Bayesian framework for multi-source data integration - application to Human extrapolation from preclinical studies
Sandrine Boulet (1), Moreno Ursino (1), Robin Michelet (2), Linda B.S. Aulin (2), Charlotte Kloft (2), Emmanuelle Comets (3), Sarah Zohar (1)
(1) Inserm, Centre de Recherche des Cordeliers, Sorbonne Université, Université Paris Cité, F-75006 Paris, France; Inria, HeKA, F-75015, Paris, France (2) Department of Clinical Pharmacy & Biochemistry, Institute of Pharmacy, Freie Universitaet Berlin, 12169 Berlin, Germany, (3) Inserm, Univ Rennes, EHESP, Irset (Institut de recherche en santé, environnement et travail) - UMRS 1085, F-35000 Rennes, France; Inserm, Université Paris Cité, IAME, F-75018 Paris, France
Introduction: In preclinical research, including in vitro studies (conducted within subcellular fractions, cell cultures, micro-organisms, organoid models, etc.), in vivo studies (animal testing in species such as mice, rats, dogs, monkeys, etc.), and in silico studies (simulations and synthetic data), the pharmacokinetic (PK), pharmacodynamic and toxicological drug features are assessed prior to progressing to a first-in-human (FIH) trial. Typically, all the studies are analyzed independently and the human dose range does not take into account the uncertainty in predictions. Incorporating all preclinical data through inferential procedures can be particularly valuable for obtaining a more precise and reliable starting dose and dose range.
Objectives: The goal of this work is to propose a Bayesian framework that takes into account differences and similarities among multi-source data to predict key quantities of interest (e.g. maximum tolerated dose - MTD, etc.) in humans. We focused on extrapolating preclinical results to human, providing a more effective utilization of all pertinent information compared to the standard methods.
Methods: We assume that several preclinical PK studies, on different species (for example, mice, rats and dogs), are run or, at least, analyzed sequentially prior to the FIH trial. Our approach is divided in four stages: (1) first, the parameters of the suitable model for each study are sequentially estimated, for example, a one-compartment PK model with one random effect for mice and two for rats and dogs; (2) second, extrapolation on PK parameters is applied via pre-specified formulas to obtain parameter distribution in humans from each study, for example on the clearance parameter of the PK model; (3) third, the coherence / commensurability of extrapolated posterior distributions is checked via the Hellinger distance, that is, we check if the extrapolated distribution from mice, rats and dogs are similar, and a selection of species is done; (4) finally, the posterior distributions selected are merged using an extension of the Bayes formula. Within the framework, some methodological innovations are proposed, such as the normalization of distributions coming from longitudinal data in stage 3 and a Bayesian formula that encompasses the Bayes theorem in special cases in stage 4. The new framework was evaluated via an extensive simulation study, inspired by the galunisertib, an inhibitor of TGF-beta signaling [1], and extrapolation approaches were applied based on a toxicity marker, specifically a threshold on the AUC targeting the MTD. The simulation study was designed using a one-compartment PK model [2] and three species (mouse, rat and dog). Two scenarios were investigated. In the first, the extrapolation of PK model parameters is correct from all animal species to human while in scenario 2 extrapolation from rat is inaccurate.
Results: The Hellinger distance used in stage 3 easily distinguishes animal species between those that exhibit similar extrapolated MTD distributions and those that show dissimilarities. In scenario 1, distances between all animal species are lower than the threshold equal to 0.5 in more than 73% of cases, showing consistency, and thus they are very often all used to calculate human MTDs in stage 4. On the contrary, for scenario 2, distances involving rat are close to one, showing consistency between mouse and dog, but not for rat, and the rat data is then discarded for stage 4.
Regarding final estimations of MTD in human in stage 4, the Bayesian approach correctly estimates it to 515 mg (standard deviation – sd- equal to 48 mg), for scenario 1, close to the true value of 502 mg, while the approach slightly overestimates it to 561 mg (sd: 296 mg) for scenario 2. Using our approach, the final predicted MTD distributions in human, obtained using all species selected in stage 3, exhibited smaller variances compared to these distributions derived from each animal species separately.
Conclusions: Compared to the standard methods, our innovative framework reduces uncertainty in predictions and potentially facilitates a more efficient dose selection. The framework offers high flexibility, since each stage of the approach could be customized, allowing for the specification of submodels (linear, generalised, mixed-effect, etc.) based on the study outcomes and these submodels can vary between studies. Moreover, stage 4 can be modified allowing downgrading for results of species not showing consistency.
References:
[1] Gueorguieva, I., Cleverly, A. L., Stauber, A., Pillay, N. S., Rodon, J. A., Miles, C. P., Yingling, J. M., and Lahn, M. M. (2014). Defining a therapeutic window for the novel TGF-β inhibitor LY2157299 monohydrate based on a pharmacokinetic/pharmacodynamic model: Early oncology development based on a pharmacokinetic/pharmacodynamic model. British Journal of Clinical Pharmacology 77, 796–807.
[2] Lestini, G., Dumont, C., and Mentré, F. (2015). Influence of the Size of Cohorts in Adaptive Design for Nonlinear Mixed Effects Models: An Evaluation by Simulation for a Pharmacokinetic and Pharmacodynamic Model for a Biomarker in Oncology. Pharmaceutical Research 32, 3159–3169.