Abstract
Hydroxychloroquine (HCQ) is a lysosomotropic autophagy inhibitor being used in over 50 clinical trials either alone or in combination with chemotherapy. Pharmacokinetic (PK) and pharmacodynamic (PD) studies with HCQ have shown that drug exposure in the blood does not correlate with autophagy inhibition in either peripheral blood mononuclear cells or tumor tissue. To better explain this PK/PD disconnect, a PBPK was developed for HCQ describing the tissue-specific absorption, distribution, metabolism, and excretion as well as lysosome-specific sequestration. Using physiologic and biochemical parameters derived from literature or obtained experimentally, the model was first developed and validated in mice, and then adapted to simulate human HCQ exposure in whole blood and urine through allometric scaling and species-specific parameter modification. The human model accurately simulated average steady-state concentrations (Css) of those observed in five different HCQ combination clinical trials across seven different doses, which was then expanded by comparison of the Css distribution in a virtual human population at this range of doses. Value of this model lies in its ability to simulate HCQ PK in patients while accounting for PK modification by combination treatment modalities, drug concentrations at the active site in the lysosome under varying pH conditions, and exposure in tissues where toxicity is observed.
Introduction
Hydroxychloroquine (HCQ) is a 4-aminoquinoline class molecule traditionally used as an antimalarial medication, but currently being investigated in over 50 cancer clinical trials either alone or in combination with chemotherapy (ClinicalTrials.gov). HCQ’s putative mechanism of anticancer activity is through the inhibition of autophagy, a naturally occurring metabolic process that involves the digestion of a cell’s own organelles, proteins, and other cellular debris in the lysosome. HCQ is the current gold standard for autophagy inhibition in a clinical setting as it exhibits similar potency yet less toxicity in long-term dosing compared with other clinically approved autophagy inhibitors, such as chloroquine (Shi et al., 2017). There is significant evidence that autophagy is heavily upregulated in certain tumor types, acting as a survival mechanism against both the harsh tumor environment and chemotherapy treatments (Yang et al., 2011; Levy et al., 2017). Inhibition of this process by augmenting the primary treatment method with HCQ has been shown to resensitize tumors to a resistant therapy regimen or enhance response to a current treatment regimen (Amaravadi et al., 2011; Carew et al., 2012; Barnard et al., 2014; Levy et al., 2014).
Mechanistically, HCQ is a weakly basic compound that alkalinizes the highly acidic lysosome, preventing the autophagosome–lysosome fusion step of autophagy (de Duve, 1983; Lu et al., 2017a). This mechanism drives its pharmacokinetics (PK), primarily through an ion-trap accumulation observed in lysosomes and other acidic compartments. Overall, HCQ PK has been well characterized in multiple species, and the PK profile appears to scale linearly with dose escalation, though exhibits a significant amount of intersubject variability (McChesney, 1983; Tett et al., 1988; Furst, 1996; Lim et al., 2009; Fan et al., 2015). Sources of variability are likely due to the fact that these studies have been done by different research groups over the course of decades as well as variance in absorption, distribution, metabolism, and elimination (ADME) factors related to the physicochemical properties that drive its PK.
HCQ is typically administered as an oral tablet, in doses ranging from 100 to 1200 mg daily, by which it is readily absorbed within 2–4 hours (Browning, 2014). The fraction absorbed is estimated to be 74% ± 13% (Tett et al., 1989; Browning, 2014). HCQ blood concentration peaks shortly after the absorption phase and falls relatively quickly due to rapid partitioning into organs. Accumulation in lysosomes appears to drive the large volume of distribution in plasma, whereas binding to melanin contributes to the long terminal half-life (t1/2) (Tett et al., 1990). It is approximately 50% bound to plasma protein in the blood (Furst, 1996). Metabolism appears to be the primary driver of HCQ clearance. It occurs in the liver through CYP3A4- and CYP2C3-driven dealkylation to form desethylhydroxychloroquine, desethylchloroquine, and bisdesethylchloroquine, the former that exhibits therapeutic activity and PK thought to be comparable to HCQ (McChesney et al., 1966; Kim et al., 2003; Browning, 2014; Qu et al., 2015). Rate of metabolism varies tremendously across species, leading to significant cross-species differences in t1/2 that ranges from hours in mice to days in humans (Tett et al., 1988). Excretion takes place primarily in the kidneys, accounting for about 22% of HCQ total blood clearance, with liver clearance assumed to account for the rest (Tett et al., 1988). Mean renal clearance from plasma is reported as three to four times greater than glomerular filtration rate (GFR) corrected for protein binding, suggesting the drug is secreted in addition to filtration (Tett et al., 1988).
