QUANTITATIVE CLINICAL PHARMACOLOGY
`
`Model-Based Development of a PPARγγ Agonist,
`Rivoglitazone, to Aid Dose Selection and
`Optimize Clinical Trial Designs
`
`Shashank Rohatagi, PhD, Timothy J. Carrothers, ScD, JinYan Jin, PhD,
`William J. Jusko, PhD, Tatiana Khariton, PhD, Joseph Walker, PharmD,
`Kenneth Truitt, MD, and Daniel E. Salazar, PhD
`
`A model-based approach was implemented for the devel-
`opment of the proliferator-activated receptor gamma
`(PPARγ) agonist rivoglitazone. Population pharmacoki-
`netic and pharmacodynamic models were developed
`using data collected from 2 phase I and 2 phase II studies
`in healthy volunteers and participants with type 2 dia-
`betes mellitus. A 2-compartment model with first-order
`absorption and elimination and an absorption time lag
`best described rivoglitazone pharmacokinetics. Modified
`indirect-response models were used to characterize
`changes in fasting plasma glucose, HbA1c, and hemodilu-
`tion as a function of rivoglitazone plasma concentrations.
`In addition, differences in hemodilution among partici-
`pants correlated with the incidence of edema. Current use
`
`of oral antidiabetic medication was a significant covariate
`for the fasting plasma glucose-HbA1c exposure-response
`model. Using a learn-and-confirm process, models devel-
`oped prior to the second phase II study were able to make
`valid predictions for exposures and response variables in
`that study. In future studies, seamless designs can be sup-
`ported by models such as those developed here.
`
`Keywords: Population pharmacokinetics; rivoglitazone;
`pharmacodynamics; learn and confirm; expo-
`sure response
`Journal of Clinical Pharmacology, 2008;48:1420-1429
`© 2008 the American College of Clinical Pharmacology
`
`Rivoglitazone
`
`thiazolidinedione
`(CS-011), a
`(TZD), increases insulin sensitivity by enhanc-
`ing insulin action in skeletal muscle, liver, and adi-
`pose tissue. By binding to and activating the nuclear
`receptor peroxisome proliferator-activated receptor
`gamma (PPARγ), which regulates the expression of
`genes involved in glucose and lipid metabolism, thi-
`azolidinediones increase glucose utilization, reduce
`hepatic glucose production, and enhance insulin
`
`From Daiichi Sankyo Pharma Development, Edison, New Jersey (Dr Rohatagi,
`Dr Walker, Dr Truitt, Dr Salazar); Pharsight Corporation, Mountain View,
`California (Dr Carrothers, Dr Khariton); and Department of Pharmaceutical
`Sciences, SUNY/Buffalo, Buffalo, New York (Dr Jin, Dr Jusko). Supplemental
`figures and the appendix are available online at http://jcp.sagepub.com/
`supplemental/. Submitted for publication March 28, 2008; revised version
`accepted July 6, 2008. Address for correspondence: Shashank Rohatagi,
`PhD, MBA, Fellow FCP, Daiichi Sankyo Pharma Development, 399 Thornall
`St, Edison, NJ 08837; e-mail: Srohatagi@dsus.com.
`DOI: 10.1177/0091270008323260
`
`sensitivity.1,2 Rivoglitazone has linear pharmacoki-
`netics in the dose range intended for future use, a
`half-life consistent with once-daily dosing, and very
`low renal clearance (CLR).3,4 In phase II dose-ranging
`trials of 6 and 26 weeks in duration, rivoglitazone
`treatment improved glycemic control in participants
`with type 2 diabetes mellitus (T2DM) and showed a
`safety profile consistent with the specific side effects
`(eg, weight gain, edema, and hemodilution) observed
`in clinical development of the currently marketed
`thiazolidinediones.5,6
`Modeling and simulation were implemented for
`rivoglitazone to enable a more complete and robust
`understanding of its benefits and risks, thus
`enabling a more informed and efficient drug devel-
`opment process (Figure 1). In the case of developing
`a “best-in-class” compound, model-based drug
`development can use the wealth of knowledge from
`predecessor drugs with a similar mechanism of
`
`1420 • J Clin Pharmacol 2008;48:1420-1429
`
`MPI EXHIBIT 1044 PAGE 1
`
`MPI EXHIBIT 1044 PAGE 1
`
`Apotex v. Novo - IPR2024-00631
`Petitioner Apotex Exhibit 1044-0001
`
`

