This study developed the first population pharmacokinetic (popPK) model of subcutaneous aqueous progesterone in Chinese healthy postmenopausal women, integrating machine learning (ML) for covariate screening. Data were obtained from an open-label, single-dose, two-period crossover bioequivalence study in 50 subjects. A two-compartment model with first-order absorption (including a lag time) and first-order elimination best described the concentration–time profiles. Prior to conventional stepwise covariate modeling, ML algorithms—including adaptive least absolute shrinkage and selection operator (LASSO), elastic net, and random forest—were employed to identify physiological predictors of pharmacokinetic (PK) variability. The ML ensemble demonstrated high predictive performance and highlighted height and weight as potential covariates. However, subsequent popPK evaluation retained the effects of height on clearance (CL) and body weight on peripheral volume (V2) and intercompartmental clearance (CL2) in the final model. While these relationships were statistically identifiable, their clinical significance was considered limited. The final model was robust and used to simulate steady-state exposure following a once-daily 25 mg regimen. The simulated mean area under the concentration–time curve at steady-state (AUCss) was 324 ± 45.9 μg h/L, and the average concentration (Cavg,ss) was 13.3 μg/L. Notably, 98.9
Translating population pharmacokinetic (popPK) models of adalimumab in inflammatory bowel disease (IBD) into clinical practice remains challenging. We aimed to develop a parsimonious popPK model using real-world therapeutic drug monitoring data and perform dosing simulations to guide model-informed precision dosing (MIPD). We conducted a retrospective, longitudinal study of adult IBD patients treated with adalimumab between 2022 and 2024. Serum adalimumab concentrations from routine therapeutic drug monitoring were analysed using nonlinear mixed-effects modelling. Covariates were selected using a stepwise likelihood ratio approach. Model performance was evaluated using goodness-of-fit (GOF) plots, prediction-corrected visual predictive checks (pcVPC), normalised prediction distribution errors (NPDE), sequential Bayesian forecasting analysis, and bootstrap analysis. Monte Carlo simulations estimated the probability of target attainment (PTA) for steady-state trough concentration thresholds of ≥ 8, ≥ 10, and ≥ 12 mg·l⁻1 across different body weight, serum albumin, and dosing scenarios. From 201 patients (56.2
The clinical utility of CYP2D6 genotyping in tamoxifen therapy remains controversial because most large studies produced inconsistent results without accounting for inhibitory co-medications, which functionally convert genetic normal metabolizers into poor metabolizers (PM). This “phenoconversion gap” represents a critical drug–drug–gene interaction (DDGI) that is hypothesized to introduce a systematic pharmacokinetic signal dilution in survival estimates. We aimed to quantify the impact of drug-induced phenoconversion on the predictive utility of CYP2D6 through a systematic audit and quantitative bias analysis (QBA). We audited 33 primary studies, identified from the most comprehensive meta-analysis of tamoxifen pharmacogenetics conducted to date, and developed a Monte Carlo simulation model (5000 iterations) to “purify” genetic normal metabolizer reference arms. Patients were stochastically reassigned to functional phenotypes across inhibitor prevalence scenarios (0–35
Published Japanese population pharmacokinetic parameter sets are used for teicoplanin dosing, but their accuracy for teicoplanin trough concentrations in patients with cancer is unclear. This study evaluated agreement between a priori predicted and measured teicoplanin trough concentrations after loading in patients with cancer and identified clinical factors associated with prediction error. We retrospectively studied 217 patients with cancer receiving fixed loading (600 mg twice on days 1 and 2 and 600 mg once on day 3) with a day 4 trough measurement. Day 4 trough concentrations were predicted a priori using three Japanese population pharmacokinetic parameter sets. Prediction error was defined as predicted minus measured (mg/l). Multivariable linear regression evaluated associations with febrile neutropenia status and creatinine clearance (CCr). Mean prediction error ranged from −5.30 to −1.85 mg/l, indicating underprediction. Underprediction was greater in patients without febrile neutropenia (−7.12 to −2.98 mg/l) than in those with febrile neutropenia (−3.21 to −0.54 mg/l). In multivariable models, febrile neutropenia (vs absence) was associated with less negative prediction error (+4.85 to +5.27 mg/l; all P < 0.001). Higher CCr was associated with more negative prediction error (−0.32 to −0.72 mg/l per 10 ml/min; all P < 0.001), indicating greater underprediction at higher CCr. Japanese parameter sets underestimated day 4 teicoplanin trough concentrations after loading in patients with cancer, and prediction error differed by febrile neutropenia status after adjustment for CCr. Early trough measurement after loading should therefore be considered when these parameter sets are used.
Varenicline is a smoking cessation drug that requires dosing over days to weeks to reach a pharmacologically effective concentration, since higher doses may result in nausea and vomiting. High interindividual variability in varenicline pharmacokinetics may also be related to these adverse effects. This study aimed to identify possible genetic causes of this variability using physiologically based pharmacokinetic (PBPK) modeling. A mean PBPK model of varenicline with mechanistic absorption, metabolism, and renal clearance was built using four patient populations. This model was then used to simulate pharmacokinetics in 20 individuals, adjusting SLC22A2 (OCT2) transporter kinetics and physiology to fit the simulations to the observed data. The mean model included parameters for renal clearance through OCT2 (Vmax 0.017 mg/s/mg-trans) and reduced duodenal and jejunal absorption. This model aligned with observed data from four studies (Cmax and AUC fold errors 0.88-1.14). In the individual models, AUC differences were mainly accounted for by OCT2 kinetics [Vmax 0.004 (4.3 × decrease from the mean) to 0.035 (2.1 × increase) mg/s/mg-trans)], but 3/20 subjects (15
