Unmixing of imaging mass spectrometry measurements using microscopy-informed constraints

Abstract Imaging mass spectrometry (IMS) provides spatially resolved molecular information of organic tissue but can be limited by pixel signals mixing contributions from adjacent biological structures, e.g. of single cells and multicellular functional tissue units (FTUs). This paper proposes computational methods to predict mass spectral profiles of biological structures on the basis of IMS data by “unmixing” pixel-level signals, leveraging microscopy-based boundary information of these structures. By modeling each biological structure as having a unique mass spectrum, we formulate a linear mixing model and solve the corresponding inverse problem that unmixes blended signals. In particular, we cover both overdetermined and underdetermined linear system scenarios and compare ordinary least squares, non-negative least squares, and singular value thresholding to a custom algorithm, coined Tissue-informed Unmixing of Labeled regions by Inverse Problem (TULIP), specifically tailored to IMS data. Validation on a synthetic in-situ single cell dataset and demonstration on a large-scale kidney FTU dataset illustrate the potential of these methods for enhanced in-situ tissue structure analysis, e.g. in cellular and tissue studies.