Automated biomarker candidate discovery in imaging mass spectrometry data through spatially localized Shapley additive explanations

• DOI: 10.1016/j.aca.2021.338522
• Publication Date: 2021-04-26T03:25:11Z
• Authors: Leonoor E.M. Tideman; Lukasz G. Migas; Katerina V. Djamb...; Nathan Heath Patt...; Richard M. Caprioli; Jeffrey M. Spraggins; Raf Van de Plas
• Keywords: Discriminative model; Biomarker Discovery; Feature (linguistics); Relevance

The search for molecular species that are differentially expressed between biological states is an important step towards discovering promising biomarker candidates. In imaging mass spectrometry (IMS), performing this manually often impractical due to the large size and high-dimensionality of IMS datasets. Instead, we propose interpretable machine learning workflow automatically identifies candidates by their mass-to-charge ratios, quantitatively estimates relevance recognizing a given class using Shapley additive explanations (SHAP). task candidate discovery translated into feature ranking problem: classification model assigns pixels different classes on basis spectra, uses as features ranked in descending order relative predictive importance such top-ranking have higher likelihood being useful biomarkers. Besides providing user with experiment-wide measure species' potential, our delivers spatially localized model's decision-making process form novel representation called SHAP maps. maps deliver insight spatial specificity highlighting which regions tissue sample each provides discriminative information it does not. also enable one determine whether relationship state interest correlative or anticorrelative. Our automated approach estimating potential characterizing user-provided class, combined untargeted multiplexed nature IMS, allows rapid screening thousands obtention broader shortlist than would be possible through targeted manual assessment. demonstrated mouse-pup rat kidney case studies.