Self-Supervised Delay-Coordinate Pretraining for Cell-Cycle State Estimation from Brightfield Cell Trajectories

Self-supervised pretraining on unlabeled brightfield trajectories can learn cellular representations without state-specific molecular or fluorescent labels, potentially reducing the paired data required for downstream biological readouts. However, each brightfield image provides only a partial observation of cellular state. Delay-coordinate embedding theory suggests that a history of partial observations can reveal state information that cannot be obtained from a single time point. Motivated by this idea, we propose a self-supervised method that predicts the evolution of delay-coordinate states derived from brightfield cell trajectories. We evaluate the learned representations through few-shot estimation of continuous FUCCI signals. Across all evaluated label budgets, delay-coordinate prediction pretraining lowers mean L1 and DTW errors relative to training without pretraining, whereas current-to-future single-frame prediction does not provide the same benefit. These results suggest that delay-coordinate prediction provides a useful dynamical inductive bias for learning cellular representations from unlabeled brightfield trajectories.