FormalPara Key Points

The population pharmacokinetic model adequately described canagliflozin pharmacokinetics in healthy volunteers and in patients with type 2 diabetes mellitus (T2DM).

The magnitude of the statistically significant covariates identified was small and therefore not considered clinically relevant.

Canagliflozin dosage adjustments are recommended for T2DM patients with renal impairment [estimated glomerular filtration rate (eGFR) ≥60 mL/min/1.73 m2: 100 or 300 mg/day; eGFR of 45 to <60 mL/min/1.73 m2: 100 mg/day].

1 Introduction

Diabetes mellitus affects nearly 347 million people worldwide and is estimated to be the seventh leading cause of death by 2030 with type 2 diabetes mellitus (T2DM) accounting for 90 % of these cases. The treatment options available for T2DM focus on reducing insulin resistance, enhancing insulin secretion, or providing exogenous insulin. The inhibition of the sodium-glucose co-transporter plays a pivotal role in increasing the excretion of glucose and lowering blood glucose. The sodium-glucose co-transporter 2 (SGLT2) inhibitors represent a novel therapeutic approach in the management of T2DM. Currently, three SGLT2 inhibitors are approved in the USA and Europe for the treatment of T2DM (canagliflozin, dapagliflozin, and empagliflozin) [1–3].

Canagliflozin, an orally active, reversible, selective SGLT2 inhibitor, is a novel oral hypoglycemic agent approved as an adjunct to diet and exercise in several countries worldwide at 100- and 300-mg doses to improve glycemic control in adults with T2DM [2]. For patients requiring the addition of canagliflozin to metformin therapy or the addition of metformin to canagliflozin, or in those patients already on metformin and canagliflozin, each dosed as separate tablets, the use of a fixed-dose combination tablet comprising metformin and canagliflozin may improve patient convenience and compliance to antihyperglycemic agent therapy [4].

The low-affinity/high-capacity SGLT2 transporter present in the early proximal convoluted renal tubule reabsorbs most of the filtered glucose [5, 6] and a very small amount of glucose is reabsorbed by the high-affinity/low-capacity glucose transporter, SGLT1. Canagliflozin decreases renal glucose reabsorption, thereby increasing urinary glucose excretion and decreasing plasma glucose levels in patients with hyperglycemia. Furthermore, the calories lost in urine (as glucose) may also reduce body weight (WT) [4, 7–10]. In a phase III study, monotherapy with canagliflozin 100- and 300-mg doses resulted in significant (p < 0.001) reductions in glycosylated hemoglobin (<7.0 %) from baseline to week 26 vs placebo in patients with T2DM [11].

The pharmacokinetic (PK) profile of canagliflozin is similar in healthy participants and patients with T2DM. Orally administered canagliflozin is rapidly absorbed achieving peak plasma concentrations within 1–2 h [2]. Dose-proportional systemic exposure to canagliflozin has been observed over a wide dose range (50–1600 mg) with an oral bioavailability of 65 %. The apparent terminal half-life was 10.6 h for the 100-mg dose and 13.1 h for the 300-mg dose. Steady state was reached after 4 days, and there is minimal accumulation observed after multiple dosing. The mean steady-state volume of distribution after a single intravenous infusion was 83.5 L in healthy participants, suggesting extensive tissue distribution of canagliflozin. Canagliflozin is extensively bound to plasma proteins (99 %), predominantly to human serum albumin. Approximately 60 and 33 % of the administered dose is excreted in the feces and urine, respectively. Canagliflozin is primarily metabolized by uridine diphosphate-glucuronosyltransferase 1 A9 (UGT1A9) and UGT2B4 to inactive O-glucuronides (M7 and M5) and minimally by cytochrome P450 3A4, and has a very limited potential to cause drug–drug interactions [2].

This article provides a comprehensive summary of the population pharmacokinetic (popPK) analysis, including a covariate search, of canagliflozin. Based on PK data from healthy volunteers and patients with T2DM, a popPK model was developed to improve the understanding of canagliflozin PK and explore possible relationships between patient covariates and PK parameters. The objectives of this analysis were: (1) to develop a popPK model that adequately describes the PK of canagliflozin after oral administration in healthy volunteers and in patients with T2DM; (2) to evaluate the effects of volunteers’ demographic characteristics and other covariates on PK parameters of canagliflozin; and (3) to evaluate the clinical relevance of statistically significant covariates and investigate predictive properties of the model.

