Integrated Longitudinal Population Dose‐Hemoglobin Response of Daprodustat Following Dose Titration in Patients With Anemia in Chronic Kidney Disease

Abstract

Daprodustat, a novel oral hypoxia‐inducible factor prolyl hydroxylase inhibitor is approved in the United States for the treatment of anemia due to chronic kidney disease (CKD) in adults receiving dialysis for at least 4 months. Pharmacodynamic dose‐hemoglobin (Dose‐Hgb) models were developed as daprodustat progressed through development. To support global phase III development, a dose‐titration algorithm, guided by simulations from the initial Dose‐Hgb model based on phase II clinical data, was implemented. This work was to update and re‐calibrate this model to support the dose titration algorithm. Data from five pivotal phase III studies in CKD patients with anemia treated with daprodustat once daily (q.d.) and/or three times a week (t.i.w.) using a titration dosing schedule were included. The data comprised 2,770 CKD patients with anemia providing 53,535 Hgb observations over a period of 6 months up to 4 years. This final Dose‐Hgb model consisted of a precursor cell compartment and 12 transit compartments to describe the red blood cell (RBC) lifespan. Treatment increased the precursor cell production rate (K _in) by a power of allometrically scaled dose. Disease progression, as an exponential decline of Hgb production rate over time, varied with dialysis status. The dose‐titration algorithm resulted in comparable response for t.i.w. dosing relative to q.d. dosing. Titration‐based visual predictive checks for Hgb target criteria for the analysis dataset and the prediction dataset showed that the model adequately predicted the observed data. This re‐calibrated Dose‐Hgb model will provide further support for the individualized dosing strategy in CKD patients with anemia treated with daprodustat.

Study Highlights.

• WHAT IS THE CURRENT KNOWLEDGE ON THE TOPIC?

Daprodustat, an oral, small‐molecule inhibitor of prolyl hydroxylase demonstrated clinical efficacy in patients with chronic kidney disease and anemia by implementing a titration algorithm based on early phase II studies using modeling and simulation.

• WHAT QUESTION DID THIS STUDY ADDRESS?

Can the phase II daprodustat longitudinal dose‐hemoglobin model be updated with phase III data to further support individualized dosing strategy in CKD patients with anemia?

• WHAT DOES THIS STUDY ADD TO OUR KNOWLEDGE?

The model‐based drug development (MBDD) of daprodustat in CKD patients with anemia adequately predicted the average longitudinal Hgb profiles to the pre‐defined Hgb therapeutic range. In addition, the heterogenous phase III data allowed the characterization of disease progression in the long‐term studies as well as demonstrated the importance of titration‐based visual predictive checks as a model diagnostic for titration‐based studies.

• HOW MIGHT THIS CHANGE CLINICAL PHARMACOLOGY OR TRANSLATIONAL SCIENCE?

This work demonstrates the value of implementing MBDD early in clinical development to optimize dosing strategy and re‐calibration of the model using emerging phase III data, where appropriate.

Anemia is a common complication of chronic kidney disease (CKD), develops early in the course of the disease, worsens as CKD progresses, and represents a significant public health challenge with ~37 million adults in the United States (US) estimated to have CKD. ¹ Anemia of CKD affects 4.8 million subjects ² in the US. The prevalence of anemia increases as renal function (glomerular filtration rate—GFR) declines ¹ , ³ and an estimated 87% of patients requiring dialysis are anemic. ⁴

The etiology of anemia of CKD is multifactorial and includes deficiency of erythropoietin (a hormone secreted by the kidneys that increases red blood cell (RBC) production), impaired ability to absorb iron, inability to utilize stored iron, blood loss (e.g., through chronic dialysis), and shortened RBC survival. ⁵ Consequences of treatment of anemia due to CKD include reduced quality of life (QoL), higher rates of cardiovascular (CV) comorbidities, hospitalizations, and mortality. ⁶ , ⁷ , ⁸ Currently, available therapeutic options include iron, erythropoiesis‐stimulating agents (ESAs), and RBC transfusions. ESAs such as epoetin alfa and darbepoetin alfa are widely used for the treatment of anemia due to CKD, they bind to and activate the human erythropoietin receptor and stimulate RBC production in bone marrow. ESA dosing is titrated based on hemoglobin (Hgb) level and can vary from three times a week to monthly, depending on the specific agent and setting. ESA use for the treatment of anemia due to CKD in patients on dialysis and not on dialysis has spanned over 30 years and none can be orally administered. Studies showed CKD patients not on dialysis and those on home dialysis would prefer an oral treatment for anemia due to the convenience of administration, especially for those in rural areas requiring transportation for in‐clinic visits, avoiding injection pain during in‐clinic visits and drug storage requirements associated with injectable administration. ⁹ , ¹⁰ Therefore, there remains a significant need for novel, accessible treatment options for this patient population for the treatment of CKD with anemia.

