Davide Ronchi1, Elena Maria Tosca1, Emmanuelle Comets2,3, Julie Bertrand2, Paolo Magni1
(1) Department of Electrical, Computer and Biomedical Engineering, University of Pavia, Pavia (Italy); (2) Université Paris Cité, IAME, INSERM, F-75018 Paris, France. (3) Univ Rennes, Inserm, EHESP, Irset - UMR_S 1085, 35000, Rennes, France.
Introduction
Preclinical studies play an essential role in cancer research [1]. Patient-derived xenograft (PDX) models that implant patient tumor tissue in immunocompromised mice are the most promising option [2]. These models allow the extraction of patient-specific cancer information, including reproducing the disease spatial structure and intratumor heterogeneity [3]. However, translating efficacy endpoints from preclinical to clinical settings remains a challenge, partly because preclinical and clinical endpoints are often inappropriately compared [4]. While preclinical studies typically focus on tumor growth inhibition (TGI) and delay, clinical trials prioritize objective response rates, progression-free survival (PFS), and overall survival (OS) as clinical endpoints of drug efficacy. Therefore, a translational approach is required to bridge the gap between preclinical knowledge and clinical trial data analysis.
Objectives
1-Setting up the translational framework
- Develop nonlinear mixed-effects TGI models in PDX mice.
- Define and validate a multi-step translational framework to predict tumor dynamics in treated cancer patients.
2-Bridging preclinical to clinical studies
- Simulate virtual patients and generate hybrid control arms using dynamic borrowing of synthetic patients.
- Integrate preclinical information as informative priors to develop PKPD nonlinear TGI-OS joint models.
Data
1-PDX mouse data
24 and 27 PDX mouse models of liver (LI) and pancreatic (PA) cancer, both including data of untreated tumor growth and response to treatment with Sorafenib (SO) and Gemcitabine (GB), respectively [5].
2-Digitized PFS data
Digitized PFS curves from clinical trials including appropriate cancer-treatment patient cohorts and digitized PFS data from the control and experimental arms of the FOHAIC-1, STAH, and SILIUS clinical trials involving LI cancer patients and SO treatment [6-8].
3-Advanced hepatocellular carcinoma (HCC) tumor size and OS data in human
Sum of the longest diameter (SLD) and OS data from the control arm of the phase III Sunitinib versus SO clinical trial in advanced HCC (NCT00699374) [9].
Methods
1-Setting up the translational framework
The Simeoni TGI model [10] combined with a nonlinear mixed-effect approach was used to characterize the distribution of the tumor exponential growth rate λ0 in mice and the anticancer potency k2 of GB and SO in PA and LI PDXs with Monolix 2021R2. Assuming exponential tumor growth in humans, λ0 and k2 were scaled up to humans based on allometric scaling according to the transitional approach proposed in [11]. Moving to human tumor trajectories, the Simeoni TGI model was simplified into the clinical Claret TGI model [12]. Tumor dynamics in PA and LI cancer patients undergoing GB and SO treatment were simulated. This process linked patient PK model with scaled TGI model. For both case studies, PFS are then computed and overlaid to digitized ones. Additionally, for PA – GB patients, we compared predicted tumor profiles to the TGI model by Garcia-Cremades et al. [13].
2-Bridging preclinical to clinical studies
First, virtual patients were generated from the translated TGI model, and their time to progression computed. These virtual patients were merged with observations from the control arm using dynamic borrowing techniques [14]. First, all real patients from the experimental arm were included. In contrast, only a subset (n = N/5) of real patients was included in the control group, and the remaining N – n subjects were replaced with virtual patients. Then, the likelihood that event times for real and virtual patients come from the same distribution was assessed by computing the maximum distance between the two cumulative distributions. If below a threshold, the algorithm ended; otherwise, an additional cohort of n real patients was enrolled in the control arm. The hybrid control arm was then compared to the experimental arm regarding hazard ratio (HR), with the obtained values compared to those from existing literature. The procedure was repeated 100 times.