`

`MODEL-BASED DEVELOPMENT OF A PPARγAGONIST, RIVOGLITAZONE
`
`Animal PK and
`exposure-response
`model on
`FPG and adiponectin
`
`Scaling and tox
`to select dose in
`healthy
`volunteers
`
`Human PK and
`effect on
`adiponectin as
`biomarker in
`healthy
`volunteers
`
`Select dose for
`patients
`
`PK and
`covariates
`in patients
`
`Edema rates
`and heart
`failure?
`
`Hemodilution
`correlated
`edema rates
`
`PK correlated
`to
`hemodilution
`
`Disease Progression (literature)
`(cid:129) Covariate on placebo data
`(cid:129) Disease progression term
`(cid:129) Naive/nonnaive
`(cid:129) Dropout
`
`PK correlated
`to FPG
`
`FPG correlated
`to HbA1c-
`clinical end
`point
`
`HbA1c and
`neuropathy
`retinopathy
`nephropathy
`
`Figure 1. Overview of rivoglitazone model-based drug development. PK, pharmacokinetic; FPB, fasting plasma glucose; Tox, toxicology.
`
`action. Beginning with first-in-human trials for the
`new drug, efficacy and safety models can be con-
`structed based on both the compound’s preclinical
`data and the predecessor’s clinical experience. As
`data are generated at each stage of clinical develop-
`ment, the models (and future projections) are
`updated, with the new drug’s properties becoming
`more definitive as the process matures. At all stages,
`the model-based process allows for the entire basis of
`relevant knowledge to be incorporated into decision-
`focused recommendations, following the learn-and-
`confirm paradigm promoted by Sheiner.7
`As part of the model-based process, initial doses
`for rivoglitazone in diabetics were chosen using
`adiponectin as a biomarker for PPARγ activity.
`Adiponectin, a hormone secreted by adipocytes, has
`been observed to reduce glucose levels in different
`animal models of obesity and diabetic mellitus by
`increasing insulin sensitivity. Circulating levels of
`adiponectin are lower in the obese and in patients
`with type 2 diabetes with respect to healthy controls
`and negatively correlate with the plasma levels of glu-
`cose, insulin, triglycerides, and insulin resistance.8,9
`Members of the TZD class are known to increase the
`
`circulating plasma levels of the active form of
`adiponectin in both healthy volunteers and diabetic
`participants.2 Based on the phase I data, up to 2 weeks
`of daily dosing with 1 to 5 mg produced increases
`from baseline in adiponectin of 66% to 527% with no
`additional effect at the 10-mg dose (see Supplemental
`Figure 1).10 Hence, the first phase IIa study in diabetic
`participants was conducted with doses of placebo or
`0.5, 1, or 5 mg, administered once daily for 6 weeks.5
`The specific objectives of the present analysis
`were to model the population pharmacokinetics (PK)
`of orally administered rivoglitazone using data col-
`lected from 2 phase I studies and 2 phase II studies.
`A pharmacodynamic (PD) model based on rosiglita-
`zone11 was applied to characterize effects on fasting
`plasma glucose (FPG), hemoglobin A1c (HbA1c), and
`hemodilution as a function of rivoglitazone plasma
`concentrations (Cp). The effects of covariates on the
`oral clearance (CL) of rivoglitazone and on the para-
`meters of the exposure-response models were char-
`acterized and quantified. For the exposure-response
`analyses, models were applied for the initial phase II
`study, confirmed by the second phase II study, and
`then updated based on the full data set.
`
`QUANTITATIVE CLINICAL PHARMACOLOGY
`
`1421
`
`MPI EXHIBIT 1044 PAGE 2
`
`MPI EXHIBIT 1044 PAGE 2
`
`Apotex v. Novo - IPR2024-00631
`Petitioner Apotex Exhibit 1044-0002
`
`

