Evaluation and Optimization of Model-based Translations in Oncology - From Xenografts to RECIST
Marcus Baaz (1,2), Tim Cardilin (1), Floriane Lignet (3) and Mats Jirstrand (1)
(1) Fraunhofer-Chalmers Centre, Gothenburg, Sweden (2) Department of Mathematical Sciences, Chalmers University of Technology and University of Gothenburg, Gothenburg, Sweden (3) Merck Healthcare KGaA, Translational Medicine - Quantitative Pharmacology, Darmstadt, Germany
Objectives:
- Evaluate currently used translational methods in oncology using preclinical (xenograft data) and clinical data (response rate based on RECIST v1.1) for several drug combinations.
- Use mathematical optimization to investigate alternative ways of scaling PD parameters from mouse to human.
Methods:
Preclinical xenograft data for different drug combinations, taken from the literature (see [1]), were used to calibrate semi-mechanistic tumor growth/inhibition models. Data for three different drug combinations were used: ribociclib/binimetinib and encorafenib/binimetinib for cutaneous melanoma, and encorafenib/cetuximab for colorectal cancer. The xenografts were generated using tissue from a large set of patients, making it more heterogeneous than typical xenograft data. Nonlinear mixed-effects modeling was used to analyze the data, in order to properly capture and quantify the considerable variability between individuals within the same treatment group.
To predict clinical response three different translational methods were evaluated:
- Replacement of PK
- Replacement of PK + allometric scaling of rate parameters (PD)
- Replacement of PK + optimized scaling of rate parameters (PD)
There are three types of rate parameters present in the model: a tumor growth rate parameter, a potency parameter for each compound, and, potentially, an interaction term to account for synergistic or antagonistic effects. All rate parameters for method 2 were scaled using the standard allometric exponent of -0.25 for rate parameters [2].
In clinical oncology studies patients are categorized based on how well they respond to a given treatment using the so-called RECIST [3]. Tables reporting response according to RECIST typically include the following three categories: Partial/Complete Response, Stable Disease, or Progressive Disease. For the three drug combinations investigated, tables reporting response according to RECIST were available in the literature [4-8].
The tumor growth/inhibition models were used to predict the results from the clinical studies using the first two translational methods. This was done by sampling from the full population model to simulate studies of the same size as the corresponding clinical studies and then categorizing the simulated individuals using the RECIST criteria.
To improve upon translational methods 1 and 2, the third translational step was applied. Instead of using the standard allometric exponent of -0.25 for each rate parameter, an optimization problem was formulated to find a set of optimal exponents. Specifically, the optimization problem was formulized as finding allometric exponents that minimize the distance between predicted and observed response according to RECIST. Different exponents were allowed for each of the three types of parameters in the models, giving a triple (A, B, C), where A is the exponent for the tumor growth rate, B the exponent for drug potencies, and C the exponent for the interaction term.
Results:
The tumor growth/inhibition models were able to describe data well for all three combinations based on standard model evaluation criteria, (VPCs, EBE plots, precision of parameter estimates, etc.).
Using translational method 1, a trend of overestimating the response for the treatment group receiving combination therapy was observed in the clinical predictions for all three drug combinations. Moreover, using translational method 2, which includes an additional ‘standard’ allometric scaling of the PD rate parameters improved the prediction of one of three drug combinations. Although translational method 3 improved the predictions for all three combinations it was not possible to fully predict the observed clinical response according to RECIST for two combinations. The optimal exponents in method 3 ranged from -0.16 to -0.24.
Conclusions: The investigated translational methods were increasingly better at accurately predicting the outcomes of the clinical studies. The result from the optimization-based approach appears promising in that the optimal exponents were similar across combinations and were also relatively close to the standard exponent of -0.25. However, the investigated methods need to be further evaluated using larger datasets before stronger conclusions can be drawn. In particular, given the nature of the optimization-based method sufficient amounts of clinical data are needed to allow for proper evaluation using cross-validation.
References:
[1] Nat Med (2015) 21, 1318–1325
[2] European Journal of Radiology (2016) 85(3), 524-533
[3] Eisenhauer, E et al. European Journal of Cancer (2009) 45(2), 228-247
[4] Kopetz, S et al. NEJM (2019) 381, 1632-1643
[5] Dummer, R et al. Lancet (2018) 19, 5:603-615
[6] Dummer, R et al. Lancet (2017) 18, 4:435-445
[7] Sosman, J et al. JCO (2014) 32, 9009
[8] Cunningham, D et al. NEJM (2004) 351:337-345