PopPK model development

A PopPK model was developed using non-linear mixed effects modeling (NONMEM) (v7.4.3) with pooled maribavir plasma concentration–time data from 667 participants in nine phase 1 studies (n = 182, single or repeated doses of 100, 200 or 400 mg), two phase 2 studies (NCT01611974, EudraCT 2010-024247-32, 400, 800, and 1200 mg BID), and one phase 3 study (SOLSTICE, NCT02931539, 400 mg BID; phase 2 and 3 combined: n = 485). The final PopPK analysis dataset comprised data from individuals who had received at least one dose of maribavir and had at least one quantifiable post-dose concentration. This model was used to describe the time course of maribavir plasma concentrations in healthy volunteers, participants with hepatic or renal impairment, stable renal transplant patients, and hematopoietic cell transplant (HCT) or solid organ transplant (SOT) recipients with CMV infection, including those with refractory CMV infection.

One- (ADVAN2 TRANS2), two- (ADVAN4 TRANS4), and three (ADVAN12 TRANS4)-compartment models with first-order absorption, an absorption lag time, and linear clearance were initially investigated. The NONMEM control stream for the final model can be found in the Supplementary Methods S1 of the Supplementary Material 1. From non-compartmental analysis it was known that concomitant administration of ketoconazole (a strong CYP3A4 inhibitor) increased maribavir area under the concentration–time curve from time 0 to infinity (AUC0-∞) and maximum plasma concentration (Cmax) by 53% and 10%, respectively [22], and concurrent administration of rifampin (a strong CYP3A4 inducer) significantly reduced plasma concentrations of maribavir, resulting in a 60% reduction in AUC, 30% reduced half-life, and 2.5-fold increased clearance [23]. Consequently, the effects of co-administration of strong CYP3A4 inhibitors and strong CYP3A4 inducers were considered for inclusion in the base model. The effect of body weight (scaled to 70 kg) on both clearance and volume terms was also considered for inclusion in the base model.

Structural and variance model parameters were estimated for the base model, and inter-individual variability (IIV) was included on all structural parameters. Runs with untransformed data were used to test the residual error models. Plots of weighted residuals were evaluated for homoscedasticity with respect to predictions and time since dose, and the structure of the base model was expanded as necessary to best reflect the characteristic shape of the observations over time. Dose and time dependency were explored for PK model parameters. Parameter estimation was performed using Monte Carlo Importance Sampling Expectation Maximization assisted by Mode a Posteriori, with MU referencing (where THETAs that define typical values of individual parameters, and are associated with random effects, are referenced in a PHI/MU format) to improve the efficiency of computations [24, 25]. The impact of post-dose observations below the lower limit of quantification (BLQ) was assessed in initial models by implementing the M3 method, which maximizes the likelihood for all the data treating BLQ observations as censored [26]. The performance of the final PopPK model was evaluated using a confidence interval prediction-corrected visual predictive check (CI-pcVPC) method. In addition, the percentage of observations outside the overall 5th and 95th percentiles of the predicted data was calculated.

Covariates

The covariates available for evaluation in the PopPK analysis included the following continuous and categorical covariates: age (years), age category (≥ 18 to < 65 years and ≥ 65 to < 80 years), and body mass index (BMI) at baseline; sex; race; health status (healthy, renal impairment, hepatic impairment, transplant recipient with CMV); study; diarrhea; vomiting; dose; disease characteristics (transplant type, baseline plasma CMV DNA, CMV category, hepatic impairment, presence of CMV mutations at baseline, baseline use of antilymphocyte antibody, episode of qualifying CMV infection, prior use of CMV prophylaxis, gastrointestinal [GI] graft-versus-host disease [GvHD]); and drug–drug interactions as categorical covariates (Yes/No) on apparent clearance (histamine H2 blockers, proton pump inhibitors [PPI]). Strong CYP3A4 inhibitors and inducers, as well as weight, were included in the base model.

Continuous covariates were obtained from observations on the first day of dosing, or at screening if this value was not available. Available covariates were evaluated and selected for inclusion in the covariate model based on one or more of the following criteria: (1) plots of individual estimates versus covariates demonstrate a correlation where the parameter estimates may increase or decrease with increasing values of a continuous covariate or particular category of a categorical covariate; (2) a statistically significant covariate effect is determined by univariate analysis of variance or by regression analysis (for categorical and continuous covariates, respectively); (3) physiological or pharmacological rationale; and (4) information from prior analyses or published sources. Parameters that showed excessive (> 30%) shrinkage in IIV were carefully reviewed as they can be ill suited for graphical assessment of covariate effects. Categorical covariates were tested and incorporated in the model as a series of index variables taking on values of zero or one.

