Spatially Informed Feature Selection and Machine Learning in Matrix‐Assisted Laser Desorption/Ionization Imaging for Cohort‐Scale Molecular Tissue Phenomics in Glioblastoma

Whereas current tissue phenomics computationally characterizes tissue states by interpreting stained histology images, label‐free mass spectrometry imaging (MSI) can provide additional untargeted spatial metabolomic/lipidomic molecular information. However, a central challenge in molecular tissue phenomics paradigm is identifying reproducible and spatially specific molecular markers across tissue specimens in patient cohorts. State‐of‐the‐art MSI feature selection methods rely mostly on ion intensity differences and ignore spatial context. This renders them susceptible to batch effects and variability in cohort‐scale studies. Here, we introduce a computational workflow, spatially informed feature selection (SIFS), which prioritizes molecules that colocalize with histo­pathology‐annotated regions like viable tumor, across tissue cohorts. When applied to a cohort of heterogeneous glioblastoma tissue specimens, SIFS delivered robust and spatially specific metabolomic/lipidomic candidate markers for viable tumor. Marker maps were integrated into histology images for intuitive, morphology‐aligned visualization and interpre­tation. SIFS reduced MSI data complexity and improved subsequent machine learning model training efficiency. The methodology reduced computation time for eight hyperpara­meter‐tuned benchmark machine learning and deep learning classifiers, while improving prediction accuracy for five of them. Taken together, SIFS could advance MSI‐based molecular tissue phenomics and provide a framework for spatial biomarker discovery and automated tissue characterization at cohort level.