Combining Desorption Electrospray Ionization Mass Spectrometry Imaging and Machine Learning for Molecular Recognition of Myocardial Infarction.

Abstract

Lipid profile changes in heart muscle have been previously linked to cardiac ischemia and myocardial infarction, but the spatial distribution of lipids and metabolites in ischemic heart remains to be fully investigated. We performed desorption electrospray ionization mass spectrometry imaging of hearts from in vivo myocardial infarction mouse models. In these mice, myocardial ischemia was induced by blood supply restriction via a permanent ligation of left anterior descending coronary artery. We showed that applying the machine learning algorithm of gradient boosting tree ensemble to the ambient mass spectrometry imaging data allows us to distinguish segments of infarcted myocardium from normally perfused hearts on a pixel by pixel basis. The machine learning algorithm selected 62 molecular ion peaks important for classification of each 200 μm-diameter pixel of the cardiac tissue map as normally perfused or ischemic. This approach achieved very high average accuracy (97.4%), recall (95.8%), and precision (96.8%) at a spatial resolution of ∼200 μm. In addition, we determined the chemical identity of 27 species, mostly small metabolites and lipids, selected by the algorithm as the most significant for cardiac pathology classification. This molecular signature of myocardial infarction may provide new mechanistic insights into cardiac ischemia, assist with infarct size assessment, and point toward novel therapeutic interventions.