Daprodustat (Jesduvroq) is the first orally bioavailable hypoxia‐inducible factor prolyl hydroxylase inhibitor (HIF‐PHI) approved in the US for the treatment of anemia due to CKD in adults who have been receiving dialysis for at least 4 months. ¹¹ It is also approved in Japan (Duvroq) for the treatment of anemia due to CKD in dialysis‐dependent (DD‐CKD) and non‐dialysis‐ (ND‐CKD) adults. ¹² It stimulates endogenous erythropoiesis and modulates iron metabolism through inhibition of HIF‐prolyl‐4‐hydroxylases. This activity results in the stabilization and nuclear accumulation of HIF‐1α and HIF‐2α transcription factors, leading to increased transcription of the HIF‐responsive genes, including erythropoietin and transferrin. ¹³ The effect of daprodustat on change in Hgb is indirect and occurs through a downstream cascade where prolyl‐4‐hydroxylase (PHD) inhibition leads to erythropoietin (EPO) stimulation and a consequent increase in Hgb. Other markers of PHD inhibition, namely, hepcidin, serum transferrin, and total iron binding capacity ¹⁴ , ¹⁵ , ¹⁶ indirectly mobilize iron utilization and metabolism thereby also contributing toward an increase in Hgb response. Hence, longitudinal dose‐ or concentration‐driven hemoglobin response with time would be more appropriate for an exposure–response assessment for daprodustat. Such lifespan models have been used to describe the Hgb response to ESAs. ¹⁷ , ¹⁸

Following availability of the phase II data in CKD patients with anemia, a semi‐mechanistic longitudinal Dose‐Hgb model ¹⁹ (herein referred to as Dose‐Hgb Model 1) contributed to the implementation of an individualized dosing scheme for the starting doses and titration algorithms applied in all subsequent phase III studies through clinical trial simulations (data in‐house), in both DD‐CKD and ND‐CKD patients either using ESA or non‐ESA treatments. This titration dosing strategy was applied successfully during the phase III program for Japan which started slightly earlier than the global phase III studies. The phase III Japan data in DD‐CKD and ND‐CKD patients were used to subsequently update Dose‐Hgb Model 1 to Dose‐Hgb Model 2. ²⁰ This manuscript focuses on the application of Dose‐Hgb Model 2 for the population analysis of daprodustat from the five pivotal, global phase III studies to show the adequacy of this titration dosing algorithm in the CKD subjects with anemia during the phase III program and will be referred to as Dose‐Hgb Model 3. A schematic of this model‐informed development of daprodustat is depicted in Figure S1 . An individualized dosing regimen of daprodustat was given either once daily (q.d., 1–24 mg) in ND‐CKD and DD‐CKD or three times a week (t.i.w., 2–48 mg) in DD‐CKD subjects. This titration dosing regimen was applied to ensure achievement and maintenance of a relatively safe target hemoglobin range of 10–12 g/dL (100–120 g/L) ²¹ depending on study objectives. It is noteworthy that during the Dose‐Hgb Model 1 development, covariate analysis showed oral clearance (determinant of AUC) of daprodustat was no better a predictor of response than dose alone. Since the lifespan of mature red blood cells (RBC) in CKD population is longer (~82 days) ¹⁸ than the elimination half‐life of daprodustat (1–4 h), ²² using allometrically scaled dose to drive Hgb response was judged suitable for this analysis especially since daprodustat dose is titrated to maintain Hgb within target range.

The objective of the present analysis was to develop Dose‐Hgb Model 3 for daprodustat using data from five pivotal, global phase III studies in DD‐CKD and ND‐CKD subjects using q.d. or t.i.w. dosing regimens, by refining and improving the characterization of system and treatment parameters, covariate effects, and disease progression of Dose‐Hgb Model 2.

METHODS

Studies and data

A summary of the five studies and patient characteristics is provided in Table 1 . The inclusion criteria of the trials mainly consisted of age limitations, disease status (i.e., CKD stage, Table S2 ), and Hgb levels at screening and randomization (Table S1 ). Dosing was titrated based on Hgb levels (Table S3 ). Four of the phase III studies included ESAs as active controls and had a target Hgb range of 10–11 g/dL. One study was placebo‐controlled with a target Hgb range of 10–12 g/dL. Blood samples for hemoglobin analysis via HemoCue (HemoCue AB, Angelholm, Sweden) and the central laboratory were collected throughout the trials. The modeling primarily used the central laboratory values which were also applied in the primary statistical analysis. Hemocue data were utilized only for missing central laboratory data.

Table 1.

Summary of description of the phase III studies in CKD patients with anemia

Study	Population	Type of study	Dosing and frequency	Planned number of patients for analysis	Study duration
205270	ESA non‐users, not on dialysis	Ph3, DB, R, PC study evaluating daprodustat in patients with anemia	daprodustat q.d. p.o. or placebo q.d. p.o., titrated to Hgb 11.0–12.0 g/dL	270 daprodustat & 270 placebo	28‐week treatment
204837	ESA users on HD	Ph3 DB, R, AC study evaluating daprodustat in patients with anemia	Daprodustat t.i.w p.o. or epoetin‐α t.i.w IV, titrated to Hgb 10.0–11.0 g/dL	240 daprodustat & 120 controla	52‐week treatment
201410	ESA non‐users, initiating HD or PD	Ph3, SB, R, AC study evaluating daprodustat in patients with anemia	Daprodustat q.d. p.o. or darbepoetin‐α Q2W/Q4W SC/IV, titrated to Hgb 10.0–11.0 g/dL	150 in daprodustat arm and 150 controla	52‐week treatment
200807	ESA users, on HD or PD	Ph3, SB, R, AC study evaluating daprodustat in patients with anemia	Daprodustat q.d. p.o. or epoetin‐α, t.i.w/Q1W IV or darbepoetin‐α Q1W/Q2W/Q4W SC, titrated to Hgb 10.0–11.0 g/dL	1,500 daprodustat and 1,500 controla	Event‐driven at least 52‐week treatment
200808	ESA non‐users (~60%) and ESA users not on dialysis (~ 40%)	Ph3, SB, R, AC evaluating daprodustat in patients with anemia	Daprodustat q.d. p.o. or darbepoetin‐α Q4W SC, titrated to Hgb 10.0–11.0 g/dL	2,250 daprodustat and 2,250 controla	Event‐driven at least 52‐week treatment

