Pharmacokinetic–Pharmacodynamic Modelling of Fostamatinib Efficacy on ACR20 to Support Dose Selection in Patients with Rheumatoid Arthritis (RA)
Abstract
R788 (fostamatinib) is an oral prodrug that is rapidly converted into the relatively selective spleen tyrosine kinase (SYK) inhibitor R406, which is being evaluated for the treatment of rheumatoid arthritis (RA). This analysis aimed to develop a pharmacodynamic model for efficacy using pooled ACR20 data from two Phase II studies in RA patients (TASKi1 and TASKi2). The goal was to describe the effect of fostamatinib as a function of fostamatinib exposure (dose and R406 plasma concentration) and other explanatory variables.
The exposure-response relationship of fostamatinib was incorporated into a continuous-time Markov model, which describes the time course of transition probabilities between three possible states at each visit: ACR20 non-responder, responder, and dropout. The probability of transitioning to the ACR20 response state was linearly related to average R406 plasma concentrations, and the onset of this drug effect was rapid. Furthermore, increasing the fostamatinib dose led to higher dropout rates and subsequent loss of efficacy.
This analysis provided a deeper understanding of the exposure-response relationship and supported the fostamatinib 100 mg twice daily (BID) dose regimen as appropriate for further clinical evaluation. By modeling the relationship between drug exposure and response, the study offered valuable insights into optimizing the dosing strategy for fostamatinib in RA patients.
Rheumatoid arthritis (RA) is a common chronic inflammatory and destructive arthropathy that currently lacks a cure, imposing significant personal, social, and economic burdens. Spleen tyrosine kinase (SYK) is an intracellular cytoplasmic tyrosine kinase that plays a crucial role as a mediator of immunoreceptor signaling in macrophages, neutrophils, mast cells, and B cells. In RA patients, SYK is present in the synovium and its activation is essential for cytokine and metalloproteinase production induced by tumor necrosis factor (TNF) α in fibroblast-like synoviocytes. R788 (fostamatinib), an oral SYK inhibitor being evaluated for RA treatment, is a prodrug rapidly converted into R406, which exhibits potent anti-inflammatory activity. This suggests that SYK inhibition could be a viable therapeutic strategy for RA. The American College of Rheumatology 20% response criteria (ACR20) is a widely used and accepted measure of treatment response and efficacy in clinical trials.
In two Phase II studies, fostamatinib was evaluated for its efficacy in RA patients. TASKi1 showed statistically significant improvements when fostamatinib was combined with methotrexate (MTX), but not when used alone. Another Phase II study in patients who had failed biologic treatment did not demonstrate statistically significant efficacy.
Building on pooled phase II data from TASKi1 and TASKi2, a longitudinal pharmacokinetic-pharmacodynamic (PKPD) analysis of ACR20 response, the primary efficacy endpoint, was conducted. The primary goal of this analysis was to understand the exposure-response relationship and provide support for the selected doses. The response was linked to patient-specific fostamatinib exposure (dosing regimen, R406 plasma concentrations) as well as other relevant patient and study covariates, while explicitly accounting for study dropout. TASKi1 included a wide range of doses but was confounded because the low dose was administered in the US and the high dose in Mexico.
The placebo response appeared to vary between these two populations, complicating the interpretation of the observed dose-response relationship. The current analysis sought to improve the understanding of the exposure-response relationship by incorporating relevant covariates and also included TASKi2, which was conducted in a similar patient population. Therefore, an integrated analysis of TASKi1 and TASKi2 was performed, aiming to enhance the understanding of the dose-response relationship by leveraging more information from the pooled studies while simultaneously providing an opportunity to describe differences between studies, baseline demographics, and disease severity through the assessment of covariate effects, which were evaluated on both the placebo response and the drug effect.
The developed model was then used to conduct model-based simulations of the expected dose-response relationship of fostamatinib under various population settings, as identified by the model. Repeated ACR20 measurements in each patient were expected to be correlated. To account for this serial correlation, a continuous time Markov model that accommodates both equally and unequally spaced observations was considered. Furthermore, a mixed-effects statistical model was employed to account for random variability in ACR20 response between patients. This Markov approach has been utilized in similar applications to capture the time course of the ACR20 response in RA and also to describe transitions between stages of sleep.
