Optimising infliximab induction dosing for patients with ulcerative colitis

Abstract

Aims

The therapeutic failure of infliximab therapy in patients with ulcerative colitis remains a challenge even 2 decades after its approval. Therapeutic drug monitoring (TDM) has shown value during maintenance therapy, but induction therapy has still not been explored. Patients may be primary nonresponders or underexposed with the standard dosing regimen. We aimed to: (i) develop a population pharmacokinetic‐pharmacodynamic model; (ii) identify the best exposure metric that predicts mucosal healing; and (iii) build an exposure–response (ER) model to demonstrate model‐based dose finding during induction therapy with infliximab.

Methods

Data were retrospectively collected from a clinical database. A total of 583 samples, from 204 patients, was used to develop a population pharmacokinetic model to generate exposure metrics for subsequent ER modelling. A subset of 159 patients was used to develop a logistic regression ER model, describing the relationship between infliximab exposure and ordered transitions between Mayo endoscopic subscore (MES) 3, 2 and ≤1 (baseline to post‐induction).

Results

A 1‐compartment population pharmacokinetic model with interindividual and interoccasion variability was found to fit the data best. Covariates influencing exposure were C‐reactive protein, albumin, baseline MES, fat‐free mass, concomitant corticosteroid use and pancolitis. The cumulative area under the infliximab concentration–time curve until endoscopy (CAUC_endoscopy) was found to be the best exposure metric for predicting mucosal healing (baseline MES >1 and post‐induction MES ≤1). The model predicted that 70% of patients will attain mucosal healing with infliximab administered at days 0, 14 and 42 and a target CAUC_endoscopy of 3752 mg/L*day at day 84.

Conclusions

TDM‐based dose individualisation targeting CAUC_endoscopy has the potential to improve the effectiveness of infliximab during induction therapy.

1.

What is already known about this subject

• Therapeutic drug monitoring based on trough concentrations has shown value during infliximab maintenance therapy for patients with ulcerative colitis, but treatment optimisation during induction therapy has still not been explored.

What this study adds

• We developed a population pharmacokinetic–pharmacodynamic model that predicts the proportion of ulcerative colitis patients with endoscopic improvement after infliximab induction therapy.

• The cumulative area under the infliximab concentration–time curve up to the time of endoscopy is the best predictor of endoscopic improvement after induction therapy.

• Our model has the potential to help optimise infliximab dosing strategies during induction therapy in patients with ulcerative colitis.

2. INTRODUCTION

Infliximab is a chimeric monoclonal antibody that neutralises tumour necrosis factor alpha. Based on the results of the landmark Active Ulcerative Colitis Trials (ACT) 1 and 2, infliximab was approved for inducing and maintaining remission in patients with moderate‐to‐severe ulcerative colitis (UC).1 In these studies, mucosal healing (defined as Mayo endoscopic subscore [MES] ≤1) was achieved in 61% (standard error 3%) of patients after administration of 3 infliximab infusions (5 mg/kg body weight, at days 0, 14 and 42; evaluation of mucosal healing at day 56). In postmarketing studies, mucosal healing rates were lower (eg 47% ±11% in Brandse et al.2), making unpredictable outcomes of infliximab induction therapy a challenge, even 2 decades after its approval.2, 3, 4, 5

In a post‐hoc analysis of the ACT data, Adedokun and colleagues showed a positive association between infliximab serum concentrations and mucosal healing, during both induction and maintenance therapy.6 In recent years, several studies have confirmed this exposure–response association.2, 3, 4, 5 Therefore, targeting infliximab to optimal exposure has the potential to improve response rates and identify primary nonresponders, thereby supporting a potential role for therapeutic drug monitoring (TDM) during infliximab induction therapy in patients with UC. Nevertheless, drug monitoring studies almost exclusively focus on maintenance therapy, whereas induction therapy is largely unexplored.7, 8, 9

It is well known that the pharmacokinetics of infliximab varies between and within patients.10 Part of this between‐ and within‐subject variability might be explained by variables such as C‐reactive protein (CRP), albumin (ALB), antidrug antibody titres and faecal calprotectin.11, 12 In a small cohort study, Brandse et al. observed lower infliximab exposure during the first 6 weeks of therapy in patients with high CRP at baseline.2

The objective of the current analysis was to develop a population pharmacokinetic–pharmacodynamic model, characterising the infliximab dose–exposure relationship, to identify the best exposure metric(s) to predict mucosal healing and to build exposure–response model(s) to describe the relationship(s) mathematically and to demonstrate model‐based dose finding during induction therapy with infliximab.

3. METHODS

3.1. Patients

Data were collected by Arias et al., who conducted a retrospective study including patients with UC who initiated infliximab therapy between January 2000 and September 2013 at the University Hospitals Leuven, Belgium.13 All patients included in the analysis had given written consent to participate in the institutional review board‐approved inflammatory bowel disease biobank (B322201213950/S53684).

All patients (i) with a diagnosis of UC, (ii) in whom therapy with infliximab was initiated (iii), with at least 1 serum sample available with a detectable infliximab concentration during induction therapy, were eligible for inclusion in the pharmacokinetic study. Of these patients, the ones who underwent colonoscopy at baseline with an assessment of MES 2 or 3, and a second assessment after 1–4 induction doses (earliest at day 31 and latest at day 127 of treatment), were included in the pharmacodynamic study.