AC, active‐controlled; CKD, chronic kidney disease; DB, double‐blind; ESA, erythropoiesis‐stimulating agent; HD, hemodialysis; Hgb, hemoglobin; IV, intravenously; PC, placebo‐controlled; PD, peritoneal dialysis; p.o., oral; q.d.: once daily; QxW: once every × weeks; R, randomized; SB, single‐blind; SC, subcutaneously; t.i.w, three times a week.

^a Control data not included in the analysis.

Furthermore, it was planned that once the overall analysis‐ready dataset was available, the evaluable subjects would be randomly split 2:1 into an analysis dataset (A) and a prediction dataset (B) (Table S4 ). Dataset A was used for the main analysis. Dataset B was used for internal validation (i.e., applied to Dose‐Hgb Model 3) for diagnostics purposes.

All studies were conducted in accordance with Good Clinical Practice Guidelines and the ethical principles that have their origin in the Declaration of Helsinki. Protocols were approved by the institutional review board or ethics committee for each site, and each participant provided written informed consent before any study‐related procedures were performed.

Daprodustat updated longitudinal population dose‐Hgb model

Dose‐Hgb Model 2

The starting point for analysis of the global phase III studies was Dose‐Hgb Model 2. ²⁰ This model was based on the publication of Lledo‐Garcia et al. ²³ and consisted of a precursor cell compartment and 12 transit compartments to describe the RBC dynamics and lifespan. Daprodustat was assumed to stimulate the precursor cell production rate via a power function of dose that was allometrically scaled with body weight. The model had patient group‐specific baselines, depending on dialysis status and prior ESA use. ESA users had an apparent baseline Hgb before treatment with daprodustat while they were still on ESA, and a true baseline, that is, the baseline after washout of the ESA effect. ESA non‐users had the same Hgb baseline before and after treatment started. In addition, the model included covariates, namely:

• CKD stage,

• Hepcidin concentration,

• Body weight on all Hgb baselines except for the apparent baseline for ESA users,

• Iron use on both Hgb baselines of ESA non‐users,

• Prior ESA dosing on the true Hgb baseline of ESA users,

• Body weight on the daily dose,

• ESA hyporesponder on Hgb baseline of ESA users, slope, and RBC lifespan

• A placebo effect was also implemented that modulated the RBC production rate and Hgb baseline after the start of treatment. Inclusion of a disease progression function based on Hgb changes was not feasible at that time as most of the studies in the model were up to 6‐month duration and only two of the studies with duration of 1 year.

Dose‐Hgb Model 3

Data from the five pivotal phase III studies were used to refine Dose‐Hgb Model 2. The initial re‐estimation of the model parameters with the new data showed some numerical instability requiring the addition of a disease progression function to maintain stability. For this stabilization step, a disease progression model affecting RBC production was added whereby red blood cell production was reduced exponentially over time. This replaced the simpler component of the previous model (called “Treatment placebo effect”) that reduced Hgb levels by a fixed fraction. Most likely, this parametrization was no longer an adequate approach due to the longer duration of two of the landmark studies included in this dataset (up to 4 years, instead of at most 1 year in the previous model). Inter‐individual variability (IIV) was included in the disease progression rate, with three separate additive IIV parameters (entered linearly in model): namely, the first, second, and third year thereafter. This was motivated by the fact that initial exploratory analysis showed that variability decreased over the course of time, and moreover, the individual disease progression rate was found to vary over time suggesting that the model fit would likely improve by resampling the individual parameter. Due to the inclusion of the disease progression parameters, estimation of the precursor and RBC lifespan estimation was challenging (unrealistically high values) and hence the values were fixed to those estimated in both Dose‐Hgb Model 1 and Model 2. These values are within the range quoted in the literature. ¹⁷ , ²⁴ However, interindividual variability in RBC lifespan was estimated with the effect of various covariates tested on this parameter.

For the optimization step, some model changes were made to improve the fit and removal of non‐relevant model components to obtain a base model. Model performance was evaluated by comparing the objective function value (OFV), parameter values and precision, standard diagnostics, and titration‐based visual predictive checks (VPC), ²⁵ , ²⁶ , ²⁷ stratified by study and previous ESA use. These titration‐based VPCs partly corrected for dropout, as simulated subjects in the VPCs either dropped out due to reaching the criteria for rescue treatment (Table S5 ), or they dropped out at the same time as in the studies. This way, correlations between dropout and covariates were retained, but correlations would be lost due to other plausible causes of dropout not captured by data. This was a reasonable approximation, without necessitating the development of a dropout model. This base model included the covariates identified from Dose‐Hgb Model 2, but did not include any of the additional covariates planned to be tested in this analysis from the phase III studies.

