Prediction of six macrophage phenotypes and their IL-10 content based on single-cell morphology using artificial intelligence

Thus, image-based machine learning using morphology-based features not only (i) classified M0, M1 and M2 macrophages but also (ii) classified M2a and M2c subtypes and (iii) predicted intracellular IL-10 at the single-cell level among six phenotypes. This simple approach can be used as a general strategy not only for macrophage phenotyping but also for prediction of IL-10 content of any IL-10 producing cell, which can help improve our understanding of cytokine biology at the single-cell level.

Using an artificial intelligence (AI) based approach, this study determined whether (i) accurate macrophage classification and (ii) prediction of intracellular IL-10 at the single-cell level was possible, using only morphological features as predictors for AI. Using a quantitative panel of shape descriptors, our study assessed image-based original and synthetic single-cell data in two different datasets in which CD14+ monocyte-derived macrophages generated from human peripheral blood monocytes were initially primed with GM-CSF or M-CSF followed by polarization with specific stimuli in the presence/absence of continuous GM-CSF or M-CSF. Specifically, M0, M1 (GM-CSF-M1, TNFα/IFNγ-M1, GM-CSF/TNFα/IFNγ-M1) and M2 (M-CSF-M2, IL-4-M2a, M-CSF/IL-4-M2a, IL-10-M2c, M-CSF/IL-10-M2c) macrophages were examined.

Phenotypes were confirmed by ELISA and immunostaining of CD markers. Variations of polarization techniques significantly changed multiple macrophage morphological features, demonstrating that macrophage morphology is a highly sensitive, dynamic marker of phenotype. Using original and synthetic single-cell data, cell morphology alone yielded an accuracy of 93% for the classification of 6 different human macrophage phenotypes (with continuous GM-CSF or M-CSF). A similarly high phenotype classification accuracy of 95% was reached with data generated with different stimuli (discontinuous GM-CSF or M-CSF) and measured at a different time point. These comparably high accuracies clearly validated the here chosen AI-based approach. Quantitative morphology also allowed prediction of intracellular IL-10 with 95% accuracy using only original data.