Model‐Informed Development and Registration of a Once‐Daily Regimen of Extended‐Release Tofacitinib

Abstract

Extended‐release (XR) formulations enable less frequent dosing vs. conventional (e.g., immediate release (IR)) formulations. Regulatory registration of such formulations typically requires pharmacokinetic (PK) and clinical efficacy data. Here we illustrate a model‐informed, exposure–response (E‐R) approach to translate controlled trial data from one formulation to another without a phase III trial, using a tofacitinib case study. Tofacitinib is an oral Janus kinase (JAK) inhibitor for the treatment of rheumatoid arthritis (RA). E‐R analyses were conducted using validated clinical endpoints from phase II dose–response and nonclinical dose fractionation studies of the IR formulation. Consistent with the delay in clinical response dynamics relative to PK, average concentration was established as the relevant PK parameter for tofacitinib efficacy and supported pharmacodynamic similarity. These evaluations, alongside demonstrated equivalence in total systemic exposure between IR and XR formulations, provided the basis for the regulatory approval of tofacitinib XR once daily by the US Food and Drug Administration.

Study Highlights.

WHAT IS THE CURRENT KNOWLEDGE ON THE TOPIC?

☑ Registration of alternative formulations, doses, and regimens to approved drugs has typically required confirmatory clinical efficacy trials, despite regulatory guidance describing the potential for well‐understood E‐R relationships as the basis for translating efficacy and safety.

WHAT QUESTION DID THIS STUDY ADDRESS?

☑ The study addressed the question of which PK parameter was most relevant for tofacitinib efficacy and whether the body of evidence from E‐R analyses supported the conclusion of similar efficacy between IR and XR formulations.

WHAT THIS STUDY ADDS TO OUR KNOWLEDGE

☑ The multidimensional analyses demonstrated that average concentration over the dosing interval is the relevant PK parameter for tofacitinib efficacy. E‐R analyses and PK studies provided the evidence to conclude that efficacy of tofacitinib XR will be similar to that of tofacitinib IR, thereby serving as the basis for registration without a phase III study.

HOW THIS MIGHT CHANGE CLINICAL PHARMACOLOGY OR TRANSLATIONAL SCIENCE

☑ Robust dose–response studies and E‐R relationships can facilitate efficient drug development and registration strategies, including providing sufficient evidence without the need for confirmatory clinical trials.

Extended‐release (XR) formulations release the active ingredient at an intentionally modified rate relative to the conventional/immediate‐release (IR) formulations in order to achieve treatment goals, which may include improved convenience and compliance through less frequent dosing and/or improved benefit:risk through modifications to the pharmacokinetic (PK) profile. A robust understanding of the PK and pharmacodynamic (PD) attributes of a drug via model‐based approaches provides the cornerstone to the development of these alternative dosage forms.1 A well‐defined exposure–response (E‐R) relationship can enable translation of efficacy and safety from one formulation to another, as has been described in the 1998 United States (US) Food and Drug Administration (FDA) guidance on clinical effectiveness.2 However, we are not aware of a previous application of model‐informed bridging between alternate regimens and formulations without a phase III trial of the new formulation in the relevant patient population. The purpose of the current investigation was to illustrate this application using a case study of tofacitinib, where E‐R relationships served as the basis for translating controlled trial data from the original IR formulation to the XR formulation to support the registration of an XR dosage form.

Tofacitinib is an oral Janus kinase (JAK) inhibitor for the treatment of rheumatoid arthritis (RA). This chronic autoimmune disease is characterized by synovial inflammation and hyperplasia, autoantibody production, and cartilage and bone destruction.3 The clinical effectiveness of tofacitinib in RA was demonstrated in several phase II,4, 5, 6, 7, 8 phase III,9, 10, 11, 12, 13, 14 and long‐term extension15 trials. Tofacitinib IR was first approved in the US in 2012 at a dose of 5 mg twice daily (b.i.d.).

For chronic conditions in some patients, a once‐daily (q.d.) dosing option offers a greater degree of compliance compared to more frequent dosing regimens.16, 17 To enable q.d. dosing with tofacitinib, an XR formulation based on extrudable core system technology was developed at a dose of 11 mg.18 A series of biopharmaceutical studies in healthy volunteers characterized the PK properties. Results from these studies demonstrated equivalence, using the standard bioequivalence (80–125%) criteria, in both area under the plasma concentration–time curve (AUC) and maximum plasma concentration (C_max) of XR 11 mg q.d. compared to IR 5 mg b.i.d. At steady state, minimum plasma concentration (C_min) for the XR formulation was 29% lower than the IR formulation.18

The primary objective of the current investigation was to determine whether a similar level of efficacy between XR 11 mg q.d. and IR 5 mg b.i.d. could be concluded on the basis of E‐R evaluations of nonclinical and clinical data from randomized controlled trials of the IR formulation. Specifically, our objectives were to characterize the PK parameter (AUC or C_max or C_min) that was most relevant for efficacy and evaluate the clinical relevance of differences in C_min between the two formulations.

