Nicolas Boespflug (1), Julie Bertrand (1), Emmanuelle Comets (1), Tina Moser (2), Ellen Heitzer (2), Audrey Delmas (3), Nolwen Guigal-Stephan (3), Magdalena Pohorecka (3), Adrien Tessier (4)
(1) Université Paris Cité, IAME, Inserm, F-75018, Paris, France, (2) Medical University of Graz, Graz, Austria, (3) Clinical Biomarkers, Translational Medicine, Servier, Saclay, France, (4) Quantitative Pharmacology, Translational Medicine, Servier, Saclay, France
Introduction: Circulating tumor DNA (ctDNA) is shed from tumors and present in blood of patients with cancer. Compared with tumor biopsy, ctDNA is collected from minimally invasive blood samples that can be frequently repeated. ctDNA is used to monitor tumor burden and clonal evolution and published data suggest that ctDNA levels are prognostic of patient outcomes (i.e. higher levels at baseline are associated with poorer survival) and predictive of benefit from treatments (i.e. on-treatment reduction in ctDNA levels and lower on-treatment levels are associated with longer survival)[1, 2]. Joint modeling of longitudinal ctDNA data and survival has promising potential to early predict response in oncology.
Objectives: This work aims at evaluating through joint modeling the link between the dynamics in longitudinal ctDNA data and the survival of patients with metastatic colorectal cancer receiving various therapies.
Methods: Plasma samples and associated clinical data were collected in 21 patients with metastatic colorectal cancer, undergoing their third line or more of treatment and treated with different combinations of targeted therapies (VEGF, EGFR or Checkpoint inhibitors) and chemotherapy. For each patient, 4-6 plasma longitudinal samples were available from the start of a new line of treatment. Time to progression (for progression-free survival, PFS) or to death (for overall survival, OS) were considered as censored when not available.
Cell free DNA (cfDNA) was extracted from plasma samples and then sequenced using an Ultra Low-Pass Whole-Genome Sequencing. The ctDNA tumor fraction (TF) was determined by analyzing large-scale copy number alterations and aneuploidies using the ichorCNA software[3]. ctDNA TF under 3% were considered below the limit of quantification (BLQ), while TF between 3 and 5% were examined individually to detect meaningful pattern over time. The TF was then converted into ctDNA concentrations (in ng/mL) using the total concentration of cfDNA.
Nonlinear mixed effects modeling was used to describe 1) the dynamics of ctDNA levels over time using empirical Tumor Growth Inhibition models (Claret, Bonate, Stein…)[4]; 2) the hazard of risk of progression or death using standard parametric functions (uniform, Weibull, Gompertz…); 3) the link between ctDNA levels and PFS or OS by testing different link functions (current value or slope of predicted ctDNA levels, model parameters such as growth rate). Model selection was performed using corrected Bayesian Information Criterion (BICc). Monolix Suite 2021R2 was used.
Results: A total of 108 ctDNA observations were available, with 58 BLQ. Out of 21 patients, 12 progressions and 7 deaths were observed, with a median of survival at 26.3 and 74.7 weeks for PFS and OS, respectively.
The ctDNA profiles were best described by a bi-exponential model estimating the level at baseline and the rates of both exponential decrease and increase of ctDNA levels, with interpatient variability on all parameters.
A Weibull hazard function was selected for both PFS and OS data.
Joint modeling of ctDNA levels and PFS data identified a link between the current value of ctDNA levels (ctDNA(t)) and the risk of progression: a higher ctDNA level resulted in a higher instantaneous hazard of progression. Other links were associated but with a lower drop in BICc: the slope (derivate) of ctDNA levels or the individual rate parameters for the exponential increase in ctDNA.
In our small sample of patients, we could not identify a link between ctDNA levels and OS. Limited drop in the -2Log-Likelihood was observed for some link functions but that was offset by the increasing complexity of the joint model.
Conclusions: Due to the multifactorial nature of OS, more data may be required to identify the link between ctDNA and OS. On the other hand, the definition of progression is based on the change in tumor size and ctDNA levels are used as a proxy of the tumor burden. Therefore, a link between the ctDNA dynamics and PFS is relevant. The use of minimally invasive blood samples and low-pass sequencing allow a measurement of the total tumor burden, unlike the quantification of a specific mutation. Such approach, when combined with joint modeling, is promising for the early prediction of response in oncology.
References:
[1] Osumi H et al. Sci Rep. 2019 Nov 22;9(1):17358. doi: 10.1038/s41598-019-53711-3
[2] Khan KH et al. Cancer Discov. 2018 Oct;8(10):1270-1285. doi: 10.1158/2159-8290.CD-17-0891
[3] Adalsteinsson VA et al. Nat Commun. 2017 Nov 6;8(1):1324. doi: 10.1038/s41467-017-00965-y
[4] Yin et al. CPT Pharmacometrics Syst Pharmacol. 2019 Oct;8(10):720-737. doi: 10.1002/psp4.12450
Reference: PAGE 32 (2024) Abstr 10949 [www.page-meeting.org/?abstract=10949]
Poster: Drug/Disease Modelling - Oncology