Exposure-response (ER) analyses assessing the efficacy of large molecule therapeutics are often susceptible to confounding, where associations between drug exposure and patient disease severity can obscure true ER signals and complicate dose selection. To better understand the pharmaceutical industry perspectives on this challenge, the IQ consortium conducted a comprehensive survey targeting clinical pharmacologists, pharmacometricians, and statisticians. The survey, completed by 125 individuals from 23 pharmaceutical companies, aimed to assess awareness, perceived prevalence, mitigation strategies, and the overall role of ER analyses in the context of confounding. Results revealed strong industry awareness, with 88.0% of respondents acknowledging the relevance of ER confounding. However, uncertainty regarding its pervasiveness persists, particularly in non-oncology settings (31.2% unsure). A consensus emerged on the value of dose-ranging study designs as an effective mitigation strategy (81.6% agreement). In contrast, the use of advanced statistical methods for causal inference is inconsistent (45.6% usage), and confidence in their reliability is mixed, with 32.8% of respondents expressing uncertainty and most others rating them only moderate or somewhat reliable. While the majority agreed that confounded ER analyses should be interpreted with caution (79.2%), opinions diverged regarding their value in decision-making when dose-ranging data is insufficient. This uncertainty, coupled with a recognized need for additional alignment with health authorities, led to a call for best-practice guidance (92.8% view as valuable). Overall, the survey findings highlight a need for an industry-wide common approach and the development of clear frameworks to manage and interpret ER confounding for large molecule therapeutics.
Semi-mechanistic modeling approaches could be useful for optimal dose selection, and one of the approaches is pharmacokinetic (PK)/receptor occupancy (RO) model for monoclonal antibodies. This study aimed to develop a semi-mechanistic PK/RO model of S-531011, a humanized monoclonal antibody against human CCR8 and under development for the treatment of solid tumors, in order to predict human RO values in tumor tissues which are not available in the ongoing clinical study. The model consists of three compartments (central, peripheral, and tumor) and includes four compartments (S-531011, CCR8, and complexes 1/2 within each compartment). The model parameters were set using available clinical data, non-clinical data, and physiological information. The human PK/RO model was developed by refinement of a mouse PK/RO model. The time courses of RO values in the tumor compartment were simulated using the model, together with a conservative condition of slow transfer of S-531011 from blood to the tumor. It was predicted that the RO value in the tumor compartment at 21 days (trough after the third dose after administration of S-531011 80 mg) would be maintained at > 90% under typical model assumptions. Under the conservative condition, the RO value in the tumor after administration of S-531011 800 mg would remain at > 90%. The developed semi-mechanistic PK/RO model could predict human RO values of S-531011 in tumor tissues. Under typical model assumptions, simulations suggested the dose range of 80-800 mg of S-531011 every 3 weeks would achieve > 90% RO in tumor tissues.
Tacrolimus dosing in liver transplantation is complicated by a narrow therapeutic index and high CYP3A5 genetic variability. While saturable Michaelis-Menten kinetics can explain nonlinearities, identifying saturable parameters from routine clinical data remains challenging. This study aimed to determine the optimal structural model and develop a precision dosing algorithm. A population pharmacokinetic analysis was conducted in 114 patients, yielding 1989 observations. CYP3A5 genotypes were determined for both recipients and donors. Using Phoenix NLME, we rigorously compared linear versus Michaelis-Menten elimination structures. Stepwise covariate modeling was conducted to quantify the impact of genetic, physiological, and pharmacological factors, followed by Monte Carlo simulations to optimize dosing. A conventional two-compartment model adequately described the data without requiring a Michaelis-Menten structure. The combined CYP3A5 genotype exhibited a distinct stepwise reduction in apparent clearance from the homozygous expressor to the non-expressor group. Fluconazole emerged as a major inhibitor, reducing clearance by 33%, whereas prednisolone showed modest induction. Hemoglobin displayed a significant inverse relationship with clearance. Crucially, incorporating the daily dose as a covariate on clearance effectively captured the apparent nonlinear disposition. Simulations confirmed that fluconazole-treated patients require substantially lower doses (1.5-3.0 mg every 12 h) compared with fluconazole-free patients (2.5-7.0 mg every 12 h). A dose-dependent two-compartment model offers a basis for model-informed dose selection, addressing reported nonlinearities through physiological covariates. We provide a model-informed dosing algorithm that accounts for combined recipient/donor genetics and drug interactions, which may improve target attainment in liver transplant populations.