RESULTS

A set of complementary E‐R analyses was performed, which consisted of: 1) identification of the PK parameter most predictive of tofacitinib efficacy in a nonclinical model of inflammation; 2) characterization of delay in the dynamics of clinical response and PK time‐course; 3) evaluation of the impact of C_min differences on clinical efficacy when the IR formulation was administered in q.d. and b.i.d. regimens; and 4) determination of the PK parameter that best described clinical efficacy.

Modeling nonclinical efficacy data

The relationship between efficacy and dosing regimen was evaluated in a murine collagen‐induced arthritis (mCIA) model, wherein vehicle and a range of fixed total daily doses of tofacitinib were administered either q.d. or b.i.d.19 Figure 1 shows the relationship from a maximum effect (E_max) model between fractional area under the severity time course and drug‐exposure predictors. For each PK parameter, the respective concentration that produced 50% of the maximal efficacy (i.e., maximum concentration producing 50% of the maximal effect (EC_max50), average concentration producing 50% of the maximal effect (EC_av50), minimum concentration producing 50% of the maximal effect (EC_min50)) was determined.

Figure 1.

Tofacitinib exposure‐response relationship in the mCIA inflammation model following q.d. and b.i.d. dosing regimens. Filled circles are mean efficacy and mean C_max, C_av, or C_min for individual cohorts (n = 10–15/dose group) of animals orally dosed q.d. (gray) and b.i.d. (black) with tofacitinib. Lines represent the best‐fit nonlinear regression for q.d. (gray) and b.i.d. (black) administrations. The EC_max50 values for the b.i.d. and q.d. regimens were 361 nM and 1,540 nM, respectively; corresponding values for EC_av50 were 102 nM and 187 nM. The EC_min50 values were 8.5 nM and 0.101 nM for the b.i.d. and q.d. regimens, respectively.95% CI, 95% confidence interval; AUC_d, area under the concentration‐severity curve for drug treatment; AUC_v, area under the concentration‐severity curve for vehicle; b.i.d., twice daily; C_av, average drug concentration in the dosing interval; C_max, maximum plasma concentration; C_min, minimum plasma concentration; EC_av50, average concentration producing 50% of the maximal effect; EC_max50, maximum concentration producing 50% of the maximal effect; EC_min50, minimum concentration producing 50% of the maximal effect; mCIA, murine collagen‐induced arthritis; nM, nanoMolar; q.d., once daily.

Good alignment of the q.d. and b.i.d. E‐R curves was observed when average drug concentration in the dosing interval (C_av, which offers the same interpretation as AUC since C_av is proportionally related to AUC as AUC/dosing interval) was used as the predictor of response. In contrast, significant divergence was observed when C_max or C_min was used as the predictor. Supporting this observation, while the mean EC_av50 (Figure 1) between q.d. and b.i.d. regimens was similar (187 nM vs. 102 nM, respectively), the corresponding values for C_max (EC_max50: 1,540 nM vs. 361 nM) and C_min (EC_min50: 0.10 nM vs. 8.5 nM) were discordant.

E‐R evaluation of clinical data

Data from five phase II studies of the tofacitinib IR formulation in RA patients were included in E‐R evaluations. These phase II data were used because they included several doses and dosing regimens, allowing better E‐R characterization. Collectively, these analyses included data from ∼1,350 patients with RA, encompassing a 30‐fold dose range of tofacitinib IR (1–30 mg b.i.d. and 20 mg q.d.) and treatment durations ranging from 6–24 weeks.

Two well‐validated clinical endpoints were included in the efficacy bridging analyses. Disease activity score using the 28‐joint count (DAS28) is a continuous composite endpoint that measures joint tenderness, joint swelling, patient global assessment, and C‐reactive protein as a laboratory marker of inflammation.20 The American College of Rheumatology (ACR)‐based responder rate is a categorical composite endpoint that includes physician‐ and patient‐assessed components, as well as a laboratory test for inflammation.21 Three ACR threshold values (ACR20, ACR50, and ACR70) are commonly used, reflecting the proportion of patients achieving 20, 50, or 70% improvement from baseline.

