Relational Graph Convolutional Network for Robust Mass Spectrum Classification.

Area of Science:

• Analytical Chemistry
• Computational Biology
• Data Science

Background:

• Supervised machine learning excels at Mass Spectrometry Imaging (MSI) data interpretation and segmentation.
• Current methods often require dataset-specific preprocessing and underutilize High-Resolution Mass Spectrometry (HRMS) chemical information.
• HRMS offers rich features like mass defects and mass differences crucial for chemical analysis.

Purpose of the Study:

• To develop a novel deep learning architecture for MSI that effectively utilizes HRMS features.
• To improve the interpretation and segmentation of MSI data by explicitly encoding chemical information.
• To enhance the robustness and interpretability of MSI analysis models.

Main Methods:

• Proposed a Relational Graph Convolutional Network (R-GCN) architecture for MSI.
• Represented spectra as graphs, encoding mass defects and mass differences to capture chemical relationships.
• Integrated Class Activation Mapping (CAM) for model interpretability.

Main Results:

• The R-GCN model demonstrated superior performance compared to conventional and deep learning baselines across diverse MSI datasets.
• The approach showed robustness against common signal variations like mass shifts and ion loss.
• CAM enabled the identification of key ion families associated with specific biological or spatial regions.

Conclusions:

• The R-GCN model offers a chemically informed approach to MSI analysis, outperforming existing methods.
• This novel architecture enhances the understanding of sample composition from MSI data.
• The method provides a more interpretable and robust tool for MSI data analysis.