Mathematical Model-Driven Deep Learning Enables Personalized Adaptive Therapy.

Area of Science:

• Computational oncology
• Artificial intelligence in medicine
• Cancer treatment optimization

Background:

• Standard cancer treatments often fail in metastatic disease due to drug resistance.
• Adaptive treatment strategies dynamically adjust therapy to combat resistant tumor populations.
• Prostate cancer shows promise for optimizing adaptive treatment protocols.

Purpose of the Study:

• To apply deep reinforcement learning (DRL) for guiding adaptive drug scheduling in cancer treatment.
• To develop personalized treatment schedules that outperform current adaptive protocols.
• To enhance the interpretability and clinical translatability of DRL-based treatment strategies.

Main Methods:

• Utilized deep reinforcement learning (DRL) to create adaptive drug scheduling protocols.
• Calibrated a mathematical model to prostate cancer dynamics for virtual patient simulation.
• Developed a five-step pathway integrating mechanistic modeling with DRL for improved interpretability.

Main Results:

• DRL-guided adaptive schedules more than doubled the time to progression in a prostate cancer model.
• DRL strategies demonstrated robustness to patient variability and monitoring schedules.
• The DRL framework generated interpretable strategies based on tumor burden thresholds, outperforming standard-of-care.

Conclusions:

• DRL can generate personalized, adaptive cancer treatment schedules that significantly improve outcomes.
• The proposed DRL framework offers a robust and interpretable approach for developing novel cancer therapies.
• This approach has the potential for clinical translation to improve treatment efficacy in complex cancer settings.