A Case Study of Model‐Informed Drug Development of a Novel PCSK9 Antisense Oligonucleotide. Part 2: Phase 2 to Phase 3

ABSTRACT

In this second part of a case study on the practical use of model‐informed drug development (MIDD), we describe the clinical development of AZD8233, a novel proprotein convertase subtilisin/kexin type 9 (PCSK9) antisense oligonucleotide, from phase 2b to the start of phase 3. The case study exemplifies the use of MIDD to answer key design questions for the phase 3 program, including the design of a pivotal phase 3 study, a head‐to‐head study, and a cardiovascular outcome study informed by model‐averaging analysis. Extensive phase 3 study simulations assessed the impact of drop‐out, readout timing, dose frequency, and analysis method on study outcomes. The final phase 3 design assumed around 1% monthly drop‐out (based on other PCSK9 inhibitor trials), used an EMA/FDA‐approved analysis method, and set the primary readout at week 16. A simulated study predicted a reduction in low‐density lipoprotein cholesterol (LDL‐C) by week 16 of −69% with AZD8233 60 mg every 4 weeks. A virtual head‐to‐head study showed AZD8233 lowered LDL‐C by 27% more than an active competitor (inclisiran) at day 270. Predicted cardiovascular relative risk reduction (RRR) for AZD8233 on top of statins ranged from 24% to 49% based on model choice; a model‐averaging approach predicted an RRR of 27% assuming 63% LDL‐C reduction from a 130 mg/dL baseline. This case study highlights the importance of cross‐functional collaboration and other key MIDD enablers to ensure that MIDD extends beyond a simple simulation exercise and is instead considered an integral part of drug development dedicated to quantitative decision making.

1. Introduction

AZD8233 is an antisense oligonucleotide that targets proprotein convertase subtilisin/kexin type 9 (PCSK9) protein synthesis and thus lowers circulating low‐density lipoprotein cholesterol (LDL‐C) [1]. The phase 2b ETESIAN study demonstrated that AZD8233 lowered LDL‐C by up to 79% in patients with hypercholesterolemia on a background of high‐dose statins [2]. This magnitude of LDL‐C reduction exceeded the range of other agents targeting the PCSK9 pathway as well as what has been observed with statin therapy alone [3, 4, 5]. Furthermore, in line with other antisense oligonucleotides, AZD8233 showed no QT prolongation based on concentration‐QT modeling of the single ascending dose data [6]. However, in September 2022, AstraZeneca decided not to advance AZD8233 into phase 3 development after the readout of a second phase 2b study (SOLANO) did not achieve the pre‐specified efficacy criterion of a 70% LDL‐C reduction using an analysis of variance (ANOVA) analysis at Week 12. Nonetheless, we believe this case study, which is a continuation of the drug development story of AZD8233 from a model‐informed drug development (MIDD) perspective [7], is of great relevance for future projects aiming to successfully employ MIDD. Part 1 of the case study focused on the key success factors in the early‐phase development of AZD8233 up to the design of the phase 2 study [7]. In this second part of the series, we present: (A) a joint PCSK9 ~ LDL‐C model developed on data from the multiple ascending dose study and the phase 2b study, (B) clinical trial simulations of the phase 3 LDL‐C program, including a head‐to‐head study versus an active control, and (C) predictions of expected major adverse cardiovascular event (MACE) reduction from a model‐based meta‐analysis incorporating model averaging. Overall, this case study focuses on using the established model based on the readout of the ETESIAN phase 2b study to design a competitive and lean phase 3 program demonstrating a best‐in‐class risk/benefit profile while ensuring a comprehensive data package for approval. As in the previous part, our focus is on describing the drug development narrative of AZD8233 from a MIDD perspective instead of a discussion of methodology. We report how we adapted our approaches and methodology as new insights emerged. Owing to the dynamic nature of drug development, it is important to stress that the real challenge for the integration of MIDD into decision making is not technical in nature.