3.2. Serum infliximab and antidrug antibody concentrations

Serum samples were collected at trough (ie right before the infliximab infusions at days 14, 42 and 98). Samples were stored at −20°C, under which condition immunoglobulin G has been shown to be stable for up to 25 years.14 Infliximab concentrations were measured, after a maximum of 14 years in storage, with an in‐house developed and clinically validated direct enzyme‐linked immunosorbent assay (ELISA), with a lower limit of quantification of 0.3 mg/L.15 Antidrug antibodies were measured using the ApDia Anti‐Infliximab ELISA (ApDia, Turnhout, Belgium), with a lower limit of quantification of 2.5 μg/L MA‐IFX10F9 equivalents.16

3.3. Mayo endoscopic subscore

Mucosal healing was defined as going from a baseline MES 2 (moderate disease) or 3 (severe disease) to a post‐induction subscore of 0 (inactive disease) or 1 (mild disease).

3.4. Software

Population pharmacokinetic and pharmacodynamic analyses were performed in NONMEM (version 7.4; Icon Development Solutions, Ellicott City, MD, USA) with a GNU Fortran 95 compiler and the Perl‐speaks‐NONMEM (PsN; version 4.7.0) toolkit and Pirana (version 2.9.7, SGS Exprimo). Simulations were performed using Simulo Expert (version 7.2; SGS Exprimo, Mechelen, Belgium). Graphics were generated using the ggplot2 package for R statistical software in the RStudio integrated development environment (version 2.2.1, RStudio, Inc., Boston, MA, USA).17, 18

3.5. Population pharmacokinetic and pharmacodynamic model development, evaluation and selection

3.5.1. Pharmacokinetic model

Different structural models were fitted to infliximab pharmacokinetic data using the Laplacian estimation method with interaction and subroutine ADVAN 6 (Appendix A). A parsimonious and stable pharmacokinetic model was developed, starting from a 1‐compartment model with intravenous infusion and linear clearance. Layers of complexity were added in a stepwise manner, and model improvement was evaluated based on standard goodness‐of‐fit diagnostic tools (residual‐ and simulation‐based plots), convergence of the minimisation criteria, precision of the parameter estimates and a drop in the objective function value of at least 3.84 points (P < .05, 1 degree of freedom, nested models). Model parameters were added to the structural model (eg increased compartmentalisation) and the random effects model (eg interindividual and interoccasion variability). All structural model parameters were assumed to be log‐normally distributed. Additive and/or proportional error models were explored for modelling the difference between observed and model‐predicted exposure and response values.

The interindividual and interoccasion variability in infliximab k_e and V were modelled as:

$${\mathrm{k}}_{\mathrm{e},\mathrm{ij}}={\mathrm{\theta }}_{\mathrm{pop}}\times {\mathrm{e}}^{{\mathrm{\eta }}_{1,\mathrm{i}}+\text{FLAG}\left(1\right)\times {\mathrm{\eta }}_{3,\mathrm{j}}+\text{FLAG}\left(2\right)\times {\mathrm{\eta }}_{4,\mathrm{j}}+\text{FLAG}\left(3\right)\times {\mathrm{\eta }}_{5,\mathrm{j}}+\text{FLAG}\left(4\right)\times {\mathrm{\eta }}_{6,\mathrm{j}}}$$ (1)

$${\mathrm{V}}_{\mathrm{ij}}={\mathrm{\theta }}_{\mathrm{pop}}\times {\mathrm{e}}^{{\mathrm{\eta }}_{2,\mathrm{i}}+\text{FLAG}\left(1\right)\times {\mathrm{\eta }}_{7,\mathrm{j}}+\text{FLAG}\left(2\right)\times {\mathrm{\eta }}_{8,\mathrm{j}}+\text{FLAG}\left(3\right)\times {\mathrm{\eta }}_{9,\mathrm{j}}+\text{FLAG}\left(4\right)\times {\mathrm{\eta }}_{10,\mathrm{j}}}$$ (2)

where θ_pop are the population values of infliximab k_e and V, η₁ and η₂ are the interindividual variabilities in k_e and V, and η₃ up to η₁₀ are the interoccasion variabilities in k_e and V. An occasion was defined as a dosing interval. The FLAG(x) dummy variables have been initialised at zero but take the value 1 at the respective dosing interval (ie FLAG(1) is 1 on the day 0 infusion, FLAG(2) is 1 on the day 14 infusion, etc).

The influence of covariates on the infliximab pharmacokinetic parameters was tested. Body weight, fat‐free mass (FFM), sex, age, disease duration, concomitant corticosteroids and thiopurines at baseline, MES at baseline, pancolitis at baseline, and time‐varying CRP and albumin were considered as potential covariates. Patients' FFM was predicted using the empirical model developed by Janmahasatian et al.19 The association between continuous covariates and the typical value of the pharmacokinetic parameters was described using power models. The effects of concomitant corticosteroid use and pancolitis at baseline were modelled as fold changes compared with the reference of no corticosteroids and no pancolitis, respectively. MES was tested on volume and elimination rate constant, assigning different typical values for each covariate level. Covariates were identified using a stepwise covariate modelling approach with forward addition (decrease in objective function value ≥6.63 units, α = 0.01, 1 degree of freedom) and backward elimination (increase in objective function value ≥10.8 units, α = 0.001, 1 degree of freedom).

Pharmacokinetic and pharmacodynamic modelling was performed sequentially. Infliximab concentrations and areas under the concentration–time curve were estimated from the empirical Bayesian estimates and individual dosing histories during estimation of the parameters of the final pharmacokinetic model (Appendix A). These exposure metrics were then used as an input for the pharmacodynamic model.