Methods
Study Patients
TASKi2 was a Phase II, multicenter (US and Mexico), randomized, double-blind, placebo-controlled, ascending dose, dose-ranging study designed to evaluate three doses of fostamatinib (50 mg BID, 100 mg BID, and 150 mg BID). A total of 186 patients who had been diagnosed with RA for a minimum of 12 months and were receiving a weekly MTX dose for a minimum of 6 months were enrolled in this 12-week study.
TASKi2 was also a Phase II, multicenter (US, Latin America, Europe), randomized, double-blind, placebo-controlled, parallel dose study aimed at evaluating two dosing regimens of fostamatinib (100 mg BID and 150 mg once daily (QD)). A total of 457 patients who had been diagnosed with RA for a minimum of 6 months and were receiving a weekly MTX dose (7.5–25 mg/week) for a minimum of 3 months were enrolled in this study for a 6-month treatment period.
ACR20 measurements in TASKi1 were available on weeks 1, 2, 3, 4, 8, and 12, while in TASKi2 they were available on weeks 1, 2, 4, 6, 8, 12, 16, 20, and 26. This extended dataset allowed for a more detailed assessment of the fostamatinib dose-response relationship over time. The design of TASKi2 facilitated a broader understanding of the drug’s efficacy across different geographic regions and dosing schedules, providing valuable insights into the optimal dosing regimen for further clinical evaluation.
Handling Missing/Censored Observations
If a patient missed an intermediate visit, the specific time point was excluded from the analysis, assuming a missing at random scenario. However, the first occurrence of any of the following events was regarded as a “dropout” point for that patient:
1. Discontinuation of fostamatinib treatment more than 2 days prior to the last planned visit (indicating early study termination),
2. Adjustment of the fostamatinib dose, including dose reductions as permitted by the study protocol,
3. Change in the selected concomitant RA medication, such as steroids or disease-modifying antirheumatic drugs (DMARDs), including methotrexate (MTX).
Changes in non-steroidal anti-inflammatory medications were not considered as a reason for dropout.
Statistical Analysis
A continuous-time Markov model was employed to describe the time course of transition probabilities between the three possible states of non-responder (0), responder (1), and dropout (2) at each visit. This model characterizes the probability of a subject being in any one particular state at a given visit, conditional on the state at a previous visit. Conceptually, the continuous-time Markov model was formulated as a three-compartmental disposition model that predicts the probabilities of the response being allocated to one of the three possible states over time. The model accounts for correlation and differences in time between successive observations and seamlessly incorporates censored data or dropout. A schematic representation of the model is provided in Figure 1. The rate constants K01, K10, K12, and K02 define the rate of change of probabilities between the three states in the indicated directions. As described in Lacroix et al., for non-responders at each visit, there are three possible transitions to the next observation: 00, 01, or 02, representing staying as a non-responder, becoming a responder, or dropping out, respectively, and the sum of the three associated probabilities equals one.
First, a base pharmacodynamic model for the ACR20 state was developed. The effect of fostamatinib exposure on the ACR20 state was investigated either as the actual daily dose or the patient-specific average steady-state R406 plasma concentration (Css) predicted from a previously developed population pharmacokinetic (PK) model. This model was a two-compartmental disposition model with first-order, delayed absorption, where clearances and volumes were scaled to body weight allometrically. Age, gender, creatinine clearance, and race (categorized as Black, Hispanic, or Caucasian) were evaluated in the PK analysis and did not influence CL/F. The CL/F of R406 (including the formation of the metabolite) was estimated to be 18.4 L/h (for a 70 kg patient) with an inter-individual variability of 40% (the ETA-shrinkage for CL/F was 13%). The average steady-state concentrations were predicted based on individual estimates of CL/F and the daily dose using the equation Css = Daily dose / CL/F.
Various functional forms were considered to evaluate the effect of time and fostamatinib exposure on the non-responder to responder transition (K01), including linear, piecewise linear, and nonlinear, asymptotic (EMAX) models. Next, the potential effect of covariates such as demographic factors, disease severity, medication history, and dosing frequency on the transition probabilities was investigated using a forward addition and backward elimination selection procedure. The goodness-of-fit of the models was based on maximum likelihood principles; the difference in the objective function value (OFV, minus twice the log likelihood) in NONMEM (the likelihood ratio test) was used to discriminate between candidate models. For the covariate selection procedure, significance levels (α) of 0.05 and 0.01 were used for the forward addition and backward elimination procedures, respectively.