2 Methods

2.1 Bioanalytical Assay

The validated assays for the quantification of total unchanged canagliflozin (JNJ-28431754) in human plasma for the clinical studies included in this pooled popPK analysis were based on liquid chromatography coupled to tandem mass spectrometry. The lower limit of quantification (LLOQ) for canagliflozin was 5.00 ng/mL, except for study 9 for which the LLOQ was 1.00 ng/mL. The bioanalytical assay range was from 5.00 to 10,000 ng/mL (all studies included in popPK analysis), except for studies 4 and 9 for which the range was from 5.00 to 5000 ng/mL and from 1.00 to 2000 ng/mL, respectively.

2.2 Study Population and Design

The study population included healthy volunteers from phase I studies and patients with T2DM enrolled in phase II and III studies. Table 1 summarizes study designs, objectives, and enrollment for each study used in this pooled popPK analysis. Each protocol was reviewed and approved by an institutional review board and informed consent was obtained from each volunteer before enrollment in the studies. The studies were conducted according to the Declaration of Helsinki, Good Clinical Practice guidelines, and other applicable regulatory requirements.

Table 1 Summary of study data included in the pooled population PK analysis
Full size table

The model was developed in steps (i.e., phase I model, pre-final model, external evaluation, and final popPK model). For the phase I model, a total of 5715 PK samples from 245 volunteers across richly sampled phase I studies were used for model development and covariate evaluation on the absorption and distribution parameters (ESM Table 2). Next, the phase I PK dataset was combined with PK data from sparsely sampled phase II and III studies. From this combined dataset, 8813 PK samples from 1526 volunteers were included in the analysis, and covariate effects were explored on the elimination rate constant (k e) resulting in the pre-final model. The external evaluation dataset contained 174 volunteers that contributed 182 PK samples [from two phase III studies (studies 13 and 14) that were not used to build the model]. The datasets used to build and to evaluate the model were then combined and the final model was re-run on the full dataset (9061 PK samples from 1616 volunteers).

2.3 Population Pharmacokinetic Analysis

The PopPK analysis was performed by non-linear mixed-effects modeling using NONMEM® (ICON plc) Version VII level 1.0, with PREDPP version V level 2.0 with gFORTRAN compiler 4.5.0 [12]. Dataset preparation, exploration, and visualization, as well as statistical and graphical analysis of all NONMEM® output, were performed using R for Windows (version 2.15.3). The first-order conditional estimation with interaction method was used for all analyses. The model was parameterized in terms of rate constants using the ADVAN4 TRANS1 option in NONMEM. The inter-individual variability (IIV) in PK parameters was regarded as random and was modeled using eta (η) variables. The individual η values were assumed to be log-normally distributed with a mean of zero and an estimated variance (ω 2). Correlations between IIV distributions were evaluated by means of graphical assessment and tested by inclusion of covariance terms between IIV parameters in the model.

Pre-dose samples and PK samples in which corresponding canagliflozin concentrations were below the LLOQ were excluded from the popPK analysis (5.6 % of all samples). Canagliflozin concentrations were excluded from the analysis if the sample date/time was missing, no or insufficient dosing information before sample collection was available, or the time since the last dose was more than 48 h (5.6 % of all samples). Additionally, missing plasma concentration values were treated as missing and were not replaced with imputed values (1.2 % of all samples). Outliers, defined as aberrant observations that significantly deviated from the rest of the observations in a particular individual, were excluded. During model development, 53 outlying data points in 39 patients were identified through the conditional weighted residuals (deviating more than ±6 units) and individual residuals (deviating more than ±2 units). These samples (0.5 % of all samples) were excluded from further analysis because of their disproportional contribution to the residual error.

2.4 Pharmacostatistical Model Development

A popPK model was developed first on the richly sampled phase I data. Extensive covariate analysis was performed on the absorption (absorption rate constant, k a and lag-time, T lag) and distribution (apparent volume of distribution of the central compartment, V c/F and distribution rate constant from the peripheral compartment, k 32) parameters. These parameters, including the covariance matrix of the random effects were subsequently fixed when fitting the richly (phase I) and sparsely (phase II and III) sampled PK data together.