The full model with backwards deletion approach was used for covariate modeling, with all covariate parameter relationships of interest entered into the model simultaneously. Highly correlated covariates were tested one at a time to avoid confounding in the estimation of covariate effects. A backwards deletion was carried out at the p = 0.001 significance level. The order of removal was based on the relative standard error and 95% confidence interval (CI) of the parameter estimate. Covariate-specific parameter estimates were compared and estimates that appeared comparable were combined and tested for significance.

Individual predicted PK parameters

The final model was used to derive individual estimates of steady-state AUC over one dosing interval (AUC0–τ), Cmax.ss, and minimum maribavir plasma concentration (Ctrough.ss) for a 400 mg BID dosing regimen for all participants in the PopPK analysis. Individual estimates of Cmax.ss, Ctrough.ss, and AUC0–τ were obtained by prediction of the concentration–time profiles (concentrations predicted at 0, 1, 2, 3, 4, 6, 8 and 12 h) after a steady-state dose of 400 mg BID for respective individuals using their individual post hoc parameter values in the absence of strong CYP3A4 inducers or strong CYP3A4 inhibitors, and zero values for residual variability. The PK parameters AUC0–τ (linear up/log down trapezoidal rule), Cmax.ss, and Ctrough.ss were determined by non-compartmental methods.

The mean (percentage coefficient of variation [CV%]), geometric mean (GM) CV%, median, and percentiles for AUC0–τ, Cmax.ss, and Ctrough.ss were obtained overall and by health status (healthy individuals, all transplant recipients with CMV infection, and transplant recipients with refractory CMV infection in the phase 3 SOLSTICE study). For transplant patients with CMV infections, summary statistics were provided by covariates of interest including age (18 to < 65 years and ≥ 65 years; 18 to < 65 years, 65 to < 75 years, and 75 to < 85 years), sex, race, ethnicity, body mass index (BMI; underweight [< 18.5 kg/m2], normal weight [18.5 to < 25 kg/m2], overweight [25 to < 30 kg/m2], obese [≥ 30 kg/m2]), concurrent PPI, concurrent H2 inhibitors, transplant type, baseline plasma CMV DNA (low and intermediate/high), CMV category, hepatic impairment (none, Child-Pugh Class A, Child-Pugh Class B), presence of CMV mutations at baseline, and baseline use of antilymphocyte antibody. For the PPI and H2 inhibitor analyses, PK parameters were computed only for patients receiving the relevant concomitant medication for all their PK observations and for those who never received the relevant concomitant medication. For healthy individuals, summary statistics were provided by the following covariates of interest: age (≤ 65 years, > 65 years), sex, race, ethnicity, and BMI.

Exposure–response analyses

The exposure–response analyses for maribavir were based on data collected from transplant recipients with CMV infections that were refractory to treatment with ganciclovir, valganciclovir, foscarnet, or cidofovir in the phase 3 SOLSTICE study (ClinicalTrials.gov: NCT02931539) [9]. SOLSTICE was a multicenter, open-label, active-controlled study in which patients were randomized 2:1 to receive maribavir 400 mg BID or IAT for 8 weeks, with 12 weeks of follow-up. The primary endpoint was confirmed CMV clearance at the end of week 8, and the key secondary endpoint was achievement of CMV clearance and symptom control at the end of week 8, maintained through week 16. TE CMV mutation conferring resistance to maribavir was assessed as an exploratory endpoint. Full details of the SOLSTICE study have been published previously [9, 15].

Posterior Bayes parameters of the final PopPK model were used to derive rich concentration–time profiles for maribavir. Actual dosing history and factors affecting the PK of maribavir (i.e., body weight and time-varying presence of strong CYP3A4 inhibitors or strong CYP3A4 inducers) were considered in the predictions. Model-derived exposure parameters of maribavir were calculated from the individual concentration–time profiles: AUC from 0 to 24 h on the day of an adverse event (AE) (AUCday); Cmax of maribavir on the day of an AE (Cmax.day); average concentration of maribavir on each study day (Cavg); AUC from 0 to 24 h at steady state on the last day of exposure (AUCss); and Cmax.ss and Ctrough.ss on the last day of exposure.

Data set preparation, exploration, visualization of the data and exposure-response analyses were performed using R® (version 4.0.5 for efficacy and resistance analyses; version 4.0.2 for safety analysis) with comprehensive R® archive network (CRAN) and Certara Strategic Consulting (CSC) package. Firth correction on logistic regression was done based on library “logistf”. Logistic regressions with log-linear exposure effect or Emax exposure effect were performed using NONMEM Version VII (version 7.4).

