Trajectory Landscapes for Therapeutic Strategy Design in Agent-Based Tumor Microenvironment Models

1. Problem Statement

The paper addresses the challenge of designing adaptive therapeutic strategies for cancer, specifically targeting the Tumor Microenvironment (TME). Key difficulties include:

• Data Limitations: Clinical data from Multiplex Tissue Imaging (MTI) provides high-dimensional spatial snapshots but is temporally undersampled and lacks longitudinal data, making it impossible to directly infer spatiotemporal dynamics or optimal intervention timing.
• Model Complexity: While Agent-Based Models (ABMs) can mechanistically simulate TME evolution (cell-cell interactions, spatial heterogeneity), their high-dimensional state spaces and uncertain parameters make direct control design and optimization intractable.
• Spatial Heterogeneity: Traditional Ordinary Differential Equation (ODE) models fail to capture critical spatial mechanisms (e.g., T-cell exclusion, compartment segregation) revealed by MTI.

The core problem is how to bridge the gap between sparse clinical observations and complex mechanistic simulations to derive state-aware, optimal treatment schedules without requiring continuous patient monitoring.

2. Methodology

The authors propose a reduced-order, simulation-driven framework that transforms complex ABM dynamics into a tractable Markov Decision Process (MDP). The workflow consists of four main stages:

A. Simulation Ensemble Construction

• Instead of calibrating a single ABM to a specific patient, the authors treat a nominal ABM (implemented in PhysiCell) as a generator of plausible dynamics.
• They perform Latin Hypercube Sampling across a 21-dimensional parameter space (covering cell motility, adhesion, and signaling).
• This generates an ensemble of trajectories representing diverse biologically plausible TME evolution regimes.

B. State Representation and Embedding

• Spatial Statistics: Raw agent configurations are converted into fixed-dimensional feature vectors ($u_k$) containing cell-type proportions, spatial mixing metrics, and neighborhood interactions.
• Time-Delay Embedding: To capture temporal dynamics from static snapshots, the authors apply Takens' theorem-inspired delay embedding. They stack a window of $w$ consecutive observations to form a vector $W_k$. This reconstructs latent dynamical structures from partial observations.
• Clustering: These embedded vectors are projected (via UMAP) and clustered to identify metastable states (discrete dynamical regimes) rather than static tissue configurations.

C. Markov State Model (MSM) Construction

• The continuous trajectory landscape is discretized into a set of metastable states ($S = \{S_1, ..., S_6\}$).
• Parameter Conditioning: A Kruskal-Wallis test identifies dominant biological parameters governing trajectory outcomes. The authors find that T-cell exhaustion rate ($r_{exh}$) and tumor self-adhesion ($r_{adh}$) partition the landscape into distinct attractor basins.
• Group-Specific Transition Matrices: Separate transition matrices are estimated for different "escape modes" (e.g., S4-prone vs. S6-prone groups) and for drug-conditioned scenarios (reduced $r_{exh}$).

D. Therapeutic Strategy Design (MDP)

• The system is formulated as a finite-horizon Markov Decision Process (MDP).
• State Space: The discrete metastable states.
• Actions: Binary control: $a=0$ (no drug) or $a=1$ (Drug A, modeled as reducing the T-cell exhaustion rate).
• Objective: Maximize the probability of absorption into the effector T-cell dominant state ($S_1$), which represents immune control.
• Solution: The optimal policy $\mu^*(s, k)$ is derived via backward induction (Bellman's equation), determining when to administer therapy based on the current state and time.

3. Key Contributions

1. Simulation-to-Control Framework: A novel pipeline that converts high-dimensional, stochastic ABM ensembles into low-dimensional, discrete-state Markov models suitable for control theory.
2. Handling Sparse Clinical Data: A method to map static clinical MTI snapshots onto a learned dynamical landscape using delay embedding, bypassing the need for longitudinal patient data.
3. Identification of Critical Drivers: Demonstrated that only two parameters ($r_{exh}$ and $r_{adh}$) primarily determine long-term TME fate, allowing for simplified patient stratification.
4. State-Dependent Policy Optimization: Showed that adaptive, state-dependent policies significantly outperform fixed schedules or immediate treatment strategies.

4. Results

A. Trajectory Landscape & Validation

• Metastable States: The simulation landscape revealed six metastable states. Three terminal attractors were identified:
  • $S_1$: Effector T-cell dominance (Immune Control).
  • $S_4$: Exhausted T-cell dominance.
  • $S_6$: T-cell exclusion (Immune Escape).
• Clinical Mapping: When mapping 360 regions of interest (ROIs) from Triple Negative Breast Cancer (MIBI) patients onto this landscape:
  • Most ROIs fell into $S_1$ or $S_6$.
  • Prognostic Value: Patients with high $S_1$ occupancy (>89%) had significantly better survival ($p=0.038$), while high $S_6$ occupancy (>22%) correlated with worse survival ($p=0.029$).
  • The mapping remained robust even when macrophage features were excluded, mimicking clinical data limitations.

B. Parameter Sensitivity

• Attractor Basins: The parameter space was effectively partitioned into three basins based on $r_{exh}$ and $r_{adh}$.
  • Low $r_{exh}$ $\rightarrow$ $S_1$ (Control).
  • High $r_{exh}$ + High $r_{adh}$ $\rightarrow$ $S_4$ (Exhaustion).
  • High $r_{exh}$ + Low $r_{adh}$ $\rightarrow$ $S_6$ (Exclusion).
• A 2-nearest-neighbor classifier achieved 99.3% accuracy in predicting the attractor basin based on these two parameters.

C. Intervention Analysis

• Time Sensitivity: The "therapeutic window" closes much faster for the $S_6$-prone group (probability of success drops below 25% by step 36) compared to the $S_4$-prone group (step 189).
• Policy Performance:
  • Optimal MDP Policy: Achieved the highest probability of redirecting trajectories to $S_1$.
  • Improvement: Compared to the best fixed-timing strategy (immediate treatment), the MDP policy increased the success rate by +14.7% for $S_4$-prone and +5.8% for $S_6$-prone trajectories.
  • Mechanism: The optimal policy withholds treatment when the system is already progressing toward $S_1$, only applying Drug A when necessary to prevent escape.

5. Significance

This work provides a computational bridge between mechanistic biological modeling and clinical decision-making. By reducing complex spatial dynamics to a tractable Markov model, the authors enable:

• Personalized Timing: Determining when to treat based on the specific dynamical regime of a patient's tumor, rather than using fixed schedules.
• Early Stratification: Identifying patients at risk of rapid immune exclusion ($S_6$) who require immediate intervention.
• Resource Efficiency: Demonstrating that simulation-grounded policies can improve outcomes without requiring continuous, invasive longitudinal monitoring of patients.

The framework suggests that future oncology strategies should move from "open-loop" fixed dosing to "closed-loop" adaptive strategies informed by the underlying dynamical state of the tumor microenvironment.