Data for this population PK analysis were pooled from eight phase 1 studies, including five studies in healthy volunteers, (Neu-2566-HV-001, Neu-2566-HV-002, Neu-2566-HV-003, Neu-2566-HV-004, Neu-2566-HV-005, ACP-2566-006, ACP-2566-007, and ACP-2566-008), four phase 2 studies (Neu-2566-Rett-001, Neu-2566-Rett-002, Neu-2566-FXS-001, and Neu-2566-TBI-001), and the phase 3 LAVENDER study of trofinetide (Supplementary Table S1). In addition to healthy adults, the selected studies included pediatric, adolescent, and adult patients (5–45 years) with RTT, adolescent and adult males aged 14–40 years with FXS, and adolescent and adult males aged 16–75 years with TBI. Trofinetide was administered orally, via gastric tube, or intravenously (bolus and infusion); dosing data for the individual studies are summarized in Table S1 and pharmacokinetic sampling details are summarized in Table S2. The pharmacokinetic analysis was confined to study participants with available dosing information and ≥ 1 measurable post-dose trofinetide concentration. All studies included in the pharmacokinetic analysis dataset were approved by an institutional review board at each study site. All participants satisfied the eligibility criteria for their respective studies and provided their informed consent prior to study entry.

The pharmacokinetic analysis dataset included information on treatment assignment, dosing (amount, route of administration, and timing), pharmacokinetic sampling (time of collection relative to dose administration, blood trofinetide concentration), demographics, and clinical laboratory values. Trofinetide concentrations were determined in lithium-heparinized whole blood using a validated liquid chromatography with tandem mass spectrometry (LC–MS/MS) assay. The lower limit of quantitation was 0.10 μg/mL. Subjects with only below the limit of quantification samples were excluded from the model.

The previous population pharmacokinetic model (based on data from nine clinical studies), a two-compartment model with first-order absorption and linear elimination, served as the initial base structural model for subsequent refinement with data from four additional studies (ACP-2566-006 [10], ACP-2566-007 [11], ACP-2566-008 [12], and LAVENDER [8]).

The updated population pharmacokinetic model, based on the expanded dataset of 13 clinical studies, was used to obtain estimates of structural model parameters, including Vc and Vp, CL, Q, F1, and ka, as well as estimates of interindividual variability in these parameters. Interindividual variability of CL, Vc, Vp, Q, and F1 was estimated using an exponential error model assuming a log normal distribution, as described in Eq. (1). Residual variability of the concentration data was estimated using two separate exponential error models (one for healthy subjects and one for subjects with Rett syndrome, FXS, or TBI). These error models are summarized as:

(1) Exponential Interindividual Variability Error Model

$$_=}_\times exp\left(_^\right)$$

where \(_\) is the individual-specific estimate of the X parameter in the Ith subject, \(}_\) is the typical value of the X parameter in the Ith subject, and \(_^\) is a random variable that represents the persistent difference between the “true” individual-specific estimate and the typical value of the X parameter in the Ith subject; the \(_^\) are independent, identically distributed statistical errors with a mean of 0 and a variance equal to \(_^\).

The exponential error model for interindividual variability assumes that the variance is constant with respect to the log of the typical value of the parameter. With this variability model, the estimates are presented as coefficients of variation expressed as a percent (%CVs) and calculated as:

$$_=\sqrt_^\right)-1}\times 100$$

(2) Log/Exponential Error Model for Residual Variability

$$log\left(_\right)=log\left(}_\right)+_$$

where \(_\) is the jth measured value in the Ith subject, \(}_\) is the jth predicted value in the Ith subject using the specified model, and \(_\) is a random variable which represents the discrepancy between the jth measured log-transformed value in the Ith subject and the log-transformation of the value predicted from the specified model; the \(_\) are independent, identically distributed statistical errors with a mean of 0 and a variance of \(^\).

The log error model for residual variability assumes that the variance is constant with respect to the log of the predicted value. This model is mathematically and numerically equivalent to the exponential error model for the estimation methods used. The estimate is expressed as a standard deviation (\(\sigma\)) in log-transformed concentration units.