3.5.2. Pharmacodynamic model

A logistic regression model was implemented to explore the relationship between infliximab exposure and the probability of achieving mucosal healing (Figure 1, Appendix B). The MES was simplified to 3 states: 3 (severe disease), 2 (moderate disease) and combined 1 and 0 (mild and inactive disease). We assumed that only ordered transitions can occur (ie patients going from 3 to 1 or 0, transitioned through state 2 and vice versa) and that the transition probabilities between states can be inversed (eg P_3➔2 = 1‐P_2➔3). The 3 observed states and 6 transitions were described using 4 parameters: the baseline proportions of patients displaying MES 3 and 2 (ie P3_baseline and P2_baseline), and the transition probabilities corresponding to going from baseline MES 3 to 2 (ie P_3➔2) and from 2 to 1 or 0 (ie P_2➔1,0). All other transition probabilities are implicit: P_3➔3 = 1‐P_2➔3, P_2➔3 = 1‐P_3➔2 and P_2➔2 = 1‐P_2➔1,0‐(1‐P_3➔2). As no patients start at MES state 1 or 0 (ie P1,0_baseline = 0), P_1,0➔2 is 0.

Figure 1.

Schematic representation of the population pharmacokinetic and pharmacodynamic model of infliximab induction therapy in patients with ulcerative colitis. The 3 observed states and 6 transitions were described using 4 parameters: the baseline proportions of patients displaying Mayo endoscopic subscores 3 and 2 (ie P3_baseline and P2_baseline), and the transition probabilities corresponding to going from baseline Mayo endoscopic subscore 3 to 2 (ie P_3➔2) and from 2 to 1 or 0 (ie P_2➔1,0). All other transition probabilities are implicit: P_3➔3 = 1‐P_2➔3, P_2➔3 = 1‐P_3➔2, P_2➔2 = 1‐P_2➔1,0‐(1‐P_3➔2) and P_1,0➔2 = 1‐P_2➔1,0. k_e, elimination rate constant; P, probability; V, volume of distribution

In the null model, all transition probabilities were equal (ie transitions occurred by equal chance, and were not affected by any infliximab exposure metric). As a result, the 6 possible transition fractions F_3➔3, F_3➔2, F_3➔1,0, F_2➔3, F_2➔2, F_2➔1,0, were identical (0.167 ± 0.005). Subsequently, the drug effect was modelled to affect the transition probabilities:

$${\mathrm{P}}_{3\to 2}={\left(\frac{\mathrm{X}}{{\mathrm{X}}_{50}^{3\to 2}}\right)}^{{\mathrm{\gamma }}_{3\to 2}}/1+{\left(\frac{\mathrm{X}}{{\mathrm{X}}_{50}^{3\to 2}}\right)}^{{\mathrm{\gamma }}_{3\to 2}}$$ (3)

$${\mathrm{P}}_{2\to 1,0}={\left(\frac{\mathrm{X}}{{\mathrm{X}}_{50}^{2\to 1,0}}\right)}^{{\mathrm{\gamma }}_{2\to 1,0}}/1+{\left(\frac{\mathrm{X}}{{\mathrm{X}}_{50}^{2\to 1,0}}\right)}^{{\mathrm{\gamma }}_{2\to 1,0}}$$ (4)

with X₅₀ ^3➔2 and X₅₀ ^2➔1,0 the values of the exposure metric X corresponding to a 50% probability of going from MES 3 to 2 and 2 to 1 or 0, respectively, and ^γ _3➔2 and ^γ _2➔1,0 the Hill's coefficients. Subsequently, the transition fractions were estimated:

$${\mathrm{F}}_{3\to 3}=\mathrm{P}{3}_{\text{baseline}}\times \left(1-{\mathrm{P}}_{3\to 2}\right)$$ (5)

$${\mathrm{F}}_{3\to 2}=\mathrm{P}{3}_{\text{baseline}}\times {\mathrm{P}}_{3\to 2}\times \left(1-{\mathrm{P}}_{2\to 1,0}\right)$$ (6)

$${\mathrm{F}}_{3\to 1,0}=\mathrm{P}{3}_{\text{baseline}}\times {\mathrm{P}}_{3\to 2}\times {\mathrm{P}}_{2\to 1,0}$$ (7)

$${\mathrm{F}}_{2\to 3}=\mathrm{P}{2}_{\text{baseline}}\times \left(1-{\mathrm{P}}_{3\to 2}\right)$$ (8)

$${\mathrm{F}}_{2\to 2}=\mathrm{P}{2}_{\text{baseline}}\times \left(1-{\mathrm{P}}_{2\to 1,0}-\left(1-{\mathrm{P}}_{3\to 2}\right)\right)$$ (9)

$${\mathrm{F}}_{2\to 1,0}=\mathrm{P}{2}_{\text{baseline}}\times {\mathrm{P}}_{2\to 1,0}$$ (10)

The exposure metrics were compared in terms of sensitivity to the probability of endoscopic improvement. We were not comparing different pharmacodynamic model structures but different pharmacokinetic input data, fitted to the same pharmacodynamic model. We compared different pharmacokinetic input data fitted to the same pharmacodynamic model, using the objective function, goodness‐of‐fit plots, convergence of the minimisation criteria and precision of the parameter estimates.

3.6. Model evaluation

Final model evaluation was performed using a visual predictive check. Standard errors for the parameter estimates were obtained from the covariance step. Furthermore, 2000 bootstrap samples were used to obtain nonparametric estimates of uncertainty in parameter estimates (the relative standard error and 95% confidence interval).

3.7. Model predictions

The selected population pharmacokinetic and pharmacodynamic models were used to perform stochastic simulations of the pharmacokinetic profile and the probability of mucosal healing, respectively. Simulations were performed for doses of 5 mg/kg and 10 mg/kg infliximab while sticking to the standard induction time schedule. Covariates were sampled from the original dataset.

