Individual virtual controls as surrogate for a control arm – application for siRNA and NA therapy in chronic hepatitis B infection
Huybrecht T'jollyn
Janssen R&D
Objectives: Clinical studies often include a control arm, either placebo or standard-of-care (SOC), to determine the (additional) effect of an investigational treatment on endpoints (e.g. predictive biomarker). Borrowing information from data sources external to the trial (by meta-analyses of control arm data (“historical borrowing”), platform trials, and/or Real-World Data1,2) offers the possibility to substitute a control arm. The current work describes a model-based approach, leveraging individual biomarker data collected during a first period with lead-in therapy, to predict the future individual biomarker response during a second period when e.g. add-on treatment is initiated. For each subject, this predicted biomarker trajectory can be used as its own ‘virtual control’ to calculate the net drug effect of an add-on treatment. The concept of Virtual Controls is illustrated by a case example in chronic hepatitis B patients, where combination treatments are being considered after lead-in with JNJ-73763989 (siRNA) + nucleos(t)ide analog (NA, SOC) to progressively reduce HBsAg (viral biomarker) and potentially increase the functional cure rate. This research describes a methodology for generating Virtual Controls in a simulation-based framework, including on- and off-treatment scenarios.
Methods: A simulation study was set up in NONMEM using a longitudinal PK-PD indirect response model with signal transduction delay to describe HBsAg dynamics during JNJ-73763989 + NA treatment, previously developed based on interim data from a Phase 2 study (REEF-1)3. To this end, 1000 virtual subjects were simulated receiving JNJ-73763989 200 mg every 4 weeks (Q4W) + NA daily for 24 weeks (on-treatment), after which siRNA treatment was stopped and follow-up under NA started for 48 weeks (off-treatment). HBsAg was sampled Q4W and only the first 12-week data was used for model fitting (assuming add-on treatment starts at week 13). After maximum-a-posteriori model fitting, the obtained Empirical Bayes Estimates (EBE) and individual-level variance-covariance estimates, were used to determine individual model parameters, predicting the individual biomarker trajectory from week 12 onwards, while the original (simulated) individual parameters were considered the ‘true’ parameters leading to ‘true’ HBsAg profiles.
Model adequacy was assessed by inspecting goodness-of-fit plots. Prediction performance was assessed, by calculating (i) the absolute prediction error (PE = log10(pred/true)) between the ‘true’ (i.e. simulated) and the model-predicted HBsAg data, and (ii) the number of ‘true’ HBsAg data points in- and outside of the individual-level prediction intervals over time, up to 48 weeks of follow-up.
Results: The prediction of HBsAg levels at week 24 (on-treatment with siRNA+NA), based on the first 12 weeks of HBsAg data, was adequate, as judged by the PE distribution, with ~85% of subjects having an acceptable PE < 0.5 log10 IU/mL difference, i.e. [-0.5;0.5]. The remaining subjects who were predicted outside the PE thresholds were mostly delayed responders (either true or apparent), where only the first 12-week data was insufficient to inform the individual parameters driving the drug effect.
Prediction of the rebound of HBsAg levels (off-treatment, after 24 weeks treatment) suffered from increased PE. Hence, to accommodate for the increased prediction uncertainty, individual-level median and 95% prediction intervals (PI) were derived for the HBsAg profiles. Apparent delayed responders had wider prediction intervals pointing to higher uncertainty. Moreover, their actual coverage of the PI containing the ‘true’ HBsAg levels decreased to only ~60%.
Conclusions: The application of Virtual Controls may obviate the need for an actual control arm, as shown in this simulation framework. The approach allows to make the best possible prediction, based on each individual’s initial response to treatment, for the expected future biomarker course. For example, VC of siRNA+NA could be used to calculate the combined drug effect of siRNA+NA and a third compound instead of enrolling an siRNA+NA+placebo arm. In this case, off-treatment predictions are associated with higher uncertainty compared to on-treatment predictions. In the REEF-1 population, ~20% of patients had a (extensive) delay in treatment effect, in which the approach (with the considered PK-PD model) performs less well and may need to be optimized.
References:
[1] Jiao F, Tu W, Jimenez S, Crentsil V, Chen YF. Utilizing shared internal control arms and historical information in small-sized platform clinical trials. J Biopharm Stat. 2019;29(5):845-859. doi: 10.1080/10543406.2019.1657132. Epub 2019 Aug 28. PMID: 31462131
[2] Su L, Chen X, Zhang J, Yan F. Comparative Study of Bayesian Information Borrowing Methods in Oncology Clinical Trials. JCO Precis Oncol. 2022 Mar;6:e2100394. doi: 10.1200/PO.21.00394. PMID: 35263169; PMCID: PMC8926037.
[3] T’jollyn H, Goeyvaerts N, et al. Understanding the dynamics of HBsAg decline through model-informed drug development (MIDD) of JNJ-3989 and JNJ-6379 for the treatment of chronic hepatitis B virus infection (CHB), Journal of Hepatology, Volume 77, S838 - S839