Integrated Prior Knowledge by Bayesian Inference for Spatial Metabolomics Annotation in Mass Spectrometry Imaging

Abstract Mass spectrometry imaging (MSI) reveals molecular distributions in tissue, but exact-mass searching often assigns multiple metabolite candidates to a single feature. We developed a Bayesian inference model for spatial metabolomics annotation (BIMSM) that prioritizes these candidates by integrating initial matching probabilities with adduct and biochemical-reaction relationships calibrated by ion-image similarity. The method was evaluated using simulated MSI data, mouse-brain MSI datasets with an external LC−MS/MS reference, and kidney-tissue MSI datasets. Among evaluable mouse-brain reference candidates, Top-1, Top-3, and Top-5 coverage increased from 6.5%, 14.4%, and 20.0% to 17.9%, 27.4%, and 32.9%, respectively. The kidney-tissue application produced less redundant candidate lists for descriptive pathway enrichment analysis. These results support BIMSM as a downstream candidate-prioritization tool that uses spatial and chemical context to guide analytical follow-up, while orthogonal validation remains necessary for definitive metabolite identification.