In the context of cancer treatment, HCQ blood exposure correlates with neither tumor exposure nor pharmacodynamic (PD) markers of autophagy inhibition in both peripheral blood mononuclear cells and tumor tissue (Barnard et al., 2014). To investigate the PK/PK and PK/PD disconnect observed in HCQ treatment, as well as the high interindividual patient variability in PK studies, a physiologically-based pharmacokinetic model (PBPK) serves as a powerful investigative tool. This model allows for simulation of patient profiles and study of relationships between drivers. Developed around key tissues as well as physiologic and biochemical properties associated with the ADME profile, this PBPK model of HCQ can simulate exposure in tissues associated with toxicity as well as therapeutic effect. It can also investigate the variability in tumor pH that may contribute to the disconnect between exposure in the blood and autophagy inhibition in the tumor, specifically the inverse relationship between HCQ uptake and extracellular pH (Pellegrini et al., 2014).
The PBPK model described for HCQ (Fig. 1) was developed using biochemical and physiologic parameters adapted from literature or obtained experimentally, with significant focus on the pH-based aspect of PK. Tissues represented in the model are involved with absorption, metabolism, excretion, and binding, or are subject to toxicity. Model output simulates single i.p. dose tissue exposure in mice. Biochemical and physiologic parameters were then adjusted accordingly to simulate whole-blood exposure in humans. Simulation output across multiple dosing regimens and across different species matches well with experimental PK data, indicating that key drivers of HCQ PK are accounted for by the PBPK model.
Materials and Methods
PK Study in Mice
Protocols for the mouse studies were approved by the Institutional Animal Care and Use Committee at Colorado State University. Female BALB/c mice were treated with a single i.p. of 20, 40, or 80 mg/kg HCQ (H0915; Sigma-Aldrich, St. Louis, MO). Tissues and whole blood were collected from three mice for each dose at 3, 6, 12, 24, 48, and 72 hours (n = 54). Levels of HCQ and desethyl-hydroxychloroquine (dHCQ) in whole blood and tissues were determined via a previously validated liquid chromatography–tandem mass spectrometry assay (Barnard et al., 2014). Briefly, 100 μl tissue homogenate was added to a microcentrifuge tube along with 10 μl Milli-Q water, 10 μl 50/50 acetonitrile/Milli-Q, 10 μl 2.5 μg/ml CQ (internal standard), and 100 μl acetonitrile. Samples were vortexed for 5 minutes and centrifuged for 10 minutes at 13,300 rpm. Eighty microliters of this mix was transferred to an autosampler vial and mixed with 120 μl Milli-Q water prior to mass spectrometer analysis. Standards and quality controls were prepared in an identical manner. Data points below the LLOQ of 1 ng/ml for HCQ and 10 ng/ml for dHCQ were excluded. Noncompartmental analysis (NCA) measured exposure, as determined by area under the drug concentration versus time curve (AUC0–inf), to both HCQ and dHCQ.
HCQ Microsomal Metabolism Studies
Metabolism rate constants for the human PBPK model were obtained through microsomal incubation. Pooled mixed gender human liver microsomes at 20 mg/ml were obtained from XenoTech (Kansas City, KS) and were split into 100 μl aliquots prior to beginning microsomal incubations. Briefly, microsome stocks were diluted to 0.5 mg/ml in 100 mM phosphate buffer and preincubated with a NADPH-regenerating system for 5 minutes. The 100× HCQ (H0915; Sigma-Aldrich) was spiked into the microsome mix and incubated for up to 3 hours. Reactions were terminated by addition of 100 μl acetonitrile, vortexed for 5 minutes, and centrifuged at 14,000 rpm for 5 minutes to pellet the microsomal protein. Prior to analysis by high-pressure liquid chromatography, the samples were mixed 50:50 with the internal standard of 500 ng/ml camptothecin dissolved in MilliQ water. Rate of metabolism was determined by rate of n-desethylhydroxychloroquine formation, the primary metabolite of HCQ (Browning, 2014). Analysis of samples was done using the liquid chromatography–tandem mass spectrometry assay mentioned in the previous section. Standards and quality control samples were prepared in an inactive microsomal homogenate matrix in the same way as samples and containing identical amounts of internal standard.
Lysosomal Quantification in Mouse Tissue
Lysosome contents of the various tissues incorporated in the model were determined by analyzing acid phosphatase enzyme activity. It was assumed that lysosome volume fraction of the tissue was directly proportional to the acid phosphatase activity in that tissue. Tissues were extracted from BALB/c mice, diluted to 100 mg/ml in 0.2 M acetic acid buffer (pH = 5.0), and homogenized in a Bullet Blender Storm 2400 at max speed for 5 minutes. Samples were centrifuged at 5000 rpm for 10 minutes at 4°C, and homogenate was collected and transferred to new tubes. Homogenate was vortexed, and 150 μl was transferred to a cuvette, followed by 600 μl acetic acid buffer to this same cuvette. The cuvette was then incubated at 37°C for 10 minutes to inactivate additional enzymes. After incubation, 250 μl 32 mM p-nitrophenol dissolved in water was added to the cuvette. Cuvette was inverted and read immediately on a DU 800 spectrophotometer at 420 nm for 5 minutes to collect absorbance data. Absorbance data at the 5-minute time point were normalized to the total protein content of the homogenate, as determined by comparison with a bovine serum albumin standard curve. Acid phosphatase enzyme activity was reported as moles/minute per milligram protein, and values for each organ were compared with a fixed fraction for kidney (Sewell et al., 1986; Kawashima et al., 1998).
