Deep learning reveals hidden variables underlying NF-κB activation in single cells

Abstract

Individual cells show great heterogeneity when responding to environmental cues. For example, under cytokine stimulation some cells activate immune signaling pathways while others completely ignore the signal. The underlying sources of cellular variability have been inaccessible due to the destructive nature of experiments. Here we apply deep learning, live-cell analysis, and mechanistic modeling to uncover hidden variables controlling NF-{kappa}B activation in single-cells. Our computer-vision algorithm accurately predicts cells that will respond to pro-inflammatory TNF stimulation and shows that single-cell activation is pre-determined by minute amounts of \"leaky\" nuclear NF-{kappa}B localization before stimulation. Theoretical analysis predicts and experiments confirm that the ratio of NF-{kappa}B to its inhibitor I{kappa}B determines the activation probability of a given cell. Our results demonstrate how computer vision can study living-cells without the use of destructive measurements and settles the question of whether heterogenous NF-{kappa}B activation is controlled by pre-existing deterministic variables or purely stochastic ones.