A foundation model of cellular dynamics from live-cell imaging
M. Halvorsen · L. Chen · Bio Group
arXiv:2606.08841
Static snapshots miss what matters most about a cell: how it changes. We assemble a set of live-cell imaging trajectories and train a model to predict the temporal response of a cell to perturbation.
The model generalises to perturbations absent from training, recovering known biology and proposing phenotypes later confirmed at the bench. We use it to prioritise experiments, screening millions of candidate interventions in silico before committing scarce wet-lab capacity.
This inverts the usual order of discovery: the model predicts, and the lab confirms.
This is a plain-language summary. The full manuscript and supplementary material are available on request.