Abstract
Recent clinical studies suggest that the efficacy of hormone therapy for prostate cancer depends on the characteristics of individual patients. In this paper, we develop a computational framework for identifying patient-specific androgen ablation therapy schedules for postponing the potential cancer relapse. We model the population dynamics of heterogeneous prostate cancer cells in response to androgen suppression as a nonlinear hybrid automaton. We estimate personalized kinetic parameters to characterize patients and employ $\delta$-reachability analysis to predict patient-specific therapeutic strategies. The results show that our methods are promising and may lead to a prognostic tool for personalized cancer therapy.
Introduction
Prostate cancer is a significant health challenge, often managed through androgen deprivation therapy. While continuous androgen suppression (CAS) is initially effective, it frequently leads to severe side effects—such as anemia, osteoporosis, and impotence—and the eventual emergence of castration-resistant cancer cells (CRCs), resulting in relapse. Intermittent Androgen Suppression (IAS) aims to mitigate these issues by cycling treatment "on" and "off" based on Prostate-Specific Antigen (PSA) levels. However, the success of IAS is highly patient-dependent; a schedule effective for one individual may fail for another. This research addresses the challenge of determining optimal, personalized PSA thresholds for IAS using formal verification techniques.
Modeling Cancer Dynamics
The core of the proposed framework is a nonlinear hybrid automaton that tracks two distinct cancer cell populations: hormone-sensitive (HSCs) and castration-resistant (CRCs). Unlike previous linear or piecewise-affine models, this approach utilizes sigmoid dynamics to describe cell proliferation and death. This mathematical choice is crucial for capturing the rapid "immediate decrease" in PSA levels observed clinically when treatment commences, which standard Michaelis-Menten kinetics often fail to reproduce.
The model also incorporates the "H3" hypothesis, which posits that physiological levels of androgen can actually inhibit CRC proliferation. This hypothesis was found to be the only one consistent with long-term non-relapse observations in clinical data.
Personalized Therapy via $\delta$-Reachability
Because exact reachability for nonlinear hybrid systems is undecidable, the framework employs $\delta$-reachability analysis. Using the dReal toolset, the system checks if a "relapse" state is reachable within a small error bound $\delta$.
The process involves two main steps:
- Parameter Synthesis: Fitting the model to a patient's historical PSA data to estimate individual kinetic parameters.
- Schedule Optimization: Using bounded model checking to identify specific PSA thresholds ($r_0$ for stopping and $r_1$ for starting treatment) that prevent relapse within a target timeframe, such as one year.
Clinical Validation and Results
The framework was validated using longitudinal PSA data from Phase II and III clinical trials. The model successfully reproduced diverse patient responses, highlighting the necessity of personalization. For example, Patient #1 shows a cyclic response that can be managed with specific thresholds, whereas others might show a more rapid progression.
Key findings include:
- High Fidelity: The model accurately matched real-world PSA trajectories across multiple patients.
- Relapse Prediction: $\delta$-reachability analysis identified "unsat" cases where no IAS schedule could prevent relapse, signaling the need for alternative therapies immediately.
- Efficiency: Despite the complexity of the underlying nonlinear ordinary differential equations (ODEs), the
dRealsolver provided results in minutes, suggesting potential for clinical decision support.
Conclusion
This work demonstrates that combining nonlinear hybrid automata with $\delta$-reachability analysis offers a powerful tool for personalizing prostate cancer therapy. By moving beyond "one-size-fits-all" guidelines, this prognostic tool can identify optimal treatment schedules that delay relapse while minimizing side effects. Future work will explore incorporating stochasticity to account for biological noise and prospective clinical trials to validate improved patient outcomes.
Summary updated