QC-GN²oMS²: a Graph Neural Net for High Resolution Mass Spectra Prediction.

Area of Science:

• Computational Chemistry
• Machine Learning
• Spectroscopy

Background:

• Mass spectra prediction is crucial for molecular identification.
• Existing methods like rules-based systems and quantum chemical (QC) modeling have limitations in accuracy and computational cost.
• Deep learning, particularly graph neural networks (GNNs), offers a promising alternative for mass spectra prediction.

Purpose of the Study:

• To enhance the predictive accuracy of GNNs for mass spectra.
• To investigate the impact of incorporating quantum chemically derived features into GNN models.
• To evaluate different types of edge features for GNN-based mass spectra prediction.

Main Methods:

• Developed and evaluated GNN models for mass spectra prediction.
• Incorporated quantum chemistry-derived features as edge features, including categorical bond order, bond force constants (from extended tight-binding, xTB), and acyclic bond dissociation energies.
• Compared models against a baseline GNN without edge features.
• Applied dynamic graph attention mechanisms.

Main Results:

• GNNs with edge features demonstrated improved predictive accuracy compared to a baseline model.
• Bond dissociation enthalpies as edge features yielded the most significant improvement, achieving a cosine similarity score of 0.462.
• Dynamic graph attention further enhanced performance and supported the inclusion of edge features.

Conclusions:

• Quantum chemically derived features, especially bond dissociation enthalpies, substantially improve GNN-based mass spectra prediction.
• Dynamic graph attention is effective for GNNs in this task.
• Further research into molecular embeddings and fragment topography recognition can advance tandem mass spectrometry prediction.