Exposure–response analysis of efficacy and treatment-emergent mutations

Maribavir exposure parameters (AUCss, Cmax.ss, and Ctrough.ss) were combined with the primary and key secondary efficacy endpoints, and with TE CMV mutations conferring resistance to maribavir from the SOLSTICE study. The following covariates were included in the dataset to identify potential influencing factors: transplant type (HCT, SOT); CMV DNA level (high, intermediate, low) at baseline; symptom status (symptomatic or asymptomatic) at baseline; CMV resistant at baseline (yes, no); donor (D) and recipient (R) CMV serostatus (D+/R−, D+/R+, or D−/R+, D−/R−); antilymphocyte use (yes, no); immune function status as measured by total white blood cells from hematology panel at baseline; CMV-specific CD4 + CD69 + and CD8 + CD69 + T cells from immune function assay as continuous variables at baseline; enrolling region (North America, Europe, Asia Pacific); prior use of CMV prophylaxis (yes, no); age (as a continuous parameter) and age group (12 to < 18 years; 18 to < 45 years; 45 to < 65 years; ≥65 years); sex (male, female); race and ethnicity; TE CMV maribavir-resistant mutations (yes, no); and temporal identification of maribavir resistance (no, on maribavir treatment, post-maribavir treatment).

The exposure–response analyses were performed using a logistic regression model whereby a response can be defined as 0 (non-responder) or 1 (responder). Logistic regression analyses were performed to explore potential associations with maribavir exposure and the probability of response. The logistic regression model had the following form, assuming that maribavir exposure is included in the model:

$$\:Ln\:\left(odds\right)=Ln\:\left(\frac_}_\:}\right)=\:\alpha\:\:+\:\left(_X}_\right)$$

(1)

Pi is the probability of the event in the ith patient and α is the baseline Ln (odds) of the event. The model assumes a linear exposure effect on logit scale, with the parameter α representing the intercept, while β1 represents the slope linking the exposure parameter (X1) of maribavir (AUCss, Cmax.ss, and Ctrough.ss) to the response. The statistical significance of α and β1 was tested in the logistic regression models with a p value of < 0.05, and 95% CIs were provided for each parameter. The odds ratio (OR) for β1 was derived for a specific unit increment of the PK parameter.

The impact of each covariate on responders was explored in the above logistic regression equation using a stepwise approach by integrating relevant covariates one by one. At each step, the covariate retained was the one with the lower Akaike information criterion (AIC) and with statistically significant effect (p < 0.05 on β). The logistic regression models including drug exposure and multiple covariates have the following form:

$$\:Ln\:\left(odds\right)=Ln\:\left(\frac_}_\:}\right)=\:\alpha\:\:+\:__+__+\dots\:__$$

(2)

Pi is the probability of the event in the ith patient and α is the baseline Ln (odds) of the event, and β1…βn is the adjusted OR characterizing the dependence of the Ln (odds) on one or more covariates (X1…Xn). For continuous covariates, ORs were derived for a specific increment. For categorical covariates, ORs were derived for the test vs. reference group (e.g., female vs. male). No interactions between covariates were evaluated in the covariate analysis. The Firth correction was implemented in the logistic regression to reduce the bias of maximum likelihood estimates when a separation problem arises in small samples with unbalanced risk factors [27].

Additional logistic regression models were tested:

Log-linear exposure effect:

$$\:Ln\:\left(odds\right)=Ln\:\left(\frac_}_\:}\right)=\:\alpha\:+\:\left(_\text\text(X}_\right))$$

(3)

Maximum effect (Emax) exposure effect [28]:

$$\:Ln\:\left(odds\right)=Ln\:\left(\frac_}_\:}\right)=\:\alpha\:\:+\:\frac_\times\:_}_+_}$$

(4)

α is the baseline Ln (odds), EC50 is the drug concentration associated with 50% maximal response, and Emax is the maximum response.

Exposure–response analysis of safety

Descriptive statistics were derived for TEAEs, TE serious AEs (SAEs), and TEAEs of special interest during the 8-week treatment phase. Based on the descriptive statistics results, key TEAEs were selected for a formal exposure–response analysis, and exposure parameters of maribavir (AUCday, Cmax.day, Cavg, AUCss, and Cmax.ss) derived from the final PopPK models were merged with retained TEAE. Logistic regression models were developed to link maribavir exposure to the above probability of TEAEs with the exposure metric resulting in the best-fitting model according to AIC selected to perform the covariate analysis. The statistical significance of covariates was tested for potential risk factors in addition to maribavir exposure in the logistic regression models with a p value of < 0.05, as in the exposure–response analysis of efficacy.