Appropriateness of the models was also assessed graphically using visual predictive checks (VPC), focusing on the exposure-response and time course of non-response, response, and dropout states. On the basis of the final ACR20 model, the expected absolute proportions of ACR20 responders and corresponding differences from placebo for various dosing regimens were predicted using simulations. Non-responder imputation (NRI) was used for simulated dropout. Simulations were based on the mean of the individual patient-specific average concentrations in a particular dose regimen. Uncertainty in the final model parameters was accounted for by repeated sampling (200 times) of fixed-effect parameters from the respective variance-covariance matrix.
Inter-individual variability was incorporated by repeated sampling (1,000 patients per dose regimen) of the random effects (h) from the respective variance-covariance matrix. The covariate distribution in the simulation dataset was based on the distribution of the same covariates in the analysis dataset. Graphical and tabular representations of model predictions were presented as dose-response relationships, although the underlying models are based on final exposure-response models that use the average plasma concentrations per dose regimen to drive the outcome. The respective average plasma concentrations used in the simulations for placebo, 50 mg BID, 150 mg QD, 100 mg BID, and 150 mg BID were 0, 219, 327, 547, and 1164 ng/mL, respectively.
The analysis was performed using NONMEM version 7.1.2 (Icon Development Solutions, Ellicott City, Maryland) in conjunction with Perl-speaks-NONMEM (PsN) version 3.5.3. The construction of the analysis datasets, exploratory graphs and tables, and summarization of simulation results were performed in R software package version 2.15.
Results
Data
Data from a total of 641 patients from the TASKi1 and TASKi2 studies were utilized in this analysis. A detailed description of the data and a summary of demographic and baseline characteristics in these studies are provided elsewhere.
For all graphical explorations of the ACR20 proportion of responders, missing observations post-dropout were defined as non-responders. The overall time course of the proportion of ACR20 responders in each dose regimen is depicted in Figure 2, with the left panel showing TASKi1 and the right panel showing TASKi2. The plots indicate an increase in response with higher doses in both TASKi1 and TASKi2. The decline in response for the 150 mg BID group in TASKi1 is primarily due to a relatively large proportion of patients being treated as dropouts (as seen in the left panel of Figure 2). The immediate effect of fostamatinib treatment is apparent, particularly in the 100 mg BID group, which showed a response rate of approximately 40% at week 1 compared to about 15% in the placebo group. The exposure-response relationship based on quantiles of the plasma concentrations is illustrated in Figure 3 (center) and suggests that response increases with higher exposure levels.
Pharmacodynamic Analysis
Based on the schematic outline of the ACR20 model, as described in the study, several models were explored to establish a reliable base model. The analysis revealed that average plasma concentration served as a better predictor for the probability of transitioning to an ACR20 response state than daily dose. This finding was supported by a 29-point drop in the objective function value (OFV), suggesting that pharmacokinetic variability accounts for a significant portion of the differences in ACR20 response among individuals.
The final base model that emerged from the analysis incorporated piece-wise linear time functions for K01PL and K10. It also included a linear effect of average plasma concentration on K01Drug, alongside a random effect on K01.
During the covariate selection process, two covariates were identified as statistically significant. These included the effect of region on drug-induced transition from non-responder to responder (K01Drug) and the impact of prior use of biological treatments on the placebo-induced transition to response (K01PL). No additional covariates were found to be statistically significant at this stage.
A visual predictive check (VPC) of the interim covariate model suggested the inclusion of two more elements: a class effect of the 150 mg BID dosing regimen (in contrast to all other dosing regimens) on K12, and a linear increase in treatment effect over time. These adjustments further reduced the OFV by 37 and 18 points, respectively. Parameter estimates derived from the final model are provided in Table 1.
According to the final parameter estimates, the effect of region on the slope between R406 plasma concentration and K01 indicates a decrease in this rate constant for patients in the USA and Europe when compared to those in Latin America. The reduction is approximately 40% and 70% for the USA and Europe, respectively. Additionally, previous use of biological therapies reduced the placebo effect on K01 by about 86%.
Furthermore, the transition rate from responder to dropout (K12) was approximately 720% higher in the 150 mg BID group compared to other dosing regimens.
