Ensemble optimal control for managing drug resistance in cancer therapies

1. Problem Statement

The paper addresses the challenge of managing drug resistance in cancer therapies, specifically focusing on long-term disease control where complete tumor eradication is not feasible.

• The Core Issue: Classical clinical approaches often rely on the Maximal Tolerated Dose (MTD), which aims to kill as many cancer cells as possible. However, MTD often eliminates drug-sensitive cells, removing the competitive pressure on drug-resistant clones and allowing the resistant population to dominate and cause relapse.
• The Uncertainty: Biological parameters (e.g., proliferation rates, initial resistant fractions) vary significantly between patients and are subject to statistical error. A single deterministic control policy may fail for specific patient profiles.
• The Goal: To derive a unified drug administration policy (control input) that optimizes outcomes across an ensemble of patients (or parameter realizations) rather than a single deterministic case, moving beyond MTD to strategies that exploit the ecological competition between sensitive and resistant subpopulations.

2. Methodology

A. Mathematical Framework: Ensemble Optimal Control

The authors formulate the problem using Ensemble Optimal Control for Ordinary Differential Equations (ODEs).

• System Dynamics: They model the tumor as a system of ODEs where the state $x_\theta(t)$ represents the size of cancer subpopulations (sensitive and resistant). The dynamics depend on an unknown parameter vector $\theta$ (e.g., growth rates, mutation rates) drawn from a compact set $\Theta$.
• Control-Affine Structure: The system is assumed to be control-affine:
  $$\dot{x}_\theta(t) = F^0_\theta(x_\theta(t)) + F^1_\theta(x_\theta(t))u(t)$$
  where $u(t)$ is the drug dosage (bounded between 0 and $D_{max}$).
• Objective Functional: The goal is to find a common control $u(t)$ that minimizes an averaged cost functional over the ensemble:
  $$J(u) = \int_\Theta \int_0^T \ell(N_\theta^u(t)) \, dt \, d\mu(\theta)$$
  where $N_\theta^u(t)$ is the total tumor burden, $\ell$ is a penalization cost, and $\mu$ is a probability measure representing the distribution of patient parameters.
• Theoretical Contributions:
  • Existence & Stability: The authors prove the existence of minimizers for $J$ and establish a $\Gamma$-convergence result. This ensures that if the continuous probability measure $\mu$ is approximated by a discrete measure (finite set of patients), the optimal controls converge to the true optimal control.
  • Gradient Computation: They derive the explicit expression for the projected gradient of the functional $J$. This allows for the use of gradient-based optimization algorithms (projected gradient descent) to solve the problem numerically, handling the constraints $0 \le u(t) \le D_{max}$.

B. Biological Model

• General Model: A multi-species Lotka-Volterra model describing competition for resources, spontaneous evolution, and drug-induced evolution.
• Specific Application (Prostate Cancer): The numerical experiments focus on a simplified two-population model (Sensitive $S$ and Resistant $R$) treated with Androgen Deprivation Therapy (ADT).
  • Dynamics: The model assumes resistant cells are pre-existing (no acquired resistance mechanism during the simulation) and compete for a carrying capacity $K$.
  • Cost Functions: Two types of running costs $\ell(n)$ are tested:
    1. Linear: $\ell_1(n) = n - n_0$ (symmetric penalty/reward).
    2. Hyperbolic: $\ell_2(n) = \sqrt{1+(n-n_0)^2} - 1 + (n-n_0)$ (asymmetric, penalizing growth more than it rewards shrinkage).

C. Numerical Strategy

• Discretization: The continuous measure $\mu$ is approximated by a discrete measure $\mu_N$ with $N=1225$ parameter combinations (sampling proliferation rates and initial resistant fractions).
• Optimization: A projected gradient descent algorithm is used to minimize the cost functionals $J_1$ (linear) and $J_2$ (hyperbolic).
• Benchmarking: The results are compared against:
  • MTD: Constant maximum dose.
  • Adaptive Therapy (AT): Specifically "On-Off" AT, where treatment is paused when the tumor shrinks below a threshold and resumed when it grows back.