Characterization of delay in dynamics of efficacy response

A longitudinal DAS28 model, incorporating a hysteresis component, characterized the delay in the dynamics of DAS28 response. Onset rate of change in DAS28 was modeled as a function of dose. Details of the final model, including parameter estimates and visual predictive check (VPC) plots, are included in Supplementary Section 1. VPC evaluation showed that the model provided an adequate fit to the data. The onset half‐life of DAS28 was estimated to be ∼1 week for doses less than IR 5 mg b.i.d. and 3 weeks for IR 5 mg b.i.d.

Based on the estimated hysteresis function, which represents the delay due to distribution of the drug to the effect site and/or modulation of the biological cascade by tofacitinib, and consistent with an indirect response mechanism,22 the DAS28 model was used to predict theoretical mediator concentrations (TMC) determining the clinical response to tofacitinib. This approach allowed visualization of the TMCs, which are in‐phase with the PD effect, for the IR and XR formulations given the expected plasma concentration–time profiles. As shown in Figure 2, the TMCs (right panel) were nearly superimposable for IR 5 mg b.i.d. and XR 11 mg q.d.

Figure 2.

Predicted steady‐state tofacitinib PK profile and TMC following IR 5 mg b.i.d. and XR 11 mg q.d. doses in RA patients. N = number of patients at each dose level. Solid and dashed lines represent XR 11 mg q.d. and IR 5 mg b.i.d., respectively. b.i.d., twice daily; h, hours; IR, immediate release; ng/mL, nanograms/milliliter; PK, pharmacokinetic; q.d., once daily; RA, rheumatoid arthritis; TMC, theoretical mediator concentrations; XR, extended release.

Evaluation of efficacy of IR q.d. vs. IR b.i.d. regimens of tofacitinib

The importance of C_min to the efficacy of tofacitinib was tested using data from a phase II study (NCT00413660) in which RA patients received tofacitinib IR 1, 3, 5, 10, 15 mg b.i.d. or IR 20 mg q.d. or placebo. With the same total daily dose, AUC over 24 h was similar between IR 20 mg q.d. and IR 10 mg b.i.d. regimens. In contrast, C_min was ∼7‐fold (86%) lower and C_max ∼2‐fold higher for IR 20 mg q.d. relative to IR 10 mg b.i.d. Comparison of efficacy measures suggested similar efficacy6 between the two regimens for both DAS28 change from baseline ((CFB) –1.72 for IR 20 mg q.d. vs. –1.82 for IR 10 mg b.i.d.) and ACR20/50/70 rates (56/36/24% for IR 20 mg q.d. vs. 58/28/12% for IR 10 mg b.i.d.).

A dose–response (D‐R) model for b.i.d. doses at Week 12 (primary timepoint)23 was constructed for DAS28, DAS28 CFB, and ACR response rates (see Supplementary Table S1 for parameter estimates). The observed data for IR 20 mg q.d. was overlaid on the b.i.d. D‐R curve to evaluate consistency in efficacy relative to the predicted b.i.d. D‐R profile (Figures 3 , 4 ). For both efficacy measures, IR 20 mg q.d. was well aligned with the b.i.d. D‐R curves.

Figure 3.

Dose–response relationship of DAS28 at Week 12 in RA patients receiving tofacitinib IR tablets as a b.i.d. regimen. X‐axis shows tofacitinib dose on a total daily dose scale. Data for IR 20 mg q.d. and IR 10 mg b.i.d. are plotted at the same total daily location, with some spacing in between for resolution. Solid and dashed lines represent typical model prediction and 80% PI, respectively. Observed data (means) are shown by filled circles, with black circles for b.i.d. doses and gray for IR 20 mg q.d. dose. Error bars on filled circles are empirical 95% CIs. N, number of patients at each dose level; PI, prediction interval; b.i.d., twice daily; DAS28‐3(CRP), disease activity score using 28‐joint counts and C‐reactive protein with three variables; IR, immediate release; mg, milligram; q.d., once daily; RA, rheumatoid arthritis.

Figure 4.

Dose–response relationship of proportion of ACR20/50/70 responders at Week 12 in RA patients receiving tofacitinib IR tablets as a b.i.d. regimen. Doses are presented on a total daily dose scale. Data for IR 20 mg q.d. and IR 10 mg b.i.d. are plotted at the same total daily location, with some spacing in between for resolution. Solid and dashed lines represent model‐predicted posterior mean and 80% PI, respectively. Observed data (means) are shown by filled symbols, with black symbols for b.i.d. doses and gray for IR 20 mg q.d. dose. 20 mg q.d. data is placed either at an x‐axis value of 20 (reflecting the same AUC₂₄ as 10 mg b.i.d.) or at an x‐axis value of 2.8 (reflecting the 86% lower C_min compared to 10 mg b.i.d.), respectively. The dashed arrow represents the comparison of responses based on AUC₂₄ and C_min values for 20 mg q.d. regimen. %, percentage of responders; ACR, American College of Rheumatology; AUC₂₄, area under the concentration–time profile from time 0 to 24 h; b.i.d., twice daily; C_min, minimum plasma concentration; E_max, maximum drug effect; IR, immediate release; mg, milligrams; q.d., once daily; SE, standard error.