Once the model built on phase I data was final, the absorption (k a, D 1, and T lag) and distribution parameters (V c/F, k 23, and k 32), including the covariance matrix of the random effects were fixed and the model was re-run on a combined phase I, II, and III dataset (pre-final model). This model was used for covariate analysis on k e and was subjected to an external evaluation. The evaluation was performed on a subset of data from two phase III studies that were not used to build the pre-final model. Once the pre-final model was evaluated, the datasets used to build and to evaluate the pre-final model were combined and the model was re-run on the full dataset to obtain the final parameters. This final popPK model was used to evaluate the clinical relevance of identified covariates.

Log-transformed plasma concentration-time data were evaluated by a sequence of compartmental models. Different absorption models were tested (e.g., zero- and first-order absorption models, sequential and parallel zero- and first-order absorption) using the first-order conditional estimation with interaction method in NONMEM®, and their results compared. IIV was described using an exponential model. Because the canagliflozin plasma concentrations were log transformed, the residual error was described using an additive error model in the log domain. An extra residual error term was added for data from phase II and III studies to account for the difference in residual variance between richly (phase I) and sparsely (phase II and III) sampled PK data. The final selection of the structural model was based on physiological considerations, goodness-of-fit plots, reduction in minimum objective function value (OFV), and decreases in residual error variance.

2.5 Covariate Analysis

The selected covariates that were assessed included sex, WT, body mass index (BMI), age, estimated glomerular filtration rate (eGFR), fed status, glycemic status (diabetic vs. non-diabetic), encapsulation (over-encapsulated vs. non-encapsulated tablets), dose, genetic polymorphism (carriers of the UGT1A9*3 allele), particle size distribution (PSD), and effect of long-term concomitant medications (substrates of UGT1A9 or UGT2B4). The effect of PSD on PK of canagliflozin was investigated as different active pharmaceutical ingredient lots with their corresponding particle sizes were used in the clinical studies included in this popPK analysis. The PSD (D90, µm) of all batches used in this popPK analysis ranged from 71.0 to 276 µm.

Covariate effects were explored graphically on the empirical Bayes or post hoc estimates of the PK parameters that were estimated with IIV, provided shrinkage was sufficiently low (<30 %). Covariates selected by this method, taking account of known physiological/pharmacological considerations, were subsequently incorporated in the popPK model.

Continuous covariates were included in the model using Eq. 1, where TVP is the typical value of the parameter, and θ x and θ y are fixed-effect parameters:

$$ {\text{TVP }} = \theta_{x} \left( {{\text{covariate}}/{\text{covariate median}}} \right)^{{\theta_{y} }} . $$
(1)

In accordance with the draft US Food and Drug Administration guidance on renal impairment, eGFR was included in the model as a continuous covariate [13].

In the case of categorical covariates, the TVP was allowed to change from category to category by using a different parameter for each category:

$$ {\text{TVP}} = \theta \left( i \right), $$

where i is the category.

Selected candidate covariates were formally evaluated for statistical significance by including them in the model one at a time. An OFV change of 10.83 units (p < 0.001, chi-squared distribution, one degree of freedom) was used as cut-off value to include a covariate in the popPK model. All covariates that were statistically significant were included in the model at the same time to construct a full model. The final significance of each covariate was evaluated by deleting it from the full model one at a time. To consider exclusion of one covariate in the full popPK model, an increase in minimum OFV of <10.83 units (p < 0.001, chi-squared distribution, one degree of freedom) was needed. The covariate resulting in the lowest increase in OFV was removed from the full model first. This process was repeated until all covariates that were not statistically significant were excluded from the model.

Evaluation of clinical relevance of the statistically significant covariates on model parameters, and hence on exposure, was based on area under the curve (AUC) values calculated based on the dose and individual CL/F. The individual CL/F was derived from the post hoc estimated individual V c/F and post hoc estimated individual k e, obtained from the final popPK model for the total study population (N = 1616), as CL/F = V c/F × k e. For each covariate, the dose-normalized AUC values (normalized to a 100-mg dose) were compared at different covariate levels (detailed information on covariate levels are provided in the footnotes of Tables 2 and 6). For each covariate level, median dose-normalized AUCs, 5 and 95 % percentiles were calculated. The ratio of median dose-normalized AUC values with their 95 % confidence intervals (CIs), computed using Fieller’s theorem [14] with the correlation between numerator and denominator set to zero (worst case scenario), were calculated for each significant covariate.

