Abstract
Accurate metabolite annotation via tandem mass spectrometry remains a major bottleneck in untargeted metabolomics. Recent implicit models that avoid molecular generation or spectra simulation have shown competitive performance by aligning spectra and molecular structures in the embedding space. Still, they overlook the detailed relationships between spectral peaks and molecular substructures that govern fragmentation. We introduce FLARE (Fine-grained Learning for Alignment of spectra-molecule REpresentations), a contrastive learning framework that leverages bidirectional peak-node alignment under learned weak supervision. Unlike models that rely solely on global embeddings, FLARE computes similarity via maxima over peak-to-atom and atom-to-peak interactions, capturing chemically meaningful local correspondences and enabling interpretable spectra-molecule matching. It achieves state-of-the-art results on MassSpecGym, with 43.15% rank@1 (mass-based) and 22.66% (formula-based), surpassing previous models by over 63%. FLARE's learned embeddings correspond with molecular classes, match fingerprint similarity, and detect differential metabolites in a breast cancer xenograft study, showcasing its translational potential.