Neural Spectral Prediction for Structure Elucidation with Tandem Mass Spectrometry

Abstract

Structural elucidation using untargeted tandem mass spectrometry (MS/MS) has played a critical role in advancing scientific discovery [1, 2]. However, differentiating molecular fragmentation patterns between isobaric structures remains a prominent challenge in metabolomics [3-10], drug discovery [11-13], and reaction screening [14-17], presenting a significant barrier to the cost-effective and rapid identification of unknown molecular structures. Here, we present a geometric deep learning model, ICEBERG, that simulates high-energy collision-induced dissociation in mass spectrometry to generate chemically plausible fragments and their relative intensities with awareness of collision energies and polarities. We utilize ICEBERG predictions to facilitate structure elucidation by ranking a set of candidate structures based on the similarity between their predicted in silico MS/MS spectra and an experimental MS/MS spectrum of interest. This integrated elucidation pipeline enables state-of-the-art performance in compound annotation, with 40% top-1 accuracy on the NIST20 [M+H]+ adduct subset and with 92% of correct structures appearing in the top ten predictions in the same dataset. It achieves 46% top-1 and 86% top-10 accuracies when tested on the open-access MassSpecGym benchmark, and outperforms SIRIUS on a recently released test set with previously uncharacterized structures. We demonstrate several real-world case studies, including identifying clinical biomarkers of depression and tuberculous meningitis, annotating an aqueous abiotic degradation product of the pesticide thiophanate methyl, disambiguating isobaric products in pooled reaction screening, and annotating biosynthetic pathways in Withania somnifera. Overall, this deep learning-based paradigm for structural elucidation enables rapid molecular annotation from complex mixtures, driving discoveries across diverse scientific domains.