkpoint are distributed through a Hugging Face Space. FLARE is a preprint not yet peer reviewed, evaluated on one benchmark and one case study, and it depends on subformula assignments the authors identify as a source of errors.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.