To further assess whether efficacy is consistent with AUC or C_min, the IR 20 mg q.d. ACR responses were plotted at two different locations on the x‐axis: one corresponding to a total daily dose of 20 mg (gray square in the figure), and the other corresponding to a total daily dose of 2.8 mg (gray circle) to reflect the 7‐fold lower C_min for IR 20 mg q.d. compared to IR 10 mg b.i.d. As shown in Figure 4, the gray squares (AUC) were within the prediction intervals (PIs) of the b.i.d. D‐R curve on all three measures and consistent with AUC as the driver. On the other hand, the gray circles (C_min) were higher than would be predicted from the b.i.d. curves, particularly for ACR50 and 70, indicating that C_min is not predictive of the efficacy of tofacitinib.

Delineation of predictive abilities of C_max, C_min, and C_av for efficacy endpoints

The predictive abilities of tofacitinib exposure metrics were compared through an E_max model for DAS28 (Table 1) and an ordered categorical E_max model (Table 2) for ACR responses (see Methods).

Table 1.

Summary of E‐R models for DAS28 at Week 12

Model IDa	Model parameters		AIC	OFV	Result
	PK predictor	PK parameter as covariate
First stage
AVG (Model 1)	Cav	NA	1,286.557	1,270.557	Model with Cav as the predictor has the lowest AIC
MAX (Model 2)	Cmax	NA	1,294.381	1,278.381	Relative to Model 1, AIC ∼8 point higher
MIN (Model 3)	Cmin	NA	1,299.139	1,283.139	Relative to Model 1, AIC ∼13 point higher
Second stage
AVG‐1 (Model 4)	Cav	Cmin as covariate on EC50	1,288.436	1,270.436	Relative to Model 1, OFV was essentially unchanged (ΔOFV = 0.121); addition of Cmin as EC50 covariate to a Cav‐only model did not show improvement.
AVG‐2 (Model 5)	Cav	Cmin as covariate on Emax	1,288.548	1,270.548	Relative to Model 1, OFV was essentially unchanged (ΔOFV = 0.009); addition of Cmin as Emax covariate to a Cav‐only model did not show improvement.
Third stage (sensitivity analysis)
MIN‐1 (Model 6)	Cmin	Cav as covariate on EC50	1,295.499	1,277.499	Relative to Model 3, ΔOFV = 5.640 decrease; addition of Cav as EC50 covariate to a Cmin‐only model showed improvement.
MIN‐2 (Model 7)	Cmin	Cav as covariate on Emax	1,299.045	1,281.045	Relative to Model 3, ΔOFV = 2.094 decrease; addition of Cav as Emax covariate to a Cmin‐only model did not show improvement.

AIC, Akaike Information Criterion; C_av, average drug concentration in the dosing interval; C_max, maximum plasma concentration; C_min, minimum plasma concentration; DAS28, disease activity score using 28‐joint counts; E_max, maximum drug effect; EC₅₀, concentration producing 50% of the maximum effect; E‐R, exposure response; ID, identification; NA, not applicable; OFV, objective function value (‐2LogLikelihood); PK, pharmacokinetic; ΔOFV, difference in OFV between reduced and test model; χ², chi‐square.

^aModel ID is provided to facilitate differentiation between models appearing in this table. Significance was assessed by comparing ΔOFV against a χ² distribution with one degree of freedom. This critical χ² value is 3.84 and is equivalent to a P‐value of 0.05.

Table 2.

