Non-invasive methods to predict treatment response are urgently needed. Here in lung cancer, the authors develop a single-cell on-chip cytometry method to metabolically phenotype disseminated tumor cells, revealing metabolic heterogeneity and predictors of therapy response and survival.