Maximum likelihood estimation in population pharmacokinetic/pharmacodynamic (PK/PD) nonlinear mixed-effects (NLME) models targets fixed-effect, interindividual-variability, and residual-variability parameters through a marginal likelihood that integrates each subject's contribution over individual random effects. Because these subject-level integrals are usually unavailable analytically, Laplace estimation approximates each contribution using a second-order Taylor expansion around the empirical Bayes estimate (EBE; conditional posterior mode of the individual's random effects), making EBE estimation and curvature evaluation recurring tasks during objective function value (OFV) and derivative evaluation. Finite differences (FD) are commonly used for these derivative calculations but are costly and step-size sensitive. We developed automatic differentiation (AD)-based methods for Laplace NLME estimation and compared three ways to account for EBE sensitivities during updates: FULL-implicit uses the EBE mode equations, FULL-unroll differentiates through the Newton steps used to find EBEs, and STOP omits EBE sensitivity during outer differentiation. In 100 matched starts for a synthetic one-compartment PK example with absorption/elimination ambiguity, FULL-implicit, FULL-unroll, and FD matched the lowest OFV within 10 - 4 OFV units, whereas STOP had maximum ΔOFV = 1.9 relative to OFV min = - 1594.62 . Median wall times were 0.045, 0.108, 0.068, and 0.758 s for FULL-implicit, FULL-unroll, STOP, and FD, respectively, making FULL-implicit about 17-fold faster than FD. In 10 runs on public warfarin PK/PD data using the ODE representation, FULL-implicit achieved a lower best OFV than FD (1624.02 vs. 1628.60) and was 36-fold faster by median wall time (0.45 vs. 16.3 min). In both scenarios, FULL-implicit provided a faster AD-based alternative to FD while reaching comparable or lower OFVs.
ABSTRACT A whole‐body mechanistic PBPK model for meropenem (MPN) was developed in PK‐Sim and validated using published plasma and tissue concentration–time data in healthy volunteers and critically ill patients. Renal elimination was implemented as glomerular filtration plus active tubular secretion (basolateral uptake via OAT3 with apical efflux), and non‐renal clearance via DHP‐mediated hydrolysis. The model was scaled to sepsis or septic shock by incorporating disease‐specific physiological changes and optimizing OAT3 activity and tissue permeability to reproduce observed variability. Simulations in a virtual septic population assessed unbound interstitial concentrations in clinically relevant tissues under ILAS‐recommended typical and maximum dosing MPN regimens. Antibacterial effect against Escherichia coli and Klebsiella pneumoniae was evaluated using a published PK/PD model driven by simulated unbound tissue concentrations. The PBPK model reproduced observed plasma profiles in healthy volunteers and captured plasma and subcutaneous interstitial exposure in sepsis. Simulations showed clear dissociation between plasma and interstitial exposure and marked tissue‐specific heterogeneity. Predicted clearance increased in sepsis (augmented renal clearance) and decreased in septic shock (impaired renal function and secretion). Although maximum dosing increased plasma and tissue exposure, PD simulations indicated effects were already near the plateau with standard dosing, yielding minimal additional antibacterial benefit from routine dose escalation. Transporter‐informed PBPK/PD modeling explains dynamic, severity‐dependent changes in meropenem clearance and tissue exposure in sepsis, highlights limitations of plasma‐only assessment, and supports individualized, mechanism‐informed optimization rather than universal dose scaling.
ABSTRACT Accurate physiologically based pharmacokinetic (PBPK) simulation of drug–drug interaction (DDI) potential requires estimation of the relative contribution of the impacted pathway. While fraction metabolized by CYP enzymes is usually estimated using dedicated clinical DDI studies with strong CYP inhibitors, this approach might not be available in some situations. In the case of the menin inhibitor, ziftomenib, a dedicated DDI study in healthy subjects was infeasible due to the potential for toxicity associated with the mechanism of action, so an alternate approach was used. AML patients are typically immuno‐compromised and hence are at a high risk of fungal infections. For this reason, co‐administration of azole antifungals (moderate and strong CYP3A4 inhibitors) was permitted in Clinical Study KO‐MEN‐001. The PK data obtained in the presence or absence of CYP3A4 inhibitors was used to refine the estimate of ziftomenib fmCYP3A4 to 60%–70%. Once refined, the model was applied to predict the victim DDI liability of ziftomenib in the presence of CYP3A4 inhibitors and inducers. Moderate or weak interaction (2.6‐fold and 1.4‐fold increase in AUC) was predicted with itraconazole and isavuconazole, respectively. Approximately 80% reduction in ziftomenib AUC was predicted with rifampicin. Ziftomenib was predicted to be a weak CYP3A4 inhibitor, causing a 1.9‐fold increase in midazolam exposure. In the absence of data from dedicated DDI clinical trials, these results were used to support regulatory interactions with the US FDA regarding concomitant administration of ziftomenib with other medications such as CYP3A4 modulators. These modeling results ultimately supported a range of DDI language in ziftomenib label.