3.8. Nomenclature of targets and ligands

Key protein targets and ligands in this article are hyperlinked to corresponding entries in http://www.guidetopharmacology.org, the common portal for data from the IUPHAR/BPS Guide to PHARMACOLOGY,20 and are permanently archived in the Concise Guide to PHARMACOLOGY 2017/18.21

4. RESULTS

4.1. Data source

The pharmacokinetic analysis was performed using 583 infliximab serum concentrations from 204 patients. The baseline patient characteristics are summarised in Table 1. Infliximab serum concentrations below the limit of quantification were excluded from the analysis as this number was small (22/605, 4% of all samples). The pharmacokinetic analysis included patients receiving different induction dosing schedules in terms of number of doses, dose size and timing of dosing. A subset of 159 patients (78%) with a diagnosis of moderate to severe disease at baseline (MES 2 or 3) and post‐induction MES evaluated between days 31 and 114 were included in the pharmacodynamic analysis. Mucosal healing was achieved in 91 (57% ± 4%) of these patients.

Table 1.

Data summary for the population pharmacokinetic and pharmacodynamic analyses

Parameter	Value
Number of patients with ulcerative colitis	204
Baseline demographics
Sex, female, n (%)	87 (43)
Age, median [IQR], years	40 [28–51]
Body weight, median [IQR], kg	72 [61–82]
Fat‐free mass, median [IQR], kg	52 [42–60]
Disease duration, median [IQR], years	4 [1–10]
Serology at baseline
C‐reactive protein, median [IQR], mg/L	6.1 [2.4–19.9]
Albumin, median [IQR], g/L	42.0 [38.8–43.9]
Concomitant medication at baseline
Corticosteroids, n (%)	89 (44)
Azathioprine, n (%)	98 (48)
Treatment
Number of doses, 2:3:4, n (%)	10:31:163 (5:15:80)
Dose size first dose, 5:10:15 mg/kg body weight, n (%)	186:18:0 (91:9:0)
Dose size second dose, 5:10:15 mg/kg body weight, n (%)	182:22:0 (89:11:0)
Dose size third dose, 5:10:15 mg/kg body weight, n (%)	172:21:1 (89:11:1:0)
Dose size fourth dose, 5:10:15 mg/kg body weight, n (%)	138:25:0 (85:15:0)
Timing second dose, median [range], days	14 [6–36]
Timing third dose, median [range], days	42 [15–92]
Timing fourth dose, median [range], days	98 [43–114]
Infliximab and antidrug antibodies
Samples available, n	605
Samples per patient, median [range], n	3 [1–6]
Samples with undetectable infliximab, n (%)	22 (4)
Samples with undetectable infliximab and antidrug antibodies, n (%)a	7 (1)
Endoscopy
Acute severe ulcerative colitis, n (%)b	34 (17)
Disease extension: E3 (pancolitis), n (%)c	123 (60)
Baseline Mayo endoscopic subscore, 0:1:2:3:N/A, n (%)	0:7:96:98:3 (0.0:3.4:47.1:48.0:1.5)
Post‐induction Mayo endoscopic subscore, 0:1:2:3:N/A, n (%)	51:56:47:41:9 (25.0:27.5:23.0:20.1:4.4)
Timing of post‐induction Mayo endoscopic subscore, median [range], days	89 [6–385]
Mucosal healing, yes:no: N/A,d n (%)	91:74:39 (45:36:19)

^aMeasured using a drug sensitive assay, only allowing the detection of antidrug antibodies when infliximab is undetectable.

^bAcute severe ulcerative colitis was subjectively assessed by the treating physician, based on Mayo endoscopic subscore, body mass index, albumin and C‐reactive protein at baseline.

^cAccording to the Montreal classification.

^dForty‐five patients were excluded from the pharmacodynamic evaluation because of a baseline Mayo endoscopic subscore of 1 (n = 7), a post‐induction Mayo endoscopic subscore evaluated before day 31 or after day 127 (n = 27), or missing baseline and/or post‐induction endoscopic evaluation (n = 11). IQR = interquartile range; N/A = not available.

4.2. Pharmacokinetic model

4.2.1. Base model

A 1‐compartment model with first‐order elimination kinetics best described the concentration–time course of infliximab (Figure 1). The model included 3 levels of random effects: interindividual variability, interoccasion variability and residual variability. Interindividual and interoccasion variability in the elimination rate constant (k_e) and volume of distribution (V) were estimated using exponential models, whereas the residual variability was modelled as additive and proportional to the observed infliximab concentration. The additive residual error was fixed to the lower limit of quantification of the assay. Covariance was not estimated. All parameters were estimated with good precision (Table 2).

Table 2.

Infliximab population pharmacokinetic parameter estimates

Parameter	BASE MODEL (OFV = 2860.2)			FINAL MODEL (OFV = 2696.5)
	Estimate	95% CI a	η‐shrinkage	Estimate	95% CI a	η‐shrinkage	Bootstrap estimate b	Bootstrap 95% CI b	Bootstrap deviation
Typical values
Elimination rate constant (/d)	0.0549	0.0521, 0.0577	N/A	–	–	–	–	–	–
Mayo endoscopic subscore 1	–	–	–	0.0521	0.0409, 0.0633	N/A	0.0529	0.0398, 0.0661	+1.5%
Mayo endoscopic subscore 2	–	–	–	0.0543	0.0500, 0.0586	N/A	0.0542	0.0501, 0.0588	−0.2%
Mayo endoscopic subscore 3	–	–	–	0.0667	0.0562, 0.0772	N/A	0.0668	0.0593, 0.0732	+0.1%
Effect of C‐reactive protein	–	–	–	0.0859	0.0438, 0.1280	N/A	0.0822	0.0412, 0.1230	−4.3%
Effect of albumin	–	–	–	−0.808	−1.426, −0.190	N/A	−0.7823	−1.7024, −0.4181	−3.2%
Volume of distribution (L)	8.44	7.83, 9.05	N/A	6.34	5.97, 7.71	N/A	6.36	5.79, 6.99	+0.3%
Effect of fat‐free mass	–	–	–	0.544	0.235, 0.853	N/A	0.556	0.236, 0.870	+2.2%
Effect of corticosteroidsc	–	–	–	1.33	1.15, 1.51	N/A	1.32	1.15, 1.51	−0.8%
Effect of pancolitisc	–	–	–	1.23	1.23, 1.23	N/A	1.22	1.08, 1.38	−0.8%
Interindividual variance
Elimination rate constant	0.152	0.104, 0.200	15%	0.0853	0.0585, 0.1121	16%	0.0817	0.0514, 0.1214	−4.2%
Volume of distribution	0.168	0.121, 0.215	28%	0.0681	0.0134, 0.1228	39%	0.0642	0.0115, 0.1255	−5.7%
Interoccasion variance
Elimination rate constant	0.0438	0.0301, 0.0575	41%	0.0343	0.0121, 0.0565	40%	0.0352	0.0042, 0.0568	+2.6%
Volume of distribution	0.0207	0.0079, 0.0335	62%	0.00	Fixedd	100%	N/A	N/A	N/A
Residual variance
Proportional error	0.0133	0.0117, 0.0149	65%	0.0362	0.0227, 0.0497	45%	0.0331	0.0106, 0.0852	−8.6%
Additive error	0.09	Fixed	65%	0.09	Fixed	45%	N/A	N/A	N/A

