Fragment-Grounded Neural Simulation of Electron Ionization Mass Spectra at Library Scale

Abstract

Electron ionization mass spectrometry (EI-MS) is a sensitive, reproducible analytical technique, yet compound identification conventionally relies on reference databases that capture only a fraction of accessible chemical space. In silico prediction of EI-MS spectra can extend these libraries, but existing data-driven models generalize poorly beyond their training distribution, and physics-based simulations are too costly for library-scale use. Here, we present a fragment-grounded neural workflow that predicts an EI-MS spectrum by exhaustively enumerating the fragmentation graph of a molecule and then assigning intensities to the resulting substructures. By relating every predicted peak back to a candidate substructure, our model generalizes more reliably to unseen scaffolds and out-of-distribution chemistry than prior fingerprint-and black-box transformer-based predictors. On a scaffold split, it improves top-1 retrieval accuracy from formula-matched compounds by 32% relative to the strongest baseline, and it matches the accuracy of quantum-chemical simulation while being roughly seven orders of magnitude faster. To test the model’s capabilities under a harder, formula-agnostic setting, we further evaluate retrieval against a library of 87.3 million compounds from