PBPK Model Development
Base Model Development.
The PBPK model for HCQ is represented by eight distinct, flow-limited tissue groupings that could describe HCQ PK based on physiologically-derived metabolism, renal elimination, absorption, macromolecule binding, and lysosomal sequestration. Key features that the model aimed to capture were melanin binding to investigate cases of retinal toxicity, and lysosomal uptake kinetics to model intracellular PK at the active site. Figure 2 depicts a schematic describing the system of these key tissues involved in HCQ PK.
Typically administered orally, HCQ absorption occurs rapidly in the gut lumen, followed by transport to the liver, where it undergoes first-pass metabolism. From there it is distributed to the rest of the body. Key tissues include the kidney, the site of renal excretion, as well as eyes and skin, major sites of melanin binding, and the heart, due to observed instances of cardiomyopathy. The rest of the body is lumped into the slowly perfused compartment, consisting of bone, fat, and muscle, as well as the rapidly perfused compartment, which is composed of the remaining viscera. Additionally, a tumor compartment can be incorporated depending on the location of the primary tumor. Each compartment of the model is further described by a lysosomal subcompartment discussed in the next subsection.
Mass balance of a typical tissue observed in the bulk portion of the model is described by eq. 1(1)where mt is the amount of HCQ within the tissue, Qt is the blood flow rate to that tissue, Ca is the unbound arterial plasma concentration carried to the tissue, Ct is the free concentration within the compartment, Pt is the partition coefficient of that tissue, At is the lysosomal surface area, and ΔJlys is the net flux between the cytosol and lysosome. The only tissue with an exception to eq. 1 is the liver, wherein arterial concentration is set to include free drug as well as bound drug due to dissociation of HCQ and protein occurring in the hepatic space (Meijer and van der Sluijs, 1989).
For tissues involved in clearance of HCQ, an additional term is added to eq. 1. For kidney, a renal clearance term is added and is represented by eq. 2. In the case of liver, a metabolism term is added and is represented by eq. 3 for mouse and eq. 4 for human.(2)(3)(4)The first term of eq. 2 accounts for renal filtration in which GFR represents glomerular filtration rate, QK represents blood flow to the kidney, and CA represents arterial blood concentration. The second term represents active secretion in which VK represents kidney volume, and Vmax and Km represent Michaelis–Menten constants for secretion. Equation 3 describes the amount of HCQ metabolized in mouse and is represented by VL and CL, the volume and concentration of liver, respectively. It is also represented by Vmax and Km, Michaelis–Menten rate constants for metabolism. Equation 4 describes the amount of HCQ metabolized in human and is represented by the linear rate constant Met. Melanin binding is the other driving factor of distribution in the model, and so the amount of free HCQ unbound to melanin, applying to skin and eyes, is represented by eq. 5.(5)where mfree is the amount of unbound drug in the cytosol; mcyt is the total amount in the cytoplasm; Tmel, a function of melanin/HCQ-binding ratio (MelBR) and melanin concentration in the tissue (Ctmel), is the binding capacity of melanin for HCQ; Kmel is the binding affinity of HCQ to melanin; and Ct is the free concentration in the cytosol. Values for Tmel and Kmel were estimated from a previous study on HCQ binding (Schroeder and Gerber, 2014), and melanin concentrations for skin and eye have been reported (Durairaj et al., 2012; Browning, 2014). The parameters for the PBPK model can be found in Table 1 along with the source of each variable.
ADME Parameters.
Tissue volumes and blood flows were standard, fixed values, as previously described (Brown et al., 1997). The only exception was eye, which came from multiple sources (Choi et al., 2012; Zhi et al., 2012).
Absorption parameters were obtained from a prior study (Tett et al., 1989). Specifically, the absorption rate constant was modeled as linear uptake that achieved near-full absorption within an average of 3 hours. Fraction absorbed was set to an average of 0.75 ± 0.13 of the dose, as described by human urine and blood data. First-pass hepatic effect was deemed insignificant in this study as well, and was calculated at only 6% of the total dose. Biliary secretion was initially considered, as fecal excretion of HCQ in rat was cited as 25% of total dose (McChesney, 1983); however, ratio of metabolite versus parent was not observed. Biliary secretion was tested in the model, assuming complete absorption, and was representative of a model lacking biliary clearance with 0.85–0.9 fraction absorbed. As this fell within the range of absorption fraction observed in humans, biliary secretion was not considered a major driving factor of HCQ PK and was omitted from the model.
Distribution of the bulk model was primarily described by partition coefficients, the concentration ratio between plasma and tissue at steady state. The partition coefficient values were derived and optimized using previously published data (McChesney et al., 1967; Tett et al., 1989; Wei et al., 1995).