2. Overview of the MIDD Strategy From Phase 2B to Phase 3

This MIDD case study starts with the readout of the ETESIAN phase 2b study and ends with the prediction of a cardiovascular outcome trial (CVOT). It is a continuation of the MIDD strategy set up in early clinical development of AZD8233 [7]. We first describe the phase 2b readout and subsequent modeling of PCSK9 and LDL‐C data. We then explain how this model was used to design a phase 3 study, including how the potential impact of study drop‐outs as well as the choice of the statistical model could affect the LDL‐C point estimate (Figure 1).

FIGURE 1.

Early clinical program and late clinical program. CVOT, cardiovascular outcome trial; FDA, Food and Drug Administration; H2H, head‐to‐head; LDL‐C, low‐density lipoprotein cholesterol; MAD, multiple ascending dose; MIDD, model‐informed drug development; SAD, single ascending dose.

3. Phase 2B Readout and Modeling

Prior to the readout of the ETESIAN phase 2b study conducted in patients with dyslipidemia, only PCSK9 data and a prior established relationship using historical data between LDL‐C and PCSK9 were used to inform dose selection and sample size calculations. The primary objective of the ETESIAN phase 2b study was to investigate the effect of AZD8233 on LDL‐C across different dose levels, namely 15, 50, and 90 mg. The 50 mg dose was predicted to be the therapeutic dose and shown to have a ~90% probability of technical success to observe a mean PCSK9 reduction of 90%, which, based on the established quantitative relationship between PCSK9 and LDL‐C, would result in a 70% reduction in LDL‐C. The observed LDL‐C reduction for the 50 mg dose in ETESIAN was −72% (95% confidence interval [CI]: −78%, −65%; Figure S1). The predicted LDL‐C reduction for the previously predicted therapeutic dose of 50 mg based solely on PCSK9 data from phase 1 studies, and the prior established relationship with AZD8233 was −69.4% (95% CI: −72.4%, −66.3%) [7].

Data from the phase 2b study enabled, for the first time in this program, the development of a kinetic‐pharmacodynamic (K‐PD) model for both PCSK9 and LDL‐C. This model would be able to confirm the previous dose prediction based on PCSK9 reduction and historical relationship, as well as assess the impact of LDL‐C baseline on LDL‐C reduction.

The point estimates for LDL‐C reduction in the phase 1 and 2b studies were based on a mixed model for repeated measures analysis on the log‐transformed change from baseline. For all subsequent studies, an analysis of covariance would be performed using relative change from baseline to a specific predefined time point. Furthermore, owing to the difference in study length between phase 2b and phase 3, it could be assumed that the effect size for LDL‐C would be affected by potential drop‐outs from the trial. In dyslipidemia, many patients typically do not experience overt symptoms and therefore do not perceive themselves as ill. Consequently, the observed drop‐out in larger phase 2 or even phase 3 trials is frequently unrelated to treatment effects. For this reason, we assumed a non‐informative drop‐out mechanism, consistent with the assumption of data missing completely at random. To model this, we applied a binomial probability distribution to assess, at each dosing occasion, the likelihood of a patient discontinuing participation in the trial. We performed extensive clinical trial simulations to help inform the team on how these clinical trial design changes, among others, would affect the LDL‐C reduction. For example, we evaluated dose and dose regimen changes, the need for a booster dose, various drop‐out scenarios, non‐adherence, and different statistical models. To account for this difference in trial conduct in subsequent clinical trials, a dose of 60 mg was chosen to be studied in the SOLANO phase 2b study and all future studies with AZD8233. Specifically, one dose level was selected for the SOLANO phase 2b study, as the primary purpose was to generate a sufficient safety database to reduce monitoring requirements and thereby reduce the burden on patients participating in the large phase 3 program. The other phase 2b study (ETESIAN) included 3 dose levels and was the primary study to support dose selection based on efficacy. A dosing frequency of once every 4 weeks (Q4W) was selected based on the K‐PD model to ensure 90% PCSK9 and 70% LDL‐C reduction throughout the dosing interval.

