Bayesian framework applied to dose escalation studies for biologics
Marion Dehez (1), Muriel Boulton (2), Astrid Jullion (2), Bruno Boulanger (2), Ruth Oliver (1), Miren Zamacona (1)
(1) Modeling and Simulation department, Global Exploratory Development, UCB, SLOUGH, UK, (2)Exploratory statistics department, Global Exploratory Development, UCB, Braine-l'Alleud, Belgium
Introduction: In phase I dose escalating studies aiming at investigating pharmacokinetics and safety of novel drugs, a range of doses has to be explored and has to be relatively high in order to define the maximum tolerated dose. As the study progresses through the cohorts, a decision has to be made concerning the escalation at the next dose level which is constrained by the tox cover from the toxicological data (NOAEL). Limited PK data are usually available to derive the AUC(0-∞) at a given dose level before selecting the next dose level, especially for long half life compounds such as biologics. This results in a difficulty to predict the rest of the concentration time profile and therefore to derive an accurate value for AUC(0-∞).
Objective: The primary objective was to define a strategy to compute the predictive probability of the AUC(0-∞) to exceed the NOAEL limit at the next dose level using prior information and limited PK data available at the time of dose escalation.
Methods and Results: FIH study type of data for biologics were used to implement this approach. A two compartment PK model was used to analyse the data expected to be available before each dose escalation step. Since using a PK model based approach, for initial doses levels, it may not be possible to estimate accurately AUC(0-∞), a linear model was also used to fit the AUC(0-∞) versus dose in which AUC(0-t) was empirically extrapolated to obtain the AUC(0-∞) . The corresponding predictive probability to exceed the tox cover at the next dose level was computed for each approach. Three sets of criteria (relative bias, coefficient of variation and predictive probability) were used to assess the quality of the estimations and to decide at which cohort it would be possible to switch from the linear model to the PK model approach. A range of residual errors and IIV values were tested. Two sets of priors were used: informative priors (in-house or literature from similar compounds) and non-informative priors. The analysis was performed in WinBUGS version 1.4.3. Simulations were performed in R2.9.
When using WinBUGS, for both models, a good prediction of the probability of the AUC(0-∞) to exceed the NOAEL level was obtained. The PK model approach can be already used after collection of the very first timepoints of the first cohort. In the case of informative prior, this prediction is more accurate compared to non informative priors. For this biologic, with only 7 days PK time points and relevant priors the full PK profile can be predicted accurately.
Conclusions: A framework has been defined to combine prior information on biologics and PK data collected during dose escalation studies to allow accurate prediction of the exposure at the next dose levels and therefore helping in the dose selection to avoid exceeding the toxicological cover.