Speaker
Description
Introduction: Macrophages are a heterogeneous population of cells. In response to microenvironmental cues macrophages shift their polarization state, alter their phenotype and adopt pro- or anti-inflammatory functions, promoting tissue inflammation or contributing to its resolution. Using a novel high throughput image-based single cell morphology and protein intensity machine learning approach, we aimed to distinguish macrophage M1, M2a vs. M2c subtypes and determine whether cell shape could predict a cell’s IL-10 immunogenic profile. This would aid in developing methods for assessing a cell population’s inflammation modulating potential using machine learning.
Methodology: Blood-derived human CD14+ monocytes were isolated and matured into macrophages using GM-CSF followed by GM-CSF/TNF-α/IFN-γ (M1 macrophages) or M-CSF followed by either M-CSF/IL-4 (M2a macrophages) or M-CSF/IL-10 (M2c macrophages). Cells were then actin stained with phalloidin and DAPI for quantification of cell area, length, width, aspect ratio, roundness, circularity, and solidity. Cells were additionally immunostained for anti-inflammatory markers IL-10 and CD163 and the pro-inflammatory marker CD80. The resulting phenotypes were confirmed via a cytokine ELISA. The image-derived single cell morphology descriptors were then used to train a machine learning algorithm to determine how accurately cell phenotype could be predicted. We then asked whether machine learning could be used to predict IL-10 production potential, quantified as IL-10 intensity per cell.
Results: A number of classification models were generated based on cell shape and immunostaining data. When only shape parameters were included, macrophage phenotype was determined with an accuracy of <50% when comparing all groups (M0, M1 control, M1, M2 control, M2a, and M2c). The accuracy increased to >80% when only assessing M1 vs. M2 macrophages. Incorporation of staining intensity and density measurements for CD163+CD80 and IL-10+CD80 improved the accuracies to 84% and 79% for M1, M2a and M2c classes, respectively. For distinguishing M2 subtypes (M2a and M2c), an accuracy of 91% was obtained with the addition of CD163+CD80 co-staining and 88% for IL-10+CD80 co-staining. Importantly, a random forest regression model was able to predict IL-10 intensity based on only cell shape data with R-squared metrics of >95% for M2c, M2a, and M1-M2a-M2c datasets. Macrophage phenotypes were confirmed by release of TNF-α by M1 and IL-10 by M2 (M2c > and M2a) macrophages into the culture supernatants.
Conclusion: The use of single cell morphology and phenotype marker expression data was able to reliably predict M1, M2a, and M2c phenotypes when applied to a machine learning model. This is the first study to use cell morphology data to accurately predict M1, M2a, and M2c macrophage phenotypes. Moreover, the incorporation of machine learning regression analysis showed, for the first time, that cell morphology is sufficient to predict IL-10 production by macrophages at the single cell level. A number of cell shape descriptors were strong indicators of IL-10 staining intensities, which point to a means of predicting IL-10 production and thereby relevant anti-inflammatory properties, based only on cell morphology. This may provide the foundation for a generalizable strategy for identifying functional subpopulations of cells and predicting the functional response of IL-10-producing cells.
52354515605
