On the Choice of Longitudinal Models for the Analysis of Antitumor Efficacy in Mouse Clinical Trials of Patient-derived Xenograft Models

Abstract

In translational oncology research, the Patient Derived Xenograft (PDX) model and its use in Mouse Clinical Trials (MCT) are increasingly described. This involves transplanting a human tumor into a mouse and studying its evolution during follow-up or until death. A MCT contains several PDXs in which several mice are randomized to different treatment arms. Our aim was to compare longitudinal modeling of tumor growth using mixed and joint models. Mixed and joint models were compared in a real MCT (N=225 mice) to estimate the effect of a chemotherapy and a simulation study. Mixed models assume that death is predictable by observed tumor volumes (data missing at random, MAR) while the joint models assume that death depends on non-observed tumor volumes (data missing not at random, MNAR). In the real data set, of 103 deaths, 97 mice were sacrificed when reaching a pre-determined tumor size (MAR data). Joint and mixed models estimates of tumor growth slopes differed significantly (0.24 [0.13;0.36]log(mm3)/week for mixed model vs -0.02 [-0.16;0.11] for joint model). By disrupting the MAR process of mice deaths (inducing MNAR process), the estimate of the joint model was 0.24 [0.04;0.45], close to mixed model estimation for the original data set. The simulation results confirmed the bias in the slope estimate from the joint model. Using a MCT example, we show that joint model can provide biased estimates under MAR mechanisms of dropout. We thus recommend to carefully choose the statistical model according to nature of mice deaths.

Statistical classification of treatment responses in mouse clinical trials for stratified medicine in oncology drug discovery

“…For each scenario, 500 simulated datasets were generated. Tumor growth kinetics were simulated using a two classes mixed model with a class-specific linear trajectory, a class-specific treatment effect and individual random intercept and slope 15 . These datasets were then analyzed with a 2-class LCMM model as described in a previous section.…”

“…For instance, it has been applied in the study of cognitive decline 14 . In a previous work, we demonstrated that linear and non-linear mixed-effects models could robustly estimate treatment effect in a MCT design 15 . Therefore, our objective was to explore the performances of the LCMM method to classify treatment trajectories in PDX models.…”