Table 2 Evaluation of the influence of BMI on k a and T lag and over-encapsulation (over-encapsulated vs encapsulated tablets) on T lag (phase I data)
Full size table

2.6 Model Evaluation

The predictive properties of the phase I and final popPK model were illustrated by prediction-corrected visual predictive checks (pcVPCs) [15]. The pcVPCs were constructed based on 1000 simulated replicates of the original dataset. The pcVPCs have time after last dose as the independent variable and are presented both on the linear and log scales. The 5th and 95th percentiles of prediction-corrected observations are depicted, corresponding to a 90 % inter-percentile range (prediction interval). As there was variability in the actual sampling times between volunteers, sampling time binning was performed. The width of each bin was made obvious in the graphs through the width of the square-shaped CIs.

The model was evaluated using an external evaluation dataset, i.e., PK data from two phase III studies that were not used to build the model. Plasma concentrations were predicted using parameter estimates of the pre-final model, fixing the inter-individual and residual random effects to their mean value of zero. The $ESTIMATION command was set as MAXEVAL = 0 and NONMEM was run [12]. A diagnostic plot of the observed versus model-predicted concentrations and a quantile–quantile (QQ) plot of the normalized prediction distribution errors (NPDEs) were generated to determine how well the model was able to predict the external evaluation dataset.

3 Results

3.1 Population Characteristics

In the population from the pooled phase I data, the median age of volunteers was 45 years, and 63 % were men and 59 % were white. The median baseline eGFR was 102 mL/min/1.73 m2 and 40 % had T2DM. The population from the pooled phase I, II, and III data was also mainly white (71 %) with a median age of 55 years, median baseline eGFR of 87 mL/min/1.73 m2, and 75 % had T2DM. The proportion of men and women in this dataset was balanced. The detailed demographic and baseline characteristics are outlined in Supplementary material.

3.2 Phase I Population Pharmacokinetic Model

A two-compartment popPK model with a lag-time and sequential zero- and first-order absorption and first-order elimination from the central compartment, with IIV on V c/F, k e, k a, k 32, and T lag, provided the best fit to the richly sampled phase I PK data (Fig. 1).

Fig. 1
figure 1

Model structure of the final population pharmacokinetic model of canagliflozin. A two-compartment disposition model with lag-time and zero-order release in a depot compartment followed by first-order absorption into and first-order elimination from the central compartment was used to describe the plasma concentration vs time profile of canagliflozin. V c/F apparent volume of distribution of the central compartment, k a first-order absorption rate constant, k e elimination rate constant, T lag lag time, D 1 duration of zero-order influx of dose into the gut compartment, k 23, k 32 distribution rate constants to and from the peripheral compartment

Full size image

The covariates on the absorption parameters were BMI on k a, and BMI and over-encapsulation on T lag. Volunteers with lower BMI had faster absorption rates than those with higher BMI. The median k a value in volunteers with BMI <24.2 kg/m2 was about two times higher than for those with BMI >28.0 kg/m2. The median T lag value in volunteers with BMI >28.0 kg/m2 was about two times higher than in volunteers with BMI <24.2 kg/m2. An approximately 93 % longer T lag was observed with the over-encapsulated tablet (median T lag of approximately 16 min) compared with the non-encapsulated tablet (median T lag of approximately 8 min) (Table 2). Although T lag was somewhat longer with the over-encapsulated tablets, over-encapsulation had no meaningful effect on the AUC of canagliflozin as the median dose-normalized AUC values were only 6 % higher compared with those of the non-encapsulated tablet (unpublished data).

WT, age, and sex on V c/F were identified as significant covariates. The effect of these covariates on V c/F and hence on AUC was assessed on the post hoc estimated median dose-normalized AUC values obtained from the final popPK model, including all significant covariates (i.e., BMI on k a, BMI and over-encapsulation on T lag, sex, and WT on V c/F, and eGFR and dose on k e). The median dose-normalized AUC values were about 22 % higher in women than in men, and approximately 29 % higher in older volunteers (age >60 years) than younger volunteers (age <50 years); and volunteers with WT <78.2 kg had about a 33 % higher AUC than those with weight >95.2 kg.

