Towards Model-Informed Precision Dosing of Voriconazole: Challenging Published Voriconazole Nonlinear Mixed-Effects Models with Real-World Clinical Data

Model-informed precision dosing (MIPD) is an emerging approach fostering the use of mathematical models in the framework of quantitative dosing decision support. This approach aims to predict, individualize, and thereby optimize dosing and therapy outcomes based on patient characteristics and data from therapeutic drug/biomarker monitoring (TDM) [1, 2]. MIPD offers the opportunity to significantly improve the efficacy and safety of drug therapies, reduce the emergence of resistance, and save costs. In brief, based on (preferably rich) pharmacokinetic (PK) and/or pharmacodynamic (PD) individual-level data, typically a nonlinear mixed-effects (NLME) model is developed, and the typical population PK and/or PD parameters are estimated. Different sources of variability between/within individuals are quantified and accounted for (e.g., interindividual, interoccasion, interstudy, or residual variability). Individual-specific factors (‘covariates’, e.g., CYP2C19 genotype or body weight) can be integrated to explain part of the variability. In the clinical setting, the values of the population parameters of the developed NLME model can be used as Bayesian uncertainty priors, if available, in combination with recorded patient-specific factors, to a priori predict the exposure likely to result from a hypothetical dosing regimen and thereby derive an adequate individual first dosing regimen (probabilistic setting). In combination with patient’s individual dosing history and drug monitoring data, individual parameters can be estimated (maximum a posteriori [MAP] Bayesian estimates). Thereby future trajectories of the individual concentration-time profiles will accurately be predicted and dosing regimens to achieve a desired efficacy or safety threshold be suggested (Bayesian forecasting setting). Prior to its application in MIPD, a candidate NLME model needs to be carefully selected, thoroughly evaluated, and must provide adequate predictive performance. Using external evaluation techniques, i.e., based on the candidate NLME model together with an external evaluation dataset, the predictive performance can be assessed using prediction-/simulation-based methods and retrospective application of Bayesian forecasting [2]. Candidate drugs for MIPD need to fulfill several criteria to be eligible for MIPD [1‐4], all of which are met by the second-generation triazole voriconazole, classified as ‘essential medicine’ by the WHO [5]: high interindividual PK variability (e.g., metabolized by polymorphic enzymes), narrow therapeutic range, potential for severe adverse drug reactions, vulnerable patient population (e.g., pediatrics, patients undergoing transplantation), and defined relationship between dose, plasma concentration and clinical response as well as safety [6‐8]. Voriconazole is commonly used for first-line treatment of invasive fungal infections [9], and as primary or secondary prophylaxis in immunocompromised patients [10, 11]. Voriconazole PK is characterized by nonlinearity and high inter-/intraindividual variability. Potential sources of variability comprise, e.g., status of metabolizing enzymes (fraction metabolized = 98% of parent compound, primarily via polymorphic CYP2C19, 3A4 and potentially 2C9), age, sex, liver or disease status, concomitant medication, and route of administration [12‐20]. The relative contribution of CYP2C19, 3A4 and 2C9 to voriconazole clearance, the (auto-) inhibitory potential of the parent compound and metabolites, as well as the mechanism of the inhibition processes were intensively discussed and only recently comprehensively investigated in systematic and quantitative in vitro investigations [21‐24]. The mean contribution of the individual isoenzymes to voriconazole N-oxide formation was 63.1% for CYP2C19, 13.2% for CYP2C9 and 29.5% for CYP3A4. Voriconazole N-oxide was the weakest and voriconazole and hydroxy-voriconazole comparably strong inhibitors of CYP2C9 and CYP3A4. CYP2C19 was significantly inhibited by voriconazole only. Time-independent (i.e., competitive and non-competitive) inhibition by voriconazole, voriconazole N-oxide and hydroxy-voriconazole was demonstrated. Voriconazole is approved to be administered as (total body) weight-adapted intravenous (IV) infusion or orally (PO) without weight adaptation (oral bioavailability of 96% reported) [25, 26]. Polymorphic CYP2C19 and corresponding genotype-predicted phenotypes significantly impact voriconazole exposure (4-fold higher area under the concentration-time curve during one dosing interval (AUC_τ) in poor than in normal metabolizers) [27, 28]. Frequently used target ranges in clinical practice for minimum plasma concentrations (C_min after 12 h) range from ≥ 1.0 mg/L up to 6.0 mg/L for effective and safe therapy, respectively [8, 29, 30]. Since its approval in the early 2000s, several voriconazole NLME models have been published and reviewed [21, 31], and externally evaluated [32‐34]. Moreover, a so-called ‘hybrid’ (aggregate) model comprising features and prior information from six previously published NLME models including nonlinear, concentration-, CYP2C19- and time-dependent elimination has been developed for MIPD (see Section 1.1 in Online Resource 1) [35]. This work aimed to provide guidance and recommendations for evidence-based application of MIPD for the model compound voriconazole by implementing a workflow for external evaluation and application of MIPD [36]. The workflow consisted of (i) externally evaluating and comparing the predictive performance of the aggregate model, versus two ‘standard’ NLME models with linear elimination (accounting [13] or not accounting [6] for CYP2C19 genotype), using data from a rich in-house database, and (ii) investigating different strategies and illustrating the clinical impact of Bayesian forecasting (Fig. 1).

