To reduce methodological bias in PBPK input parameters, this work mostly used permeability, solubility, surface pH, and dissolution data generated using pre-defined methods by this working group. However, if literature data used comparable methods, values were not re-measured. This information is indicated in Table I.
A comprehensive compound list with clinically observed PK changes in the presence of food was collated through a detailed literature search and curation (Supplementary Table 1). The information collected included the outcome of a food effect study, the compound type (i.e., acid, base, or ampholyte), and proposed mechanism of food effect. To focus on absorption-related mechanisms of food effect and to reduce variability in modeling where there is low confidence in the disposition of a given compound, all compounds lacking clinical intravenous (IV) PK data or population PK-based data, as well as compounds with high hepatic extraction, were excluded from this list. Furthermore, prodrugs and compounds whose absorption is known to be limited by active transport were excluded, though these compounds are not expected to make up a large subset of clinical compounds displaying food effect (Supplementary Table 1). The compound list was subsequently refined to 30 compounds for final modeling and analysis while ensuring equal distribution of compound, BCS, and food effect type (Table I, Fig. 1). Food effect (FE) type was defined based on AUC and/or Cmax ratios of fed to fasted using BE criteria (i.e., within 0.8–1.25). FE definitions were based on the drug label and set to positive if the ratio of fed > fasted, negative if ratio fasted > fed, or none if no significant change in AUC and Cmax with food.
Permeability Measurement in MDCK Cells
Wild-type Madin-Darby canine kidney (MDCK-WT) cell line was obtained from NKI (Amsterdam, The Netherlands) and modified to knockdown endogenous canine P-glycoprotein (P-gp). Permeability through a cell monolayer was determined with a Transwell™ system. Cells were plated on the apical side of 96-well Transwell plates 4–7 days prior to the experiment and were cultured at 37°C under a 5% CO2 atmosphere. All compounds were dissolved in Hanks’ balanced salt solution (HBSS) plus 80 mM Lucifer Yellow (LuY) and 10 μM cyclosporin A, which was added to the apical wells. The corresponding receiver (basolateral) wells were filled with HBSS plus 10 μM cyclosporin A, a P-gp Inhibitor.
Permeation rates of the compounds, including reference compounds, were measured in the apical-to-basolateral (AB) direction. The donor and receiver wells were sampled immediately after application of the compound to the donor well to determine baseline concentrations, and again after 1 h. Quantification was done using high-performance liquid chromatography combined with mass spectrometry (HPLC-MS/MS) analysis and monolayer integrity was verified by analyzing the receiver samples for LuY fluorescence in a plate reader. Control compounds were run in parallel to test compounds and were used to scale the apparent permeability (Papp × 10−6 cm/s) to an effective human permeability (Peff,man × 10−4 cm/s) using the software’s built-in calibration curve (21).
Solubility Measurement in Aqueous Buffer Solutions and Biorelevant Media
The solubility of the drug substances was determined in biorelevant media as well as in aqueous buffers at different pH. Fasted state simulated gastric fluid (FaSSGF) (37), fasted state simulated intestinal fluid (FaSSIF-V2) (38), and fed state simulated intestinal fluid (FeSSIF-V2) (39) were prepared according to the instructions provided by Biorelevant (Biorelevant.com Ltd., London, UK). Hydrochloric acid pH 2, citrate buffer pH 4, and phosphate buffer pH 7, as well as additional buffer solutions if required, were prepared according to the standard buffer solutions described in the United States Pharmacopeia (USP) [USP 41, buffer solutions, 5748–5749]. For neutral compounds, solubility was determined at pH 2, pH 4, and pH 7. For ionizable compounds, two additional solubility data points, one pH unit above and below the pKa value(s), were collected.
Excess of drug substance was equilibrated in the media on a magnetic stirrer (200 rpm) at 37°C (biorelevant media) and at room temperature (aqueous buffer). The concentration of dissolved drug and the medium pH were determined after 1, 2, 6, and 24 h. The equilibrium solubility was interpreted as the concentration measured after a plateau was reached and was at the latest measured after 24 h. For freely soluble compounds, the extent of solubilization was measured only up to 10 mg/mL.
