Hybrid Deep Learning Model for EI-MS Spectra Prediction.

Area of Science:

• Computational Chemistry
• Spectroscopy
• Artificial Intelligence

Background:

• Electron ionization (EI) mass spectrometry (MS) is crucial for compound identification.
• Limited reference spectral libraries hinder the analysis of novel molecules.

Purpose of the Study:

• To develop a deep learning model for predicting EI-MS spectra directly from molecular structures.
• To augment existing spectral libraries and improve compound identification accuracy.

Main Methods:

• A hybrid deep learning model combining a Graph Neural Network (GNN) encoder and a Residual Neural Network (ResNet) decoder was developed.
• The model incorporates cross-attention, bidirectional prediction, and probabilistic, chemistry-informed masks for refinement.
• Training was performed on the NIST14 EI-MS database.

Main Results:

• The hybrid GNN-ResNet model achieved strong library matching performance with Recall@10 ≈ 80.8%.
• High spectral similarity was observed between predicted and experimental spectra.
• The model successfully generated high-quality synthetic EI-MS spectra.

Conclusions:

• Data-driven models show significant potential for augmenting EI-MS spectral libraries.
• The developed model can reduce the cost and effort associated with experimental spectrum acquisition.
• Further research is needed to address challenges in model generalization and spectral uniqueness.