Covariate Testing: The covariates of interest were pre‐defined before the analysis, based on biological/pharmacological/physiological plausibility as well as based on prior knowledge from the previous dose‐Hgb population models. The first step in the covariate analysis was a graphical evaluation of covariate distributions to investigate correlations between covariates, differences in distribution between studies, treatment groups, dialysis status, ESA‐use and covariate range in the studied population, taking into account the effect of shrinkage on the ability to graphically assess covariate relationships. The covariates that were selected for the formal analysis are summarized in Table S6 . Thus, following development of the base model and graphical assessment of covariate vs. inter‐individual random effects (η) relationships, the selected covariates were included in a model‐based evaluation using stepwise covariate modeling (SCM). ²⁸ This technique tests for covariate parameter relationships (Supplementary Methods—covariate selection) in a forward fashion (P < 0.01 and ∆OFV > 6.63 for one parameter, χ², 1 degree of freedom (df)) to build up the full model, which in turn are evaluated in the backward elimination step (P < 0.001 and ∆OFV > 10.8 for one parameter, χ², df = 1).

Target Range Statistics: Statistics on fractions within, above, and below the Hgb target range were used to compare predictions and simulations to corresponding observations. Predictions were obtained with the final Dose‐Hgb model. Simulated values were obtained from the titration‐based VPC. Another diagnostic compared observed mean dose over time with titration‐based VPCs stratified by study and ESA use.

Dropouts: The potential impact of selective dropout was monitored for diagnostic purposes during model development, in particular by checking that the 5^th and 95^th percentile predictions in the titration‐based VPCs matched observations for the full duration of the studies. No formal dropout model was undertaken in this analysis.

Internal validation

To assess whether the model is predictive for data not used in the analysis dataset (A), the final model was utilized for the analysis of the prediction dataset (B). Titration‐based VPCs of the final model on dataset B were compared with observations for selected groups defined by study, treatment, hyporesponder, and identified covariates. The comparison analyzed selected percentiles (5^th, 50^th, 95^th) of observations and simulations, and compared target statistics between observations and simulations.

Computation

Non‐linear mixed‐effects modeling was undertaken using NONMEM (version 7.5.0, Icon Development Solutions, Ellicott City, Maryland USA), ²⁹ in combination with PsN (version 5.0.0). ³⁰ , ³¹ Diagnostic graphics, exploratory analyses, and post‐processing of NONMEM output were performed using R (version 4.0.3 and 4.1.1, The R foundation for Statistical Computing), ³² and RStudio (version 1.3.1093 and 2021.09.0 Build 351, RStudio, Boston, USA). ³³

RESULTS

Demographic and baseline characteristics

An overview of the demographics and baseline characteristics of patients in this analysis is presented for analysis dataset (A) in Table S7 . Briefly, 53,535 Hgb observations from 2,770 subjects contributed to the Dose‐Hgb Model 3 analysis. The non‐dialysis CKD group represented 54% of the subjects. The overall median (min–max) age in the CKD population was 63 years (17–96) and body weight was 73 kg (36–194).

Dose‐Hgb modeling

The Hgb observations of analysis dataset (A) were predicted using the Dose‐Hgb Model 2. The standard VPCs showed a large discrepancy between observations and predictions (Figure S2 —top). This is readily explained in part due to ignoring of the dose titration scheme during VPC simulations ²⁴ , ³⁴ which need to account for dose adjustment rules used in the generation of data because with adaptive (titration) dosing, dose becomes a dependent variable based on the evaluated Hgb response. Therefore, a titration‐based VPC was applied, where the doses of simulated subjects were titrated according to their (simulated) Hgb levels and the titration schedule from the respective study in which they participated thus providing a better diagnostic (Figure S2 —bottom). Hence, all subsequent VPCs were titration‐based unless stated otherwise. It was also noted in Figure S2 —top, that the predicted variability of the recent ESA group in study 200807 decreased over time compared with the observed variability. This may potentially be related to influence of disease progression, which was not included in the previous Dose‐Hgb Model 2 (except in the form of placebo effects) and therefore could cause inflated variability between subjects.

Dose‐Hgb Model 3: refinement

During optimization, the following changes were made to the previous Dose‐Hgb Model 2 to obtain the base model:

• Implement additional Hgb baseline parameters for subjects that were just starting or had just started dialysis at the start of the study (study 201410).

• Inter‐individual variability of the apparent Hgb baseline (before daprodustat treatment) was estimated separately for studies 205270 and 201410, which improved the fit statistically.

• Residual error was estimated separately by study.

• The covariate “Hyporesponder on slope” was removed because it was no longer significant.

