Statistical analysis of comparative tumor growth repeated measures experiments in the ovarian cancer patient derived xenograft (PDX) setting

Ann L Oberg,Krista M Goergen,Matthew J Maurer,Xiaonan Hou,Melissa C Larson,Marc A Becker,Andrea E Wahner Hendrickson,Paul Haluska,S John Weroha,Mariam M Al Hilli,Jeanette E Eckel-Passow,Ethan P Heinzen,Scott H Kaufmann,Rachel M Hurley

doi:10.1038/s41598-021-87470-x

Abstract

Repeated measures studies are frequently performed in patient-derived xenograft (PDX) models to evaluate drug activity or compare effectiveness of cancer treatment regimens. Linear mixed effects regression models were used to perform statistical modeling of tumor growth data. Biologically plausible structures for the covariation between repeated tumor burden measurements are explained. Graphical, tabular, and information criteria tools useful for choosing the mean model functional form and covariation structure are demonstrated in a Case Study of five PDX models comparing cancer treatments. Power calculations were performed via simulation. Linear mixed effects regression models applied to the natural log scale were shown to describe the observed data well. A straight growth function fit well for two PDX models. Three PDX models required quadratic or cubic polynomial (time squared or cubed) terms to describe delayed tumor regression or initial tumor growth followed by regression. Spatial(power), spatial(power) + RE, and RE covariance structures were found to be reasonable. Statistical power is shown as a function of sample size for different levels of variation. Linear mixed effects regression models provide a unified and flexible framework for analysis of PDX repeated measures data, use all available data, and allow estimation of tumor doubling time.

Highlights

Abbreviations patient-derived xenograft (PDX) Patient derived xenograft poly (ADP-ribose) polymerase inhibitor (PARPi) Polymerase inhibitor analysis of variance (ANOVA) Analysis of variance MK MK-4827, niraparib IACUC Institutional Animal Care and Use Committee AIC Akaike information criterion BIC Bayesian information criterion Maximum Likelihood (ML) Maximum likelihood
Statistical analysis of the data from these types of xenograft studies is commonly carried out by computing average percent change from baseline for each study arm at each timepoint. These values are plotted as a function of time for each arm, and analysis of variance (ANOVA) or t-tests are performed at each measurement time point so that the assumption of independent observations is met
Statistical power is reduced because tests use data at only one time point, and power changes at each time point because of varying sample sizes caused by mouse attrition

Summary

Introduction

Abbreviations PDX Patient derived xenograft PARPi Polymerase inhibitor ANOVA Analysis of variance MK MK-4827, niraparib IACUC Institutional Animal Care and Use Committee AIC Akaike information criterion BIC Bayesian information criterion ML Maximum likelihood. Patient-derived xenograft (PDX) models have been established with the goal of more closely representing the genetics and biological behavior of patient tumors[1,2,3,4,5,6] This has enabled a more diverse array of tumors to be studied. Statistical analysis of the data from these types of xenograft studies is commonly carried out by computing average percent change from baseline for each study arm at each timepoint These values are plotted as a function of time for each arm, and analysis of variance (ANOVA) or t-tests are performed at each measurement time point so that the assumption of independent observations is met. The classic implementation of this analysis found in common laboratory software packages is severely limited by the requirement of spaced and complete data for each mouse over the entire study period, as well as the assumptions of equal variance at all time points and equal correlation between repeated measurements regardless of the length of time (lag) between the two observations

Objectives

Methods

Results

Discussion

Conclusion