3. Key Results

A. Linear Cost vs. Hyperbolic Cost

• Linear Cost ($J_1$): The optimal control derived from the linear cost function closely resembles the MTD strategy. It prescribes high doses for long periods, leading to the extinction of sensitive cells and rapid dominance of resistant cells. The Time-To-Progression (TTP) is similar to MTD.
• Hyperbolic Cost ($J_2$): The optimal control exhibits a distinct "Off-On" behavior:
  • Initial Phase: The therapy is initially turned OFF (or reduced), allowing the tumor to grow. This promotes the proliferation of sensitive cells, which then compete with resistant cells for resources.
  • Intervention: Once the tumor approaches a progression threshold, therapy is initiated to reduce the burden.
  • Outcome: This strategy significantly delays tumor progression (increases TTP) compared to MTD and standard AT for most initial tumor sizes ($n_0 = 0.5, 0.75$).

B. The "Off-On" Adaptive Therapy

Motivated by the numerical evidence of the hyperbolic cost strategy, the authors propose a new clinical protocol called "Off-On" Adaptive Therapy:

1. Vacation (Off): Start with no treatment until the tumor grows to a specific threshold (e.g., 1.2x initial size). This builds up the sensitive population.
2. Treatment (On): Administer maximum dose until the tumor shrinks to 50% of the initial size.
3. Cycle: Repeat.

Performance:

• TTP Improvement: "Off-On" AT outperforms both MTD and standard "On-Off" AT in terms of mean Time-To-Progression.
  • For $n_0 = 0.75$, "Off-On" AT achieved a mean TTP of 589 days vs. 462 days for standard AT.
• Robustness: While the pure optimal control (hyperbolic) sometimes allowed the tumor to exceed the progression threshold dangerously in worst-case scenarios, the "Off-On" heuristic successfully controlled the tumor size while maintaining the benefits of the initial "Off" phase.

C. Minimax vs. Averaged Control

The paper also explored Minimax control (optimizing for the worst-case patient). The results showed that Minimax strategies tended to revert to MTD-like behavior, sacrificing the performance on likely cases to protect the worst-case scenarios. This supports the authors' choice of the averaged ensemble approach, which aligns better with clinical practice of treating the majority of patients effectively.

4. Significance and Contributions

1. Theoretical Advancement: The paper provides a rigorous mathematical framework for ensemble optimal control in cancer, proving existence, stability under measure approximation, and deriving gradients for constrained control problems. This bridges the gap between abstract control theory and clinical application.
2. Novel Strategy ("Off-On"): It introduces and validates a new therapeutic paradigm that combines Active Surveillance (delayed treatment initiation) with Adaptive Therapy (treatment vacations). This challenges the dogma that immediate, aggressive treatment is always superior.
3. Ecological Insight: The results reinforce the ecological theory of cancer: preserving sensitive cells to compete with resistant ones is a more effective long-term strategy than total eradication.
4. Clinical Relevance: The proposed "Off-On" strategy offers a concrete, interpretable protocol for prostate cancer (and potentially other solid tumors) that could improve patient outcomes by delaying resistance. It suggests that for certain patient profiles (e.g., low-to-moderate tumor burden), delaying treatment initiation can be mathematically optimal.

5. Limitations

• Model Simplification: The specific application uses a two-population model without acquired resistance mechanisms (resistance is pre-existing).
• Fixed Horizon: The optimization is performed over a fixed time horizon, whereas clinical goals are often open-ended.
• Parameter Uncertainty: The study assumes a specific distribution of parameters; real-world parameter estimation from patient data is complex and not fully integrated into the current simulation.

In conclusion, the paper successfully demonstrates that ensemble optimal control can generate sophisticated, non-intuitive drug scheduling policies that outperform standard clinical protocols by leveraging the competitive dynamics of tumor subpopulations.