Drug Combinations offer increased therapeutic efficacy and reduced toxicity compared with single agents. Understanding a drug combination's mechanisms of action (MoA) can provide important insights into therapeutic efficacy. The MoA of many FDA-approved drugs, however, often remains unclear. To decipher the underlying molecular mechanisms of drugs used alone and in combination, we investigated the combination of a statin (atorvastatin or simvastatin) plus ezetimibe using drug-treated RNA-seq transcriptome data from the human hepatocyte-like SOAT2-only-HepG2 cells and from liver biopsies of non-obese normolipidemic patients with uncomplicated cholesterol gallstone disease in the Stockholm Study. We proposed a novel Boolean logical modeling framework to simulate the MoA of a drug combination using fourteen two-variable Boolean models. Thereafter, a pattern matching approach was applied to associate drug-induced differentially expressed genes with the idealized differential expression templates derived from Boolean models. We found 1560 and 565 genes differentially expressed in at least one treatment condition in SOAT2-only-HepG2 cells and liver biopsies, respectively. Our analysis revealed both expected and novel combinatorial modes of the statins and ezetimibe. We mapped the downstream genes of each combinatorial mode to the human protein-protein interactome and obtained underlying pathways, which are important for understanding the therapeutic effects of the drug combinations. Functional enrichment and disease-association analyses of the downstream genes also provide critical insights into the additional therapeutic actions of the drugs. Our study demonstrates that drug-induced transcriptomes, integrated with the human interactome, are informative in deciphering the MoA of drug combinations using Boolean logical modeling.
The standard Douglass ternary binding model (TBM) for three-body equilibria assumes a well-mixed, three-dimensional solution. When applied to bispecific T-cell engagers (BiTEs), however, the productive trimeric complex forms not in bulk solution but within a nanoscale membrane synapse with finite receptor copy numbers. We present a generalization of the TBM to membrane-confined systems that replaces the macroscopic bulk volume with a coarse-grained reactive contact volume defined by synapse geometry and microvillus topology, and extends the deterministic equilibrium to a stochastic description via the chemical master equation. The framework preserves the original algebra while restoring its representational capacity for the regime in which therapeutic activity occurs. A key finding is that conventional bulk mapping places the system in the affinity-limited regime, where antigen density is mathematically inert and the TBM predicts identical dose-response regardless of target expression. Membrane confinement shifts effective antigen concentration by six orders of magnitude-from ∼ 10 - 3 nM to ∼ 10 3 nM-restoring antigen density as a governing variable for trimer formation. Using blinatumomab (anti-CD19 BiTE) as a case study, we introduce the absolute formation dose d N * : the drug concentration required to produce a fixed number of ternary complexes sufficient for T-cell activation. This metric replaces the conventional T F 50 , which normalizes each cell line to its own maximum, erasing the density dependence that confinement rescues. For NALM-6 and HAL-01 cell lines (CD19 density ratio ≈ 2.5 ), the framework predicts a corresponding ≈ 2.5 -fold difference in required dose-a prediction structurally invisible to the bulk formulation.