`

`ROHATAGI ET AL
`
`METHODS
`
`Trial Design and Participants
`
`The pharmacokinetic data and population analysis
`were derived from 60 healthy volunteers (58 men, 2
`women) enrolled in 2 phase I studies (CS0011-A-
`U103 and CS0011-A-U104) with full-profile sam-
`pling and 461 diabetic participants (265 men, 196
`women) enrolled in 2 phase II studies (CS0011-A-
`U202 and CS0011-A–U203). Only data from the
`phase II studies were used to conduct the pharma-
`codynamic analysis (n = 461 on active treatment and
`n = 184 on placebo). Phase I studies were conducted
`at MDS Pharma Services (Neptune, New Jersey), and
`the phase II studies were conducted at multiple clin-
`ical sites in the United States. An investigational
`review board approved each study protocol. Written
`informed consent was obtained from each study par-
`ticipant before any study-specific procedures or
`assessments. Studies were conducted under the
`principles of
`the World Medical Assembly
`Declaration of Helsinki and its most recent amend-
`ments, the US Code of Federal Regulations, and
`good clinical practice.
`All participants received either oral rivoglitazone
`or a matching placebo. Trial design, population,
`treatment regimens, duration, and sampling sched-
`ules are available online in Supplemental Table 1.
`
`Analytical Methods for Rivoglitazone
`Measurement in Plasma Samples
`
`A validated liquid chromatography/tandem mass
`spectrometry (LC/MS/MS) method was used for
`measuring rivoglitazone in plasma for the 2 phase II
`studies (CS0011-A-U202 and CS0011-A-U203) and 2
`phase I studies (CS0011-A-103 and CS0011-A-
`U104). An aliquot of human plasma (EDTA) con-
`taining the analyte and internal standard was
`extracted using an automated TOMTEC solid-phase
`extraction procedure. The extracted samples were
`analyzed by a high-performance liquid chromatogra-
`phy (HPLC; Polaris 50 × 2-mm, 5-μm column, 60/40
`methanol/10 mM ammonium acetate [pH 4.0]
`mobile phase, isocratic mode) system equipped with
`an ABI/MDS Sciex API 3000 mass spectrometer.
`Positive ions were monitored in the multiple-reaction
`monitoring (MRM) mode. The lower limit of quan-
`tification (LLOQ) of the validated method was set at
`0.5 ng/mL, whereas the dynamic ranges were 0.5 to
`1000 ng/mL (for the CS0011-A-U103 study) and
`
`1422 • J Clin Pharmacol 2008;48:1420-1429
`
`0.5 to 500 ng/mL (for the CS011-A-U104, CS011-A-
`U202, and CS011-A-U203 studies). The MRM transi-
`tion was 398.3/176.1 for rivoglitazone and 404.3/182.1
`for the internal standard (2H6-Rivoglitazone). Human
`plasma (EDTA), free of significant interference, was
`used to prepare calibration standard and quality con-
`2 weighted linear regression
`trol (QC) samples. A 1/x
`model was then used to calculate slope, intercept,
`and correlation coefficient. Back-calculated results
`of QCs and study samples were then obtained by fit-
`2 weighted regres-
`ting the peak area ratio to the 1/x
`sion equation for the relevant standards. For QC
`samples, between-batch precision (% coefficient of
`variation [CV]) ranged from 2.1 to 8.2, and accuracy
`(%Bias) ranged from –4.2 to 8.7. For calibration
`standards, between-batch precision (%CV) ranged
`from 1.4 to 7.0, and accuracy (%Bias) ranged from
`–4.0 to 6.0.
`
`Pharmacokinetic Sampling
`
`Intensive pharmacokinetic sampling was conducted
`after both single and multiple oral doses in the
`phase I trials, up to 72 hours postdose. In the phase
`II trials, pharmacokinetic and biomarker samples
`were drawn at the time of trough steady-state con-
`centration—that is, at predose; at weeks 0, 2, 4, and
`6 in the CS0011-A-U202 study; and at weeks 0, 4, 8,
`12, 16, 20, and 26 in the CS0011-A-U203 study. In
`study CS0011-A-U203, intensive PK sampling was
`performed after the last dose in the study in a subset
`of subjects who volunteered for intensive sampling.
`
`Pharmacokinetic Data Analysis
`and Model Development
`
`All data preparation and graphic representations
`were performed using S-PLUS software, Version 6.2.
`All pharmacokinetic and pharmacodynamic analy-
`ses were implemented within the NONMEM soft-
`ware program, Version V, Level 1.1. The development
`of the population pharmacokinetic model was based
`on the Food and Drug Administration’s (FDA’s)
`Guidance for Industry Population Pharmacokinetics.12
`Further details are provided online in supplemental
`materials.
`Change in FPG concentrations was modeled as a
`function of Cp via an indirect-effect model on the
`assumption that rivoglitazone reduces glucose by
`increasing the removal rate of glucose from the
`plasma compartment (Figure 2) as developed by
`Benincosa and Jusko11:
`
`MPI EXHIBIT 1044 PAGE 3
`
`MPI EXHIBIT 1044 PAGE 3
`
`Apotex v. Novo - IPR2024-00631
`Petitioner Apotex Exhibit 1044-0003
`
`