Goodness-of-fit criteria were applied to determine the appropriateness of the base structural model in characterizing the pharmacokinetics of trofinetide in the pediatric and adult populations. Goodness-of-fit criteria and/or considerations included: (1) convergence of the estimation and covariance routines; (2) estimation of ≥ 3 significant digits for each fixed and random effect parameter; (3) size of gradients associated with each parameter at the final iteration of estimation; (4) reasonable parameter estimates based upon the expected relationship; (5) adequate precision of parameter estimates as measured by the %RSE; (6) agreement in scatterplots of measured versus predicted and individual predicted observations assessed visually; (7) lack of trend or pattern in scatterplots of conditional weighted residuals versus predicted observations and time assessed visually; (8) lack of trend or pattern in scatterplots of individual weighted residuals versus individual predicted observations assessed visually; and (9) estimates of interindividual variability and residual variability for the specified model versus comparator models.

Following development of the base structural model, sources of variability in CL, Vc, and ka were explored by testing the statistical significance of covariate–pharmacokinetic relationships using univariate stepwise forward selection followed by backward elimination procedures. Covariates of interest comprised body weight, age, sex/disease state [male healthy volunteers, female healthy volunteers, patients with RTT (female only), patients with FXS (male only), male patients with TBI, and female patients with TBI], body mass index, creatinine clearance, aspartate aminotransferase, alanine aminotransferase, and total bilirubin. For forward selection, based on α = 0.01, covariates contributing a change in the minimum value of the objective function (VOF) of at least 6.64 (P < 0.01, one degree of freedom) and resulting in a decrease of ≥ 5% in interindividual variability of the parameter of interest were considered significant. During the backward elimination step, a covariate was considered statistically significant if it resulted in an increase in the VOF of at least 10.83 (P < 0.001, one degree of freedom for chi-square distribution) when removed from the model. The least statistically significant covariate (i.e., the covariate with the highest P > 0.001) was removed from the model first, and backward elimination was repeated until all remaining covariates were statistically significant at α = 0.001. The reduced multivariable model, with all significant covariates, was evaluated for any remaining biases in interindividual variability and residual variability error models. The predictive capability and performance of the final model was evaluated using a simulation-based, prediction-corrected visual predictive check (pcVPC) methodology [13]. The final population pharmacokinetic model was used to simulate 500 replicates of the analysis dataset to achieve ≥ 10,000 subjects overall. The results from pcVPC provided 95% confidence intervals of the simulated data for prediction percentiles, and plots of the 5th, 50th (median), and 95th percentiles of the distributions of simulated and observed trofinetide concentrations were overlaid to visually assess concordance between the model-based simulated data and the observed data. In addition, forest plots of the geometric mean ratios and 90% confidence intervals of exposure measures for covariates of interest were constructed to illustrate the relative differences in impact of various covariates. A significant covariate was considered to have a clinically meaningful impact on trofinetide pharmacokinetic exposure if the lower or upper bounds of the 90% confidence interval of the geometric mean ratio exceeded the bioequivalence acceptance limits of 0.8 and 1.25 relative to the reference group.

Using the updated population pharmacokinetic model, stochastic simulations were performed to predict the range of trofinetide exposures that would be achieved among virtual subjects aged 5–20 years receiving the weight-banded dosing regimen used in the LAVENDER study. Covariate data from the pharmacokinetic dataset from LAVENDER study subjects (n = 92) were used to estimate individual steady-state trofinetide exposures. Nonlinear mixed effects modelling was used to generate individual measures of daily trofinetide exposure according to the weight-banded dosing regimen in LAVENDER via integration of the predicted concentration time profile for each subject based on interindividual random effects in the final population pharmacokinetic model and individual empiric Bayesian estimates of pharmacokinetic parameters. Predicted concentration–time profiles were used to compute corresponding steady-state AUC0–12 estimates and to compare these estimates against the target exposure range (steady-state AUC0–12) of 800‒1200 μg h/mL.

All data analyses and presentations of data were performed using SAS Version 9.4 (SAS Institute, Cary, NC, USA) and KIWI Version 4 202111 or R Version 3.4.3 (Simulations Plus, Buffalo, NY, USA). Population modeling was performed using the computer program NONMEM, Version 7, Level 3.0 (ICON Development Solutions, Hanover, MD, USA).