A Literature-Calibrated Chemotherapy Digital Twin for Dose and Schedule Optimization

Chemotherapy scheduling requires a balance between tumor control and treatment-related toxicity, and an optimization method can appear effective if one side of that tradeoff is considered without the other. We developed a literature-calibrated digital-twin simulator for docetaxel in metastatic breast cancer that combines a three-compartment drug-exposure model, tumor growth, and neutrophil dynamics. The model was calibrated on 250 virtual twins to two outcomes reported for docetaxel 100 mg/m² every 21 days: a 30.0% objective response rate and 93.1% grade 3/4 neutropenia. The calibrated population reproduced these outcomes at 29.6% and 93.2%, respectively. We then compared the standard regimen with Bayesian optimization and short-horizon model predictive control in 30 separate virtual twins under a predefined toxicity constraint. The standard regimen was feasible for 63.3% of twins, compared with 80.0% for Bayesian optimization and 76.7% for model predictive control. Bayesian optimization also lowered the mean final tumor ratio from 0.891 to 0.706 and increased the simulated response rate from 23.3% to 46.7%. Model predictive control improved feasibility but produced weaker tumor control, showing that protection against near-term toxicity can conflict with the longer treatment objective. Two virtual twins had no feasible schedule among 135 tested dose-interval combinations. These results support an optimization framework in which tumor response, toxicity, and infeasibility are reported together rather than forcing a treatment option for every simulated case. The simulator is a research model and does not provide clinical dosing guidance.