Metabolism values for mouse were obtained and optimized from liver PK data by determining the Michaelis–Menten relationship that represented the compartmental clearance of this region. The human metabolism rate constant was obtained by incubating HCQ with activated human liver microsomes, as described in the microsome section of Materials and Methods, and optimized to best fit human PK data. Intrinsic scaling of in vitro to in vivo metabolism for CYP450 enzymes, as previously described (Chiba et al., 2009), was not considered because moderate optimization of the microsome-derived metabolism rate yielded clearance that fit the clinical data well, and satisfied the hepatic clearance portion observed in human patients (Tett et al., 1988). Intestinal metabolism was considered negligible due to the significantly lower CYP3A and 2C levels observed in intestine compared with liver as well as low first-pass metabolism (Thelen and Dressman, 2009).
Excretion was modeled as filtration and secretion because mean renal clearance was found to be 3 to 4 times greater than GFR (Tett et al., 1988, 1989, 1990). GFR was set to 0.125 for mouse (Qi and Breyer, 2009) and 0.11 for human of the unbound plasma flow to the renal tubules. The Km and Vmax for renal secretion were optimized using the Nelder–Mead method to best fit the data, accounting for 3 to 4 times GFR as secretion as well as the ∼3:1 ratio of nonrenal clearance (assumed to be liver clearance) to total renal clearance upon scaling from mouse to human (Tett et al., 1990; Furst, 1996).
Lysosomal Compartment Development.
Within each tissue compartment was a lysosomal subcompartment that was mathematically described as separate from its parent compartment. The subcompartment was modeled to represent the ion trap effect that causes HCQ accumulation to millimolar concentrations in acidic regions, primarily the lysosome. A schematic of lysosomal sequestration is represented in Fig. 3, which depicts HCQ and HCQ+ moving freely across membranes at physiologic pH, but, upon entering the acidic lysosome, a majority of HCQ is in the 2+ state, where it can no longer permeate the membrane and becomes trapped.
The mass balance of the lysosomal subcompartment for each tissue is represented by the second term in eq. 1 and is described by eq. 6:(6)where At represents the lysosomal surface area and ΔJlys represents the net flux between the cytosol and lysosome compartments. The flux portion was adapted from the cell model proposed by Trapp et al. (2008) that describes lysosomal drug uptake on a cellular pharmacokinetic level. It was validated using a select handful of drugs with a variety of pKa and lipophilicity values, one of which was chloroquine (Trapp and Horobin, 2005; Trapp et al., 2008). The overarching principle in their cell model was application of Fick’s first law of diffusion with the Nernst–Planck equation, which incorporates electric charge into the movement of molecules through a biologic environment. Fick’s first Law of Diffusion, which only applies to the neutral form of HCQ (Jn), is depicted by eq. 7, and the combined Fick’s–Nernst–Planck version of diffusion (Jd) is eq. 8. These equations are combined to yield eq. 9:(7)(8)(9)where fn represents the fraction of freely dissolved neutral HCQ, described by eq. 10 [or reference eq. 6 from Trapp et al. (2008) for a more derived description]. Pn and Pd represent the permeabilities of neutral and ionic HCQ states, Dz is the Henderson–Hasselbach activity ratio, C is the total HCQ concentration in the compartment (applying to either cytosol or lysosome), and N is the Nernst equation. Equations 7 and 8 represent HCQ flux in the neutral and ionic states and are combined to give the net flux between both compartments in eq. 9. The fraction of freely dissolved HCQ in this equation is represented by fn and is described by eq. 10:(10)which represents the fraction of HCQ in the system that is available to move. This description has been well described previously (Trapp et al., 2008), but briefly it represents the fraction available to move by taking into account lipid binding, sorption, and ionic activity. W represents water fraction in the compartment, L represents lipid fraction in the compartment, Kn and Kz are sorption coefficients of neutral and ionic HCQ, and γ represent the ionic activity coefficients for the compartment.
The effect of charge on lipophilicity of a molecule, which sets the framework for the ionic sorption coefficents (Kz) and ionic permeability, has been previously described (Trapp et al., 2008). The equation for charged lipophilicity is shown in eq. 11:(11)where z is the charge HCQ. Equation 11 is applied to the permeability (eq. 12) and sorption (eq. 13) to represent these properties of charged HCQ:(12)(13)Additionally, eq. 8 incorporates the Henderson–Hasselbach activity ratio, Dz, which represents the ratio between neutral and ionic fractions of the molecule. Dz is represented by eq. 14 for the 1+ state and eq. 15 for the 2+ state.(14)(15)All values of physicochemical properties of HCQ and lysosomal properties used in equations related to this section of the model can be found in Table 2 along with the source of these values (Kawashima et al., 1998; Ishizaki et al., 2000; Warhurst et al., 2003; Trapp et al., 2008).