The schematic workflow of the external model evaluation framework comprising deterministic simulations, prediction-/simulation-based model performance evaluation and different strategies for Bayesian forecasting is illustrated in Fig. 1.

Several in silico and clinical studies have shown the benefit of MIPD in supporting and optimizing dosing regimen selection for various drugs [1]. Our study demonstrated the need for thorough external model evaluation before applying a candidate NLME model in the context of MIPD. We identified and selected voriconazole (representative of compounds with complex PK and narrow therapeutic range) as an ideal compound to implement a workflow for model evaluation and MIPD. A published aggregate model for voriconazole comprising features and prior information from six previously published NLME models, was compared with two additional ‘standard’ NLME models. The aggregate model of voriconazole showed better predictive performance than the ‘standard’ models but overall, the predictive performance of all evaluated models was only modest. This suggests the following knowledge gaps: (i) important PK processes (particularly after PO administration) were not sufficiently implemented in the structural sub-model, (ii) magnitude and sources of interindividual variability were not exhaustively captured, and (iii) interoccasion variability was not adequately accounted for. Our results indicate the need for a comprehensive (pre-)clinical database as basis for NLME model development for MIPD application, to (i) identify and implement critical PK processes in the structural model, (ii) identify sources of interindividual variability, and (iii) adequately quantify interoccasion variability. Successful examples for this (semi-)mechanistic approach comprise rifampicin [53‐57], linezolid [58, 59], isoniazid [60], itraconazole [61, 62], and voriconazole [63]. Adequate quantification of interoccasion variability was demonstrated to be an important step for MIPD [39].

Since its approval in the early 2000s, several voriconazole NLME models have been published and reviewed. Voriconazole NLME models are highly variable in structure, including covariates, and parameter values, as the study design, underlying population, and sampling scheme have a significant impact on the selected model structure. Given the available data, a structural model comprising a lower number of compartments/parameters is often preferred over a more complex (semi-)mechanistic structural model. Consequently, although voriconazole has been characterized as a compound demonstrating complex nonlinear PK [37, 64], inclusion of simple first-order (linear) absorption and elimination processes was selected for most of the NLME models. Autoinhibition of relevant CYP-mediated elimination pathways, which was determined in vitro [22, 65] and in vivo [21, 66, 67], has only been included in a few NLME models. This process was implemented using an empirical time-dependent function (elimination decreases over time) [68], or an empirical inhibition compartment (elimination decreases with respect to the drug concentration in an empirical inhibition compartment) [69]. Furthermore, significant differences in PK were identified in (i) children and adults, (ii) different CYP2C19 genotypes and genotype-predicted phenotypes, and (iii) healthy volunteers and patients (mainly due to inflammatory processes) [21, 31]. Most influential covariates identified were age, body size descriptors such as total or lean body weight, liver function markers and co-medication [31]. When available, CYP2C19 genotype was included primarily as a covariate, such that voriconazole elimination would decrease in the presence of loss-of-function allele CYP2C19*2 (*1/*2 or *2/*2). Interestingly, only one NLME model reported a significant influence of gain-of-function allele CYP2C19*17 on voriconazole clearance [70].

Kallee et al. identified seven population PK models for voriconazole (without pharmacogenetic information) during a systematic literature review [34]. A total of 66 measured voriconazole plasma concentrations from 33 critically ill patients were available for external evaluation. The results indicated that the precision of all models was low, hence the authors concluded that the use of the models is not recommendable and that it should be investigated if models that incorporate pharmacogenetic information would yield better predictive performance. Comparable to our results, the integration of the first TDM sample improved the predictive performance of all models. Huang et al. identified six candidate models for voriconazole suitable for their external evaluation [33]. The published population PK models differed significantly in prediction performance, and none of the models could adequately predict voriconazole concentrations in the external dataset from patients with hematological malignancy. In accordance with findings from our study, Bayesian simulations showed that predictive performance and accuracy of all models were significantly improved when one or two prior concentrations were available.

In our analysis, differences in the ability to accurately predict and forecast voriconazole concentrations were observed between the three investigated NLME models. There was no model demonstrating overall high predictive performance. Implementation of a linear, non-saturable clearance in models M1 and M2 neglected the reported dose dependency and autoinhibition. Voriconazole is almost exclusively eliminated via hepatic metabolism with less than 2% of the dose excreted unchanged in the urine [71]. The activity of metabolizing enzymes can be highly impacted by polymorphisms or co-medication. In the aggregate model, these mechanisms were accounted for through separation of clearance into two pathways: a constant linear part (representing CYP3A4, CYP2C9 and FMO-mediated metabolism and renal excretion of unchanged drug) and a concentration- and time-dependent nonlinear part (Michaelis-Menten, where V_max is decreasing over time due to CYP2C19-mediated autoinhibition). In a sensitivity analysis, McDougall et al. showed that inclusion of nonlinear clearance was the most important factor for utility of the aggregate model in MIPD [72]. The aggregate model was developed specifically for personalized dosing, and we confirmed the overall best predictive performance. Yet, particularly individual predictions for C_min, were highly variable and forecasted with low precision. All investigated models tended to overestimate observed concentrations, which could lead in the clinical setting to the design of suboptimal dosing regimens resulting in higher risk of treatment failure but lower risk of toxicity. In their recent publications, Wicha et al. identified underlying population size, sampling design, and accurate documentation of dosing and sampling time for model development as main drivers for differences in predictive performance [2, 44, 73].