In addition, the placebo effect of the previous Dose‐Hgb Model 2 was removed and replaced by a disease progression function in Dose‐Hgb Model 3 that reduced the production of red blood cell precursor exponentially over time, of the form.

$$\text{Production}=\max \left\{\left(k_{\text{in}}-k_{\text{in},\min }\right)e^{-\mathit{Rt}}+k_{\text{in},\min },0\right\}$$ (1)

where t (day) is the time from the start of treatment, k _in and k _in,min (g/L/day) are the production rates at the start of treatment and time infinity, respectively, and R (/day) is the disease progression rate. The rate k _in,min was allowed to be negative, but the modeled production rate was limited to non‐negative values throughout the study period. A negative k _in,min value corresponds to a faster than exponential decline of Hgb precursor production. With the fitted parameter values and study duration, the production rate remained positive. The parameter R was allowed to be negative to include the possibility of spontaneous improvement for a subset of subjects. Additive IIV was included on R, with separate estimates for the first year, the second year, and subsequent years, to allow for variability between subjects and over time. The parameter k _in was constrained to be positive. Model parameters were adequately estimated (Table 2 ), and VPCs stratified by study and ESA use showed an adequate model fit for the base model (Figure S3 ).

Table 2.

Parameter estimates of the base and final Dose‐Hgb Model 3

Type	Variable	Base model		Final model
		Parameter value	RSE (%)	Parameter value	RSE (%)
Structural	Baseline ND ESA non‐users (g/L)	93.3	0.5	93.3	0.4
	Baseline ND ESA users (g/L)	103	0.4	103	0.3
	Baseline HD ESA users (g/L)	105	0.3	105	0.2
	Baseline PD ESA users (g/L)	104	0.7	103	0.7
	Baseline just started or starting HD in 201410 (g/L)	96.1	1.0	95.5	1.1
	Baseline just started or starting PD in 201410 (g/L)	101	2.6	101	2.6
	True baseline ND ESA users (g/L)	94.3	0.9	95.3	0.8
	True baseline HD/PD ESA users (g/L)	96.4	0.7	94.9	0.7
	Relative effect of Iron use on the apparent baseline for ESA non‐users	−0.0156	34.1
	Relative effect of Prior ESA on True baseline (−)	−0.00791	7.2	−0.00797	6.8
	Relative effect of ESA Hyporesponders on the apparent baseline for ESA users (−)	−0.0336	12.9	−0.0307	13.3
	Coefficient power model STAGE on True and apparent baseline (−)	−0.116	12.0	−0.107	12.9
	Coefficient power model Body Weight on True and apparent baseline (−)	0.0408	21.5	0.0520	16.8
	Coefficient power model BHEP on True and apparent baseline (−)	−0.0192	14.6	−0.0170	16.0
	Relative effect of Race = Asian on True baseline (−)			−0.0467	21.1
	Relative effect of Race = Black on True and apparent baseline (−)			−0.0376	17.5
	Precursor lifespan (days)	3.19	(Fixed)	3.19	(Fixed)
	RBC lifespan (days)	82.8	(Fixed)	82.8	(Fixed)
	Relative effect of ESA Hyporesponders on RBC lifespan	0.164	27.2
	Relative effect of prior ESA use on RBC lifespan (−)			0.100	33.0
	Coefficient power model baseline BW on RBC lifespan (−)			0.271	20.7
	Relative effect of HD on RBC lifespan (−)			0.138	30.3
	Relative effect of just starting/started HD or PD on RBC lifespan (−)			0.309	41.3
	Slope (/mg)	0.0511	2.0	0.0580	1.9
	Correlation Slope * Gamma (−)	16.5	3.0	14.9	2.8
	Relative effect of Race = Black on Slope (−)			−0.128	11.1
	Allometric coefficient body weight on Dose (−)	0.750	(Fixed)	0.750	(Fixed)
	Factor on CDOSE for t.i.w. dosing in 204,837 (−)	0.426	10.8	0.422	9.5
	Relative effect of Clopidogrel use on Dose (−)			−0.102	16.7
	Disease progression (/1,000 day)	−0.0462	6.0	−0.0851	8.3
	Minimum KIN for Disease progression (g/L/day)	−0.575	10.4	−0.678	11.1
	Relative effect of prior ESA use on Disease progression (−)			−0.318	32.7
	Relative effect of HD or PD on Disease progression (−)			−0.573	17.7
	Relative effect of just starting/started HD or PD on Disease progression (−)			0.968	34.2
	Study 205270 on omega 1 (−)	0.758	7.0	0.729	7.2
	Study 201410 on omega 1 (−)	1.10	9.3	1.13	9.3
IIV	Variance Baseline (additive)	51.6 [7.19]	3.5	50.1 [7.08]	3.5
	Covariance between apparent and true baselines			40.3 [0.497]	6.1
	Variance True baseline (additive)	134 [11.6]	4.4	131 [11.4]	4.3
	Variance Slope daprodustat effect (exponential)	0.0340 [0.184]	4.7	0.0280 [0.167]	5.2
	Variance TRBC (exponential)	0.250 [0.500]	4.7	0.206 [0.454]	4.8
	Variance Disease progression rate first year (additive)	0.0460 [0.214]	9.3	0.0504 [0.224]	11.3
	Covariance between disease progression first and second year			0.0212 [0.707]	11.8
	Variance Disease progression rate second year (additive)	0.0126 [0.112]	8.6	0.0178 [0.134]	11.2
	Covariance between disease progression first year and third year and beyond			0.0121 [0.528]	13.5
	Covariance between disease progression second year and third year and beyond			0.0108 [0.793]	11.9
	Variance Disease progression rate third year and beyond (additive)	0.00725 [0.0852]	9.6	0.0104 [0.102]	11.6
Residual	Standard deviation additive residual error study 205270 (g/L)	5.73	1.0	5.72	1.0
	Standard deviation additive residual error study 204837 (g/L)	6.22	0.9	6.28	0.9
	Standard deviation additive residual error study 201410 (g/L)	7.30	1.1	7.24	1.2
	Standard deviation additive residual error study 200807 (g/L)	7.32	0.2	7.27	0.2
	Standard deviation additive residual error study 200808 (g/L)	6.43	0.2	6.39	0.2

The table presents the values and relative standard errors (RSE). The RSE column displays“(Fixed)” for parameters that were fixed. For IIV variance parameters, standard deviations are shown in square brackets. For IIV covariance parameters, Pearson's correlation coefficients are shown in square brackets.

