Neural Spectral Prediction for Structure Elucidation with Tandem Mass Spectrometry.

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Biochemistry

Background:

• Untargeted tandem mass spectrometry (MS/MS) is crucial for scientific discovery.
• Differentiating isobaric structures in MS/MS is a significant challenge in various scientific domains.

Purpose of the Study:

• To present ICEBERG, a geometric deep learning model for simulating MS/MS fragmentation.
• To improve the accuracy and speed of molecular structure elucidation and compound annotation.

Main Methods:

• Developed ICEBERG, a deep learning model that simulates collision-induced dissociation (CID) in mass spectrometry.
• Generated chemically plausible fragments and their relative intensities, considering collision energies and polarities.
• Integrated ICEBERG predictions into a pipeline to rank candidate structures based on MS/MS spectral similarity.

Main Results:

• Achieved state-of-the-art performance in compound annotation, with 40% top-1 accuracy on the NIST'20 [M+H]+ adduct subset.
• 92% of correct structures were identified within the top ten predictions in the same dataset.
• Demonstrated successful applications in identifying biomarkers, pesticide degradation products, reaction screening, and biosynthetic pathways.

Conclusions:

• ICEBERG provides a chemically interpretable, deep learning-based approach for structural elucidation.
• This paradigm enables rapid molecular annotation from complex mixtures, driving scientific discovery.
• The model significantly overcomes challenges in differentiating isobaric structures using MS/MS data.