When cancer treatment moves cells—and changes how fast they multiply

A controlled cancer-cell population can be steered toward a desired distribution with an error of about $7.87\times 10^{-4}$. That is the reported $L^2$ discrepancy between a simulated terminal population and its target—not evidence that a treatment works in patients, but a striking demonstration of a new mathematical framework. The framework tackles a problem that ordinary transport models miss: a therapy can change where cells are in a biological state-space while also changing how many of them survive or reproduce.

That coupling matters. A drug may push cells toward a drug-tolerant phenotype even as it kills many of them. A model that treats movement and population growth as separate controls can miss the trade-off at the heart of the treatment problem. Dongjun Wu and Anders Rantzer develop a mathematical approach in which the same control input governs both processes, then use it to ask what terminal cell distributions are possible and how to compute a low-cost route toward one.

The paper’s numerical example shows a simulated control steering an initial cell distribution toward a prescribed target. The target and starting distributions are shown in

; the subsequent evolution and the reported terminal discrepancy appear in

. Those pictures are an illustration of a model, not a clinical result. Their broader significance is that they make population-level treatment design a question of constrained control: which distributions can a therapy reach, given that the therapy changes both cell state and cell abundance?

The Science

Optimal transport is a way to describe how one distribution can be transformed into another at minimum cost. Imagine a cloud of particles, each with a location and a price for moving. In biology, the “location” might instead be a cell’s phenotype: a collection of traits that influence how it behaves, including how sensitive it is to a drug. Dynamical optimal transport adds time and rules for how each particle may move.

The classic version conserves mass. If there are a hundred units in the starting distribution, there are still a hundred at the end; only their locations change. Unbalanced optimal transport allows mass to be created or destroyed along the way. That is useful for populations that grow, die, enter, or leave. But in many standard formulations, the model chooses movement and growth or loss independently.

Wu and Rantzer focus on systems where that independence is unrealistic. Their governing equation describes a measure ${\mu }_t$, a distribution of mass over states at time $t$. The state moves with velocity $F(t,x,u)$, while its local mass grows or decays at rate $G(t,x,u)$. Here, $x$ is a state, and $u$ is a control such as treatment intensity. Crucially, the same $u$ appears in both $F$ and $G$: adjusting the treatment can alter phenotype and cell survival at once.

A compact way to see the model is

$${\partial }_t\mu +\nabla \cdot (F\mu )=G\mu .$$

The left side tracks how the distribution shifts through state-space. The right side adds or removes mass. If $G$ is positive in some region, the population grows there; if negative, it shrinks. The paper assumes the dynamics and growth rate are affine in the control—roughly, a baseline effect plus control-weighted effects—and assumes a convex treatment-cost function. These conditions make it possible to build a mathematically tractable relaxation.

The challenge is that the original problem is nonlinear. A control field must guide all parts of a distribution, while also affecting their weights, and must land the resulting population on a specified endpoint measure. The endpoint might have a different total mass from the starting population. The authors’ key move is to represent the evolution as a collection of weighted trajectories.

Each trajectory has two parts. One, $\gamma (t)$, records where a particle or subpopulation travels in phenotype space. The other, $h(t)$, records its changing contribution to the total population. If the trajectory passes through conditions where cells die, its weight falls; where they proliferate, it rises. A probability distribution over controls can also be assigned along a path. This is called a relaxed control: rather than insisting on one precise action at every instant, it allows a mixture of actions, which captures rapid switching and helps expose convex structure.

The resulting object is a measure over weighted trajectories. The starting distribution fixes where those trajectories begin; their weighted endpoints must reproduce the target distribution. The cost is accumulated along each path, weighted by the number of cells represented there. In this representation, both the objective and endpoint requirements are linear in the path measure. The authors prove that, under their stated assumptions, this relaxation is tight: it has the same optimal value as the original controlled population problem, rather than merely giving a looser approximation.

The relaxed trajectory picture is also a bridge to the paper’s second central tool: duality. The authors derive a Kantorovich-style dual problem, closely related to the dual formulations used in optimal transport. Its central object is a Hamilton–Jacobi–Bellman (HJB) inequality. An HJB equation is a way to express, across all possible states and times, the best value obtainable under a control system. Here, the inequality has an extra ingredient: the value function itself is multiplied by the growth-or-decay rate, because the mass attached to a path changes too.

This dual is not just a different mathematical language. It can certify that a target is reachable, or prove that it is not. The paper establishes strong duality using smooth subsolutions: functions that obey the required HJB inequality and can be used to bound the achievable endpoint reward. Smoothness is not automatic, so the authors’ proof includes a result showing that restricting attention to smooth functions does not lose the optimal value under their assumptions.

What They Found

The first result is an exact reformulation. The authors show that the original Eulerian problem—described by a distribution evolving in time—has the same optimal cost as the linear program over weighted relaxed trajectories. This matters because linear optimization over measures has a well-developed theory, while the original model’s coupled dynamics are more difficult to analyze directly.

The second result is strong duality between that path-space problem and a smooth HJB dual. “Strong duality” means the best lower bound supplied by the dual equals the best achievable cost in the original problem; there is no gap between the two formulations. The dual can therefore do more than give a convenient estimate. It characterizes the value of the underlying control problem.