Summary of E‐R Models for ACR response rates at Week 12

Model IDa	Model parameters		AIC	OFV	Results
	PK parameter	Secondary parameter as covariate
First stage
AVG (Model 1)	Cav	NA	2,371.987	2,337.987	Model with Cav as the predictor has the lowest AIC
MAX (Model 2)	Cmax	NA	2,374.001	2,340.001	Relative to Model 1, AIC ∼2 points greater
MIN (Model 3)	Cmin	NA	2,399.066	2,365.066	Relative to Model 1, AIC ∼27 points greater
Second stage
AVG‐1 (Model 4)	Cav	Cmin as covariate on EC50	2,373.986	2,337.986	Relative to Model 1, OFV was essentially unchanged (ΔOFV = 0.001); addition of Cmin as EC50 covariate to a Cav‐only model did not show improvement.
AVG‐2 (Model 5)	Cav	Cmin as covariate on Emax	2,373.908	2,337.908	Relative to Model 1, OFV was essentially unchanged (ΔOFV = 0.079); addition of Cmin as Emax covariate to a Cav‐ only model did not show improvement.
Third stage (sensitivity analysis)
MIN‐1 (Model 6)	Cmin	Cav as covariate on EC50	2,390.001	2,354.001	Relative to Model 3, ΔOFV = 11.065 decrease; addition of Cav as EC50 covariate to a Cmin‐only model showed improvement.
MIN‐2 (Model 7)	Cmin	Cav as covariate on Emax	2,393.041	2,357.041	Relative to Model 3, ΔOFV = 8.025 decrease; addition of Cav as Emax covariate to a Cmin ‐only model showed improvement.

AIC, Akaike Information Criterion; ACR, American College of Rheumatology; C_av, average drug concentration in the dosing interval; C_max, maximum plasma concentration; C_min, minimum plasma concentration; E_max, maximum drug effect; EC₅₀, concentration producing 50% of the maximum effect; E‐R, exposure‐response; ID, identification; NA, not applicable; OFV, objective function value (‐2LogLikelihood); PK, pharmacokinetic; ΔOFV, difference in OFV between reduced and test model; χ², chi square.

^aModel ID is provided to facilitate differentiation between models appearing in this table. Significance was assessed by comparing ΔOFV against a χ² distribution with one degree of freedom. This critical χ² value is 3.84 and is equivalent to a P‐value of 0.05.

For both efficacy measures, in the first stage of evaluation univariate analysis favored the C_av model compared with C_max or C_min models based on lowest Akaike Information Criterion (AIC) values.24 Goodness‐of‐fit plots using C_av, C_min, or C_max as exposure metrics also confirmed better alignment of the endpoints with C_av as the predictor (Supplementary Section 2 for ACR response rates; figure not shown for DAS28). Because of C_min differences between XR and IR formulations of tofacitinib, further analyses focused on evaluating the added value of C_min over and above C_av.

In the second stage, the objective function value (OFV) for models in which C_min was additionally included as a covariate on E_max or concentration producing 50% of the maximum effect (EC₅₀) were not significantly different from C_av‐only model; 90% confidence interval (CI) for the C_min covariate effect also included zero (Supplementary Table S2 and Supplementary Section 2). These results indicated that adding C_min as a covariate did not offer any additional improvement over a C_av‐only model. Finally, in the sensitivity analyses (3rd stage), the addition of C_av as a covariate on E_max or EC₅₀ to a model with C_min as the predictor yielded significant improvements, compared with C_min alone for ACR response rates (Table 3), even with the observed correlation of 0.79 between C_av and C_min. For DAS28, inclusion of C_av as a covariate on EC₅₀ decreased OFV significantly, while the addition of C_av as a covariate on E_max did not (Table 2).

Table 3.

Characteristics of tofacitinib IR phase II studies in RA included in the E‐R analyses

Clinical Trials.gov identifier	Pfizer study number	Treatments/tofacitinib IR doses	Number of patients	Treatment duration (weeks)	Analysis
					Characterization of delay in dynamics of efficacy response	Evaluation of efficacy of IR q.d. vs. IR b.i.d. regimens of tofacitinib	Delineation of predictive abilities of Cmax, Cmin, and Cav for efficacy endpoints
NCT00147498	A3921019	PBO and 5, 15, or 30 mg b.i.d.	264	6	✓	–	–
NCT00413660	A3921025	PBO and 1, 3, 5, 10, or 15 mg b.i.d. or 20 mg q.d.	507	24	✓	✓	✓
NCT00550446	A3921035	PBO and 1, 3, 5, 10, or 15 mg b.i.d.	384	24	✓	–	✓
NCT00603512	A3921039	PBO and 1, 3, 5, or 10 mg b.i.d. (Japanese patients)	161	12	✓	–	✓
NCT00687193	A3921040	PBO and 1, 3, 5, 10, or 15 mg b.i.d. in (Japanese patients)	317	12	✓	–	✓

b.i.d., twice daily; C_av, average drug concentration in the dosing interval; C_max, maximum plasma concentration; C_min, minimum plasma concentration; E‐R, exposure‐response; IR, immediate release; PBO, placebo; q.d., once daily; RA, rheumatoid arthritis.