^a95% confidence intervals were calculated as estimated mean ± 1.96 × standard error.

^bMedian and 95% confidence interval estimated based on 1170 bootstrap replicates.

^cThe corticosteroid and pancolitis effects were modelled as a fold change compared with the reference of no corticosteroids and no pancolitis, respectively.

^dInteroccasion variability of volume of distribution was fixed to zero after being estimated <0.0001.

CI = confidence interval; N/A = not applicable; OFV = objective function value.

The η‐shrinkage of the empirical Bayesian estimates of the interindividual random effects was sufficiently low—15% for elimination rate constant and 28% for volume of distribution, respectively. This allowed us to use the empirical Bayesian estimates for the exploration of covariate effects.

4.2.2. Final model

The pharmacokinetic parameters with significant covariate dependencies in the final model were modelled as:

$${\mathrm{k}}_{\mathrm{e},\mathrm{ij}}=\left({\mathrm{\theta }}_{\mathrm{MES}1}\times \mathrm{MES}1+{\mathrm{\theta }}_{\mathrm{MES}2}\times \mathrm{MES}2+{\mathrm{\theta }}_{\mathrm{MES}3}\times \mathrm{MES}3\right)\times {\mathrm{e}}^{{\mathrm{\eta }}_{1,\mathrm{i}}+\text{FLAG}\left(1\right)\times {\mathrm{\eta }}_{3,\mathrm{j}}+\text{FLAG}\left(2\right)\times {\mathrm{\eta }}_{4,\mathrm{j}}+\text{FLAG}\left(3\right)\times {\mathrm{\eta }}_{5,\mathrm{j}}+\text{FLAG}\left(4\right)\times {\mathrm{\eta }}_{6,\mathrm{j}}}\times {\left(\frac{\mathrm{ALB}}{42.0}\right)}^{{\mathrm{\theta }}_{\mathrm{ALB}}}\times {\left(\frac{\mathrm{CRP}}{6.1}\right)}^{{\mathrm{\theta }}_{\mathrm{CRP}}}$$ (11)

$${\mathrm{V}}_{\mathrm{ij}}={\mathrm{\theta }}_{\mathrm{pop}}\times {\mathrm{e}}^{{\mathrm{\eta }}_{2,\mathrm{i}}+\text{FLAG}\left(1\right)\times {\mathrm{\eta }}_{7,\mathrm{j}}+\text{FLAG}\left(2\right)\times {\mathrm{\eta }}_{8,\mathrm{j}}+\text{FLAG}\left(3\right)\times {\mathrm{\eta }}_{9,\mathrm{j}}+\text{FLAG}\left(4\right)\times {\mathrm{\eta }}_{10,\mathrm{j}}}\times {\mathrm{\theta }}_{\mathrm{CS}}\times {\mathrm{\theta }}_{\text{PANC}}\times {\left(\frac{\mathrm{FFM}}{52}\right)}^{{\mathrm{\theta }}_{\mathrm{FFM}}}$$ (12)

where MES(x) is a dummy variable initialised at zero but taking the value 1 at the respective MES. ALB, CRP and FFM represent the time varying ALB, CRP and FFM (backward step interpolation) and CS and PANC are either zero (no concomitant corticosteroids and no pancolitis at baseline, respectively) or 1(concomitant corticosteroids and pancolitis at baseline, respectively).

Parameter estimates from the final and bootstrap models for infliximab are summarised in Table 2. A total of 1170 (58.5%) bootstrap replicates converged successfully. The point estimates from the original data set and the median parameter estimates from the bootstrap data sets were similar (no more than 8.6% difference between the corresponding estimates), indicating robustness of the final model and the parameter estimates. The goodness‐of‐fit plots show that the model described the observed data adequately (Figure S1). The visual predictive check shows good agreement between simulated and observed data (Figure 2).

Figure 2.

Prediction‐corrected visual predictive check. The solid line connects the observed median prediction‐corrected infliximab serum concentrations (mg/L) per bin. The dashed lines connect the 5th and 95th percentiles of the prediction‐corrected observations. Shaded areas indicate the 95% confidence interval of the median and 5th and 95th percentiles of the simulated values. Outlying percentiles of the real data are highlighted with a red asterisk. The observed prediction‐corrected infliximab concentrations are represented by yellow circles

Infliximab elimination is higher with higher CRP (+13% from 5 mg/L to 20 mg/L), lower ALB (+23% from 45 g/L to 35 g/L) and higher MES (23% higher when MES 3 compared with 2). The volume of distribution is 33% larger when on concomitant corticosteroids at the start of induction therapy and 23% larger when the patient has pancolitis at the start. Furthermore, the volume of distribution is larger with FFM (+25% from 40 kg to 60 kg). The elimination half‐life of infliximab for the typical patient (CRP 6.1 mg/L and ALB 42 g/L) was calculated to be 10.4 days and 12.8 days with MES 3 and 2, respectively. Figure 3 shows the empirical Bayesian estimates of the infliximab elimination rate constant and volume of distribution vs the withheld covariates. Figure S2 shows the covariate correlations. The unexplained variability (ie combined interindividual and interoccasion variability) reduced by 51% after introduction of the 6 covariates.