The VPC plots illustrating the probabilities of non-response, response, and dropout in relation to average R406 plasma concentrations confirmed the accuracy of the final model. These plots showed a close match between model-predicted and observed values. Time-based VPCs, stratified by regimen and trial, also demonstrated that the model adequately captured the time course of these outcomes.
Simulation of ACR20 Response
Based on the final model, predictions for the various dosing regimens were carried out at 12 and 26 weeks, corresponding to the study durations of TASKi1 and TASKi2, respectively. A graphical representation of the predicted ACR20 proportion of responders for the different dosing regimens and the active dosing regimen differences from placebo at 12 and 26 weeks is provided in Figure 4. The results indicate that the response increases with the dose up to 100 mg BID. For the 150 mg BID dose, a smaller difference from placebo is observed due to the higher dropout rate (as seen in Figure 2). Figure 4 demonstrates that, based on the confidence intervals of the predicted ACR20 difference versus placebo, all doses (except the 150 mg BID) performed better than placebo at both 12 and 26 weeks. Excluding the 150 mg BID group, which is influenced by dropout, a clear dose-response relationship becomes evident in Figure 4.
The model-predicted exposure-response estimates stratified by region are presented in Table 2. The findings suggest that the lowest fostamatinib effect is anticipated in Europe, with higher efficacy expected in the USA and the highest in Latin America. It was also noted that the model-predicted response in the placebo group is lower for patients with prior use of biologicals compared to those without such use. These results provide valuable insights into regional differences in fostamatinib efficacy and the impact of previous biological treatment on placebo response, which could inform future clinical trial designs and regional adaptation of therapeutic strategies.
Discussion
A pharmacodynamic model was developed to characterize the time course of the exposure-response relationship of fostamatinib for the clinical endpoint ACR20, based on pooled data from two Phase II studies (TASKi1 and TASKi2). While the two studies could have been analyzed separately to account for inter-individual and inter-trial variability, the approach taken here was an integrated analysis of both studies. This integrated approach was considered more advantageous as it provided a better opportunity to describe and understand the dose-response relationship while simultaneously capturing differences between the studies in baseline characteristics and disease severity through covariates. It is worth noting that several efficacy outcomes (e.g., ACR20/50/70 and Disease Activity Score for 28 joints [DAS28]) and safety variables should ideally be analyzed, and their results synthesized for the purpose of dose selection. However, this paper focuses solely on ACR20, which was the primary endpoint in the individual trials.
The main objective of this analysis was to better characterize the relationship between fostamatinib exposure and response by employing an integrated analysis that incorporates the time course of the drug effect. This enabled the assessment of the onset of the drug effect. The binary ACR20 response was analyzed using a continuous-time Markov model, formulated as a three-compartmental model with transitions between non-responder, responder, and dropout states. Serial ACR20 measurements within a patient are expected to be correlated. When the correlation is moderate, mixed-effects logistic regression models are generally sufficient.
However, when the correlation between ACR20 measurements is substantially higher, as expected for serial observations, a Markov model is better suited to account for this serial correlation. The model also accommodates both equally and unequally spaced measurements. Furthermore, non-random study dropout related to the ACR20 response state, as observed in the study data, can be easily integrated into the Markov model, which defines probabilities of transitioning between various states (response, non-response, and dropout). Although not explicitly stated in the analysis, non-random dropout due to lack of efficacy and safety/tolerability concerns was implicitly captured in the transitions to the dropout state from the non-responder and responder states, respectively, as illustrated in Figure 1.
Another commonly used approach for modeling ACR20/50/70 data is the latent variable approach. In this method, a latent variable—an underlying (unobserved) continuous variable that determines the disease condition, such as inflammation—is assumed. This latent variable can then be mapped onto a binary or categorical outcome using a threshold, allowing the discrete responses to be related to exposure based on the postulated mechanism. This approach offers a parsimonious way to jointly model multiple outcomes, such as ACR20, ACR50, and ACR70.
However, the latent variable approach does not handle dropout, which can be effectively managed by the Markov model. Additionally, as mentioned by Hu et al., the Markov model has the advantage of accounting for within-subject correlations of observations. In their application of the latent variable approach, Lacroix et al. reported that the latent variable model alone did not seem to capture the serial correlation in the data. A larger drop in the objective function value was observed when a Markov element was included in the model.