ABSTRACT Reliable physiologically based pharmacokinetic (PBPK) modeling depends on tissue‐specific expression profiles that reflect protein activities governing drug disposition. In PK‐Sim, existing expression databases rely on transcriptomics data. However, mRNA levels often exhibit limited correlation with protein abundance, frequently necessitating empirical expression modification to align bottom‐up simulations with clinical observations. To address this limitation, we developed ProteinDB as a proteomics‐based expression database for PK‐Sim, using proteomics data primarily from PaxDb v6.0. Raw proteomics data were mapped to gene identifiers and standardized into absolute concentrations (μmol/L tissue) before integration into PK‐Sim. Cross‐platform comparisons were performed for hepatic protein abundance across PBPK platforms, while cross‐omics comparisons were made of relative tissue distributions with transcriptomics‐based PK‐Sim databases. The performance of ProteinDB in PBPK modeling was evaluated using the probe substrates midazolam, digoxin, rifampicin, and tizanidine, with associated drug–drug interactions. Cross‐platform comparisons showed strong agreement for most hepatic enzymes and transporters, while revealing divergences for proteins with greater inter‐individual variability, lower abundance, or limited evidence base. Cross‐omics analyses demonstrated tissue‐dependent discrepancies between transcript‐ and protein‐based expression patterns, with higher consistency observed for kidney and small intestine, particularly with the RT‐PCR and Bgee databases. For PBPK modeling, ProteinDB showed consistently comparable or superior predictive performance for systemic exposure and other clinical endpoints compared with transcriptomics‐based baseline and empirically modified library profiles. By providing a direct physiological basis for system parameterization, ProteinDB offers a robust alternative to current transcriptomics PK‐Sim databases and reduces the reliance on empirical expression modification, thus improving the reliability of prospective PBPK modeling.
Apparent permeability ( P app ) of a drug molecule serves as a principal indicator of drug absorption for pharmacokinetic modeling, and accurate prediction of P app values using machine learning would avoid time-consuming and expensive In Vitro experiments. In this study, four different feature representations were compared for predicting log P app values, with each representation serving as input to three subsequent regression models. The best predictive model achieved an R 2 value of 0.68 and RMSE of 0.42 on the test data. Furthermore, the SHAP values corresponding to individual molecular descriptors were analyzed to interpret the influence of these descriptors on the model prediction. Some key properties like octanol-water partition coefficient and number of basic atoms were found to have a strong influence on the P app value. Additionally, some specific molecular substructures were identified that contribute either positively or negatively to the model output, thereby inferring effects on drug permeability. Finally, the uncertainty in the predictions was quantified through a distribution-free and model-agnostic method, jackknife+, used for the estimation of confidence intervals. At the optimal experimental setting, the confidence intervals derived using the overall best performing model covered 96.12% and 86.73% of the original log P app values with more than 95% confidence for test and independent data, respectively. With the same confidence, the confidence intervals contained 99.87% and 100% of the predicted log P app values for the test and independent data, respectively, underscoring the reliability of the model. Overall, the presented ML provides the most robust informed P app prediction among compared models presented to date.
ABSTRACT Selecting the most relevant models for model‐informed precision dosing (MIPD) remains challenging, particularly when numerous population pharmacokinetic (popPK) models are available, making external evaluation of all candidates time‐ and resource‐consuming. We aimed to develop and assess an adaptable relevance‐based scoring methodology, illustrated with tacrolimus in adult transplant recipients. Candidate criteria were derived from published guidelines and external evaluations. A 23‐item relevance score was developed, covering training dataset, study design, evaluation methodology, covariates, and parameter precision. For tacrolimus, a systematic literature review identified 28 models which were scored for relevance and externally evaluated using an independent dataset from Rennes University Hospital (86 patients). Predictive performances were assessed at individual and population levels using bias and imprecision metrics. Predicted and observed AUCs were compared against therapeutic thresholds. Models were then ranked using an external evaluation score (maximum 14 points) based on predefined acceptance limits for bias and imprecision. Correlations between relevance‐based and external scores and ranks were assessed using Pearson and Spearman correlations. The mean relevance score was 22.3/56 (range: 6–38). Higher scores were associated with larger and richer datasets, prospective or multicentric designs, inclusion of clinically relevant covariates, and precise parameter estimates. Individual predictions showed acceptable bias and imprecision, whereas population predictions were consistently inaccurate. AUC concordance (56%–94%) was mainly limited by overestimation. Highest external score was 6/14. Correlation between relevance‐based and external scores was moderate (Pearson = 0.56; Spearman = 0.61). This work provides a structured basis for popPK model selection and identified tacrolimus models suitable for MIPD applications.