Figure 3.

Empirical Bayesian estimates of the infliximab elimination rate constant vs Mayo endoscopic subscore, albumin and C‐reactive protein, and the volume of distribution vs fat‐free mass, concomitant corticosteroid use and pancolitis at baseline. Empirical Bayesian estimates for the first occasion for all 204 patients are represented here. Comparisons in A, E and F, were carried out using the Wilcoxon rank‐sum test. In B, 3 data points are not presented (albumin:k_e 10.7:0.0875, 22.5:0.1540 and 24.9:0.0948), although they are considered for the locally weighted smoothing (dashed line)

4.3. Pharmacodynamic model

The estimated cumulative area under the infliximab concentration–time curve up to the time of endoscopy was identified as the best predictor of mucosal healing (lowest objective function), followed by the estimated trough concentration at day 14 (Table 3). A good correlation was observed between the estimated cumulative area under the curve up to the time of endoscopy and the estimated trough concentration at day 14 (Spearman's ρ = 0.83) (Figure S3).

Table 3.

Overview of pharmacodynamic models and their final parameter estimates

Model	OFV	Parameter	Estimate	95% CIa
Null model; no effect of infliximab	569.94	–
Cumulative dose up to the time of endoscopy (CDendoscopy) drives response	524.21	CDendoscopy,50 3➔2 (mg)	184	106, 262
		CDendoscopy,50 2➔1,0 (mg)	723	482, 964
Infliximab cumulative area under the concentration–time curve up to day 14 (CAUC14) drives response	509.79	CAUC14,50 3➔2 (mg/L*day)	66	39, 93
		CAUC14,50 2➔1,0 (mg/L*day)	282	188, 376
Infliximab trough concentration at day 14 (C14) drives response	502.76	C14,50 3➔2 (mg/L)	2.8	1.6, 4.0
		C14,50 2➔1,0 (mg/L)	13.1	8.8, 17.4
Infliximab cumulative area under the concentration‐time curve up to the time of endoscopy (CAUCendoscopy) drives response	495.38	CAUCendoscopy,50 3➔2 (mg/L*day)	305	174, 436
		CAUCendoscopy,50 2➔1,0 (mg/L*day)	1400	934, 1866

Hill's coefficients were fixed to one.

^a95% confidence intervals were calculated as estimated mean ± 1.96 × standard error. CI = confidence interval; OFV = objective function value.

Aiming at mucosal healing in 55% of patients, the model estimated a target cumulative area under the curve of 2,015 mg/L*day (Figure 4). Aiming at mucosal healing in 70% of patients, the model estimated a target cumulative area under the curve of 3752 mg/L*day. Thus, a 15% higher response rate is estimated to increase the drug cost by 86%, while aiming at 4792 mg/L*day (75% of patients achieving mucosal healing) requires drug expenditures to increase by 138%. Similarly, the model estimated an infliximab trough concentration at day 14 of 18.8 mg/L, 35.1 mg/L and 49.8 mg/L to predict mucosal healing in 55%, 70% and 75% of patients, respectively.

Figure 4.

Goodness‐of‐fit plot of the logistic regression exposure–response model. Observed (tiles) fraction of patients attaining post‐induction Mayo endoscopic subscore states 3 (red), 2 (yellow) and ≤1 (green), and predicted (lines) fraction of patients attaining post‐induction Mayo endoscopic subscore states 3 (red) and >1 (green) as a function of the cumulative area under the infliximab concentration–time curve up to the time of endoscopy (7 data bins). Grey numbers indicate the number of patients in each bin

4.4. Simulations

The simulated pharmacokinetic profiles, predicted cumulative area under the curve at day 84 and the probability of mucosal healing are represented in Figure 5. Administration of 5 mg/kg at days 0, 14 and 42 resulted in a median [90% prediction interval] cumulative area under the curve at day 84 of 2187 [929–4773] mg/L*day. The mean predicted probability of mucosal healing was 56%. Administration of 10 mg/kg at days 0, 14 and 42 resulted in a median [90% prediction interval] cumulative area under the curve at day 84 of 4366 [1808–9572] mg/L*day. The mean predicted probability of mucosal healing was 72%. In the 10 mg/kg simulation, 22% of patients have a higher cumulative area under the curve than the 95th percentile of the original dataset, implying extrapolation from the pharmacodynamic model (Figure S4).

Figure 5.

The population distribution of pharmacokinetic profiles A, B, the cumulative area under the infliximab concentration–time curve at day 84 (CAUC₈₄) C, D, and the probability of mucosal healing E, F, following 5 mg/kg (left) and 10 mg/kg (right) infliximab infusions at days 0, 14 and 42. N = 50 000 stochastic simulations. Covariates were sampled from the original dataset. Time‐varying covariates were evaluated using next observation carry backward. Missing baseline Mayo endoscopic subscores were replaced by sampling from the population distribution

5. DISCUSSION

Infliximab is the most effective therapy currently available for treating patients with UC.1 However, more than 40% of these patients do not achieve mucosal healing after receiving the established induction dosing regimen.2, 3, 4, 5 In our population, 55 ± 4% of the patients experienced mucosal healing. The differentiation of these patients into primary nonresponders (lack of effect despite “adequate exposure”) vs underexposed patients will in all likelihood improve these numbers.22 The adequate infliximab exposure required during induction therapy is not well defined. Papamichael et al. observed a relationship between trough concentrations and mucosal healing.4 After the start of treatment, responders had significantly higher trough concentrations than nonresponders. However, in the absence of a pharmacokinetic–pharmacodynamic model linking dose, exposure and response, they were not able to target a certain probability of mucosal healing, let alone recommend the associated dose. We recommended exposure metrics corresponding to different probabilities of mucosal healing and the accompanying doses, thereby reducing the number of underexposed patients during infliximab induction therapy.

