Shringi Sharma and Anita Mathias
Gilead Sciences, Inc., Foster City, CA, USA
Objectives: In oncology drug development, tumor growth related metrics (such as, tumor size ratio or time to tumor growth) are being increasingly used to predict clinical outcomes such as progression free survival (PFS) and overall survival1. Modeling was conducted to simultaneously characterise the longitudinal tumor (i.e. lymph node, LN) growth and PFS in subjects with relapsed/refractory chronic lymphocytic leukemia (CLL). The ability of the model to predict PFS (i.e. simulate PFS distributions) was evaluated.
Methods: A phase 3 placebo-controlled study which included 207 subjects treated with idelalisib (IDELA) in combination with bendamustine (B) and rituximab (R), and 208 subjects treated with B/R was included in the analysis. The final dataset included 1337 tumor measurements from IDELA +B/R and 1053 measurements from B/R treatment arms (measured at baseline and Weeks 12, 24, 36 and 48).
Longitudinal LN data was analysed using a tumor growth inhibition (TGI) model2, as shown below:
yi (tij) = y0,i . exp[KLi . tij – KDi/ λi . (1-e– λi.tij)] + epsij
where y is the LN size; KL and KD are the LN growth rate and LN growth inhibition rate, respectively; λ is the rate constant that accounts for a decrease in LN growth inhibition rate (KD) over time (t); y0 is the LN size at baseline (BSLN); j is the observation of an individual i.
PFS was described using a single time-to-event model with the hazard (h) defined as a function of longitudinal LN growth, as shown below:
h(t) = λ • exp[β*log(t)] • eyhaz • y(t)]
where, β is the baseline hazard, λ is the shape parameter for Weibull distribution, and yhaz is the estimated link between model predicted LN size at time t, y(t), and the hazard. The estimation of LN size and PFS parameters was conducted simultaneously, which allows estimation of model parameters from a joint likelihood that combines uncertainty in parameter estimates. A non-linear mixed method approach implemented in NONMEM 7.4.0 was used, assuming lognormal inter-individual variability and an additive residual error. Simulations (N=1000) were conducted, based on the final model, to predict the PFS distributions and compare with observed data.
Results: Longitudinal LN growth was well characterized by the TGI model with inter-individual variability estimated on all parameters; mean (%CV) values were y0=5220 mm2 (65), KL = 0.03 month-1 (151), KD = 0.64 month-1 (35), and λ = 0.24 month-1 (43). Larger BSLN size was associated with a greater tumor growth inhibition rate (KD); ~±30% change in KD at the 5th and 95th %ile of BSLN, relative to median BSLN values. The parameter estimates for the PFS model were β=0.0017 month-1, λ=1.2 and yhaz= 6.8e-05 mm-1. Based on the predicted Kaplan-Meier distributions, the joint model showed good performance in predicting PFS.
Conclusions: The LN growth dynamics and the observed PFS were well characterized by the joint model in subjects with CLL. By linking the full time course of LN growth as a predictor of PFS in a parametric time to event model, the present model was successful in predicting PFS. This model based framework could be leveraged for predicting clinical outcomes when performing clinical trial simulations to support clinical study designs or alternative dosing regimens.
References:
[1] Claret L, et al. J Clin Oncol; 31, 2110-4, 2013
[2] Claret L, et al. J Clin Oncol; 27, 4103-8, 2009
Reference: PAGE 27 (2018) Abstr 8588 [www.page-meeting.org/?abstract=8588]
Poster: Drug/Disease Modelling - Oncology