An additional component to consider at the cellular scale is the effect of lysosomal HCQ accumulation on pH. Previous studies have demonstrated the pH neutralization effect of weak base accumulation in lysosomes (Ohkuma and Poole, 1978), and this has been taken into account in more recently developed cellular models (Ishizaki et al., 2000; Kornhuber et al., 2010). Equation 16 was used to describe the dynamic pH of the lysosome compartment:(16)where Clys is the free lysosomal concentration and β is the buffering capacity, as previously described (Ishizaki et al., 2000). Dynamic pH affects the free fraction of HCQ in each ionic state, which causes net flux to approach an equilibrium as lysosome pH becomes more neutral. Dynamic pH based on buffering capacity was not considered for the cytosolic compartment. Initial pH values for lysosome and cytosol compartments were the same for all tissues and were set to 5.0 and 7.2, respectively.
Volume of the lysosome compartment for each tissue was determined by total acid phosphatase activity in the tissue. Acid phosphatase activity was assumed to correlate directly with total lysosomal volume fraction of the tissue. Acid phosphatase activity/lysosome volume fraction was compared between tissues on a ratio basis using a set value for kidney lysosome volume fraction, and optimized to fit the mouse model (Sewell et al., 1986; Kawashima et al., 1998). Lysosome size was set to a constant spherical volume using a radius of 275 nm.
Data Analysis
The ability of the model to accurately predict concentration–time profiles was analyzed by comparing PBPK simulation pharmacokinetic variables with those of actual data, as well as through calculation of the median absolute performance error (MAPE%) and the median performance error (MPE%). The performance error (PE) of the model was calculated as the difference between the measured and simulated concentrations normalized to the simulated concentration, as shown in eq. 17:(17)The MAPE%, which measures prediction accuracy, was calculated by eq. 18:(18)where N represents the total number of samples for the given tissue. The MPE%, which is a measure of positive or negative prediction bias, was calculated by eq. 19:
(19)Computer Simulation and Software
PBPK model development and simulation were done in Advanced Continuous Simulation Language Xtreme (acslX), version 3.1.4.2, from Aegis Technologies Group (Huntsville, AL). Pharmacokinetic evaluation of simulated and clinical data was done on Phoenix 64 build 7.0.0.2535.
Results
HCQ Model Simulations in Mice.
PK data from mice were generated following a single i.p. dose of 20, 40, or 80 mg/kg HCQ. Tissues and whole blood were collected from mice at 3, 6, 12, 24, 48, and 72 hours after dosing. The PBPK model was developed from the following tissues in this study: whole blood, liver, kidney, and gut. Figure 4 shows the HCQ PK profile from these mice compared with PBPK-generated simulation data. In addition to whole-tissue PK, the contributions of lysosomal subcomponents were examined and are found in Supplemental Fig. 1. Lysosomal HCQ PK profiles were similar between tissues, and simulated concentrations peaked between 20 and 40 mM, which are consistent with lysosomal concentrations of lysosomotropic agents reported in previous studies (Browning, 2014).
The ability of the model to accurately predict the concentration–time profile of HCQ was determined through calculation of MAPE% and MPE% presented as a percentage for tissues and whole blood, shown in Table 3. MAPE% is a method of determining model accuracy, whereas MPE% indicates under- or overprediction bias (Sheiner and Beal, 1981). Overall, MAPE% for each dosing cohort were within similar ranges. The ranges were 27%–73%, 19%–71%, and 24%–44% for 20, 40, and 80 mg/kg, respectively. Upper bounds of this range were skewed by the deviation of model simulation from actual PK data for liver after 24 hours due to the concentration–time profile for HCQ PK dropping to zero after 24 hours. The model prediction follows the actual data well for the first 24 hours, after which it predicts a slower decline at these later time points. Removing the liver MAPE% drops the ranges to 27%–39%, 19%–27%, and 24%–37%. Comparing this range of MAPE% with the variation in actual PK data, determined by the range of average coefficient of variation for each dosing cohort, is a strong indicator of whether model prediction is within the dispersion of the actual data. Coefficient of variation values for the PK cohorts are 16%–43%, 9%–20%, and 8%–28% for 20, 40, and 80 mg/kg, respectively, which are quite similar to the MAPE% for the respective cohorts. This suggests that the model accounts for a majority of natural variation within the clinical data. Taking into account MPE%, the model tends to slightly overpredict blood and liver concentrations in each dose cohort. Kidney and gut are overpredicted in the 20 mg/kg cohort, but slightly underpredicted in the 40 and 80 mg/kg mice, indicating no significant bias in over- or underprediction of the model. An additional method of determining the model’s predictive capability was through comparison of simulation versus actual pharmacokinetic variables generated through NCA, which are presented in Table 4.
The NCA variables compared for blood and each tissue were terminal t1/2 and area under the curve (AUC). Values of each were determined over the 72-hour interval and compared ratiometrically (actual/simulated). Overall, the AUC ratios were very close to one, indicating a high degree of accuracy within the model. The t1/2 were also relatively close, except for liver 20 and 40 mg/kg cohorts, which, as noted with MAPE%, are due to the lack of a terminal phase in the actual liver data. Although t1/2 ratios were low for each dosing cohort, the ratios near one for AUC indicate that the model predicts overall liver exposure with a high degree of accuracy.