The parameter estimates presented in Table 3 show that all PK parameters could be estimated from the phase I dataset with adequate precision as expressed by their low relative standard errors (RSE %), ranging from 2 % to 22 %. The IIV on CL/F, V/F, and k 32 was 20, 15, and 35 %, respectively. The high IIV observed on k a and T lag was most probably owing to the limited PK data collected around the time to maximum plasma concentrations. The pc-VPCs of the final phase I model revealed acceptable predictive performance of the absorption phase and an over-prediction was observed in the elimination phase (Fig. 2).

Table 3 Parameter estimates of the phase I population PK model
Full size table
Fig. 2
figure 2

Prediction-corrected visual predictive checks with 90 % inter-percentile range for the final phase I model. The solid red line represents the median prediction-corrected observed plasma concentrations. The dashed red lines represent the observed 5 and 95 % percentiles. The 95 % confidence intervals for the corresponding model predicted percentiles are shown as gray fields. DV dependent variable

Full size image

3.3 Pre-Final Population Pharmacokinetic Model

The pre-final model constituted the basis for final covariate model development. Shrinkage on IIV for k e was acceptable for graphical exploration of covariate effects (~21 %), but was unacceptably high for k a, T lag, k 32, and V c/F (>55 %). A summary of the covariate testing and the difference in OFV is provided in Table 4. The covariates eGFR, dose, and genetic polymorphism had a significant effect on k e. No significant effect of age, sex, fed status, race, glycemic status, and concomitant medications on k e was observed. The effects of PSD and genetic polymorphism on canagliflozin PK were evaluated in the final model, after including all other significant covariates, by forward addition. As shown in Table 4, PSD had no significant effect on k e, while genetic polymorphism was a significant covariate. As volunteers carrying the UGT1A9*3 allele only represented ~1 % of the total number of volunteers (N = 1616) included in this popPK analysis, and as no visual trends could be detected, this covariate was not included in the final popPK model.

Table 4 Covariate model development phase I, II, and III
Full size table

3.4 Model Evaluation

The pre-final model, including all statistically significant covariates, was externally evaluated using a subset of data that was not used to build the model. The diagnostic plot of the observed vs model-predicted concentrations and the QQ plot of the NPDEs are provided in Fig. 3. These plots demonstrate that the popPK model of canagliflozin predicts the evaluation dataset adequately, as the trend line of observed vs model-predicted plasma concentrations lies around the line of unity and points in the QQ plot fall approximately along the line of identity.

Fig. 3
figure 3

Goodness-of-fit diagnostic plots for external evaluation. a The log of the observed concentrations vs the log of the model-predicted concentrations derived from the population mean parameter estimates. Dashed line (black) reference identity line, full line (gray) lowess smoother. Observed and model predicted concentrations for a typical patient (i.e., population prediction) are log-transformed concentrations. b The quantile–quantile (Q–Q) plot of the normalized prediction distribution errors (NPDEs)

Full size image

3.5 Final Population Pharmacokinetic Model

The parameter estimates of the final popPK model are described in Table 5.

Table 5 Parameter estimates of the final population PK model
Full size table

The popPK parameters could be estimated with adequate precision as expressed by their low RSE %, ranging from 1 % to 16 %. k e was estimated at 0.145 h−1, resulting in a CL/F of 14.4 L/h in men and 12.0 L/h in women. The estimated IIV on k e was low (23 % CV). An additive error model was used to describe the residual variability for the log-transformed data and the residual error was 20.2 % for phase I PK data and 55.9 % for phase II and III sampled PK data.

In addition to the covariates identified on the absorption and distribution parameters, eGFR, dose, and genetic polymorphism were found to be significant covariates on k e. The additional covariate effect of eGFR accounts for the effect of differences in renal function on canagliflozin clearance, after having corrected for the effect of sex, age, and WT differences on V c/F. Volunteers with mild (60–89 mL/min/1.73 m2), moderate (30–59 mL/min/1.73 m2), and severe (15–29 mL/min/1.73 m2) renal impairment had higher median dose-normalized AUC values (~11, 40, and 29 %, respectively) than volunteers with normal renal function (≥90 mL/min/1.73 m2) (Table 6).

