A deep learning approach for time-consistent cell cycle phase prediction from microscopy data

The cell cycle is a series of regulated stages during which a cell grows, replicates its DNA, and divides. It consists of four phases - two growth phases (G1 and G2), a replication phase (S), and a division phase (M) - each characterized by distinct transcriptional programs and impacting most other cellular processes. In imaging assays, the cell cycle phase can be identified using specific cell-cycle markers. However, the use of dedicated cell-cycle markers can be impractical or even prohibitive, as they occupy fluorescent channels that may be needed for other reporters. To address this limitation we propose a method to infer the cell cycle phase from a widely used fluorescent reporter: SiR-DNA, thereby bypassing the need for phase-specific markers while leveraging information already present in common experimental setups. Our method is based on a Variational Auto-Encoder (VAE), enhanced with two auxiliary tasks: predicting the average intensity of phase-specific markers and enforcing temporal consistency through latent space regularization. The reconstruction task ensures that the latent space captures cell cycle-relevant features, while the temporal constraint promotes biological plausibility. The resulting model, CC-VAE, classifies cell cycle phases with high accuracy from widely used DNA markers and can thus be applied to high-content screening datasets not specifically designed for cell cycle analysis. CC-VAE is freely available, along with a new, publicly released dataset comprising over 600,000 labeled HeLa Kyoto nuclear images to support further development and benchmarking in the community.