Application of Ensemble Machine Learning to Metabolomic Data Identifies Metabolites Associated with Macrophage Polarization

Abstract

Towards developing quantitative models of anti-tumor activities of macrophages, we evaluated the effects of cytokines, tumor exosomes, and polarization states of macrophages in a tumor microenvironemnt using a system of differential equations. We modeled the non-linear dynamics of macrophage polarization states (M0/M1/M2), tumor cell killing by macrophages, and evasion of macrophage mediated killing by tumor originated extracellular vesicle decoys. Solving these coupled differential equations using numerical approaches, showed that the rate of macrophage polarization into the M1 state is the critical determinant of anti-tumor activity mediated by M1 polarized macrophages. To determine what metabolomic factors correlate with the polarization of naive macrophage into anti-tumor M1 or pro-tumor M2 phenotypes, we performed LC/MS-based untargeted metabolomic analysis. Statistical analysis using Python-Scikit-learn was performed on the metabolomic data from naive, M1 or M2 polarized murine macrophages followed by multiple feature selection methods. Application of ensemble machine learning methods to both secreted and cell associated metabolites revealed novel molecules of fatty acid metabolism to be the main mediators of polarization. Integration of ensemble machine learning feature-ranking tools into our analysis of metabolomic data identified new potential targets in macrophage metabolism for enhancing anti-tumor activities.