The pharmacokinetic model was informed by trough concentrations at 14, 42 and 98 days after initiation of therapy. As there were no concentrations available immediately after administration, only a 1‐compartment model was identifiable, as confirmed by others.23, 24, 25 As expected, the trough concentrations were adequately captured by the model. To assess the potential impact of missing the distribution (alpha) phase on the calculation of the area under the infliximab concentration–time curve, we calculated its contribution to the total area under the curve in the 2‐compartment model by Fasanmade et al.26 The contribution of the alpha phase to the area under the curve was 2%, arguing against a relevant systematic error and supporting our use of a more parsimonious 1‐compartment model.

The pharmacodynamic model was informed by 2 readouts of the MES per patient, as used in the pivotal ACT trials.1 A logistic regression model was used to estimate transition probabilities given chance alone (null model), cumulative dose and exposure (cumulative area under the curve up to day 14, estimated day 14 trough concentration and cumulative area under the curve up to the time of endoscopy). As the major improvement in model fit occurred between the null model and the model using dose as predictor of probability of mucosal healing, drug administration was most relevant to treatment success. However, substituting dose with cumulative area under the curve (the best exposure metric) further improved the model fit, supporting a role for the optimisation of exposure (eg by TDM). Although a specific area under the curve can be obtained by many combinations of dose sizes and dosing intervals, we would like to caution against changing dosing intervals during the induction therapy based on our results alone. All dose optimisations were based on the existing time schedule of 0, 14 and 42 days.

Covariates identified for the pharmacokinetic model were time‐varying CRP, ALB and FFM, and baseline MES, concomitant corticosteroid use and pancolitis. Using FFM and thereby intrinsically accounting for sex differences on the physiologically most meaningful descriptor of size for antibodies apparently obviates the need to account for both body weight and sex. Thirty‐four patients (17%) in our cohort with acute severe UC, characterised by a high MES, low body mass index, low ALB and high CRP, had a lower infliximab exposure and may thus benefit from higher infliximab dosing during induction therapy.13

Contrary to Fasanmade et al.,26 we demonstrated that patients with endoscopic assessment of ulceration and/or spontaneous bleeding (ie MES 3) have a significantly higher infliximab elimination than patients with MES 2. Furthermore, not only disease severity, but also disease extent (absence/presence of pancolitis) was associated with reduced infliximab exposure. The effect of corticosteroid use at baseline on the infliximab volume of distribution might reflect the physician's anticipation of higher disease activity (leaky gut) in selected patients.26 Furthermore, higher CRP and lower ALB reflect higher disease activity, possibly explaining the increased infliximab elimination by a higher target load.11, 12 Although CRP is often in the normal range (≤5 mg/L) in patients with active UC, it was elevated in 112/204 (55%) patients in our cohort. As demonstrated by Brandse et al.,2 we confirmed that CRP predicts lower infliximab serum concentrations.

Antidrug antibodies were detected in 1% of the samples. As the assay used was unable to detect antidrug antibodies in the presence of quantifiable infliximab concentrations, this is probably an underestimation. Although antidrug antibodies are shown to have a great impact on infliximab pharmacokinetics during maintenance therapy, their occurrence during induction therapy is less well characterised, but presumably lower than during maintenance therapy, given the expected protective effect of concomitant immunosuppressive drugs and high infliximab exposure.3, 4, 27 We decided to omit the 4% samples with unquantifiable infliximab from the analysis owing to a lack of pattern (22 samples from 17 patients at variable time points).

Our exposure–response model can be used to optimise the probability of achieving mucosal healing on a population level. As mucosal healing occurred in 55% of our population, which consisted of an equal number of patients with moderate (MES 2) and severe (MES 3) UC, on the standard dosing regimen, targeting this percentage with our model should yield similar dose sizes (given unchanged times of dosing). As would be expected, targeting this probability of mucosal healing yielded a cumulative area under the curve target of approximately 2100 mg/L*day, which can be achieved by administering 5 mg/kg at days 0, 14 and 42 to the typical patient. In populations with a higher proportion of patients with severe UC, higher doses are needed, as these patients need additional exposure to go from MES 3 to 2.

Dose optimisation during infliximab induction therapy can be guided by TDM, thereby targeting all patients to a predefined “sufficient” exposure level (eg 3752 mg/L*day). Model‐based infliximab dose optimisation during induction therapy therefore facilitates the decision to continue/discontinue infliximab maintenance therapy. Justification of the increased drug expenditure depends on available resources, although improved outcomes and lower indirect costs should be balanced against this decision. Next to measuring infliximab concentrations (ie TDM), patient covariates can also be used to reduce pharmacokinetic variability and maximise mucosal healing rates.

