Colocalization Features for Classification of Tumors Using Desorption Electrospray Ionization Mass Spectrometry Imaging

Abstract

Supervised modeling of mass spectrometry imaging (MSI) data is a crucial component for the detection of the distinct molecular characteristics of cancerous tissues. Currently, two types of supervised analyses are mainly used on MSI data: pixel-wise segmentation of sample images and whole-sample-based classification. A large number of mass spectra associated with each MSI sample can represent a challenge for designing models that simultaneously preserve the overall molecular content while capturing valuable information contained in the MSI data. Furthermore, intensity-related batch effects can introduce biases in the statistical models. Here we introduce a method based on ion colocalization features that allows the classification of whole tissue specimens using MSI data, which naturally preserves the spatial information associated the with the mass spectra and is less sensitive to possible batch effects. Finally, we propose data visualization strategies for the inspection of the derived networks, which can be used to assess whether the correlation differences are related to coexpression/suppression or disjoint spatial localization patterns and can suggest hypotheses based on the underlying mechanisms associated with the different classes of analyzed samples.

PICA: Pixel Intensity Correlation Analysis for Deconvolution and Metabolite Identification in Mass Spectrometry Imaging

“…As such, correlation analysis can be also applied in MSI to quantify the colocalization of different ions. Indeed, colocalization analysis has been applied to cluster mass features and classify sample tissues using colocalized mass features to represent the entire molecular information in MSI. , Figure presents a general workflow of correlation analysis for MALDI imaging. A “pizza slice-shaped” region of a tomato fruit section was selected and analyzed by MALDI imaging.…”

Emerging Computational Methods in Mass Spectrometry Imaging

“…[142] In the early spatial data compression methods, 2D ion images are converted into 1D vectors and several similarity measurements of vectors including Euclidean distance, Pearson correlation coefficient, cosine similarity, and hypergeometric similarity are used to determine molecular colocalizations. [143][144][145][146][147][148] Based on the pairwise similarity measures, an adjacency matrix is computed for constructing a molecular network, which enables the detection and visualization of colocalized ions. [145,146] In addition, unsupervised classification approaches for ion images have been developed using dimensionality reduction (PCA and UMAP) and clustering (k-mean, HC, GMM, and HDBSCAN).…”

“…[143][144][145][146][147][148] Based on the pairwise similarity measures, an adjacency matrix is computed for constructing a molecular network, which enables the detection and visualization of colocalized ions. [145,146] In addition, unsupervised classification approaches for ion images have been developed using dimensionality reduction (PCA and UMAP) and clustering (k-mean, HC, GMM, and HDBSCAN). [83,95,[149][150][151][152][153] However, these methods cannot effectively correlate high-level spatial features based on 1D vectors making them disproportionately sensitive to the experimental artifacts and noise.…”

Ratiometric Mass Spectrometry Imaging for Stain-Free Delineation of Ischemic Tissue and Spatial Profiling of Ischemia-Related Molecular Signatures

“…Molecular changes may precede morphometric alterations detectable by histology, potentially providing a more sensitive platform to characterize disease states at an early onset (Woolman et al, 2021). However, histology-independent MSI analysis usually uses the entire MSI dataset to identify specific regions, which may need to develop additional algorithms for reducing dimension and clustering of the huge MSI data (Abdelmoula et al, 2016;Inglese et al, 2019;Guo et al, 2021).…”