To our knowledge, the so-called ‘hybrid’ model approach by McDougall et al. has not yet been systematically evaluated or implemented for other compounds. Other approaches addressing the challenge of selecting the most accurate NLME model as prior for MIPD have been described in the literature. Uster et al. derived an automated model averaging/selection approach to find the best model or combination of models for each patient using vancomycin as a case study [74]. Guo et al. showed that adaptive MAP estimation (which handles data over multiple iterations) outperformed standard MAP-based Bayesian forecasting in predictive performance [75]. Recently, Bayesian data assimilation and reinforcement learning methods were also proposed for MIPD, overcoming major limitations of MAP-based approaches by enabling accurate uncertainty quantification and propagation, and complex patient state/dose combinations, respectively [76‐78]. A sequential Bayesian hierarchical modelling framework for continued learning across patients during MIPD (i.e., updating the population parameters using posterior samples) was introduced by Maier et al. using the example of neutrophil-guided dosing of paclitaxel [79]. Hughes et al. demonstrated how a continuous learning approach could be used to adapt a population model to a local environment with a case study in pediatric dosing of vancomycin [80].

To assess the influence of prior observations on model predictivity, different strategies within a Bayesian forecasting setting were investigated. Accuracy and precision of model predictions improved substantially by including voriconazole concentrations using Bayesian forecasting compared to a priori model predictions. In line with expectations, the most recent voriconazole concentration was found to be the most informative. In our analysis, the “all data” method was used, which bases each forecast on the original prior updated with all concentration data for the individual as it becomes available. Finally, we illustrated an MIPD workflow and demonstrated the clinical impact of Bayesian forecasting for voriconazole (Fig. 6).

Our investigation was primarily limited by the available clinical data, which came mainly from healthy volunteers from two study centers. Due to the very diverse study designs and inclusion of clinical patient data (22.0%), a comprehensive evaluation of voriconazole was still feasible. However, it should be noted that the ‘predictive performance evaluation’ and the ‘Bayesian forecasting’ subsets of the database comprised monocentric data of 31 patients and 8 healthy volunteers, respectively. In some instances, interpretation of the published NLME models was complicated when methodologies were not exhaustively described, and details were missing in the publications. Publication of the full variance-covariance matrix, complete model code, or output files (e.g., '*.lst' file for NONMEM^®) as supplementary materials is highly recommended. In addition, sponsors are highly encouraged to publish the model that was used to analyze the data from pivotal clinical trials (i.e., for the submission), including the clinical trial characteristics.

Despite these limitations, our analysis demonstrates the importance of externally evaluating NLME models before application in the context of MIPD. Our analysis identified a comparatively better model, while confirming the main factors explaining the high interindividual variability. Due to the complex PK of voriconazole, a comprehensive database pooling in vitro and in vivo data from different sources and across multiple clinical studies is needed for development of a more generalizable NLME model. Amalgamating knowledge from translational bottom-up and top-down modelling is a promising approach (‘middle-out’) to elucidate complex PK and to enable development of semi-mechanistic NLME models [22, 81], and has been already successfully applied for other compounds with complex PK, e.g., ciprofloxacin [82, 83], propofol [84], and hydrocortisone [85]. Further inclusion of metabolite data (voriconazole N-oxide) in NLME model development and MIPD can improve the assessment of in vivo CYPC19 activity and thereby prediction of exposure during Bayesian forecasting, especially if the CYP2C19 genotype of a patient is unknown [86]. The MIPD framework for voriconazole should be expanded in the future by inclusion of saliva PK, pathogen susceptibility and relevant PD markers (specific biomarkers, e.g., galactomannan, or non-specific biomarkers, e.g., C-reactive protein) [87‐90].

Carefully designed clinical trials are needed to evaluate the clinical benefit of MIPD [2, 91, 92].

This work presents recommendations for evidence-based model development and application of MIPD for compounds with complex PK, illustrated by voriconazole. An aggregate model provided comparatively better predictive performance than two selected ‘standard’ NLME models using external data from a rich in-house clinical database. Ultimately, with further elucidation of the complex PK of voriconazole [21, 22], more (semi-)mechanistic or ’middle-out’ modelling approaches are warranted for successful implementation of compounds with complex PK in MIPD. Predictive performance of MAP-based MIPD improved using even a single individual concentration, thus at least one PK sample should be obtained. The results of our assessment can guide future implementation of MIPD for voriconazole to improve individual dosing strategies, enhance clinical efficacy, reduce the risk for adverse events, and prevent the spread of antifungal resistance [93].