The tumor microenvironment evolves stochastically through states that can be simulated but not directly observed in patients. Clinical imaging gives you one snapshot — a frozen frame from a movie you can't watch. The temporal dimension is absent from the data.
The researchers build agent-based models (ABMs) of tumor evolution, then systematically perturb biologically plausible parameters to generate ensembles of simulated trajectories. From these, they construct a low-dimensional trajectory landscape — a map of where tumors can go and the transitions between metastable states. The map is built from simulation but anchored to reality by projecting patient imaging snapshots onto it.
The key move is converting this landscape into a Markov State Model, then formulating treatment scheduling as a Markov Decision Process. The therapy doesn't target the tumor's current state — it targets the transition probabilities. Conditioning the transition model on dominant governing parameters creates group-specific treatment policies. Different patients at the same point on the landscape might need different interventions because their underlying parameter regime generates different transition probabilities.
The structural insight: you can design therapies for a system you cannot observe longitudinally by simulating the space of possible trajectories and placing clinical observations within it. The single snapshot isn't a limitation if you know the landscape it sits on. The simulation provides the temporal dimension that the data lacks.
(arXiv:2603.18333)