`

`MODEL-BASED DEVELOPMENT OF A PPARγAGONIST, RIVOGLITAZONE
`
`Dose
`
`C3,V3
`
`Q
`
`GI
`
`ka
`
`Cp, V2
`
`CL
`
`kin
`
`kg
`
`Smax, SC50
`kout
`
`kin
`
`IL
`
`Plasma
`Volume
`
`kout
`
`Glucose
`
`kd
`
`HbA1c
`
`the function for the inhibition of koutp by Cp, and PV
`is the plasma volume. Pretreatment baseline values
`were set as the initial conditions. The potential rela-
`tionship between edema and hemodilution was
`explored first through exploratory graphical analysis
`and subsequently
`through
`logistic
`regression
`models.
`Predicted subject-specific plasma concentrations
`of rivoglitazone, available from the population phar-
`macokinetic model, were used for pharmacody-
`namic model-fitting purposes. Stimulation and
`inhibition functions in the indirect-response models
`+
`were investigated with saturable (ie, Smax*Cp/(Cp
`SC50)) and linear models (ie, 1 + slope*Cp). If the
`relationship showed linearity without saturation, or
`if the saturable model parameters could not be esti-
`mated, the linear model was used.
`
`RESULTS
`
`Figure 2. Depiction of FPG-HbA1c and hemodilution exposure-
`response models. FPG, fasting plasma glucose; GI, gastrointesti-
`nal tract.
`
`Participant Demographics
`
`dFPG/dt = kin – kout*S(Cp)*FPG(t) FPG(0) = FPG0,
`
`(1)
`
`where kin is the zero-order glucose production rate,
`is the first-order glucose removal rate from
`kout
`plasma, S(Cp) is the function for the stimulation of
`kout by Cp, and FPG is the fasting plasma glucose con-
`centration. Changes in HbA1c were modeled as sec-
`ondary to changes in FPG in a first-order process:
`
`dHbA1c/dt = kg*FPG(t) – kd*HbA1c(t)
`HbA1c(0) = HbA1c0,
`
`(2)
`
`where kg is the pseudo first-order HbA1c production
`rate constant, and kd is the first-order HbA1c degra-
`dation rate constant.
`For hemodilution, the inverse of hemoglobin con-
`centration, denoted PV for plasma volume, was
`modeled as a function of the plasma concentration
`of rivoglitazone via an indirect-effect model on the
`assumption that rivoglitazone increases plasma
`volume by a linear inhibition of the loss of plasma
`volume:
`
`dPV/dt = kinp – koutp*I(Cp)*PV(t) PV(0) = PV0,
`
`(3)
`
`where kinp is the zero-order plasma volume produc-
`tion rate, koutp is the first-order removal rate, I(Cp) is
`
`Data were available for 518 unique participants for
`the pharmacokinetic analysis and 461 unique partic-
`ipants for the pharmacodynamic analyses. A sum-
`mary of the demographic characteristics of the data
`sets is presented online in Supplemental Table 2.
`
`Pharmacokinetic Model
`
`Population Pharmacokinetic Analysis
`Explorations of 1- and 2-compartment models deter-
`mined that the pharmacokinetics of rivoglitazone
`were best described by a 2-compartment linear
`model (CL, V2, V3, Q) with first-order absorption/
`elimination (ka) and an absorption time lag (tlag).
`The parameters for the final covariate model for
`the compartmental pharmacokinetics of rivoglita-
`zone are shown in Table I. Clearance was signifi-
`cantly affected by sex, body weight, renal function
`(as measured by SCr), and subject/healthy volunteer
`status (P < .05) as follows:
`
`CLi (L/h) = 1.15 × (SCri/1)–0.246 × (WTi/191)0.347
`– (0.163*SEXi) + (0.17*Healthyi).
`
`(4)
`
`The model indicates that a male patient with base-
`line SCr and body weight equivalent to the median
`of the study population, 1 mg/dL and 191 lb, would
`have a CL estimated to be 1.15 L/h. A healthy male
`volunteer with similar SCr and body weight had a
`CL of 1.32 L/h. Female patients had approximately
`
`QUANTITATIVE CLINICAL PHARMACOLOGY
`
`1423
`
`MPI EXHIBIT 1044 PAGE 4
`
`MPI EXHIBIT 1044 PAGE 4
`
`Apotex v. Novo - IPR2024-00631
`Petitioner Apotex Exhibit 1044-0004
`
`

