Speaker
Description
BACKGROUND: Adaptive therapy delays drug resistance by modulating treatment instead of continuously applying the maximum tolerated dose \cite{1}. While Deep Reinforcement Learning (DRL) can optimize adaptive therapy in non-spatial, well-mixed deterministic tumor models \cite{2}, extending it to spatial models is challenging because tumor dynamics become stochastic and clinically observable data are limited.
METHODS: We combined DRL with an existing spatial Agent-Based Model \cite{3} and addressed the challenge of stochastic spatial dynamics by introducing a memory mechanism based on historical tumor responses to treatment and holiday periods, using clinically observable tumor burden. We also developed a transfer-learning framework to stabilize learning in stochastic spatial environments.
RESULTS: We found that the memory mechanism can significantly improve time to progression relative to memory-free agents. Besides, we show that decision-relevant information is stored in the memory, allowing the agent to infer latent tumor composition and spatial structure from treatment history. Memory therefore provides a biological interpretation for decision-making by acting as a proxy for hidden tumor state. The learned strategies remained robust under treatment perturbations, and dynamic threshold therapy emerged from memory-informed control.
CONCLUSION: Historical response memory provides an interpretable and robust mechanism for controlling tumors under limited access to spatial data.
Bibliography
@article{1,
title={Integrating evolutionary dynamics into treatment of metastatic castrate-resistant prostate cancer},
author={Zhang, Jingsong and Cunningham, Jessica J and Brown, Joel S and Gatenby, Robert A},
journal={Nature communications},
volume={8},
number={1},
pages={1816},
year={2017},
publisher={Nature Publishing Group UK London}
}
@article{2,
title={Mathematical Model-Driven Deep Learning Enables Personalized Adaptive Therapy},
author={Gallagher, Kit and Strobl, Maximilian AR and Park, Derek S and Spoendlin, Fabian C and Gatenby, Robert A and Maini, Philip K and Anderson, Alexander RA},
journal={Cancer Research},
volume={84},
number={11},
pages={1929--1941},
year={2024},
publisher={AACR}
}
@article{3,
title={Spatial structure impacts adaptive therapy by shaping intra-tumoral competition},
author={Strobl, Maximilian AR and Gallaher, Jill and West, Jeffrey and Robertson-Tessi, Mark and Maini, Philip K and Anderson, Alexander RA},
journal={Communications medicine},
volume={2},
number={1},
pages={46},
year={2022},
publisher={Nature Publishing Group UK London}
}
