Inferring single-cell dynamics of a molecular reporter from unlabeled live-cell light microscopy data analyzed with delay embedding

Quantification of the temporal sequence of molecular behavior in live individual cells holds promise for improving causal and mechanistic models of cell biology. In recent years, different methods for inferring molecular labeling from microscopy data have been developed, especially in the context of "virtual pathology", but less effort has been directed to the context of single-cell dynamics and live-cell imaging. We demonstrate that phase-contrast live-cell imaging of MCF10A cells, without labeling data, is predictive of dynamical, single-cell behavior of the cell-cycle reporter Human DNA Helicase B (HDHB) - in particular, of the nuclear vs. cytoplasmic localization of this fluorescent reporter of cyclin-dependent kinase activity. Prediction quality improves substantially when temporal sequences of images are combined in a "delay embedding" framework. When different featurizations of the imaging data are examined, we find that features derived from a variational auto-encoder (VAE) outperform "classical" image features derived from shape and texture. We find the best performance, with Pearson R[~] 0.9 on test data, using VAE features augmented by categorical predictions, all within the delay-embedding framework - in comparison to R [~] 0.5 based on "ordinary" regression with VAE features.