`

`ROHATAGI ET AL
`
`Table I Rivoglitazone Population Pharmacokinetic Parameter Estimates, Standard Errors of Estimates, and
`Variability Estimates Before and After Inclusion of CS0011-A-U203 Data Set
`
`Parameter
`
`Estimate
`
`SEa
`
`Estimateb
`
`SEc
`
`Population Mean
`
`Intersubject Variability
`
`1.10
`19.1
`3.02
`1.93
`0.295
`0.323
`
`–0.159
`–0.210
`0.288
`0.121
`0.046
`
`1.15
`22.1
`2.92
`1.87
`0.273
`0.280
`
`Before CS0011-A-U203
`Fixed effects
`CLTYP, L/h
`V2, L
`V3, L
`ka, per h
`tlag, h
`Q, L/h
`Effect of covariates on CL
`Gender, CLSEX
`Renal function, CLSCR
`Weight, CLWT
`Health status, CLHV
`Residual variabilityd
`After CS0011-A-U203
`Fixed effects
`CLTYP, L/h
`V2, L
`V3, L
`ka, per h
`tlag, h
`Q, L/h
`Effect of covariates on CL
`–0.163
`Gender, CLSEX
`–0.246
`Renal function, CLSCR
`0.347
`Weight, CLWT
`0.172
`Patient status, CLPS
`0.067
`Residual variabilityd
`a. Coefficient of variation of the estimates (100 × SE estimate/estimate).
`b. Estimates of variability expressed as appropriate percent coefficient of variation (%CV) 100√Ω.
`c. Percent square root of the relative standard error of the coefficient of variation 100√SEETAestimate / ETA estimate.
`d. Residual intrasubject variability.
`
`3.0
`9.0
`13
`12
`10
`19
`
`26
`41
`21
`44
`7.6
`
`3.2
`8.4
`15
`10
`8.7
`31
`
`23
`31
`15
`16
`8.2
`
`22
`42
`65
`35
`42
`—
`
`—
`—
`—
`—
`—
`
`36
`76
`79
`74
`55
`—
`
`—
`—
`—
`—
`—
`
`44
`64
`57
`66
`61
`—
`
`—
`—
`—
`—
`—
`
`56
`68
`62
`96
`70
`—
`
`—
`—
`—
`—
`—
`
`14% lower CL rates than men. Each 10% decrease in
`body weight from the data set median of 86 kg trans-
`lated to a 3.4% decrease in CL, and each 10%
`increase in SCr translated to a 2.5% decrease in CL.
`Goodness-of-fit plots for the final 2-compartment
`model showed
`the model
`to be appropriate
`(Supplemental Figure 2). The individual- and popu-
`lation-predicted plasma concentrations matched the
`observed plasma concentrations, demonstrating that
`the model adequately described the data. Supple-
`mental Figure 2 also shows the residuals and weighted
`residuals versus the population predictions. Covariate
`plots for the final model showed no remaining
`patterns between the covariates and rivoglitazone
`clearance.
`Prediction of CS0011-A-U203 results from the
`best model developed prior to inclusion of that
`
`study’s data was robust and validated the learn-and-
`confirm process. Supplemental Figure 3 depicts
`model prediction plots for rivoglitazone trough con-
`centrations in each of the 3 CS0011-A-U203 dose
`arms, overlaid with observed concentrations. Across
`the plots, 82% to 94% of the data were within the
`95% confidence interval (CI) of the model predic-
`tions, supporting
`the validity of
`the model.
`Supplemental Figure 4 depicts the observed concen-
`tration versus time profiles data and prediction
`intervals for the 8 intensively sampled participants,
`indicating the appropriateness of the predictions.
`
`Pharmacodynamic Analyses
`The proposed structural model for the PK/PD
`models fit the data well for both study CS-011-202
`alone and
`together with
`study CS-011-203.
`
`1424 • J Clin Pharmacol 2008;48:1420-1429
`
`MPI EXHIBIT 1044 PAGE 5
`
`MPI EXHIBIT 1044 PAGE 5
`
`Apotex v. Novo - IPR2024-00631
`Petitioner Apotex Exhibit 1044-0005
`
`

`

`MODEL-BASED DEVELOPMENT OF A PPARγAGONIST, RIVOGLITAZONE
`
`Table II Rivoglitazone FPG-HbA1c Pharmacodynamic Parameter Estimates, Standard Errors of Estimates,
`and Variability Estimates Before and After Inclusion of CS0011-A-U203 Data Set
`
`Population Mean
`
`Intersubject Variability
`
`Parameter
`
`CS0011-A-U202 Model
`Fixed effects
`kout, 1/h
`FPGss, mg/dL
`SC50, ng/mL
`Smax (-)
`kd, 1/h
`HbA1c,ss, %
`Effect of covariates
`Nonnaive on FPGss, mg/dL
`Nonnaive on HbA1c,ss, %
`Residual variabilityd
`FPG, mg/dL2
`HbA1c, %2
`CS0011-A-U202 ++ CS0011-A-U203 Model
`Fixed effects
`kout, 1/h
`FPGss, mg/dL
`SC50, ng/mL
`Smax (-)
`kd, 1/h
`HbA1c,ss, %
`Effect of covariates
`Nonnaive on FPGss, mg/dL
`Nonnaive on HbA1c,ss, %
`Residual variabilityd
`FPG, mg/dL2
`HbA1c, %2
`
`Estimate
`
`0.00235
`160
`91.8
`0.81
`0.000525
`7.66
`
`48
`2.3
`
`9.6
`2
`
`0.00259
`160
`126
`0.43
`0.000664
`7.57
`
`48
`2.3
`
`12
`3
`
`SEa
`
`8.9
`2.6
`30
`13
`6.1
`2.2
`
`Fixed
`Fixed
`
`10.5
`8.5
`
`12
`3
`25
`36
`8
`3
`
`Fixed
`Fixed
`
`12
`20
`
`FPG, fasting plasma glucose; HbA1c, hemoglobin A1c.
`a. Coefficient of variation of the estimates (100 × SE estimate/estimate).
`b. Estimates of variability expressed as appropriate percent coefficient of variation (%CV) 100√Ω.
`c. Percent square root of the relative standard error of the coefficient of variation 100√SEETAestimate / ETA estimate.
`d. Residual intrasubject variability.
`
`Estimateb
`
`SEc
`
`—
`20
`193
`—
`—
`21
`
`—
`—
`
`—
`—
`
`—
`16
`316
`—
`—
`15
`
`—
`—
`
`—
`—
`
`—
`34
`54
`—
`—
`33
`
`—
`—
`
`—
`—
`
`—
`19
`80
`—
`—
`35
`
`—
`—
`
`—
`—
`
`Population parameters for the FPG-HbA1c and
`hemodilution models are listed in Tables II and III.
`The FPG decrease depended on drug concentra-
`tion. For this relationship, a saturable pharmaco-
`logic model for stimulation of kout gave a better fit
`than a similar linear model. One covariate relation-
`ship on model parameters was found, that of prior
`medication status on steady-state FPG and steady-
`state HbA1c. This difference, representing the impact
`of prior medications on nonnaive patients, was
`48 mg/dL and 2.3%. To achieve model convergence,
`we estimated this effect through model iterations
`wherein one or more model parameters were fixed.
`
`For hemodilution, a linear function for inhibition
`of koutp was preferable to a saturable model, and
`gender had a significant influence on baseline
`plasma volume, with women having a higher base-
`line plasma volume. No other covariate relation-
`ships on model parameters were observed.
`As seen in Figure 3, individual predictions plot-
`ted against observed values in residual plots for the
`final PK/PD models showed the models to be appro-
`priate. Simulations of CS0011-A-U203 based on the
`CS0011-A-U202 data generally showed the actual
`data to fall within the 90% prediction interval for
`each of the 3 response variables (Figure 4).
`
`QUANTITATIVE CLINICAL PHARMACOLOGY
`
`1425
`
`MPI EXHIBIT 1044 PAGE 6
`
`MPI EXHIBIT 1044 PAGE 6
`
`Apotex v. Novo - IPR2024-00631
`Petitioner Apotex Exhibit 1044-0006
`
`