The third result concerns reachability. In this setting, reachability means that some admissible control can transform a specified initial positive measure into a specified terminal one. The authors characterize the reachable set through its support function and HJB inequalities. In practical terms, a smooth dual function can separate a target that the system cannot reach from the set of targets it can: it provides a mathematical certificate of impossibility, not just a failed numerical attempt.

They also derive sufficient conditions that make reachability easier to check, relating the measure-valued problem to controllability of an augmented finite-dimensional system. The excerpt supplied for this digest does not include the full conditions or their numerical thresholds, so the result should not be read as a universal test that applies to every biological model. The paper also notes a direct characterization in the balanced case, when total mass is conserved.

Finally, the authors develop a numerical method based on an augmented-Lagrangian algorithm. Such methods alternate between enforcing constraints and updating variables associated with those constraints. Their formulation starts from the dual, enforces a nonlinear condition with pointwise projections, and reduces the update for the dual potential to a sparse linear system. Through Hamiltonian conjugacy—a mathematical relationship between a system’s dynamics and its cost—the algorithm’s multiplier can be interpreted as the discrete mass, transport flux, and reaction flux. That interpretation also lets the method recover the control input.

The example is a phenotype-structured cancer-treatment model in which therapy simultaneously drives phenotypic adaptation and population growth or death. The initial and desired distributions are presented in

. The simulated evolution reaches the target with a reported $L^2$ error of about $7.87\times 10^{-4}$, as shown in

. Snapshots of the computed control actions appear in

; the figure displays actions above a stated visibility threshold of $10^{-3}$. The paper excerpt provides no table of numerical outcomes beyond that error and threshold, so there is not enough verified numerical data to construct meaningful charts without inventing values.

Why This Changes Things

The conceptual change is to treat treatment as a coupled intervention on state and population, rather than as a steering wheel for one and a separate growth dial for the other. That is closer to the biology described in the paper: therapy can suppress proliferation or induce death while also selecting for, or driving movement toward, phenotypes that tolerate the drug. A model that optimizes only cell movement may prescribe a route that looks good in phenotype space but ignores the population consequences. A model that separately optimizes killing may miss how the surviving cells adapt.

The work also changes the kind of question mathematical models can answer. “What treatment minimizes cost?” is only one question. Before optimization can guide a decision, it helps to know whether the desired outcome is attainable at all. A target that asks for a particular mixture of phenotypes and a particular total abundance may be incompatible with the available controls. The reachability theory provides a way to identify that incompatibility, and potentially to distinguish a biologically impossible target from a computational method that simply has not found a path.

That distinction could matter beyond cancer. Similar coupling appears whenever an intervention moves a population through states while changing its size: cells responding to therapy, microorganisms under environmental stress, or populations of agents that reproduce or disappear as they move through a system. The authors’ framework is general at the level of control-affine dynamics, though each application would need its own defensible state variables, growth laws, costs, and constraints.

There is also a practical computational point. The path-space formulation is infinite-dimensional, so it is not directly a ready-made software solution for a clinic. But duality gives the problem structure that an algorithm can exploit. The paper’s augmented-Lagrangian scheme turns a difficult coupled model into a sequence of operations that include sparse linear solves, and its primal-dual interpretation helps connect numerical quantities back to physical ideas: how much mass is present, how it moves, and where it grows or decays.

The potential benefit is not a promise of an optimal cancer regimen. Rather, it is a foundation for asking more coherent optimization questions. If a model can represent how treatment changes both cell phenotype and population size, then one can compare candidate goals and treatment costs without pretending those effects are independent. In principle, that could help researchers explore schedules or control policies that balance immediate population reduction against the possibility of adaptation. The current paper demonstrates a mathematical and computational route to that exploration, not its efficacy in humans.

What’s Next

The most immediate next step is biological validation. A real phenotype-structured model needs data to estimate how treatment changes phenotype, proliferation, and death, and how those effects vary across cells and over time. It also needs a cost that reflects what “better” means. A mathematical optimum is only as meaningful as the dynamics, measurements, and objective supplied to it.

The simulation’s small terminal error is encouraging as a numerical demonstration, but it should be interpreted narrowly. It measures agreement between a computed endpoint and a target distribution in the model. It does not establish that the target is clinically desirable, that the control is experimentally implementable, or that the model predicts patient outcomes. The paper excerpt does not report clinical testing, comparisons with existing therapies, or evidence that the computed control can be delivered safely.

There are further mathematical and computational questions. The results rely on assumptions including compact, convex control choices, affine control effects in the dynamics and growth rate, and convexity of the running cost. Biological systems may violate these simplifications, and realistic treatment constraints can be more complicated. Scaling the numerical method to high-dimensional cell-state descriptions, incorporating uncertainty in the model, and imposing practical constraints such as dose limits and treatment timing are important challenges.

Still, the framework opens a useful door. It gives researchers a way to ask not only how to move a population, and not only how to grow or shrink it, but how a single intervention couples those outcomes. In a disease where surviving cells can change as they are being treated, that is the difference between optimizing a picture of the population and modeling the process that produces it.