BHEP, baseline hepcidin; BW, body weight; CDOSE, allometrically scaled dose; ESA, erythropoietin stimulating agents; HD, hemodialysis; KIN, precursor production rate; ND, non‐dialysis; PD, peritoneal dialysis; RBC, red blood cell (count); RSE, relative standard error; STAGE, chronic kidney disease stage; t.i.w. three times a week; TRBC, transit RBC rate constant.

As a sensitivity analysis, it was assessed whether a concentration‐response model would perform better than this dose‐response model. As expected, the longitudinal concentration‐response model did not provide any additional insight to that obtained from Dose‐Hgb Model 3 (data not shown).

Covariate analysis

Major covariate effects on the relevant Dose‐Hgb model parameters are depicted in Table S8 with corresponding estimates of the magnitude of the covariate influence. The main effects were decreasing true Hgb baseline with increasing prior ESA dose; allometric body weight on dose; slower disease progression in chronic hemodialysis (HD) and peritoneal dialysis (PD) patients (−57.3%) vs. non‐dialysis (ND) patients, and faster disease progression in patients starting dialysis (+96.8%); slower disease progression in ESA users (−31.8%) than ESA non‐users; longer RBC life span in HD patients (+13.8%) and patients just starting dialysis (+30.9%) compared with ND or PD patients. All remaining covariate effects were less than 13%.

Final Dose‐Hgb Model 3

In the final step, correlations were introduced between IIV parameters, namely, between the two baseline IIV parameters and among the three disease‐related IIV parameters. All were significant gaining 900 points in OFV for four additional parameters. A schematic of this final model is depicted in Figure 1 and the final model parameters are presented in Table 2 alongside the base model parameters. Where both base and final models had the same structural parameters their values were similar. The condition number was acceptable at 1005. The titration‐based VPC shows an adequate fit (Figure 2 ) and is similar to that of the base model (Figure S3 ). The fit is also adequate for VPCs stratified by dialysis population or ESA hyporesponder category, with a slight overprediction of the variability of non‐hyporesponders. The value for the factor of 0.426 on the allometrically scaled dose (CDOSE) for t.i.w. dosing in study 204837 (Table 2 ) was close to its theoretical value of 3/7 ≈ 0.429. A random selection of individual fits per study and ESA use group is displayed in Figures S8 and S9 showing an adequate fit of the individual observations. The highest shrinkage observed was for IIV on RBC lifespan (35.8%) so Empirical‐Bayes Estimate (EBE) plots would typically be interpreted with caution. However, RBC lifespan acts primarily as a delay on the drug effect with little to no impact on the steady‐state Hgb response and was thus deemed relatively unimportant. The bias (≤1%) and shrinkage (≤8.9%) of residual error was acceptable.

Figure 1.

Schematic of Dose‐Hgb Model 3. K _in: Production rate of precursor cell; K _prec: precursor transit rate constant; K _RBC red blood cell transit rate constant.

Figure 2.

VPC with titration of the final Dose‐Hgb Model 3, stratified by study and ESA use (dataset A). Dots and lines: median, 5^th and 95^th percentile of the observed values. Gray areas: 95% prediction intervals of the same for simulated values, simulating each subject 250 times. no ESA = ESA non‐user, hist ESA = historic ESA user (last use was more than 28 days before study start), recent ESA = recent ESA user (last use was at most 28 days before study start). The historic ESA groups are small, hence the prediction intervals are large and results should be interpreted with caution.

Target statistics

Table 3 presents a comparison of observations and individual predictions above, within and below the target range of Hgb values. Over all time points, the Hgb predictions match the observations in 69.1% of the cases. The frequencies of over and underpredictions were approximately equal (14.9% vs. 15.8%). The predictions were more often within target than the observations (55.5% vs. 42.5%). This may be because the predictions do not include residual error, and therefore are less volatile than observations. The additional metrics comparing simulations (based on titration‐based VPCs) and observations (Figure 3 ), and VPC of mean dose over time (Figure S4 ) also support the adequacy of Dose‐Hgb Model 3 when comparing at a group level defined by study, ESA use, and time period.

Table 3.