HCQ Model Simulations in Humans.
After validating the simulation in mice, the model was scaled to human patients by adjusting appropriate physiologic parameters, which are compared in Table 1. Concentration–time data for individual humans were used (Tett et al., 1988, 1989) and included both oral and i.v. infusion doses of 200 and 400 mg HCQ sulfate. Whole-blood data were available for two patients (patients 1 and 4) from this study, and cumulative urinary excretion of HCQ was available for another (patient 5). The concentration–time profile between simulation and actual data was compared over a 72-hour time period and is presented in Fig. 5. In addition to whole-blood PK, tissue concentrations were simulated following these dosing schemes as well and can be found in Supplemental Fig. 2. Simulated PK profiles for each of the dosing routes were notably different from each other, specifically in that liver and gut peak concentrations were approximately 2 to 3 times higher, and kidney much lower, in patients receiving oral versus i.v. infusion dosing due to the absorption phase. Lysosomal PK for these human tissues was also simulated following these dosing schemes and can be found in Supplemental Fig. 3. Peak concentrations are similar to those in mice, at 20–30 mM in liver and gut, and as low as 6 mM in kidney following 200 mg oral dosing.
As individual physiologic data were not available for the human patients, generic human physiologic parameters were assumed (see Table 1) for patients 4 and 5 (Brown et al., 1997). After fitting the model to these patients through optimization of absorption, metabolism, and renal clearance parameters, the model was fit to patient 1 by simply increasing body weight from 70 to 85 kg. For the human model, the MAPE% range, presented in Table 3, for whole blood was between 19% and 42%, and for urine was between 16% and 21%, indicating strong predictive power of the model. As this was individual patient data, the coefficient of variation was not reported for comparison with these patients. MPE% indicated a slight underprediction of the actual PK profile for all of those compared except patient 4 i.v. infusion, which was a slight overprediction.
NCA parameters were also compared for human whole-blood data and can be found in Table 4. Overall, the actual/simulation ratios of t1/2 and AUC indicated strong predictive power of the model.
To investigate the toxicity-related aspect of HCQ regarding retinopathy, HCQ exposure was simulated in human eyes at 200, 600, and 1200 mg/d oral dosing and can be found in Supplemental Fig. 4. Single-dose HCQ at these levels generate relatively low concentrations in the eye relative to the other tissues shown in Supplemental Fig. 2, although eyes continue to retain HCQ as blood levels decline. After 30 days of once-daily dosing, it is clear that 1200 mg/d causes significantly higher accumulation of HCQ within the eyes, approximately 10-fold more than 200 mg/d. It is of note that even 2 months after dosing is discontinued the concentration of HCQ in the eyes decreases only slightly, indicating an extremely long t1/2. Sixty days of once-daily dosing show a similar trend as 30 days, although concentration buildup begins to taper off. Concentrations in the human eyes at 1200 mg/d are close to those observed in the eyes of hooded rats dosed at 40 mg/kg for 6 days/wk over the same time frame (McChesney, 1983).
Simulated Steady-State Concentration and Clinical Trial Comparison.
After the model was validated for single-dose HCQ in human patients, the dosing interval was extended out to investigate predictive capability over a much longer period of time. Using generic human PBPK parameters as listed in Table 1, the model was used to simulate once-daily oral dosing of 100, 200, 400, 600, 800, 1000, and 1200 mg HCQ for a 30-day period. The steady-state concentration (Css) was taken as the average whole-blood concentration occurring after 20 days. Simulated Css for each dosing regimen was compared with the patient Css distribution of the corresponding dosing regimen in five different cancer clinical trials using HCQ in combination with another treatment in Fig. 6 (Mahalingam et al., 2014; Rangwala et al., 2014a,b; Rosenfeld et al., 2014; Vogl et al., 2014).
The trials reported reaching the concentration distribution depicted by the following days: temsirolimus at 15 days; vorinostat at 20 days; bortezomib at 21 days; temozolomide glioblastoma did “not distinctly represent steady state” and the time was not reported; and temozolomide solid tumor/melanoma at 16 days. The depicted range in all trials represents concentrations falling within the second and third quartile of patient data, except for vorinostat, which represents the mean ± S.D. Of all trials examined, only the temozolomide glioblastoma patient population had a regression line statistically significant from the PBPK simulation. The regression line for this trial had a mean slope of 3.3 with 95% confidence interval bounds of 2.3–4.3, compared with the regression line of the PBPK simulation that had a mean slope of 2.2. The vorinostat patient population did not undergo regression analysis due to only two daily doses of HCQ (400 and 600 mg) used in the trial, but the slope of this line was 1.7 compared with the PBPK simulation of 2.2. Of all trials tested against the PBPK simulation, average Css from the model human fell within the bounds of the second and third quartile for all dosing regimens except for the temozolomide glioblastoma 800 mg and temsirolimus 400 mg.