Ignoring dropout, especially non-random, exposure-related dropout (e.g., due to lack of efficacy or safety/tolerability issues), could lead to biased results, loss of statistical power to detect drug effects, and potentially inappropriate conclusions. The final model identified that dropout was dose-dependent, resulting in a “bell-shaped” relationship between drug exposure and ACR20 response (Figure 4). This indicates that simply increasing the fostamatinib dose does not necessarily lead to a higher ACR20 response. Instead, exposure-related dropout due to safety/tolerability concerns, followed by non-responder imputation (NRI), results in lower response rates at the highest fostamatinib dose.
Accounting for this dropout improved the model fit, as evidenced by the numerical drop in the objective function value (OFV) and a better visual representation of the response profile. Dropout was defined here as either discontinuation of treatment, adjustment of the randomized fostamatinib dose, or change in concomitant medication. The purpose of this analysis was to characterize the relationship between fostamatinib drug concentration and ACR20 response (per protocol analysis). Therefore, this definition of dropout differed somewhat from the definition used in the reported clinical results of these studies (intention-to-treat analysis). As a result, the reported ACR20 response rates may vary between the two types of analyses.
The average steady-state plasma concentration of fostamatinib was found to be a better predictor of efficacy (ACR20 response and dropout) than daily dose. This implies that a portion of the variability in the ACR20 response can be explained by variability in fostamatinib plasma concentration. Furthermore, the onset of response, in terms of difference from placebo, was rapid and complete from the first assessment onward.
In TASKi1 and TASKi2, the lowest dose of 50 mg BID was investigated only in the USA population, while the highest dose of 150 mg BID was evaluated only in the Mexican population. By explicitly assessing the effect of covariates, including region, on both the placebo effect and the drug effect, the model allowed for adjustments to account for these region-dependent dose allocations. It is worth noting that not all doses were investigated in all regions.
For example, the 50 mg BID and 150 mg BID treatments have not been evaluated in the European population, so predictions for this group were based on extrapolations of regional effects at other doses (Table 2). The model accounted for the effect of statistically significant covariates on the probability of transitioning from the non-responder to responder state. A region-dependent effect of fostamatinib was identified, and the placebo response tended to depend on the use of biologicals. However, it is important to note that the majority of patients in both TASKi1 and TASKi2 had no prior use of biologicals, while almost all patients had prior use of disease-modifying antirheumatic drugs (DMARDs), including methotrexate (MTX).
Therefore, this result should be interpreted cautiously. This analysis suggests a difference in therapeutic effect between geographical regions, which could be due to variations in patient selection across the regions. No difference in CL/F between ethnic groups was found in the pharmacokinetic (PK) analysis. However, region is correlated with dose (certain doses/regimens were evaluated in mutually exclusive geographical regions or ethnic groups), and thus with Css. This means the regional difference could theoretically be due to confounding between region and exposure. Subsequent predictions of dose-response were based on the final pharmacodynamic model, which related ACR20 response and dropout to fostamatinib concentration.
These model-based predictions illustrated the relationship between fostamatinib dose and ACR20 response at weeks 12 and 26 (Figure 4). Although the probability of response increases monotonically with average concentration/daily dose, the apparent dropout rate at higher doses (150 mg BID) and the NRI for dropout result in an overall lower response rate at this highest dose (Figure 2). This “bell-shaped” relationship between fostamatinib dose and expected ACR20 response identified maximum efficacy at a total daily dose of 200 mg. Consequently, the current analysis supports the appropriateness of the 100 mg BID dose for testing in Phase III studies.
The higher dropout rates in the 150 mg BID regimen could be linked to adverse events. The two most common adverse events in the overall safety population were diarrhea and neutropenia, which occurred more frequently in the active drug groups. Other gastrointestinal adverse events included nausea and gastritis, which also occurred more commonly in the active drug groups compared to placebo.
In these Phase III confirmatory trials, a statistically significant effect of fostamatinib on ACR20 response compared to placebo was observed. However, the size of the fostamatinib treatment effect was not large enough to justify continuing its further development for the treatment of patients with RA.
In summary, this integrated analysis clarified the relationship between R406 exposure in plasma and ACR20 response. The findings supported the notion that a fostamatinib regimen of 100 mg BID offers an optimal ACR20 efficacy profile, aligning with the expectations derived from the Phase II data. Despite this, the overall treatment effect was deemed insufficient to proceed with the drug’s advancement in RA therapy.