`

`ROHATAGI ET AL
`
`Table III Rivoglitazone Hemodilution Pharmacodynamic Parameter Estimates, Standard Errors of
`Estimates, and Variability Estimates Before and After Inclusion of CS0011-A-U203 Data Set
`
`Population Mean
`
`Intersubject Variability
`
`Parameter
`
`Estimate
`
`SEa
`
`Estimateb
`
`SEc
`
`103
`32
`53
`
`—
`—
`
`73
`28
`47
`
`—
`—
`
`CS0011-A-U202 Model
`Fixed effects
`koutp, 1/h
`PV0, dL/mg
`Slope, mL/ng
`Effect of covariates on CL
`Gender on PV0
`Residual variabilityd
`CS0011-A-U202 ++ CS011-A-U203 Model
`Fixed effects
`koutp, 1/h
`PV0, dL/mg
`Slope, mL/ng
`Effect of covariates on CL
`0.00663
`Gender on PV0
`3.1
`Residual variabilityd
`a. Coefficient of variation of the estimates (100 × SE estimate/estimate).
`b. Estimates of variability expressed as appropriate percent coefficient of variation (%CV) 100√Ω.
`c. Percent square root of the relative standard error of the coefficient of variation 100√SEETAestimate / ETA estimate.
`d. Residual intrasubject variability.
`
`0.00293
`0.0671
`0.000809
`
`0.00579
`2.9
`
`0.00220
`0.0669
`0.000786
`
`11
`5.5
`6.2
`
`12
`6.3
`
`11
`0.4
`4.6
`
`7.7
`5.2
`
`26
`7.1
`45
`
`—
`—
`
`94
`7.3
`58
`
`—
`—
`
`An additional analysis was performed to look at
`the correlation of hemodilution with edema. In
`study CS-011-203, 50 participants developed edema.
`As shown by the histograms of percent change in
`hemoglobin concentrations in Figure 5, patients
`with edema had, on average, a greater drop in their
`hemoglobin concentrations. A logistic relationship
`(Figure 5) fit this relationship well:
`
`logit(probability of edema) = –2.47 – 0.117*
`(% Δ baseline of Hb).
`
`(5)
`
`Finally, the performance of adiponectin as a bio-
`marker for PPARγ activity was confirmed by its cor-
`relation with the respective changes from baseline of
`each response variable (Supplemental Figure 5).
`
`DISCUSSION
`
`The broader objectives achieved by this phar-
`macometric analysis for rivoglitazone included
`(1) establishing a quantitative understanding of key
`dose-exposure-biomarker-endpoint relationships (both
`safety and efficacy), (2) understanding the impact of
`intrinsic and extrinsic factors on the pharmacokinet-
`ics and exposure-response relationships to deter-
`mine whether there was a need for dose adjustment
`
`in special populations, and (3) having a fuller
`understanding of the impacts of trial design (eg,
`length of washout period for participants on prior
`antidiabetic medications, inclusion/exclusion crite-
`ria, participant baseline characteristics, and discon-
`tinuation criteria), for both better explanation of
`results within the rivoglitazone trial sequence as
`well as to enable a more direct comparison to his-
`torical trials with different designs and/or partici-
`pant characteristics.
`For the pharmacokinetics of rivoglitazone, a 2-
`compartment model with log residual error and a lag
`time for absorption was found to be the best structural
`PK model. Clearance was found to be affected by
`gender, serum creatinine, weight, and patient status,
`with lower clearance (and higher exposure) for women,
`diabetic patients, and those with higher serum creati-
`nine and/or lower weight. As rivoglitazone has very
`low renal clearance, the relationship with serum crea-
`tinine may reflect changes in protein concentrations
`and subsequent changes in free (unbound) rivoglita-
`zone. Correlated reductions in renal and hepatic clear-
`ance have been seen in patients with lower renal
`function.13 This relationship will need to be clarified
`further in future studies because fraction bound/
`unbound was not assessed during the 2 phase II studies
`comprising this analysis.
`
`1426 • J Clin Pharmacol 2008;48:1420-1429
`
`MPI EXHIBIT 1044 PAGE 7
`
`MPI EXHIBIT 1044 PAGE 7
`
`Apotex v. Novo - IPR2024-00631
`Petitioner Apotex Exhibit 1044-0007
`
`

