This study applies a deep reinforcement learning (DRL) agent to design systemic treatment schedules for metastatic non-small cell lung cancer (NSCLC). The DRL agent is informed by a mathematical two-population tumour growth model and trained on a cohort of virtual patients parameterized from clinical data of NSCLC patients treated with erlotinib. The aim is to produce treatment protocols that not only extend tumour control but also are robust to operational delays in care and that account for patient preferences regarding quality of life (QoL).
Virtual patients were generated using parameters previously fitted to data from patients with NSCLC treated with erlotinib. The underlying tumour model distinguishes two populations, allowing the simulation of eco-evolutionary interactions between drug-sensitive and drug-resistant cells. The DRL agent learns policies by interacting with these simulated patient trajectories.
The study compares three approaches:
Comparisons were made across multiple metrics and under different operational constraints, such as varying decision intervals (the time between dosing adjustments).
Beyond the commonly used time to progression (TTP), the authors introduce a robustness metric called margin-to-failure (MTF) that measures how tolerant a policy is to delayed treatment restart. They also evaluate quality-adjusted survival (QAS) to capture how different patient preference profiles for QoL affect outcomes.
Across all tested decision intervals, the DRL policy achieved greater median TTP, larger MTF, and higher QAS compared with both the Zhang et al. adaptive protocol and MTD. The findings indicate that DRL can balance preserving drug-sensitive cells—which helps suppress resistance—with the risk of unsafe tumour regrowth.
Decision interval, defined as the time between dosing adjustments, was varied to reflect differences in monitoring frequency. As decision intervals lengthened, TTP under the DRL policy declined gradually toward the TTP observed under MTD. By contrast, the Zhang et al. protocol performed inconsistently across longer decision intervals and in some simulations led to premature progression. These results highlight that monitoring frequency and the timing of dosing decisions materially affect comparative protocol performance.
The investigators trained a population-level DRL policy on a cohort of virtual patients and tested it on previously unseen simulated patients. This population-level policy produced an interpretable treatment rule: it extended TTP for most unseen patients and suggested that treatment should be resumed at a lower tumour burden when patient monitoring is less frequent. This interpretable behavior links operational constraints (monitoring cadence) to actionable treatment thresholds.
The study explored reward shaping within the DRL framework to incorporate patient QoL preferences into the learned policy. By adjusting the reward function, the DRL agent produced treatment strategies that reflect different quality-of-life trade-offs. This demonstrates a mechanism for producing QoL-aware treatment schedules that can prioritize individual patient values alongside tumour control and robustness.
In simulations based on a two-population tumour model and virtual patients parameterized from erlotinib-treated NSCLC data, the DRL-derived policy outperformed the Zhang et al. adaptive protocol and MTD across TTP, MTF, and QAS. The DRL approach maintained benefits as the decision interval increased, although TTP advantage diminished gradually toward MTD at very long intervals. A population-level policy generalized to previously unseen virtual patients and produced interpretable recommendations commensurate with monitoring frequency. Reward shaping enabled the encoding of patient preferences for QoL into the learned strategies.
These results provide a biologically informed approach to designing robust, patient-centered evolutionary therapies for rapidly growing cancers such as NSCLC. The authors note code and data availability in a GitLab repository linked in the source. The paper declares no competing interests and reports funding sources as indicated by the authors.