ABSTRACT Aging is a systems‐level process linking metabolic dysfunction, inflammation, impaired repair, frailty, and multimorbidity, whereas existing pharmacological strategies usually optimize disease‐specific endpoints such as weight loss or HbA1c rather than aging‐related trajectories. We developed an SBML‐compliant quantitative systems pharmacology (QSP) model in which aging is represented as a dynamic, pharmacologically modifiable endpoint. The model integrates four coupled layers: metabolic/pharmacodynamic responses to GLP‐1 receptor agonism, SGLT2 inhibition, metformin and rapamycin; adverse‐event dynamics; aging states including damage accumulation, repair capacity, frailty and biological age gap; and biomarker outputs including GDF15, cystatin C, leptin, adiponectin and estimated glucose disposal rate. The semaglutide submodel was calibrated against published STEP trial endpoints, and Bayesian hierarchical meta‐analysis, global sensitivity analysis, practical identifiability analysis and internal consistency checks were used to assess model behavior. The calibrated model reproduced semaglutide‐associated weight loss, HbA1c reduction and transient nausea within pre‐specified error benchmarks. Bayesian meta‐analysis confirmed strong metabolic effects for semaglutide, moderate glycaemic effects for SGLT2 inhibitors and metformin, and a near‐zero HbA1c effect for rapamycin. Sensitivity analysis revealed largely orthogonal metabolic and aging parameter spaces. Combination simulations identified two mechanistically distinct optima: GLP‐1 receptor agonist plus SGLT2 inhibitor plus metformin for metabolic improvement, and GLP‐1 receptor agonist plus SGLT2 inhibitor plus rapamycin for aging‐related benefit. Metabolic optimisation and aging optimisation are therefore mechanistically distinct objectives that do not converge on the same drug combination. These predictions are hypothesis‐generating and require external validation against independent longitudinal datasets and clinical safety evaluation before translation to treatment recommendations.
ABSTRACT Physiologically based pharmacokinetic (PBPK) modeling is increasingly used to predict infant exposure to medications via breast milk. Existing guidance on reporting of PBPK modeling analyses is unstructured and nonbinding to researchers. We evaluated reporting practices and modeling approaches in published PBPK modeling studies. A systematic search was conducted across three online databases. Studies published through March 2025 that leveraged PBPK to evaluate milk drug excretion and infant exposure were included. We developed a 22‐item structured data abstraction tool—Lactation PBPK Model Reporting Checklist (LAC‐PBPK). Study findings were reported using descriptive and narrative analyses. Twenty‐six published studies modeling 54 medications met the inclusion criteria. Using the data abstraction tool, only 4 (18%) of reporting domains were reported in all studies. These include the drug modeled, choice of mammary distribution model, context of PBPK modeling (model purpose), and the modeling platform. Most studies (88.4%) implemented whole‐body PBPK model structure, with one‐half (53%) assuming perfusion‐rate‐limited milk excretion. Assumptions reflect modeling limitations, primarily due to inadequate reporting in clinical lactation studies and knowledge gaps in infant and mammary physiology. PBPK models can support mechanistic assessment of drug excretion into breast milk, but confidence in model outputs depends on intended use, data availability, and transparent reporting. This highlights the need for a prioritized, fit‐for‐purpose lactation PBPK reporting framework. From a regulatory perspective, transparent reporting and cautious selection of approaches for estimating infant milk intake, model performance evaluation, physiological extrapolation, and parameter uncertainty analyses are essential to ensure reproducibility and confidence in PBPK‐informed infant exposure assessment.
ABSTRACT Large language models can execute pharmacometric workflows, but they make consequential domain‐specific errors when task instructions lack adequate details. This tutorial teaches pharmacometricians how to define self‐contained tasks that embed domain‐specific rules, verification criteria, and worked examples into each step of a pharmacometric workflow. These individual tasks are then organized into a structured task library where each task runs in a fresh LLM instance (with clean context), with information passed between tasks through shared workspace files. The tutorial covers context engineering, controlling what information reaches the LLM at each decision point, along with verification layers and methods to iteratively refine the task library. We demonstrate the approach on a synthetic population PK/PD scenario and provide the task library and implementation guide in the Supplementary Material.