Simulation of HCQ in a Virtual Population.
Once model-simulated Css was found to be in good agreement with clinical trial Css distributions, the model was tested in a virtual population of humans. Two hundred and fifty males and 250 females were generated from the physiologic parameters for PBPK-modeling (P3M) database using PopGen (McNally et al., 2015). Briefly, P3M is a massive database of US patient records containing their associated physiologic data, and was developed by the Lifeline Group (Annandale, VA). Using PopGen to extract from this database, a virtual population of individuals was generated using available demographic data from the bortezomib/HCQ combination phase I clinical trial (Vogl et al., 2014). This was assumed to generate a population representative of a general cancer clinical trial. Css was then simulated for this population at each dosing regimen, and the simulated distribution was compared with the actual distribution observed in the clinical trials referenced. Each dose from 100 to 1200 mg dosed once daily was simulated in each of the 500 total patients to generate a realistic range of Css. The results of this simulated patients compared with the actual distribution from each trial are shown in Fig. 7.
The model captured the distribution of whole-blood Css in the simulated patient population extremely well at low doses. The higher doses in the clinical trial population exhibited much more distributional variability between trials. This was expected due to the more prominent effect that variability in physiologic parameters will have at higher doses. Additionally, the clinical trial patient populations were relatively small, especially in comparison with the 500 virtual patients simulated by the model, which is another likely contributor to the distributional variability between trials.
Discussion
PBPK modeling is a relatively well-established concept, dating back to the first modern model of methotrexate in 1971 (Bischoff et al., 1971). Cancer therapy is a major field for development of these models due to the ability to simulate multiple dosing regimens in a relatively short period of time with the purpose of maintaining therapeutic efficacy while minimizing toxicity. This includes PBPK models for some drugs that have been used in clinical trials in combination with HCQ, including three of those discussed in this work (Gustafson et al., 2002; Ballesta et al., 2014; Zhang and Mager, 2015). Although the utility of PBPK models appears increasingly clear, their use in a clinical setting is not well established. With treatment including, but not limited to malaria, lupus, rheumatoid arthritis, and now cancer, HCQ is a prime candidate for use of PBPK models in a personalized-medicine setting. Due to the widespread use of this drug, accounting for sources of PK variability from easily monitored driving parameters, such as gastrointestinal absorption, CYP3A4 and CYP2C3 metabolism, and renal clearance, could be a quick method of accounting for the broad range of patient PK profiles observed in a clinical setting. This would concurrently improve prediction of exposure in areas that are more difficult to monitor, such as melanin binding in the eye or variation in lysosomal content of a target site.
The described PBPK model for HCQ successfully predicts blood and tissue disposition in both mice and humans in single-dose oral, i.p., and i.v. scenarios. More so, it is capable of establishing a population distribution at doses ranging from 100 to 1200 mg for long-term daily oral dosing in a human population representative of a general cancer clinical trial. This model was developed based on flow-limited uptake for the bulk compartments with perfusion-based uptake for the lysosomal compartment and accounts for specific binding to melanin and nonspecific binding to plasma protein. It is built based off physiologic and biochemically relevant parameters for eight compartments, three of which are validated by clinical PK data. These validated compartments consist of gut, liver, and kidney, along with whole blood. Gut and liver were included due to absorption and metabolism, respectively. Kidney is involved in renal clearance through filtration and secretion. There is evidence of reabsorption, but that mechanism was not considered in this model due to the lack of clinical data available relating urine pH to renal clearance of HCQ. Melanin binding was the primary reason for including eyes, specifically in regard to retinopathy observed in higher doses of HCQ used in cancer clinical trials. Accumulation of HCQ in the eyes was simulated to investigate the difference that dosing levels may have on retinal exposure to this drug. Although actual concentrations associated with retinopathy are unknown, patients classified as high risk are categorized as those receiving >6.5 mg/kg per day for >5 years (Pandya et al., 2015). The model can be used to simulate and avoid HCQ retinal accumulation thresholds observed within these high-risk dosing levels. Heart was included due to rare incidence of irreversible cardiomyopathy that could possibly be exacerbated during long-term dosing of combination therapies that also exhibit some level of cardiovascular toxicity (Al-Bari, 2015). Treatments associated with cardiomyopathy are extremely rare, with only 42 known cases as of 2014 and reported mean cumulative doses of 1843 g (Yogasundaram et al., 2014). Remaining tissues were grouped into the rapidly perfused compartment, consisting of remaining viscera, and the slowly perfused compartment that consists of bone, muscle, and adipose.