Second, PDX data were reanalyzed using Bayesian inference implemented in Stan. Informative prior distributions were defined for λ0 and k2, which were then used to develop a nonlinear TGI-OS joint model of the data in the control arm of the NCT00699374 trial [9]. A drug effect resistance term in the SLD model was included to improve data fitting. The impact of a PDX-based prior was assessed by comparing it with two alternative literature-based priors: the first with a highly informative prior for λ0 and weakly informative prior for k2, and the second with weakly informative priors for both λ0 and k2. Models were evaluated based on goodness of fit, Watanabe-Akaike information criterion (WAIC) and relative standard error (RSE) of the model parameter estimates. Then, a retrospective interim analysis was conducted, simulating scenarios including 25, 33, 50, 67, and 75% of total events to highlight the advantages of using a PDX-based prior.
Results
1-Setting up the translational framework
The simulated and digitized PFS curves for both case studies were consistent throughout the entire time course, with median PFS agreement between observed and predicted outcomes. Comparison of PA-GB tumor dynamics with [13] revealed notable overlap between curves.
2-Bridging preclinical to clinical studies
In the 3 clinical trials, the final hybrid control arm included a median of 40%, 80% and 60% of real patients, respectively. The median mean distance between PFS distributions in the final hybrid control arm was 0.038, 0.039, and 0.039. Comparison with the experimental arms revealed HR of 0.4 (95% CI [0.3-0.52]), 0.68 ([0.51-0.92]), and 0.63 [0.50-0.79]), respectively, contrasting with observed HR of 0.45 ([0.34 – 0.59]), 0.75 ([0.57 – 1.00]), and 0.73 ([0.58 – 0.91]).
In the NCT00699374 trial control arm, we chose the lognormal distribution for its superior fit with survival data. The link function that best connected longitudinal and survival data emerged as a linear combination of individual patient characteristics (log(λ0), log(k2) and initial tumor size). The PDX-based prior model outperformed literature-based prior models on the full dataset (WAIC values of 17223.5, 17243.5 and 17287.2, respectively). In a retrospective interim analysis, we found the impact of k2 on the survival model to be accurately characterized using a PDX-prior scenario with only 50% of observed events (βk2 RSE of 47.5%) while at that stage using literature-based priors led to βk2 RSE values of 88.5% and 112.1%, respectively.
Conclusions
Our framework connects preclinical data to clinical trials. Using this approach, we successfully predicted tumor growth in patients based on PDX models. This showcases the value of translational modeling in oncology, facilitating the design of more efficient clinical trials and allowing for earlier decision-making in advanced cancer trials.
References:
[1] Yates JWT, Fairman DA. How translational modeling in oncology needs to get the mechanism just right. Clin Transl Sci. 2022 Mar;15(3):588-600. doi: 10.1111/cts.13183. Epub 2021 Nov 12. PMID: 34716976; PMCID: PMC8932697.
[2] Cancer models for reverse and forward translation. Nat Cancer 3, 135 (2022). https://doi.org/10.1038/s43018-022-00346-5
[3] Liu, Y., Wu, W., Cai, C. et al. Patient-derived xenograft models in cancer therapy: technologies and applications. Sig Transduct Target Ther 8, 160 (2023). https://doi.org/10.1038/s41392-023-01419-2
[4] Yates JWT, Byrne H, Chapman SC, Chen T, Cucurull-Sanchez L, Delgado-SanMartin J, Di Veroli G, Dovedi SJ, Dunlop C, Jena R, Jodrell D, Martin E, Mercier F, Ramos-Montoya A, Struemper H, Vicini P. Opportunities for Quantitative Translational Modeling in Oncology. Clin Pharmacol Ther. 2020 Sep;108(3):447-457. doi: 10.1002/cpt.1963. Epub 2020 Jul 25. PMID: 32569424.