Taken together, the results demonstrated that C_av (or AUC) is the most predictive drug‐exposure measure of tofacitinib efficacy and that C_min did not provide additional predictive value over and above that of C_av.

DISCUSSION

The purpose of the current investigation was to use an E‐R modeling approach to inform the development and US registration of the tofacitinib XR formulation administered q.d. From a mechanistic standpoint, tofacitinib as a JAK inhibitor blocks signaling through the common gamma chain of the surface receptors for several cytokines that are central to the pathogenesis of RA, including interleukins (IL)‐7, ‐15, and ‐21.25, 26 It also attenuates signaling by proinflammatory cytokines such as IL‐6 and interferon. Because cytokine signaling promotes disease through the recruitment and activation of effector cells at sites of pathologic inflammation,27 the pharmacological effect of tofacitinib on clinical endpoints resulting from inhibition of cytokine signaling is indirect. This provides a sound scientific basis to expect that the clinical endpoints would not be significantly influenced by short‐term fluctuations in plasma concentrations within the dosing interval, but instead would be dependent on the overall average exposure over a period of time (e.g., weeks to months), as measured by AUC (or C_av).1 This hypothesis is supported by the clinical data from the IR RA development program as well as the nonclinical data, as described below.

Results from the nonclinical analyses showed concordance of E‐R curves and EC₅₀ values using C_av (ratio of EC₅₀ values (q.d./b.i.d.) ∼1.8), and divergence with either C_max (∼4.3) or C_min (∼84), supporting the relevance of C_av in predicting nonclinical antiinflammatory activity. The mCIA model has proven useful in the screening and development of new therapies for treatment of RA28 and has allowed quantitative E‐R analyses for a number of approved RA therapies with a wide array of mechanisms (methotrexate,29 abatacept,30 anakinra,31 etanercept,32 glucocorticoids33, 34, 35, 36). The dose fractionation technique that was employed in this experiment has been successfully used in other therapeutic areas37, 38 to delineate the relative effect of various PK parameters on response.

E‐R characterization from clinical data provides further evidence of the importance of C_av to tofacitinib efficacy. From the DAS28 hysteresis model in RA patients, a 3‐week onset half‐life was estimated for IR 5 mg b.i.d., indicating achievement of PD steady state in weeks compared to achievement of PK steady state within 24–48 h. Consistent with the delay, the TMC profiles were essentially superimposable between the XR and IR formulations, suggesting that within‐day fluctuations in the PK profile of tofacitinib are unlikely to confer differential effectiveness.

Clinical data from a phase IIB double‐blind study, which included tofacitinib IR 20 mg q.d. and several tofacitinib IR b.i.d. doses (1–15 mg), support the indirect response dynamics. The efficacy of IR 20 mg q.d. was similar to IR 10 mg b.i.d. across various clinical domains and consistent with the b.i.d. D‐R profiles across these endpoints. Importantly, the two 20 mg total daily dose IR regimens achieved similar AUC over 24 h but with large differences in the shape of the concentration–time course, resulting in 7‐fold (86%) lower C_min and 2‐fold higher C_max for IR 20 mg q.d. compared to IR 10 mg b.i.d.

Efficacy characterization from this study was limited by the use of a 2‐fold higher dose (IR 20 mg q.d.) compared to the approved dose (IR 5 mg b.i.d.), the relatively shallow shape of the b.i.d. dose response in this dose range, and the 2‐fold higher C_max compared to IR 10 mg b.i.d. However, if C_min were the driver, the resulting efficacy would have been similar to that of ∼IR 3 mg b.i.d., a dose with a low probability of achieving clinically meaningful responses, particularly on stringent measures of efficacy such as ACR50 and ACR70.39 Since the ACR data showed similar efficacy between IR 20 mg q.d. and 10 mg b.i.d., it can be concluded that C_max or C_min differences were not meaningful.

The efficacy bridging was substantiated by comparison of goodness‐of‐fit characteristics between PK parameters of tofacitinib. For both DAS28 and ACR response rates, models using C_av had the lowest OFV and AIC values, and the addition of C_min as a covariate on E_max or EC₅₀ showed no added value compared to the C_av‐only model. In contrast, the addition of C_av as a covariate of EC₅₀ to a model using C_min as a predictor yielded statistically significant improvements for both efficacy measures. Additionally, improvements in OFV were also noted when C_av was added as a covariate on E_max for the ACR model. Taken together, the three‐stage analysis established C_av as the better predictor of efficacy than C_min or C_max.

