(M1030-10-64) Concentrations and Causal Confounding: Choosing Exposure Measures with Care When Constructing an Exposure-Response Analysis

Purpose: In drug development, a summary measure of drug exposure is often linked to an effect (and used to predict an outcome) in order to determine the benefit-risk profile of a treatment in a specific population. A variety of exposure measures can be used to drive the exposure-response (ER) relationship, such as maximum concentration (Cmax) or average concentration within a dosing interval (Cavg). A logical and attractive exposure measure for ER analysis of adverse events (AE) is average drug concentration until time of the event (CavgTE). This measure has been used in publications and submissions for new therapies [1–4]. By using this measure, all available exposure data up until the event is utilized and this is especially attractive in scenarios where the dose is modified (up-titrated or reduced) during the course of treatment. Although this seems like an appropriate and compelling way to link drug exposure to a response, there are problems with this approach. The purpose of this analysis is to demonstrate, through simulation, that the choice of exposure measure in an ER analysis should not be driven only by goodness-of-fit, but by careful analysis of causality and avoiding exposure measures that are influenced by outcome.

Methods: Individual exposure measures were simulated from a population pharmacokinetics (PK) model, emulating a clinical trial (N=160) with variability in PK parameters (CV on CL = 44%). All subjects received a loading dose of 100 mg and maintenance dose of 50 mg every 3 weeks. One cycle = 3 weeks (C1). Then, hypothetical events (AEs) were simulated, using a weibull distribution, for 2 scenarios: (1) one in which there is no relationship between exposure and events, and (2) one in which there is a direct relationship between exposure and events (probability of event increases by 20% for increasing quartiles of AUCss). For each scenario, logistic regression was used to model the ER relationship using five exposure measures (CavgTE, CavgC1, AUCC1, CmaxC1, CminC1). Finally, exposure measures were evaluated in terms of its ability to predict outcome using graphical assessment and Akaike's Information Criterion (AIC).

Results: Simulated subjects had a median CavgC1 of 259 (124, 569; 2.5-97.5 percentile) ng/mL. A median concentration profile is shown in Figure 1. In the scenario with no drug-AE relationship, using CavgTE as the exposure measure results in a lower AIC than the model using CavgC1, which may suggest the CavgTE model would make better predictions. However, it falsely identified an exposure-response relationship (Figure 2 a & b). This is because early events (e.g., within the first dosing cycle) were associated with shorter treatment duration, hence higher CavgTE. In this example, subjects have high CavgTE because they had an event, not the other way around. Alternatively, an exposure metric that is fixed with respect to time (CavgC1) accurately predicted the lack of exposure-response relationship in scenario 2.1 (Figure 2 c & d). In the scenario where there was an exposure-driven response, C1 exposure measures were predictive of events, even when the majority of events occurred in later cycles (Figure 3).

Conclusion: Although the best exposure metric to use in an ER analysis depends on many factors, such as the specific outcome, study design, and mechanism of the study drug, often an early exposure measure, like CavgC1, is predictive of later events. This may not be adequate when doses are reduced during the course of study due to adverse events or there is significant dropout in the study and a time-varying exposure measure may be selected. It may still be sufficient depending on the estimand, questions of interest, and why the events occur. Additionally, using an exposure metric like CavgTE causes problems when simulating into new paradigms, where exposure is needed to simulate AEs but event time is needed to generate exposure, creating a circular dependency. Finally, if CavgTE is used to describe exposure, the definition of “high exposure” changes depending on the AE, making it difficult to make comparisons between efficacy and safety. In ER analyses, an exposure metric that depends on the event should not be used to predict the occurrence of the event.

References: 1 Tran BD, Li J, Ly N, et al. Cabozantinib exposure-response analysis for the phase 3 CheckMate 9ER trial of nivolumab plus cabozantinib versus sunitinib in first-line advanced renal cell carcinoma. Cancer Chemother Pharmacol 2023;91:179–89.
2 Hamuro L, Hu Z, Passarell J, et al. Exposure-Response Analysis to Support Nivolumab Once Every 4 Weeks Dosing in Combination with Cabozantinib in Renal Cell Carcinoma. Clin Cancer Res 2022;28:1603–13.
3 Center for Drug Evaluation and Research, US Food and Drug Administration. Clinical Pharmacology and Biopharmaceutics Review: Application Number 761040Orig1s000; Inotuzumab Ozogamicin. 2016.https://www.accessdata.fda.gov/drugsatfda_docs/nda/2017/761040Orig1s000MultidisciplineR.pdf (accessed 1 May 2023).
4 Center for Drug Evaluation and Research, US Food and Drug Administration. Clinical Pharmacology and Biopharmaceutics Review: Application Number 212306Orig1s000; Selinexor. 2019.https://www.accessdata.fda.gov/drugsatfda_docs/nda/2019/212306Orig1s000MultidisciplineR.pdf (accessed 1 May 2023).

Median concentration profiles of 160 simulated subjects for 6 cycles

Kaplan-Meier (KM) curves (a) and regression fit (b) for scenario 1, using CavgTE and KM curves (c) and regression fit (d) for scenario 1 using CavgC1 as the exposure metric

KM curves (a) and regression fits for scenario 2, using exposure metrics CavgC1 (b), AUCC1 (c), CmaxC1 (d), CminC1 (e), and CavgTE (f)