Elena Maria Tosca1, Ilaria Gaiffi1, Paolo Magni1
1Department of Electrical, Computer and Biomedical Engineering, University of Pavia
Introduction: Assessment of tumor dynamics is crucial in cancer therapy evaluation. For solid tumors, RECIST [1] classify tumor dynamics into derived metrics, like objective response (tumor shrinkage) and progression disease (PD, tumor growth or appearance of new lesions), which are further summarized into overall response rate (ORR) and time-to-progression (TTP), key early efficacy endpoints in phase II and III clinical trials. Predicting ORR and TTP in advance holds significant value for oncology drug development, enhancing decision-making. Tumor growth inhibition (TGI) studies in patient-derived xenografts (PDX) mice provide valuable insights on treatment efficacy and are typically used to complement clinical evidence. However, translating TGI in PDX mice into clinical endpoints remains challenging. Objective: A translational modeling framework was proposed to predict ORR and TTP in solid tumors from TGI data in PDX mice. The approach is here validated on the case studies of Gemcitabine and Sorafenib treatment of pancreatic (PA) and hepatocellular (HCC) cancers, respectively. Data: TGI data in a panel of PDXs (n=27 for PA, n=24 for HCC) were sourced from HuBase [2]. Clinical data from the control arm of two phase III trials in advanced HCC (NCT00699374) and metastatic PA cancer (NCT00574275) were obtained from Project Data Sphere [3]. Longitudinal tumor data, RECIST classifications, administration protocols and covariates in PA (n=216) and HCC (n=373) patients receiving Gemcitabine or Sorafenib, respectively, were derived. Methods: For each case study, a multistep approach was applied. I) A population TGI model [4] was developed to characterize the exponential tumor growth rate (TGR) and anticancer potency (K2) distributions in PDXs. II) TGR and K2 were allometrically scaled from mice to humans [5]. Inter-PDX variability was propagated obtaining a multivariate lognormal distribution. III) A Human TGI model, assuming an exponential tumor growth and a direct anticancer effect driven by plasma concentration, was informed by scaled parameters and linked to literature pop PK models [6,7]. IV) Virtual cancer patients (n=216 for PA, n=373 for HCC) were generated each defined by a set of individual TGI model parameters (sampled from the scaled multivariate lognormal distribution), and covariates (sampled from clinical covariate distributions). To enable comparison, baseline tumor size and administration protocols were sampled from clinical data. Tumor dynamics in the virtual population were simulated for the corresponding study period. V) ORR and TTP were derived from simulated tumor dynamics, applying RECIST definition. To account for parameter estimation uncertainty, a Monte Carlo simulation framework (500 replicates of steps II-V) was employed. Medians and 90% prediction intervals (95%PIs) across the replicates were reported. Results: For both case studies, simulated tumor diameter profiles closely matched observed longitudinal data of sum of longest diameters (SLD) of target lesions. Observed ORR fell within the 95%PI of the model, highlighting an outstanding agreement. Observed and predicted ORR were 11.11% vs 10.6% with 95%PI=[7.41,15.3]% for Gemcitabine/PA, and 6.7% vs 4.2% with 95%PI=[1.88,7.2]% for Sorafenib/HCC, respectively. Predicted Kaplan Meier (KM) curves for TTP were overlaid to the clinical ones, showing a slight overestimation of TTP in the early phases. An analysis of the discrepancies revealed that the initial drop in the clinical TTP curves was mainly due to PD events driven by the appearance of new lesions which were not modeled in our framework. Removing these PD from the clinical TTP data, the agreement between observed and predicted curves significantly increased with medians TTP aligned closely. Observed and predicted medians TTP were 198 days (95%CI=[172,239]days) vs 191days (95%PI=[167,277]days) for Gemcitabine/PA and 127 days (95%CI=[103,131]days) vs 140 (95%PI=[121,174]days) for Sorafenib/HCC. Conclusions: The translational modeling framework accurately predicted ORR and TTP in solid tumors from TGI data in PDX mice. Despite the approach was here validated on two case studies, its generalizable design supports broader applicability. Ongoing validation in other solid tumors and newer targeted therapies further assesses its robustness. With further validation, this approach could become a valuable tool for clinical decision-making and oncology drug development.
[1] Eisenhauer, E. A., et al. “New response evaluation criteria in solid tumours: revised RECIST guideline (version 1.1).” European journal of cancer 45.2 (2009): 228-247. [2] HuBase database, Crownbio Bioscience Inc., https://www.crownbio.com/. [3] Project data Sphere platform, https://data.projectdatasphere.org/projectdatasphere/html/home [4] Simeoni M, et al. “Predictive pharmacokinetic- pharmacodynamic modeling of tumor growth kinetics in xenograft model after administration of anticancer agents”. Cancer Res. 2004; 64:1094–101. [5] Tosca, E. M., et al. “Predicting Tumor Volume Doubling Time and Progression-Free Survival in Untreated Patients from Patient-Derived-Xenograft (PDX) Models: A Translational Model-Based Approach.” The AAPS Journal 26.5 (2024): 92. [6] Zhang, L., et al. “Model-based drug development: the road to quantitative pharmacology.” J. Pharmacokinet. Pharmacodyn. 33, 369–393 (2006). [7] Jain, L., et al. Population pharmacokinetic analysis of sorafenib in patients with solid tumours. Br. J. Clin. Pharmacol. 72, 294–305 (2011).
Reference: PAGE 33 (2025) Abstr 11705 [www.page-meeting.org/?abstract=11705]
Poster: Drug/Disease Modelling - Oncology