`

`MODEL-BASED DEVELOPMENT OF A PPARγAGONIST, RIVOGLITAZONE
`
`18
`
`16
`
`14
`
`12
`
`10
`
`8
`
`18
`
`16
`
`14
`
`12
`
`10
`
`8
`
`Population Predicted HGB [mg/dL]
`
`Individual Predicted HGB [mg/dL]
`
`0
`
`0
`
`0
`
`0
`
`12
`
`10
`
`8
`
`6
`
`Population Predicted HbA1c [%]
`
`12
`
`10
`
`8
`
`6
`
` [%]
`
`Individual Predicted HbA1c
`
`0
`
`og
`
`0
`
`0
`
`0
`0
`
`0
`
`oo
`
`@
`0
`
`/j
`0 0
`
`0
`
`0
`
`0
`
`400
`
`300
`
`200
`
`100
`
`400
`
`300
`
`200
`
`100
`
`Population Predicted FPG [mg/dL]
`
`Individual Predicted FPG [mg/dL]
`
`100
`
`200
`300
`Observed FPG [mg/dL]
`
`400
`
`6
`
`8
`10
`Observed HbA1c [%]
`
`12
`
`8
`
`10
`
`12
`14
`Observed HGB [mg/dL]
`
`16
`
`18
`
`Figure 3. Goodness-of-fit plots for the final exposure-response models based on CS0011-A-U202 and CS0011-A-U203. FPG, fasting
`plasma glucose; HbA1c, hemoglobin A1c; and HGB, hemoglobin.
`
`The mixed-effects indirect-response model of
`FPG and HbA1c fit the data adequately for both stud-
`ies, with excellent fits at the individual level. For
`nonnaive patients, the removal of their prior med-
`ications was evident in the model parameters of
`FPGss and HbA1c,ss as their glucose concentrations
`returned
`to
`their nonmedicated equilibrium.
`Rivoglitazone counteracted this rebound, especially
`at higher doses. This impact of prior medications
`identified in the model is roughly in line with the
`effects of metformin and sulfonylureas on FPG and
`HbA1c. The exact impact would have been easier to
`quantify more precisely had there not have been
`such frequent discontinuations in the study, espe-
`cially in the placebo arm.
`A drawback of this model is that the population
`predictions showed a bias when plotted against the
`observed data. Fortunately, the individual predic-
`tions were generally able to achieve a much closer
`fit. However, this was accomplished by an unusually
`large degree of interindividual variability in model
`
`parameters. Here again, the frequent dropouts in the
`placebo and lower dose arms made parameter esti-
`mation difficult, as many patients did not remain in
`the study long enough to achieve a new equilibrium
`for FPG, much less HbA1c. Data from future studies
`may be needed to further evaluate the stability of
`exposure-response parameters as well as the poten-
`tial effects of covariates.
`The indirect-response model of changes in hemo-
`globin concentrations fit the data quite well for both
`studies, with a clear PK/PD relationship identified.
`In addition, the CS011-A-U202 model predicted the
`CS011-A-U203 results well. As women have lower
`baseline hemoglobin levels, the model accounts for
`this difference. Changes in hemoglobin were also
`shown to correlate with the incidence of edema.
`The exposure-response data set from CS0011-A-
`U202 and CS0011-A-U203, plus corresponding fast-
`ing serum insulin (FSI) values from the sample
`points, was used to extend a disease progression model
`of pioglitazone14
`to rivoglitazone. (Details on the
`
`QUANTITATIVE CLINICAL PHARMACOLOGY
`
`1427
`
`MPI EXHIBIT 1044 PAGE 8
`
`MPI EXHIBIT 1044 PAGE 8
`
`Apotex v. Novo - IPR2024-00631
`Petitioner Apotex Exhibit 1044-0008
`
`