ABSTRACT Model‐informed precision dosing is often constrained by the limited generalizability of traditional population pharmacokinetic models, especially in critically ill patients. A hybrid machine learning‐population pharmacokinetic framework is proposed to improve a priori pharmacokinetic predictions by integrating real‐world clinical data. This approach was applied to vancomycin trough concentration prediction. Two widely used two‐compartment population pharmacokinetic models provided individual pharmacokinetic parameter estimates. Maximum a posteriori Bayesian estimation was used to adjust population parameters for individual patients based on drug administration records, therapeutic drug monitoring values, and patient‐specific covariates from the MIMIC‐IV database. The resulting clearance and central volume of distribution estimates served as training targets for XGBoost and symbolic regression models. Machine learning‐predicted parameters were reinserted into the original pharmacokinetic equations to generate a priori vancomycin trough concentrations without reliance on therapeutic drug monitoring input. The hybrid models demonstrated improved prediction accuracy over traditional population pharmacokinetic covariate models and reduced vancomycin trough concentration prediction error by up to ~20%. XGBoost generally provided the highest predictive performance, while symbolic regression produced interpretable mathematical expressions revealing associations between non‐traditional clinical predictors and pharmacokinetic parameters, highlighting a trade‐off between accuracy and interpretability. This framework illustrates the potential of combining machine learning with population pharmacokinetic modeling to refine pharmacokinetic parameter estimation and support more precise, individualized dosing. The workflow is adaptable to other drugs and patient populations, offering a generalizable methodological strategy to identify non‐traditional predictors and enhance existing pharmacometric model performance in real‐world clinical settings.
ABSTRACT Numerous clinical trials (CTs) and population pharmacokinetic (PK) models have been published on dapagliflozin, an approved SGLT2 inhibitor used to treat type 2 diabetes, heart failure, and chronic kidney disease. This study proposes a Bayesian workflow for the development of minimal physiologically‐based PK models (mPBPK) that integrates all available PK information, to support uncertainty quantification and informed drug development. First, a systematic collection of published dapagliflozin PK data and models was performed. A mPBPK model was then developed, and posterior parameter distributions were obtained based on data from parallel‐design CTs. Three Bayesian modeling packages: NIMBLE (v1.1.0), MCSim (v6.2.0), and Torsten (v0.89.0), and three alternative prior specifications were evaluated. Model validation was performed using PK data from crossover CTs and urinary recovery data, followed by sensitivity analysis. 18 studies reporting PK data and 10 reporting PK models were identified. Posterior distributions were comparable across programs: 95% credible intervals overlapped. Torsten showed superior sampling efficiency, as compared to NIMBLE and MCSim (5.7 and 10.0 times higher in tails and central posterior regions, respectively); however, the average effective sample size per hour was comparable for Torsten and MCSim. The predicted urinary recovery was 2.2%–4.4% (mean 3.2%), while the observed values were in the 0.8%–4.0% range (mean 2.0%). Glomerular filtration rate and fraction unbound were the main contributors to inter‐trial variability in urinary recovery, while volume of distribution and clearance had the highest influence on maximum concentration and area‐under‐the‐concentration curve, respectively. The proposed Bayesian workflow is flexible and transferable to other mechanistic model types.
ABSTRACT pH‐dependent drug–drug interactions (DDIs) commonly occur when acid‐reducing agents (ARAs) are co‐administered with weakly basic drugs. Although physiologically based pharmacokinetic (PBPK) modeling effectively evaluates such DDIs, its use is limited by reliance on costly commercial software. This study developed a PBPK‐informed machine learning model to support early assessment of pH‐dependent DDI risk in drug development. PBPK models were built for 14 representative weakly basic drugs using literature‐derived parameters to identify eight key determinants (e.g., solubility and pKa). Based on these distributions, virtual drugs were generated and simulated under varying gastric pH conditions; compounds with DDI AUC ratios < 0.1 were excluded, yielding 4339 virtual drugs. An extreme gradient boosting (XGBoost) algorithm was used to develop the machine learning model, and an external validation set comprising clinically observed data from an additional eight drugs was employed. The XGBoost model showed excellent internal performance (training: R2 = 1.00, MAPE = 0.99; test: R2 = 0.98, MAPE = 2.64). When evaluated using an external validation set comprising clinically observed data from eight drugs, 100% of the predicted values fell within the 0.5–2.0‐fold range of the observed clinical values. For DDI AUC risk classification, the XGBoost model achieved an accuracy of 87.5% (7/8). This PBPK‐informed ML framework enables efficient screening of pH‐dependent DDI risk for weakly basic drugs co‐administered with ARAs. The freely accessible web tool (https://ddi‐antacid.xy3yx.com/), integrating structure‐based ADMETlab3.0 estimation, offers a practical complement to conventional PBPK modeling for early drug development.