PBPK Modeling Approach
An aligned decision tree was defined by working group members prior to modeling, as outlined in Fig. 2. In short, PBPK models were built for all compounds in Simcyp V17.1 (Certara, USA, Inc.) and/or GastroPlus V9.5 (Simulations Plus, Inc.). A software comparison was not the aim of this working group. However, if the results of the two software platforms (i.e., model 1 vs. model 2) showed any large discrepancies, this was reported (Table II, footnotes), and where possible, the underlying mechanisms were investigated and described (Table III, Fig. 5). For GastroPlus, individual, population-representative simulations were conducted as best practice. For Simcyp, all simulations were run in the healthy volunteer population using the default system parameters and with the clinical trial design and doses matched to the reported studies. Published and measured values for physicochemical properties, permeability, solubility, and dissolution were utilized as input parameters, respectively, to build mechanistic, bottom-up models for absorption. In order to reduce the uncertainty and variability and narrow the analysis of food effect predictions to absorption-related mechanisms, clearance and disposition were modeled based on published clinical IV PK and/or population PK data. This was not done because IV data is required or recommended by the working group for PBPK model success, but rather to simplify the modeling approach and subsequent analyses, such that model outcomes could be interpreted in the context of absorption parameters only. Furthermore, the focus of this work was to study FE related to absorption mechanisms and not, for example, hepatic first pass or metabolism changes.
The decision tree outlined specific criteria for a model to be considered verified, as well as potential steps to optimize a model where necessary (Fig. 2). The decision tree was followed by all modelers and used to determine the degree of success in predicting food effect. Parameters that were optimized were limited to clearance, precipitation time, and/or permeability. Due to uncertainty in the bio-relevance of in vitro solubility data, solubility was not optimized once the relevant solubility input was evaluated by comparing the different in vitro measured solubility values. In some cases where the decision tree did not lead to a successful model even after optimization, additional steps were taken to optimize the model to enable hypothesis testing; these examples are further discussed in an accompanying manuscript in this issue (e.g., discussion on pazopanib).
PBPK models were developed based on the decision tree outlined in Fig. 2, initially using a bottom-up approach and subsequently using a middle-out approach for cases where verification based on the decision tree criteria was not successful. Compounds were assigned to a pair of modelers with one modeler building the model and the other reviewing it for accuracy and goodness of fit. Success was defined based on visual inspection of the PK profile overlay (i.e., if there was a Tmax or Cmax shift), as well as quantitative assessment of Cmax and AUC ratios (verification range defined in Fig. 1 and described in more detail below).
Model performance was evaluated in the context of the stage of drug development (i.e., purely bottom-up vs. middle-out) using two key criteria: confidence in predicting the likelihood of food effect (i.e., risk assessment) and confidence in predicting the direction and extent of food effect.
The first criterion was assessed using a qualitative yes/no categorization in answer to the question: was the food effect captured correctly in the absence of model optimization with clinical data?
The second criterion was quantitative in nature and involved evaluation of observed versus predicted AUC and Cmax ratios of fasted and fed.
When the bottom-up model could accurately capture the fasted and fed PK parameters and profile within 2-fold of observed, and visual inspection indicated good overlay of the PK profiles without the need for optimization of absorption parameters, the model was considered to have high confidence with respect to bottom-up success (confidence category: high confidence bottom-up). Second, where the bottom-up model could accurately capture the fasted and fed PK parameters and profile within 0.8–1.25 range, but only after optimization of absorption parameter(s) as defined in the decision tree, the model was considered high confidence with respect to middle-out success, e.g., for informing food effect of new formulations and dose strengths (confidence category: high confidence middle-out). Third, where the model could capture the fasted and fed PK parameters and profile following optimization using fasted data, though it fell outside the conservative criteria defined above, but within 2-fold of observed PK parameters, the model was considered to have moderate confidence. Finally, where the model failed to capture the fasted and/or fed PK parameters and profile even after optimization as described in the decision tree, it was categorized as low confidence (Fig. 2). While modeling the latter subset of compounds using a broad, pre-defined decision tree around optimization was not found to be suitable, deviating from the general workflow helped improve the accuracy of some of the models; these examples are captured in an accompanying manuscript in this issue focusing on low confidence predictions.