Relational Graph Convolutional Network for Robust Mass Spectrum Classification

Supervised machine
learning methods have shown impressive
performance
in interpreting mass signals and automatically segmenting spatially
meaningful regions in Mass Spectrometry Imaging (MSI). Such segmentation
generates maps that provide researchers with valuable insights into
sample composition and serve as a foundation for downstream statistical
analyses. However, these models often require data set-specific preprocessing
and do not fully exploit the rich mass features available in high-resolution
mass spectrometry (HRMS). Unlike low-resolution mass spectrometry,
HRMS reveals additional features such as mass defects and repeated
mass
differences that carry important chemical information. In this work,
we propose a novel deep learning architecture based on a Relational
Graph Convolutional Network (R-GCN) that captures and leverages those
HRMS mass features. Our model explicitly encodes structural features
such as mass defects and known mass differences to represent each
spectrum as a graph, enabling the learning of associations between
chemically related ion families. To the best of our knowledge, no
existing deep learning models for MSI classification incorporate this
level of chemically informed mass structure. Most existing methods
treat spectra as flat vectors or image-like inputs, thereby ignoring
the underlying mass relationships. We evaluate our R-GCN approach
against several conventional machine learning and deep learning baselines
across diverse MSI data sets, demonstrating its robustness to common
signal variations (e.g., mass shift, ion loss). Finally, we integrate
Class Activation Mapping (CAM) to enhance model interpretability,
enabling the identification of ion families that are relevant to specific
biological or spatial regions.