Biology-informed virtual staining enables label-free detection of cell state transitions at the single-cell level

Assessing cell viability and death pathways is central to biomedical research, yet conventional fluorescence-based live/dead assays are invasive, labour-intensive, and limited in capturing dynamic transitions. Deep learning-based virtual staining has emerged as a noninvasive alternative, but its accuracy remains restricted by the inconsistencies and artifacts inherently present in chemically stained training datasets. Here, we introduce a biology-informed augmented staining framework that integrates human-in-the-loop curation with generative adversarial networks to translate phase-contrast microscopy images into reliable virtual fluorescence representations. By imposing biologically motivated constraints — such as replacing unreliable labels with controlled synthetic noise — we mitigate biases in conventional datasets and improve the representation of live, unhealthy (early apoptotic), and dead cells. Using human endothelial cells subjected to oxidative stress in microfluidic devices, we demonstrate that our approach recapitulates population-level fluorescence viability dynamics while also enabling analysis at single-cell resolution. Transition points between viability states are automatically detected by comparing smoothed intensity traces and quantifying change points with associated transition bands. In an independent experiment comprising 20 cells and 2000 frames, assessed by three annotators blinded to the predictions, biology-informed training increased the macro-averaged F1 score from 0.57 to 0.78 and Cohen’s kappa from 0.34 to 0.65 compared with training on raw fluorescence labels. This analysis also reveals substantial heterogeneity in apoptotic onset and progression, consistent with morphological cues. Our results establish biology-informed virtual staining as a strategy to extend the scope of image translation methods, enabling label-free, longitudinal, and accurate single-cell monitoring of viability states. This approach addresses the key bottleneck of training data reliability and provides a framework for integrating domain knowledge into machine learning methods for biological imaging.