Nan Zhang (1), Balaji Agoram (2,5), Imke Bartelink (2), Jing Li (2), Yanan Zheng (2), Chaoyu Jin (2), Paolo Vicini (3), Lorin Roskos (3), Raffaell Faggioni (2), Rada Savic (1)
(1) University of California San Francisco, San Francisco, CA, USA (2) MedImmune, Mountain View, CA, USA (3) MedImmune, Gaithersburg, MD, USA (4) MedImmune, Cambridge, UK (5) FortySeven Inc, Menlo Park, CA, USA
Objectives:
The identification of biomarkers that can predict subjects to respond to a therapy is an important objective of what is popularly called “precision medicine”. Some genetic markers, particularly for mutational drivers of tumor growth, have been used to provide a relatively straightforward rationale for therapy response or lack thereof due to their dichotomous nature. However, pathway activation markers as predictors of response to therapy tend to be more complex for various reasons. Qualifying these biomarkers as predictive and fit for purpose during early phase development presents many challenges from a data analysis point of view, including presence of data from sub-optimally effective doses tested during phase 1 and 2, variation of placebo effect as disease severity changes, and cut-point determination for optimal separation of effect. So far, all assessment of the value of biomarkers has only been done using pair-wise comparison of the effect at a specific time point in the biomarker-high vs biomarker-low groups [1]. Therefore, we developed a pharmacometric framework for the analysis of biomarker data to assess whether a complementary longitudinal data analysis using pharmacokinetic/pharmacodynamic (PK/PD) and biomarker modelling could add benefit in the interpretation of predictive biomarker data, as well as to evaluate therapeutic drug monitoring and adaptive phase 3 design.
Methods:
The simulations in this work were performed in context of standard Phase 2B design for a biological therapy for the treatment of inflammatory bowel disease and it was based on two different dose ranges, A) 300 mg, 1000 mg, 2000 mg (all above ED50) or placebo, and B) 50 mg, 300 mg, 2000 mg (all doses at ED50 and above) or placebo, and three different study settings: at drug range A or B, a placebo-controlled parallel study with biomarker effect on drug effect (Setting 1 and 2), or on placebo effect (Setting 3 and 4), or on both drug effect and placebo effect (Setting 5 and 6). PK of this virtue drug is described using a linear 2-compartment model, whereas its PD is described as an indirect inhibition. The biomarker effect is implemented as a function on drug effect (Emax) and/or Placebo using range of functions including Emax model with g of 1.5, 2, 3, 5, or 50, linear model and dichotomous model. The biomarker covariate is sampled using the mean and standard deviation of C-reactive protein level from a clinical dataset following a log normal distribution.
In each setting, we assess the power of identifying biomarker effect using modelling versus standard statistical approaches (group comparison) at end of the study. Stochastic simulations and estimations were used to compute model parameter precision and accuracy as well as to power.
Results:
All studied scenarios with biomarker effect on model parameters had a power more than 80% for the biomarker detection except the linear biomarker effect model when biomarker effect is on Emax and/or placebo effect and the alternative model is drug effect only model or when both Emax and placebo effect is dependent on biomarker level and the alternative model is biomarker-dependent drug effect model. Overall, the order of power for detecting such an effect is linear model < Emax model (g=1.5) < Emax model (g=2) < Emax model (g=3) ≈ Emax model (g=5) ≈ Emax model (g=50). The power for dichotomous model for both two dose ranges is either between that of Emax models and linear model or above that of all Emax models depending on biomarker effect relationship.
Model parameters could be estimated with reasonable precision and bias with small sample sizes, except the precision of EC50, placebo effect and its inter-individual variability, and gamma, as well as the bias of PD residual error and placebo effect.
Additionally, the power for detecting biomarker effect using statistical analysis is not consistent with that using PK/PD modeling method, and the power differs depending on the cutoff value and doses.
Conclusions:
A framework has been developed for analyzing continuous predictive biomarker. Multiple factors, such as biomarker-dependent drug effect and/or placebo effect and dose ranges, have been evaluated in affecting the power of detecting biomarker effect. This platform has improved the ability to characterize both exposure-response and the predictive value of the biomarker by simultaneous modeling of PK/PD data from all dose cohorts while including a predictive biomarker covariate.
References:
[1] Buyse et al; Biomarkers and Surrogate End Points-the Challenge of Statistical Validation. Nature Reviews Clinical Oncology 7, 309-317 (2010)
Reference: PAGE 27 (2018) Abstr 8574 [www.page-meeting.org/?abstract=8574]
Poster: Methodology - Study Design