Percentage of observed and predicted Hgb values below, within, and above target

Time period	N	Observed Hgb	Predicted Hgb
			Below target (%)	Within target (%)	Above target (%)
All	48,333	Below target	19.2	10.7	0.2
		Within target	4.6	33.9	4.0
		Above target	0.3	10.9	16.0
Month 1–3	13,494	Below target	32.7	9.2	0.2
		Within target	5.6	26.6	3.4
		Above target	0.2	8.1	14.1
Month 4–6	6,932	Below target	17.0	9.2	0.4
		Within target	3.8	33.9	4.6
		Above target	0.1	9.8	21.1
Month 7–12	11,697	Below target	11.7	10.6	0.1
		Within target	3.4	39.9	4.4
		Above target	0.1	11.9	17.8
Month 13	16,210	Below target	14.4	12.8	0.4
		Within target	5.0	35.6	4.1
		Above target	0.4	13.1	14.2

The target range is 11–12 g/dL for study 205270, and 10–11 g/dL for all other studies. The column “N” shows the number of observations per time period (for dataset A). The percentages are shown for different time periods, and for all data combined. For each time period they add up to 100%.

Figure 3.

VPC with titration for analysis dataset (a) and prediction dataset (b) using the final Dose‐Hgb Model 3 showing the fraction Hgb above, within, and below target, stratified by study, and ESA use. Bars: fractions above, within, and below the target for the simulated values, simulating each subject 250 times. Lines: averages of the same for observations. no ESA = ESA non‐user, hist ESA = historic ESA user (last use was more than 28 days before study start), recent ESA = recent ESA user (last use was at most 28 days before study start). The historic ESA groups are small, hence the results should be interpreted with caution.

Disease progression

Based on the final Dose‐Hgb Model 3, disease progression negatively affected the production rate of the RBC precursor. Typical profiles of the disease progression over time are shown in Figure 4 for the subpopulations that were modeled to have different disease progression (the covariates on disease progression were dialysis and prior ESA use). The weakest disease effect was observed for dialysis subjects, and the strongest for subjects that just started dialysis, while the ND subjects were in between. Simulations with the final model of typical subjects (that is, without IIV or residual error) with dose titration are shown in Figure S5 . For each combination of dialysis status and prior ESA use present in the studies, two typical subjects were simulated, one with zero dose (untreated) and one with dose titration (treated), starting at a dose equal to the median for that subpopulation, and at their typical Hgb baseline. Historic and recent ESA users were modeled the same in the final model, and hence combined in the subpopulation of ESA users. All untreated subjects experienced a decrease in Hgb level over time, as expected. The treated subjects experienced a gradually increasing dose, to keep the Hgb level in its target range (10–11 g/dL). This increase in dose level was faster for ND than for dialysis subjects, in accordance with the faster disease progression of the former group (Figure S5 ). Hyporesponders (to ESA) had an initial decrease in dose because they were not hyporesponders to daprodustat in the model.

Figure 4.

Time profiles of the disease progression (DP) for typical subjects relative to the study start. DP is displayed as a reduction factor in the production rate of the RBC precursor. The “Just started HD/PD” group is shown for 1 year only, as there is no data beyond that timepoint to support the model.

Internal model validation

Application of the final model to the prediction dataset (B) showed that this dataset was predicted adequately by the model. This is further supported by the VPCs of Hgb (Figure 3 ), target Hgb statistics (Figure S6 ) and mean dose (Figure S7 ) which show results similar to those for analysis dataset (A).

DISCUSSION

The present work reflects a continuation of the model‐based drug development (MBDD) of daprodustat from the early clinical phase through the successful efficacy outcome of the phase III program. ³⁵ , ³⁶ , ³⁷ , ³⁸ , ³⁹ This analysis was undertaken to further refine and expand the semi‐mechanistic Dose‐Hgb model framework from the five pivotal phase III studies with daprodustat. The designs of these studies were driven by the individualized dosing algorithm and dose selection implemented through modeling and simulation of Dose‐Hgb Model 1 built from early phase studies.

The final Dose‐Hgb Model 3, developed on the analysis dataset (A), provided an adequate description of the data, as shown by model diagnostics including titration‐based VPCs for Hgb and dose. This model consisted of a precursor cell compartment and 12‐transit compartments to describe the red blood cell (RBC) lifespan. Treatment was modelled as the stimulation of the precursor cell production rate by a power of dose (see supplement for details). The transit compartment model framework which considers variation in life‐span in the RBC population was shown to better describe RBC survival data than the random destruction or homogenous life‐span models. ²³ In particular, it was shown that the 12‐transit compartment model was superior to the 4‐ or 10‐transit compartment or the homogenous life‐span model. ²³ In addition, the internal validation (prediction) dataset (B) was adequately characterized by this model. For the titration‐based VPCs, simulated subjects were titrated according to their simulated Hgb values, mimicking as well as possible the titration applied in the studies. Some differences with the actual titration algorithms were unavoidable as not all information was available or could be simulated (for example, where the titration algorithm called for resampling, which could not be simulated, or where simulations were restricted to scheduled visits thereby reducing the number of opportunities for dose adaptation). Sensitivity checks were undertaken to ensure no significant bias, demonstrating the adequacy of these titration‐based VPCs. In addition, the successful outcome of the phase III studies in achieving the target Hgb range provides confidence in the adequacy of implementing the individualized titration schedule for daprodustat in CKD patients with anemia.