ABSTRACT Respiratory syncytial virus (RSV) infection can result in a range of disease severity from mild upper respiratory tract symptoms to lower respiratory tract illness, with clinical manifestations dependent on age, co‐morbidities, and immune status. At present, there is no antiviral approved for adults with RSV disease, and only one treatment option with limited utilization is available for pediatric patients. While several RSV antiviral candidates have demonstrated virological efficacy in healthy adult challenge studies, to date they have subsequently failed to show efficacy in patients. Mechanistically understanding the translation of efficacy from preclinical to challenge to patient studies is essential in addressing this unmet medical need in children and adults with RSV disease. This work uses a quantitative systems pharmacology (QSP) approach to mathematically represent the underlying pathophysiology of RSV infection to simulate and predict the effects of investigational direct acting RSV antiviral therapeutics. We developed virtual populations that capture viral load dynamics in viral challenge studies, where the participants are healthy adult volunteers, upon treatment with the F‐protein inhibitor sisunatovir and the N‐protein inhibitor zelicapavir. Due to underlying differences in dynamics between key populations, we also developed a pediatric patient virtual population to match the results from the zelicapavir pediatric Phase 2 study. This work projects that for a minimum coverage > 1× free EC90 at Cmin, a treatment window of 5 to 7 days post‐symptom onset is feasible for observing virological efficacy in pediatric patients and may lead to potential success in a phase 3 pivotal RSV antiviral drug trial.
ABSTRACT In this work we describe the development of a web‐based application for static drug–drug interaction (DDI) risk assessment in accordance with the International Council for Harmonization (ICH) M12 guidance. The app was built using the Shiny for Python framework and it employs a modular mathematical modeling structure that incorporates models for the assessment of enzyme inhibition, enzyme induction, and transporter inhibition at intestinal, hepatic, and renal levels. In addition, a “net effect” model estimates the combined impact of inhibition and induction on victim drug exposure, expressed as the area under the curve ratio (AUCR). Matrix‐specific approaches for estimating unbound fractions in microsomes and hepatocytes (fu,mic and fu,hep), as indicated in regulatory alignment, are implemented. The application provides a dynamic interface that supports flexible parameter input, real‐time evaluation across multiple DDI scenarios, and automated risk categorization using predefined thresholds based on the ICH M12 guidelines, with visual cues to aid risk interpretation. Additional features include integrated equation display for transparency and interpretation, export capabilities to PDF and Excel formats and a glossary with links to guidance documents and resources from major regulatory authorities (FDA, EMA, PMDA and NMPA). This DDIapp is validated against Certara's drug–drug interaction calculator under ICH M12 evaluation conditions. DDIapp offers a user‐friendly platform for assessing the risk of pharmacokinetic drug interactions as a perpetrator involving metabolic enzymes and transporters, supporting both scientific research and regulatory decision‐making.
ABSTRACT When study designs are evaluated using clinical trial simulations for their ability to identify covariate effects in population pharmacokinetic (PopPK) modeling, it is typically assumed that the true model will be known at the data analysis stage. In this study, this was compared with the more realistic assumption that the PopPK model needs to be built on the data generated by the planned study. Three approaches were compared: (i) stochastic simulation and re‐estimation (SSE) with the simulation model, (ii) automated model development (AMD) with exploratory covariate search (AMD‐exploratory), and (iii) AMD forcing the covariate effect into the model from the start and reevaluating it in the end (AMD‐structural). With a simulated covariate effect (a hypothetical pregnancy effect on clearance), we assessed (i) the type 1 error (T1E) and the power of covariate identification and (ii) covariate parameter accuracy. The T1E rate was controlled in SSE and AMD‐exploratory but 20% inflated for AMD‐structural. The power of covariate identification in rich, medium, and sparse designs was (i) 99%, 100%, and 79% in SSE, (ii) 74%, 72%, and 41% in AMD‐exploratory, and (iii) 92%, 93%, and 80% in AMD‐structural. Sparse designs amplified power differences between strategies, with AMD‐exploratory often selecting alternative or no covariates. The rRMSE of covariate parameter estimates was lowest in SSE (27%, 22%, and 42%), followed by AMD‐exploratory (34%, 26%, and 49%) and then AMD‐structural (42%, 47%, and 60%). SSE provides optimistic power estimates as model building is data‐driven, while AMD‐based approaches incorporate model uncertainty and reflect real‐world analysis conditions.