The model was initially developed in mice and began with gathering of relevant biochemical and physiologic parameters. After this step, parameter effect on model fit to actual data was assessed. This was followed by optimization to maximize model accuracy across dosing regimens. Scaling to humans involved modifying relevant physiologic parameters and biochemical parameters, as seen in Table 1, and optimizing the model to fit human data. Validating against multiple types of human PK data offers the predictive power to simulate HCQ PK in patients while taking into account PK modification by combination treatment modalities, drug concentrations at the active site in the lysosome under varying pH conditions, and exposure in tissues where toxicity is observed. Interindividual variability in response is one of the primary concerns in treatments with HCQ. From a PK perspective, of all parameters the ones that the blood portion of the model are most dependent on are body weight, rate of metabolism, and blood partitioning and protein binding. Contribution of these factors to model performance is not dose dependent, i.e., linear with dose escalation. Patients with impaired liver or kidney function would likely be exposed to higher concentrations of HCQ, which would need to be accounted for in application of this model.
Primary value of this model is found through accurate simulation of HCQ PK in mice and humans while taking into account exposure in tissues, especially those where toxicity or clinical efficacy may be observed. Mice were the species chosen to begin model development due to the option of PK/PD investigation in whole organs. Between the mouse PK data, PD data (not yet published), and the likelihood that mice will be used in future preclinical studies with HCQ or next-generation analogs, this model will prove a valuable tool to make connections between exposure and response in a preclinical setting. This is especially relevant in the context of lysosomal uptake and the effect of these various factors on the PK/PD of HCQ and next-generation lysosomotropic autophagy inhibitors. Being a major driver of HCQ PK, it is important to note that lysosomes are a dynamic system undergoing effects such as swelling, biogenesis, pH modulation, and turnover that is not well characterized from a quantitative perspective in the presence of HCQ (Zhitomirsky and Assaraf, 2014; Lu et al., 2017). As pH is the driver for lysosomal accumulation, it plays an important role, especially in a tumor context where acidic conditions can dramatically affect HCQ uptake (Pellegrini et al., 2014; Zhitomirsky and Assaraf, 2015). The cellular model of lysosomal uptake applied in our PBPK model (Trapp and Horobin, 2005; Trapp et al., 2008; Kornhuber et al., 2010) serves as a strong base of modification to begin describing whole-tissue PK based on the pH-lysosome relationship. Further advancement of this PBPK model will require stronger prediction power in a tumor setting. Achieving this involves further characterization of these natural and HCQ-influenced lysosomal dynamics, and how they are related to intra- and extracellular pH in a tumor setting.
In addition to its use as a tool for further characterization of autophagy dependence and preclinical study, this model has proven its utility in predicting exposure in human patients as well. The model was validated in humans by accurately simulating blood exposure and urinary clearance of HCQ over a 72-hour period following oral and i.v. administration of two different doses. Additionally, it captured the distribution of Css at seven different dosing levels in a larger population of patients. Overall, it is able to simulate both early time points and steady-state values of patients at all concentrations used in a cancer clinical trial setting. This validation, in conjunction with the physiologic nature of the model, makes it a powerful tool in predicting patient exposure in the 50+ clinical trials currently administering HCQ.
Acknowledgments
We thank Dr. Andrew Thorburn at the University of Colorado Anschutz Medical Campus for guidance with autophagy and hydroxychloroquine research, and also Dr. Stefan Trapp at the Technical University of Denmark for assistance with the lysosomal component of this model.
Authorship Contributions
Participated in research design: Collins, Jackson, Gustafson.
Conducted experiments: Collins, Jackson.
Contributed new reagents or analytic tools: Collins.
Performed data analysis: Collins.
Wrote or contributed to the writing of the manuscript: Collins, Gustafson.
Footnotes
- Received October 10, 2017.
- Accepted February 6, 2018.
This work was supported by National Institutes of Health National Cancer Institute [Grant R01CA190170 Therapeutic Targeting of Autophagy-Dependent Cancer] and by Department of Defense Congressionally Directed Medical Research Program (CDRMP) Breast Cancer Research Program [Grant BC130103P1 Identifying and Targeting Autophagy Dependence to Eliminate Metastatic Breast Cancer]. Support for animal studies was provided by the University of Colorado Cancer Center Pharmacology Shared Resource and the associated Cancer Center Support [Grant P30CA046934].
Part of this work was presented as follows: Collins KP, Jackson KM, and Gustafson DL (2017) A physiologically-based pharmacokinetic (PBPK) model of hydroxychloroquine [1503]. American Association of Cancer Research 77 (Suppl 13).
↵This article has supplemental material available at jpet.aspetjournals.org.
Abbreviations
- ADME
- absorption, distribution, metabolism, and elimination
- AUC
- area under the curve
- Css
- steady-state concentration
- dHCQ
- desethyl-hydroxychloroquine
- GFR
- glomerular filtration rate
- HCQ
- hydroxychloroquine
- MAPE%
- median absolute performance error
- MPE%
- median performance error
- NCA
- noncompartmental analysis
- PBPK
- physiologically-based pharmacokinetic modeling
- PD
- pharmacodynamics
- PE
- performance error
- PK
- pharmacokinetics
- t1/2
- half-life
- Copyright © 2018 by The American Society for Pharmacology and Experimental Therapeutics