Table 6 Ratios of dose-normalized AUC values with 95 % CIs for covariates identified on V c/F and k e potentially affecting its plasma concentrations
Full size table

An additional significant covariate effect of total daily dose on k e was observed. The median dose-normalized AUC value at 300 mg was approximately 16 and 7 % higher than at 50 and 100 mg, respectively (Table 6). PSD did not appear to affect the PK of canagliflozin, while a significant drop in OFV was observed for the effect of genetic polymorphism on k e (Table 4). The impact of genetic polymorphism on AUC is presented in Table 6. From this popPK analysis, volunteers (N = 21) carrying the UGT1A9*3 allele had somewhat higher (about 26 %) exposure (median dose-normalized AUC values) than volunteers not carrying this allele (N = 700). No genotype information was available for N = 895 volunteers.

In a 12-week, phase II, dose-ranging study in patients with T2DM, administration of canagliflozin at doses ranging from 50 to 300 mg once daily (qd) and 300 mg twice daily (bid) were deemed to be safe and were generally well tolerated. The AUC τ,ss of 300 mg bid was 93 % higher than the highest recommended dose of 300 mg qd. Given the small magnitude of the effects of sex, age, WT, presence of the UGT1A9*3 allele, or renal function on canagliflozin exposures, they are expected to be well within that of the dose range shown to be safe and well tolerated. Therefore, no dosage adjustment is required on the basis of these factors.

The goodness-of-fit plots suggest that the final model adequately described the observed concentration data (Fig. 4). Overall, the final popPK model characterized well the observed concentrations of canagliflozin in the studies, with good precision on the parameter estimates. The pc-VPCs of the final model (Fig. 5) indicate that the model captured the temporal profile adequately for the whole set of observations.

Fig. 4
figure 4

Goodness-of-fit diagnostic plots for the final population pharmacokinetic model. Dotted line (black) reference identity line; full line (gray) lowess smoother. Observed concentrations, population and individual predictions are log-transformed concentrations. CWRES conditional weighted residuals

Full size image
Fig. 5
figure 5

Prediction-corrected visual predictive checks with 90 % inter-percentile range for the final population PK model. The solid red line represents the median prediction-corrected observed plasma concentrations. The dashed red lines represent the observed 5 and 95 % percentiles. The 95 % confidence intervals for the corresponding model predicted percentiles are shown as gray fields. DV dependent variable

Full size image

4 Discussion

A two-compartment popPK model with lag-time and sequential zero- and first-order absorption and first-order elimination with IIV on V c/F, k e, k a, k 32, and T lag best fitted the data. The popPK model was parameterized in terms of rate constants instead of primary PK parameters (CL/F and V/F). In fact, whether the NONMEM control stream is specified using rate constants (k e) or primary PK parameters (CL/F, V/F), the associated NM-TRAN translator routines (TRANS1 and TRANS3, respectively) ensure that the internal model fitting procedures compute the kinetic equations in terms of rate constants [16–18]. In other words, internally NONMEM always computes the kinetic equations in terms of k e and, when asked, TRANS3 returns CL/F as the product of k e and V/F. It can be demonstrated that, as a result, any covariates evaluated on V/F will internally also be evaluated (inversely) on k e, although that is not apparent from the TRANS3 control stream. This allows for computational dependencies between the covariate effects estimated on V/F and those estimated on CL/F, which remain hidden for the unsuspecting end user. For the popPK analysis of canagliflozin, we opted for maximal statistical transparency during model development using TRANS1 and explicitly performing the transformation of k e to the pharmacologically more meaningful CL/F, rather than letting TRANS3 perform this transformation implicitly and hidden to the end user.

The popPK model was developed first on the richly sampled phase I data (N = 245), which contained the most information on the PK profile of canagliflozin, especially for the absorption and distribution phases. A strong partial inverse correlation was observed between k a and T lag that might be caused by either a real relationship between k a and T lag or possibly by insufficient information in the absorption phase to estimate IIV on k a and T lag independently. Attempts to use a block structure for the IIVs of k a and T lag failed, but otherwise the model performed well on the richly sampled data.