[5] HuBase database, Crownbio Bioscience Inc., https://www.crownbio.com/).
[6] Lyu N, Wang X, Li JB, Lai JF, Chen QF, Li SL, Deng HJ, He M, Mu LW, Zhao M. Arterial Chemotherapy of Oxaliplatin Plus Fluorouracil Versus Sorafenib in Advanced Hepatocellular Carcinoma: A Biomolecular Exploratory, Randomized, Phase III Trial (FOHAIC-1). J Clin Oncol. 2022 Feb 10;40(5):468-480. doi: 10.1200/JCO.21.01963. Epub 2021 Dec 14. PMID: 34905388.
[7] Park JW, Kim YJ, Kim DY, Bae SH, Paik SW, Lee YJ, Kim HY, Lee HC, Han SY, Cheong JY, Kwon OS, Yeon JE, Kim BH, Hwang J. Sorafenib with or without concurrent transarterial chemoembolization in patients with advanced hepatocellular carcinoma: The phase III STAH trial. J Hepatol. 2019 Apr;70(4):684-691. doi: 10.1016/j.jhep.2018.11.029. Epub 2018 Dec 6. PMID: 30529387.
[8] Kudo M, Ueshima K, Yokosuka O, Ogasawara S, Obi S, Izumi N, Aikata H, Nagano H, Hatano E, Sasaki Y, Hino K, Kumada T, Yamamoto K, Imai Y, Iwadou S, Ogawa C, Okusaka T, Kanai F, Akazawa K, Yoshimura KI, Johnson P, Arai Y; SILIUS study group. Sorafenib plus low-dose cisplatin and fluorouracil hepatic arterial infusion chemotherapy versus sorafenib alone in patients with advanced hepatocellular carcinoma (SILIUS): a randomised, open label, phase 3 trial. Lancet Gastroenterol Hepatol. 2018 Jun;3(6):424-432. doi: 10.1016/S2468-1253(18)30078-5. Epub 2018 Apr 7. PMID: 29631810.
[9] https://doi.org/10.34949/2dm4-k327
[10] Simeoni M, Magni P, Cammia C, Nicolao G De, Croci V, Pesenti E, et al. Predictive pharmacokinetic- pharmacodynamic modeling of tumor growth kinetics in xenograft model after administration of anticancer agents. Cancer Res. 2004; 64:1094–101.
[11] PAGE 31 (2023) Abstr 10576 [www.page-meeting.org/?abstract=10576]
[12] Claret L, Girard P, Hoff PM, Van Cutsem E, Zuideveld KP, Jorga K, Fagerberg J, Bruno R. Model-based prediction of phase III overall survival in colorectal cancer on the basis of phase II tumor dynamics. J Clin Oncol. 2009 Sep 1;27(25):4103-8. doi: 10.1200/JCO.2008.21.0807. Epub 2009 Jul 27. PMID: 19636014.
[13] M. Garcia-Cremades, C. Pitou, P. W. Iversen, e I. F. Troconiz, «Predicting tumour growth and its impact on survival in gemcitabine-treated patients with advanced pancreatic cancer», Eur. J. Pharm. Sci. Off. J. Eur. Fed. Pharm. Sci., vol. 115, pp. 296–303, mar. 2018, doi: 10.1016/j.ejps.2018.01.033.
[14] Louis Dron, Shirin Golchi, Grace Hsu, Kristian Thorlund. Minimizing control group allocation in randomized trials using dynamic borrowing of external control data – An application to second line therapy for non-small cell lung cancer, Contemporary Clinical Trials Communications, Volume 16, 2019, 100446, ISSN 2451-8654, https://doi.org/10.1016/j.conctc.2019.100446
Reference: PAGE 32 (2024) Abstr 11049 [www.page-meeting.org/?abstract=11049]
Poster: Drug/Disease Modelling - Oncology