In summary, application of E‐R approaches from different data sources demonstrated consistently that tofacitinib AUC (or C_av) is the most relevant drug‐exposure parameter for efficacy and that the 29% lower C_min with tofacitinib XR is not clinically important to the efficacy of tofacitinib. This is supported by the observed delay in clinical response dynamics relative to PK; the similar clinical efficacy observed with IR 20 mg q.d. and IR 10 mg b.i.d. doses, despite a large difference in C_min; the improved goodness‐of‐fit characteristics with C_av compared to C_max or C_min; and the corroborative evidence from nonclinical dose fractionation data.

Given the PK profile of the XR formulation, which included equivalence on AUC and C_max, and ∼29% lower C_min relative to the IR formulation, the primary focus of our analyses was to bridge efficacy between the XR and IR formulations. Considering these PK characteristics, the safety profile of the XR formulation is also expected to be consistent with the IR formulation, given the similar or slightly lower systemic exposure parameters. Additionally, the expected duration of steady‐state plasma concentrations above the in vitro, whole‐blood concentration producing 50% of the maximum inhibition for JAK 1/3 signaling (17 ng/mL) is ∼12–13 h for both formulations over a 24‐hour period (Pfizer data on file), suggesting a similar level of target enzyme inhibition over the dosing interval.

On the basis of the demonstrated equivalence in AUC between the two formulations and the evidence from E‐R relationships that AUC is the relevant parameter for clinical response, tofacitinib XR 11 mg q.d. was granted regulatory approval by the FDA without the need for a phase III trial. Our analyses illustrate the potential of robust D‐R studies and E‐R relationships to not only facilitate efficient drug development for alternative formulations/doses/regimens but also provide evidence sufficient to obviate the need for confirmatory clinical trials.

METHODS

Modeling nonclinical efficacy data

Fixed total daily doses of tofacitinib (1, 3, 10, 30, and 100 mg/kg) and vehicle control were administered orally in q.d. and b.i.d. dosing frequencies to male Harlan Sprague‐Dawley mice (n = 10–15/treatment group) after a boost of collagen, as described previously by Dowty et al.19 Arthritis severity scores were measured at the start of the study and postboost of collagen on Day 21. Sparse PK samples were collected from all treatment groups.

Efficacy was assessed by area under the severity score time‐course (AUEC). Mean AUEC for each treatment group was subtracted from that of the respective mean vehicle control and standardized with the vehicle control to yield a fractional AUEC. Utilizing a one‐compartment PK model, AUC from time 0 to 24 h and PK parameters including C_av, C_max, and C_min were determined. Mean fractional AUEC for each treatment group was modeled as a function of mean PK parameters using a standard E_max model (GraphPad Prism, GraphPad Software, La Jolla, CA).

E‐R evaluation of clinical data

Clinical studies and endpoints

Analyses of clinical efficacy utilized data from up to five phase II randomized, placebo‐controlled, double‐blind studies from the tofacitinib IR development program.4, 5, 6, 7, 8 All studies were approved by the Institutional Review Boards (IRBs) and/or Independent Ethics Committees of each investigational center or a central IRB. The studies were conducted in compliance with the Declaration of Helsinki and the International Conference on Harmonization Good Clinical Practice Guidelines. All patients provided written informed consent. Pertinent features of the five phase II studies included in the E‐R analyses are provided in Table 3.

Characterization of delay in dynamics of efficacy response

Pooled data from five phase II studies (Table 3) were used for longitudinal DAS28 modeling. Using an E_max structure form, the model was parameterized on an exponential scale to ensure individual predictions were >0, as outlined below:

$$DAS28(t)=\text{exp}\left(Base+Placebo(t)+\frac{E_{\text{max}}\cdot S(t)}{S(t)+E_{50}}\right)+g\cdot \epsilon$$

where Base is the baseline function of parameters; Placebo (t) represents the nondrug function of parameters when exposure is 0; S(t) refers to the mediator concentrations that are in‐phase with the drug effect at time t; E_max represents the maximum drug effect; E₅₀ is the S(t) that achieves ½ E_max; g represents the residual variance function by study; and ε is the residual random effect assumed to be normally distributed with mean of 0 and variance of 1. Additive intersubject variability was included on baseline, placebo, and drug‐effect parameters. Onset of drug effect or hysteresis was evaluated by inclusion of a dose‐dependent rate constant. NONMEM versions 7.2.0 and 7.3 (ICON Development Solutions, Ellicott City, MD), using the Laplace approximation was implemented. The population mean and 10th and 90th percentiles of simulated DAS28 were computed for each replicate, followed by PIs for each of these three statistics across the trial replicates via VPC.40 Model performance was determined by agreement between the 80% PIs and the observed statistics across studies and doses.

