迈向个性化轮换多药疗法以克服胰腺癌治疗耐药：一个小鼠虚拟试验框架

Introduction. Pancreatic ductal adenocarcinoma (PDAC) is highly lethal in part because tumors rapidly evolve resistance to potent regimens. Rotational, multi-agent schedules have emerged as a promising approach to outpace this adaptive escape. To attack this problem, we propose a mechanistic “virtual-trial” framework that couples an ordinary differential equation model with patient-specific data to quantify responses to three first-line chemotherapies (cisplatin, paclitaxel, gemcitabine), stromal-modulating agents (calcipotriol, losartan), and an immune-checkpoint inhibitor (anti-PD-L1). Using an estimated dynamic resistance, our model provides an in-silico testbed for generating and ranking rotational-therapy hypotheses before clinical translation, supporting more adaptive treatment design for pancreatic cancer. Methods. Longitudinal tumor volume measurements for five distinct combinations of therapy agents were acquired in 49 mice over 14 days. Our mathematical model captures key physiological features such as tumor proliferation, drug efficacy, and temporal treatment resistance to emulate the progression and regression of pancreatic tumors to predict variation in tumor growth. Bayesian calibration of model parameters is derived on data from in vivo experiments conducted on mice with a genetically engineered model (GEM) of pancreatic cancer (KPC). We use adaptive optimization to develop personalized rotational therapy regimes across a 2-week simulation of 1000 patients. Results. The model successfully mimics tumor growth in both control and treatment cases, with an average concordance correlation coefficient (CCC) of 0.99 ± 0.01 when comparing observed and predicted changes in tumor volumes. We extend our analysis by conducting leave-one-out predictions (average CCC = 0.7 ± 0.06), mouse-specific predictions (average CCC = 0.75 ± 0.02), and group-informed, mouse-specific predictions (CCC = 0.85 ± 0.04). Group-informed, mouse-specific predictions show an 82.17 ± 15.07% accuracy in discerning responders from non-responders. Our optimization predicts that switching to a personalized, adaptive schedule would cut median tumor burden by 30.5% and shrink final tumor volume by a median 65.9% relative to any fixed protocol in simulated mice. Conclusion. Our modeling framework reproduces the experimental tumor-growth data and demonstrates strong predictive power for how pancreatic tumors respond to varied therapeutic combinations. By correctly classifying most responders versus non-responders and by forecasting sizable reductions in tumor burden with individually optimized rotational schedules, the approach offers a practical in-silico tool for designing adaptive treatment regimens. Our framework lays the groundwork for adaptive clinical trials poised to finally outmaneuver PDAC resistance and improve outcomes.