To conclude, we hereby present the first exposure–response model of infliximab in UC. We identified the cumulative area under the curve up to the time of endoscopy and the trough concentration at day 14 as the best individual predictors of mucosal healing. Furthermore, we evaluated dosing regimens aiming at a predefined 70% probability of mucosal healing, proposing a first dose of 10 mg/kg for all patients. Ideally, the day 14 dose is then adjusted based on the day 14 trough concentration, requiring a rapid assay for point of care TDM.28 However, there were several limitations to our study, mainly with regard to the sparsely available pharmacokinetic and pharmacodynamic data, which resulted in high shrinkage of interoccasion variability, a slight misprediction in the visual predictive check at day 98 and a large number of failed bootstrap runs. Nevertheless, interoccasion variability was retained in the model to improve the description of individual trough concentrations. Interoccasion variability has previously been used to capture the time‐varying kinetics of infliximab that was not explained by covariates.29 The pharmacokinetics and pharmacodynamics of monoclonal antibodies are strongly interrelated. Although higher infliximab clearance may be the cause of therapeutic failure, it might as well be the mere consequence of more severe disease. The selected pharmacokinetic model predicted the individual concentrations well and is therefore able to aggregate to relevant metrics of exposure. Titrating individual patients to a target exposure will require model‐based application of Bayesian methodology (ie a dedicated software tool). Although our analysis revealed a clear association between exposure and treatment success, we must acknowledge that the model was built on retrospective data, with all the associated shortcomings. Therefore, infliximab dose optimisation strategies for improving mucosal healing rates in a population of patients with UC should be evaluated prospectively, ideally using clinical trial simulation and design using a pharmacokinetic–pharmacodynamic model.

COMPETING INTERESTS

G.V.A., M.F. and S.V. are senior clinical researchers at the Research Foundation Flanders (FWO). G.V.A. received financial support for research from Abbott and Ferring Pharmaceuticals; lecture fees from Janssen, MSD and Abbott; and consultancy fees from PDL BioPharma, UCB Pharma, Sanofi‐Aventis, Abbott, AbbVie, Ferring, Novartis, Biogen Idec, Janssen Biologics, Novo Nordisk, Zealand Pharma A/S, Millenium/Takeda, Shire, Novartis and Bristol Mayer Squibb. M.F. received financial support for research from Takeda; lecture fees from MSD, Janssen, AbbVie, Boehringer‐Ingelheim, Ferring, Chiesi, Tillotts, Zeria and Mitsubishi Tanabe; and consultancy fees from MSD, Janssen, AbbVie, Boehringer‐Ingelheim, and Ferring. S.V. received grant support from MSD, AbbVie, Pfizer and Takeda; lecture fees from AbbVie, MSD, Ferring Pharmaceuticals, Takeda; Hospira and consultancy fees from AbbVie, Takeda, Pfizer, Ferring Pharmaceuticals, Shire Pharmaceuticals Group, MSD, Hospira, Mundipharma, Celgene, Galapagos, Genentech/Roche. A.G. has served as a speaker for MSD, Janssen Biologicals, Pfizer, Takeda and AbbVie; as a consultant for UCB; and has received Investigator Initiated Research Grants from Pfizer and a national grant from Takeda. R.F. is an employee of SGS Exprimo NV. ED and TB declare that they have no conflicts of interest.

CONTRIBUTORS

E.D., A.G. and T.B. designed the research. E.D., G.V.A., M.F. and S.V. generated and assembled the data. E.D., R.F. and T.B. carried out the research and analysed the data. E.D. wrote manuscript. R.F., G.V.A., M.F., S.V., A.G. and T.B. carried out critical revision of the manuscript. All authors approved the final version of the manuscript.

Supporting information

Figure S1 Goodness‐of‐fit plots for the final pharmacokinetic model. (A) Observed infliximab concentrations vs population‐predicted infliximab concentrations. (B) Observed infliximab concentrations vs individual predicted infliximab concentrations. (C) Conditional weighted residuals vs time after dose. (D) Conditional weighted residuals vs population‐predicted infliximab concentrations. Dashed lines represent the locally weighted smoothing (LOESS) of the represented data. Due to a 45% ε‐shrinkage, the power of individual predictions (B) to detect model misspecification is reduced, while conditional weighted residuals (C and D) are not affected.

Figure S2 Correlations between the covariates withheld in the final pharmacokinetic model. Pairwise plots between the Mayo endoscopic subscore at baseline (MPRE), pancolitis at baseline (PANC), concomitant corticosteroids at baseline (CS), C‐reactive protein (CRP), albumin (ALB) and fat‐free mass (FFM) for patients informing the pharmacokinetic model.

Figure S3 Correlations between the evaluated exposure metrics. Pairwise scatterplots (under the diagonal), density plots (diagonal) and Spearman's rank correlation coefficients (above the diagonal) between the estimated infliximab trough concentration at day 14 (C₁₄), the estimated cumulative area under the curve up to day 14 (CAUC₁₄), the estimated cumulative area under the curve up to the time of endoscopy (CAUC_endoscopy), the cumulative dose to the time of endoscopy (CD_endoscopy) and time to endoscopy (T_endoscopy), for patients having post‐induction Mayo endoscopic subscore 0 or 1 (green), 2 (yellow) or 3 (red).

Figure S4 The simulated impact of the infliximab dose (same dose at days 0, 14 and 42) on the probability of mucosal healing (based on the cumulative area under the infliximab concentration–time curve up to day 84). N = 50 000 stochastic simulations. The solid line represents the median probability, the grey shaded area is the 90% prediction interval, crosses represent the predicted arithmetic mean probability, the dashed lines mark the probability range of mucosal healing corresponding to the predicted exposure range in the original dataset. The stochastic simulation of 10 mg/kg resulted in 22% of subjects in a cumulative area under the infliximab concentration–time curve at day 84 (CAUC₈₄) higher than the maximum predicted CAUC₈₄ of 6385 mg/L*day in the original dataset. For these subjects, the model predictions are based on an extrapolation beyond the observed exposure range, so it is not clear whether the model still holds.

APPENDIX A.

Pharmacokinetic model

APPENDIX B.

Exposure–response model

Dreesen E, Faelens R, Van Assche G, et al. Optimising infliximab induction dosing for patients with ulcerative colitis. Br J Clin Pharmacol. 2019;85:782–795. 10.1111/bcp.13859