As the richly sampled data from phase I was pooled with increasing amounts of sparsely sampled data from phase II and III, which included only pre-dose trough concentrations, the model started to show more and more signs of instability (reported by NONMEM as singular non-positive semi-definite R or S matrices) and eventually failed to minimize altogether. This was accompanied by increasing shrinkage of the random effects on all PK parameters except k e, which could be attributed entirely to the sparsely sampled patients. This suggested that the lack of information with regard to the absorption and distribution characteristics of canagliflozin for the sparsely sampled patients, increasingly inhibited the model in obtaining unambiguous estimates for the parameters describing these characteristics as more sparsely sampled patients were added to the dataset. Therefore, extensive covariate analyses were performed for the absorption (k a, T lag, and D 1) and distribution (V c/F, k 23, and k 32) parameters on the richly sampled phase I data. The estimates for these parameters as obtained on the richly sampled data, including their covariate relationships and the covariance matrix of their random effects, were subsequently fixed when fitting the pooled richly and sparsely sampled PK data together. This resulted in a stable model fit with a successful covariance step, and allowed a covariate search for k e on the pooled richly and sparsely sampled data. An additional covariate search for V c/F on the pooled data was not feasible because of recurring minimization problems. As noted above, the high shrinkage observed for the random-effects distribution for V c/F, but not for k e, when evaluated on the sparse phase II and III PK data, indicated that these data were much less informative for V c/F than for k e. Moreover, estimating V c/F on the sparse data did not significantly improve the model fit, which is to be expected when trying to estimate two PK parameters on just a single, pre-dose, PK time point.

The popPK parameters of the final popPK model could be estimated with adequate precision as expressed by their low RSE %, ranging from 1 to 16 %. The high variability observed on the absorption parameters [k a (123 % CV) and T lag (79 % CV)] is most probably owing to the limited PK data collected around the time to maximum plasma concentrations. The residual error was 20.2 and 55.9 % for richly (phase I) and sparsely (phase II and III) sampled PK data, respectively. The higher residual error observed for sparsely sampled PK data vs richly sampled PK data was expected as greater variability is commonly observed for data obtained in multicenter phase II and III studies when compared with tightly controlled, single-center phase I studies. The predictability of the model was evaluated by pc-VPC, and the pc-VPC of the final phase I model revealed an acceptable predictive performance in the absorption phase. An over-prediction was observed in the elimination phase but this misspecification disappeared following inclusion of the phase II and III data and the covariate effects on the elimination.

Results from this analysis indicated that participants with WT <78.2 kg are likely to have about 33 % higher median dose-normalized AUC values than participants with WT >95.2 kg. This is to be expected as the volume of plasma and tissue in which canagliflozin can distribute may tend to increase with body size and therefore with WT [19]. Results from this analysis also indicated that dose-normalized AUC values were 29 and 22 % higher in volunteers aged older than 60 years and in women, respectively, owing to the reduced V c/F [19, 20]. These increases in exposure are not considered to have a clinically relevant effect on the PK of canagliflozin as explained below. Because of the limitations of comparing observed PK parameters between studies conducted in demographically different populations of small numbers of patients, the final popPK model, which incorporates covariates shown to affect canagliflozin PK, was used to predict steady-state canagliflozin AUC24 h values for a typical population with T2DM for the highest proposed to-be-marketed dose (300 mg qd) and the highest dose studied in long-term clinical studies (300 mg bid). The final popPK model predicted that the mean steady-state canagliflozin AUC24 h values for the 300-mg qd and 300-mg bid doses were 24,941 ng·h/mL (95 % CI: 24,311–25,571 ng·h/mL) and 54,258 ng·h/mL (95 % CI: 50,623–57,893 ng·h/mL), respectively. Thus, based on these predictions, increases in canagliflozin steady-state AUC24 h by up to 118 % compared with that for a 300-mg qd canagliflozin dose would not be considered to pose any safety concerns. The absence of safety signals in the 12-week, phase II, dose-ranging study in patients with T2DM treated with canagliflozin 300 mg bid supports this assessment [21]. However, because patients aged 65 years and older may be more sensitive to adverse reactions related to reduced intravascular volume by canagliflozin, correcting this condition prior to initiation of canagliflozin is recommended in patients with volume depletion [2].