PK profiles for the XR and IR formulations were predicted using a population PK model (clearance (20.6 L/h), volume (90.2 L)) in RA patients and absorption parameters (0.189 h and 0.34 h for XR and IR formulations, respectively).41 A 9% estimated reduction in relative bioavailability (F _rel = 0.912) for the XR formulation relative to the IR was applied. The hysteresis parameter estimated from the DAS28 model was linked to the PK model to yield the TMC profile for each formulation.

Evaluation of efficacy of IR q.d. vs. IR b.i.d. regimens of tofacitinib

A D‐R model at Week 12 was constructed for DAS28 and ACR efficacy measures to evaluate the consistency of the efficacy of IR 20 mg q.d. relative to the b.i.d. dose–response curves. For DAS28, a nonlinear E_max model with additive residual error was fitted to DAS28 and DAS28 CFB. Maximum likelihood estimation, as implemented in PROC NLMIXED (SAS, Cary, NC), was used. A previously described Bayesian ACR longitudinal dose–response model was implemented for ACR response rates.42 The model is an extension of the three‐parameter E_max model with parameters (maximum drug effect: E_max; dose producing 50% of the maximum effect: ED₅₀ (measure of potency); placebo response: P0) that can change as functions of time. For both endpoints, data from b.i.d. doses were used for developing the model; IR 20 mg q.d. was overlaid on the model‐predicted b.i.d. dose–response profile.

Delineation of predictive abilities of C_max, C_min, and C_av for efficacy endpoints

To characterize the most relevant tofacitinib PK parameter for efficacy and assess the predictive value of C_min over and above C_av, efficacy data at Week 12 from four phase II studies were analyzed. Steady‐state subject‐specific PK parameters (C_av, C_max, and C_min) were predicted from a population PK model in RA patients.41 For DAS28, a standard E_max model was developed. The three ACR endpoints (20/50/70) were jointly modeled using a four‐category response model (ACR20 nonresponder, ACR20 but not ACR50 responder, ACR50 but not ACR70 responder, and ACR70 responder) represented by values of 0, 1, 2, and 3, respectively (see Supplementary Section 2). The ordered categorical approach utilizes information from ACR20, ACR50, and ACR70 endpoints in a simultaneous approach compared to individual binary models resulting in improved precision of common parameters (e.g., EC₅₀) of the E‐R model.43 The models were fitted using maximum likelihood estimation of the NLMIXED procedure in SAS v. 9.3.

Model testing was conducted in three stages. In Stage 1, each exposure metric was individually tested as the predictor in the E‐R model. In Stage 2, C_min was added to the E‐R model as a covariate of E_max or EC₅₀ to a model having C_av as the predictor, to evaluate if C_min had any added predictive value over and above C_av. In Stage 3, a sensitivity analysis was conducted by reversing the order of testing, whereby C_min was set as the predictor and C_av was included as a covariate on E_max or EC₅₀. The general form of the equation for Stages 2 and 3 is shown below:

$$D={\theta }_D\cdot [1+{\theta }_P(P-median(P))]$$

where D is the drug‐effect parameter θ_D (E_max or EC₅₀); θ_P is the fractional change in D; P is the PK parameter (C_av or C_min). Models were compared based on AIC (Stage 1, comparing non‐nested models) and OFV (Stages 2 and 3, comparing nested models) at a significance level of 0.05 (Δ OFV ≥3.84). Consideration was also given to goodness‐of‐fit plots, and CIs of the covariate effect on parameters.

CONFLICT OF INTEREST/DISCLOSURE

M.L., M.E.D., C.N., T.S., J.C., and S.K. are employees and shareholders of Pfizer Inc. M.H. has acted as a consultant for Pfizer Inc. D.C. was an employee of Pfizer Inc at the time these analyses were conducted. D.E.F. has received research support and is a consultant for Pfizer Inc. A.D. is a consultant and member of speakers' bureaus for AbbVie and Pfizer Inc.

AUTHOR CONTRIBUTIONS

S.K., M.L., M.M.H., D.E.F., A.D., M.E.D., D.C., T.S., C.N., and J.C. wrote the article; S.K., M.L., M.M.H., D.E.F., A.D., M.E.D., D.C., T.S., C.N., and J.C. designed the research; S.K., M.L., M.M.H., D.E.F.,, A.D., M.E.D., D.C., T.S., C.N., and J.C. performed the research; S.K., M.L., M.M.H., D.E.F., M.E.D., D.C., T.S., C.N., and J.C. analyzed the data.

Supporting information

Supporting Information

Supporting Information

Supporting Information

Supporting Information