Another important aspect for the design of the phase 3 studies was the impact of LDL‐C baseline on the LDL‐C reduction, so we performed a subgroup analysis predicting the LDL‐C reduction for the dose regimen of 60 mg Q4W based on the empirical Bayes estimates for each individual. The model predicted negligible differences in LDL‐C reduction based on LDL‐C baseline groups (Supporting Information S1).

4. Predicting Pivotal Phase 3 Study

The K‐PD model allowed us to predict the expected LDL‐C reduction for AZD8233 in a pivotal phase 3 study. Assuming a drop‐out rate of approximately 1% per month based on studies with other PCSK9 inhibitors [8], the predicted LDL‐C lowering was −69% at Week 16 (primary readout of phase 3 endpoint recommended by the European Medicines Agency and the US Food and Drug Administration). Furthermore, it could be seen that the effect of AZD8233 is fast and sustained (Figure 2).

FIGURE 2.

Clinical trial simulation of pivotal phase 3 study (AZD8233 60 mg Q4W vs. placebo, n = 1200) with drop‐out rate of 1% per month. Error bars represent 95% CIs. CI, confidence interval; LDL‐C, low‐density lipoprotein cholesterol; Q4W, every 4 weeks.

5. Predicting Pivotal Head‐To‐Head Study Versus Active Control

To further evaluate the differentiation and best‐in‐class profile of AZD8233, the team needed to understand the expected LDL‐C difference versus an active comparator, in this case inclisiran. The only information available at the time was a conference poster presenting a structurally similar model to our phase 2b PCSK9‐LDL‐C model [9]. To perform the simulations, we decided to extract the model parameters from the conference poster given that model files or complete outputs were unavailable to us. This poses limitations for the clinical trial simulations, including being unable to account for important covariates and parameter uncertainty. Despite these limitations, we decided to proceed with the clinical trial simulations to optimize the design of the head‐to‐head study and to illustrate the potential outcomes (Figure 3). A newly available K‐PD model for inclisiran using all clinical development data [10] could potentially affect these results should this exercise be repeated.

FIGURE 3.

Clinical trial simulation of pivotal head‐to‐head study (AZD8233 60 mg Q4W versus inclisiran 300 mg Q6M, n = 300 per arm in 1000 clinical trial simulations) assuming drop‐out rate of 1% per month (A) summarized over time and (B) distribution of LDL‐C mean change from baseline (%) at day 270. Error bars represent 95% CIs. CI, confidence interval; LDL‐C, low‐density lipoprotein cholesterol; Q4W, every 4 weeks; Q6M, every 6 months.

6. Prediction of Mace for CVOT Using Model Averaging

Silverman et al. quantified the relationship between LDL‐C lowering and cardiovascular risk reduction using a systematic review and meta‐analysis [11]. In this analysis, the data for PCSK9 inhibitors was limited and thus a difference in MACE effect could not be predicted. Recent studies with PCSK9 inhibitors have shown a lower‐than‐anticipated MACE effect [3, 4]. The reason for this difference has not been identified; plausible hypotheses are difference in study length (5 years vs. 2 years), a potential pleiotropic effect of statins [12], or the evolving standard of care for these patients. To predict the expected relative risk reduction with AZD8233, we used data from the meta‐analysis by Silverman et al. [11] and added recent studies for PCSK9 inhibitors, summarized in Figure S8 [3, 4]. Owing to differences in study population and/or assumptions about the pleiotropic effect of statins, selection of the correct mathematical model can become highly uncertain. This uncertainty was considered by using model averaging.

We used the relative penalized likelihood factor introduced by Buckland et al. [13] to assign weights for each of the four linear models included in the model averaging analysis:

$$w_i=\frac{\exp \left(-0.5\mathrm{AI}{C}_i\right)}{\sum _i\exp \left(-0.5\mathrm{AI}{C}_i\right)}$$

The predicted relative risk reduction ranged from 24% to 49% for the chosen linear models; more details can be found in the Supporting Information S1. The model average of these four linear models predicted a relative risk reduction for AZD8233 on top of statins of 27% (Figure 4).

