Why does knocking down the same target cause one cell to undergo apoptosis, while another cell in the same tissue rewires its pathways and survives?
In a recent Cell paper, Yusuf Roohani, Patrick Hsu, and the Arc Institute team present State, training models on 167 million cells to predict perturbation responses across unseen cellular contexts.
Evaluating on unobserved cell lines is an essential step forward. In perturbation modeling, the hardest test is generalizing across genuine distribution shifts.
Yet correlation metrics can be deceptive. A model can achieve strong nominal scores simply by reflecting basal expression rather than authentic perturbation dynamics. Testing against conditioned statistical baselines is where rigorous evaluation begins.
From a systems biology perspective, the ultimate test of a virtual cell is whether predicted shifts map to functional endpoints: cell viability, pathway rewiring, and fate commitment.
Living cells are non equilibrium systems governed by epigenetic competence. Biological specificity lives in non linear interactions: chromatin accessibility, transcription factor cooperativity, and compensatory circuit buffering.
Grounding large scale perturbation models in biophysical mechanisms is where predictive computation transforms discovery.
Original paper in Cell: https://www.cell.com/cell/fulltext/S0092-8674(2600921-9
When evaluating perturbation models across unseen contexts, what readout confirms to you that the model captured an authentic biological decision?