SPHERE-MS: Hybrid Fragment–Vocabulary Modeling for Accurate and Efficient Tandem Mass Spectra Prediction

Abstract Identifying known compounds and elucidating unknown chemical structures from mass spectra remains a significant challenge due to the limited size of experimental reference libraries. Enhancing these libraries with predicted mass spectra through computational methods offers an opportunity to resolve unmatched spectra. In this paper, we present SPHERE-MS (Secondary-loss Powered Hybrid Edge Removal Engine for predicting Mass Spectra), a graph neural network model designed to predict mass spectra from molecular structures. The hybrid approach, combining bond-breaking and vocabulary-based strategies, enables SPHERE-MS to efficiently simulate multilevel fragmentation processes without the need to enumerate all possible bond combinations. Evaluations on the NIST-20, MSnLib, CASMI2022, and MassSpecGym data sets demonstrate that SPHERE-MS outperforms baseline models that rely on single fragment generation strategy, showing higher prediction accuracy and library retrieval success rates. Notably, SPHERE-MS also maintains reasonable accuracy even when applied to molecules with low structural similarity to the training set. This strong generalization capability makes SPHERE-MS a valuable tool for mass spectra prediction across large libraries, including those containing uncommon structures.