A further useful diagnostic was a comparison of the Hgb target ranges, pre‐defined for each study (and implicit in the titration algorithms). Incidence of observations within, below, and above target were compared with individual predictions or simulated values. Individual predictions could be linked to observations, and so target statistics of predictions and observations could be compared point‐by‐point. Generally, the model was found to describe the data well. ³

Another important addition to the Dose‐Hgb model was disease progression over time, enabled by the broader, more heterogeneous characteristics of CKD patients with anemia and Hgb observations up to 4 years duration in two of the pivotal studies. The development of a placebo model was not pursued given the limited duration of placebo data over 6 months only in the ND ESA non‐users already on standard of care (such as iron treatment) and thus was limited to subjects in early stages of the disease, where progression is expected to be relatively slow. Disease progression was modeled as an exponential decline of Hgb production rate over time (rather than in terms of CKD stage) and differed between baseline patient categories. Table S9 lists the effects of disease progression on steady state Hgb level for typical subjects in these subpopulations, in the absence of daprodustat treatment. These results do not mean that dialysis and ESA user status caused changes in disease progression, but only that these were correlated in the studies under consideration, for example, patients would experience a more rapid disease progression at the time that they are initiating dialysis (but not because of it).

Despite fixing the RBC lifespan parameter, no bias was observed in the EBE distribution, indicating that the fixed parameter was consistent with the data. In addition, the patient population effect on RBC life span (Table S8 ) suggests the typical life spans were 94 days for HD, 108 days for patients initiating dialysis, and 82.8 days for ND or PD patients. This only affected the delay between daprodustat dose and its effect, and not the steady state Hgb levels.

Most of the identified covariate relationships cluster around a central theme of disease severity (where dialysis dependency, higher ESA doses, high hepcidin concentrations, and high CKD stage would prevail). Separate Hgb baselines were estimated for different patient groups (depending on dialysis status and ESA use), reflecting disease status‐dependent differences in Hgb production. An increase in prior ESA dose was associated with a decrease (by 14%) in true Hgb baseline (i.e., Hgb baseline after washout of prior ESA treatment). Comparing the q.d.  vs. t.i.w. dosing schedule, the estimated factor of 0.426 is close to the value 3/7 ≈ 0.429 (i.e., three doses during a 7‐day interval) that one would expect if the effect was driven by the average daily dose. This suggests that the t.i.w. and q.d. dose frequencies are equally effective.

While daprodustat is primarily metabolized by cytochrome P450 CYP2C8, ⁴⁰ the most commonly co‐administered moderate CYP2C8 inhibitor, clopidogrel, was not a relevant covariate in presence of daprodustat in the current Dose‐Hgb analysis. This is consistent with the clinical data in these phase III studies suggesting that following 4 weeks of daprodustat dosing, hemoglobin changes from baseline were similar in subjects with and without concomitant use of clopidogrel. ¹¹ A population PK analysis with daprodustat based on a subset of CKD patients from the phase III studies showed there was a statistically significant increase in daprodustat AUC ratio of 1.59‐fold (90% CI:1.39–1.82) for CKD patients on clopidogrel to those not on clopidogrel ⁴¹ which would not be deemed clinically relevant due to the titration dosing schedule. However, following FDA regulatory review based on the totality of the data, it was recommended to reduce the starting dose of daprodustat by half in CKD patients with anemia when initiating treatment with moderate CYP2C8 inhibitors such as clopidogrel. ⁴¹

The potential impact of selective dropout and of rescue treatment was qualitatively monitored during model development by visually checking that 5^th and 95^th percentile predictions in the VPCs matched with the observations. Total dropout (informative or non‐informative) was high (range 22–52%), but the VPCs did not indicate that informative dropout was a relevant factor in this analysis. In addition, the VPCs are already partly corrected for dropout by removing simulated data according to the criteria for rescue treatment and using observed dropout information from the study.

The aim of this work was to update a previous population longitudinal Dose‐Hgb model with phase III data and further support the recommended titration dosing schedule for daprodustat in CKD patients with anemia. This semi‐mechanistic Dose‐Hgb model has been adequately re‐calibrated with data from pivotal phase III studies in a broader CKD patient population with anemia. The individualized treatment algorithm with daprodustat is well supported given the concordance of the model‐based simulations and corresponding hemoglobin observations. This lifespan based indirect response model linking with the transit (age) compartment stages as implemented with daprodustat effect on Hgb has effectively contributed towards the dose individualization algorithm of the drug which was prospectively applied successfully in the global phase III program in CKD patients with anemia. Some other examples of applications of the life‐span model concept implemented by using serial transit compartment include HbA1c in diabetes, ⁴² thrombopoietic stimulation in platelet function, ⁴³ erythropoietin in chronic kidney disease ¹⁷ as well as modeling of chemotherapy‐induced myelosuppression in oncology. ⁴⁴

FUNDING

All clinical studies mentioned in this manuscript were funded by GSK.

CONFLICT OF INTEREST

K.M.M., S.V., and S.Y. are employees of GSK and hold GSK shares. T.M.P., M.v.N., and P.v.d.B. are Consultants at LAP&P, Netherlands.

AUTHOR CONTRIBUTIONS

K.M.M., S.Y., S.V., M.v.N., P.v.d.B., and T.M.P. designed the research. M.v.N., P.v.d.B., and T.M.P. performed the research. M.v.N. and P.v.d.B. analyzed the data. All authors wrote the manuscript.

Supporting information

Data S1.