FIGURE 4.

Predicted relative risk reduction for AZD8233 based on model average. Open circles represent observed relative risk outcomes for completed trials; triangles denote the predicted relative risk for AZD8233. Circle size is adjusted based on the number of participants in the study. LDL‐C, low‐density lipoprotein cholesterol; PCSK9, Proprotein convertase subtilisin/kexin type 9.

7. Discussion

In this case study, we continue the MIDD story for AZD8233 that demonstrated an accurate prediction of the readout of the phase 2b study showing best‐in‐class potential for this oligonucleotide. Here, we include the phase 2b PCSK9 and LDL‐C modeling, in which for the first time we used both biomarkers in the model development as opposed to making use of the historical PCSK9‐LDL‐C relationship as in the first time in human study. We showcase how the K‐PD model was used to set the strategy for the phase 3 plan for AZD8233, including a head‐to‐head study versus an active comparator. Focusing on cross‐functional embedding and application of MIDD, we demonstrated how drop‐out, dosing regimen, and statistical analysis method would affect the phase 3 readout. Finally, in this case study, we showcase how model assumptions can impact the readout of a CVOT, making use of model‐averaging to reduce bias. Despite the decision to stop development of AZD8233, this case study served as an internal example at AstraZeneca of what is possible. Since the first AstraZeneca/Food and Drug Administration MIDD meeting, several other AstraZeneca projects have pursued the same path and relied on MIDD to make drug development decisions.

Due to general safety concerns for antisense oligonucleotides with regards to thrombocytopenia, besides the ETESIAN dose finding study, an additional phase 2b study (SOLANO) was designed to generate a robust safety database to reduce monitoring requirements and thereby reduce burden for patients participating in a large phase 3 program. Based on the available data at the time, a dose was chosen that was slightly higher than the therapeutic dose to allow for safety margins while maintaining around 70% LDL‐C reduction. The SOLANO study did not meet its pre‐specified efficacy endpoints. Notably, the 60 mg dose in SOLANO resulted in substantially lower‐than‐anticipated drug exposure, which was consistent with the observed lower reductions in LDL‐C and PCSK9 levels. Population pharmacokinetic model predictions (based on phase 1 and phase 2 data [1] and including the baseline characteristics of the SOLANO study population) substantially overpredicted the observed exposures (data not shown), highlighting a key discrepancy. The reason for this lower than anticipated exposure is unknown.

A key modeling aspect of the MIDD strategy for AZD8233 was the choice to model pharmacokinetics and pharmacodynamics separately. This decision was based on the observation that the variability in the core biomarker PCSK9 was lower than the observed variability in AZD8233 plasma concentrations. Therefore, the K‐PD approach was deemed more reliable, as it did not introduce additional variability. A population pharmacokinetics analysis for AZD8233 found that body weight explains most of the observed variability [14]. While body weight was also found to be a significant variable in the K‐PD model, the impact was not clinically relevant (see Figure S6).

The key enablers for a fully integrated and impactful MIDD are cross‐functional collaboration and buy‐in as well as strong organizational support. This includes the functional support to provide the additional resources needed to enable the pharmacometrician to work as a fully integrated team member in the wider project team.

It is our ambition that parts 1 and 2 of this case study will serve as a real‐life example of the application of MIDD, showing how data are analyzed as they emerge, including a transparent account of discussions when theories are challenged by unexpected observations. We also conclude that, with cross‐functional collaboration and organizational support, the value and impact of MIDD is greatly increased with the ability to reduce timelines and cost, and to increase certainty in decision making.

Conflicts of Interest

B.C., A.H., P.J., T.R.B., and B.H. are employees of AstraZeneca. J.K., C.N., and D.R. were employees of AstraZeneca at the time this research was conducted.

Supporting information

Data S1.