`

`ROHATAGI ET AL
`
`Patients without edema
`
`150
`
`100
`
`50
`
`0
`
`15
`
`10
`
`5
`
`0
`
`Count
`
`Count
`
`1.0
`
`0.8
`
`0.4
`0.6
`P(Edema)
`
`0.2
`
`0.0
`
`0 0
`
`200
`
`–100
`100
`FPG CFB [mg/dL]
`
`0
`
`–200
`
`4
`
`–2
`HbA1c CFB [%]
`
`2
`
`0
`
`,
`·~.
`
`0
`0
`
`0
`
`0
`
`t--,:;oo-~-
`
`0
`
`e __ ~
`
`o o
`
`:'-i0o----,-\-~-----!
`~ 0
`
`0
`
`li/t"~-F-+---1--~.S:------"-~'
`
`0
`
`8
`o °o
`o o o
`oo
`- a....
`.. - e- - - - - - - _q - - - - - - - - - - - - - - - - - - - - - - - o
`
`00
`
`0
`
`0
`
`–4
`
`18
`
`10
`16
`Hemoglobin [mg/dL]
`
`14
`
`12
`
`8
`
`0
`
`5
`
`10
`
`15
`
`20
`
`25
`
`30
`
`Time [weeks]
`
`Figure 4. Results from study CS0011-A-U203 were well pre-
`dicted by prior exposure-response models. CFB, change from
`baseline; FPG, fasting plasma glucose; HbA1c, hemoglobin A1c.
`
`modeling steps are included as an online appendix.)
`Figure A1 in the online appendix shows that the
`model was successfully adapted from its previous
`dose-response form to an exposure-response form
`for rivoglitazone. As the literature model used studies
`
`1428 • J Clin Pharmacol 2008;48:1420-1429
`
`–20
`
`–10
`
`0
`% CFB Hemoglobin
`
`10
`
`20
`
`Figure 5. Edema was shown to be related to hemodilution in
`both the graphical analysis (top) and subsequent logistic regres-
`sion model (bottom). CFB, change from baseline.
`
`up to 2 years in length versus the 6 months of data
`in CS0011-A-U203, some model parameters in the
`rivoglitazone analysis were fixed to the literature
`values. The disease progression model separated the
`short- and long-term effects of oral antidiabetic ther-
`apy on beta-cell function and insulin sensitivity,
`enabling simulations (Figure A2) of the predicted
`long-term effectiveness of rivoglitazone and distin-
`guishing it from older therapies such as metformin
`and the sulfonylureas.
`The results of these modeling analyses were com-
`municated regularly at all phases of the program,
`thus enabling the continuous improvements of the
`models and testing of a variety of “what-if” scenarios
`for study populations and study designs. Pharmacometric
`knowledge informed phase III design decisions for
`
`__ .1. __
`
`–30
`
`–20
`
`0
`–10
`PCT CHG HEMO
`
`10
`
`20
`
`Patients with edema
`
`–30
`
`–20
`
`0
`–10
`PCT CHG HEMO
`
`10
`
`20
`
`logit(p) = –2.47 – 0.117*(%CFB Hemoglobin)
`
`MPI EXHIBIT 1044 PAGE 9
`
`MPI EXHIBIT 1044 PAGE 9
`
`Apotex v. Novo - IPR2024-00631
`Petitioner Apotex Exhibit 1044-0009
`
`

`

`MODEL-BASED DEVELOPMENT OF A PPARγAGONIST, RIVOGLITAZONE
`
`selection of doses, timing of samples, inclusion/
`exclusion criteria, length of washout, and definition
`of discontinuation criteria. Future development of
`similar compounds can use models such as these to
`enable seamless phase II designs, with the promise
`of faster, cheaper, and more informative trials.
`
`CONCLUSIONS
`
`An iterative learn-and-confirm, model-based process
`for development of a “best-in-class” compound was
`successfully implemented during the development of
`rivoglitazone
`to support development decision
`making. Successful use of adiponectin as a biomarker
`for phase IIa dose selection was verified by the subse-
`quent activity of the drug in the selected dose range
`and by the correlation of changes in adiponectin with
`changes in clinical response variables. Models of
`rivoglitazone population PK/PD based on the phase I
`and initial phase IIa studies were able to make valid
`predictions for the second, longer term phase IIb study.
`In an example of sharing and adopting common dis-
`ease progression models, the exposure-response model
`developed for rivoglitazone was successfully linked to
`a published disease progression model for a similar
`compound, enabling predictions of long-term clinical
`response and differentiation from other classes of oral
`antidiabetic compounds. The rivoglitazone modeling
`experience provides an example of both the challenges
`and promise of model-based drug development.
`
`The authors thank Ling He, PhD, Smita Kshirsagar, PhD, and
`Lars Lindbom, PhD, for their contributions to the work described
`in this article.
`Financial disclosure: None declared.
`
`REFERENCES
`
`1. Krentz AJ, Bailey CJ. Oral antidiabetic agents: current role in
`type 2 diabetes mellitus. Drugs. 2005;65:385-411.
`
`2. Diamont M, Heine RJ. Thiazolidinediones in type 2 diabetes
`mellitus: current clinical evidence. Drugs. 2003;63:1373-1405.
`3. Walker JR, Triscari J, Dmuchowski CF, et al. Single and multi-
`ple dose pharmacokinetics and pharmacodynamics of rivoglita-
`zone (CS-011) in healthy male volunteers. J Clin Pharmacol.
`2006;46:1069.
`4. Walker JR, Dmuchowski CF, Samata N, Salazar DE. The steady-
`state pharmacokinetics of