A clear relationship between baseline renal function categories and canagliflozin dose-normalized AUC was observed. The median dose-normalized AUC values were about 11, 40, and 29 % higher in volunteers with mild, moderate, and severe renal impairment, respectively, compared with the normal renal function group. These observations are generally consistent with a phase I study conducted in participants with impaired renal function [22]. Similar observations were also made with dapagliflozin. The concentrations of dapagliflozin were increased by 4, 6, and 9 % and its metabolite D3OG by 20, 37, and 52 %, respectively in patients with mild, moderate, and severe renal impairment compared with those with normal kidney function [23]. Empagliflozin showed an increase of 18, 20, and 66 % in patients with mild, moderate, and severe renal impairment [24]. Although systemic exposure to canagliflozin is increased in patients with renal impairment relative to those with normal renal function, the efficacy is reduced in patients with renal impairment owing to the reduced filtered glucose load. In addition, the rate of adverse events (osmotic diuresis, urinary tract infections, and reduced intravascular volume) was slightly higher with 300 mg qd compared with 100 mg qd in participants with stage 3 chronic kidney disease [22, 25]. Therefore, canagliflozin is recommended at a dose of 100 or 300 mg qd for T2DM patients with an eGFR ≥60 mL/min/1.73 m2, and canagliflozin 100 mg qd is recommended in T2DM patients with an eGFR of 45 to <60 mL/min/1.73 m2.

The systemic exposure (C max and AUC∞) of canagliflozin increased dose proportionally following single doses of 50, 100, and 300 mg canagliflozin in healthy volunteers [26]. The final PK parameters were used to estimate the exposure (AUC) for each volunteer who participated in the dose-proportionality study [26] and the AUC values were compared with those calculated by Non-compartmental analysis (NCA) (unpublished data). The mean AUC values estimated by the final model were in good agreement with the AUC values calculated by NCA and therefore, although a significant effect of dose on k e was observed, the impact of dose on systemic exposure of canagliflozin is expected to be minimal (data on file).

As canagliflozin is glucuronidated by UGT1A9 and UGT2B4 enzymes, its PK potentially could be influenced by the polymorphism in the genes encoding these drug-metabolizing enzymes. The participants carrying the UGT1A9*3 allele only represented 1 % of the total number of participants included in the popPK analysis and it was observed that participants carrying the UGT1A9*3 allele had higher (about 26 %) dose-normalized AUCs than those without this allele. These findings are consistent with the exposure-pharmacogenomic analysis results, supporting a role for UGT1A9 in the metabolism of canagliflozin [27]. In addition, when considering the safety data from UGT1A9*3 carriers, there appeared to be no increase in the overall incidence of adverse events or in study drug-related adverse events in participants treated with canagliflozin 100 and 300 mg relative to placebo-treated participants. Given the small effect size of this covariate on the PK (AUC) of canagliflozin, the low frequency of the variation in the population, and exposure data from a 300-mg bid cohort in the phase II study (NCT00642278), it was not deemed to be of clinical relevance.

Glycemic status, race, fed status, long-term concomitant medications (substrates of UGT1A9 or UGT2B4), and PSD were not identified as statistically significant covariates affecting the PK of canagliflozin. Therefore, no dose adjustments of canagliflozin are needed based on these factors.

5 Conclusion

The developed popPK model adequately described canagliflozin PK in healthy volunteers and in patients with T2DM. No dose adjustments of canagliflozin are needed based on BMI, body weight, sex, race, and genetic polymorphisms (with respect to the UGT1A9*3 allele). The covariate age (in patients with age-appropriate renal function) and eGFR also had no clinically relevant effect on the PK of canagliflozin. However, because patients aged 65 years and older and patients with impaired renal function (eGFR <60 mL/min/1.73 m2) may be more sensitive to adverse reactions related to reduced intravascular volume, correcting this condition prior to initiation of canagliflozin is recommended in patients with volume depletion. Based on a phase III program, canagliflozin is recommended at a dose of 100 or 300 mg qd for T2DM patients with eGFR ≥60 mL/min/1.73 m2, and canagliflozin 100 mg qd is recommended in T2DM patients with an eGFR of 45 to <60 mL/min/1.73 m2.