Population Pharmacokinetic, Exposure‐Response Efficacy and Safety Analyses of Favezelimab in Patients With Solid Tumors

ABSTRACT

Favezelimab (MK‐4280) is a humanized monoclonal antibody (mAb) targeting lymphocyte activation gene‐3 (LAG‐3). Favezelimab population pharmacokinetics (popPK) was evaluated using data from 2 trials; a Phase 1 study evaluating favezelimab across a range of doses (7 to 800 mg administered intravenously [IV] every 3 weeks [Q3W]) in patients with advanced solid tumors, and a Phase 3 study evaluating 800 mg IV Q3W favezelimab combined with pembrolizumab (200 mg Q3W) in patients with programmed death ligand 1 (PD‐L1) positive colorectal cancer (CRC). Prespecified covariates of clinical interest were evaluated using a full model. Favezelimab exposure measures from the final popPK model were used to evaluate the exposure‐response (E‐R) relationships with objective response rate (ORR), Grade ≥ 3 drug‐related adverse events (AE), and drug‐related AE of special interest (AEOSI) occurrence in gastric cancer patients evaluated across 2 randomized doses. The final popPK model was a 2‐compartment target‐mediated drug disposition (TMDD) model and included the effect of antidrug antibody status on linear clearance (CL). Population volume and clearance parameters, and corresponding random effects were estimated with < 15% relative standard error. Lower body weight, as well as female sex and Asian race (covariates correlated with lower body weight), were associated with higher exposure. Lower albumin levels were also associated with lower exposure. Increasing favezelimab exposures were associated with increased ORR and had no significant effect on safety. No covariates were associated with ORR. This analysis was used to characterize variability in favezelimab pharmacokinetics and support the choice of Phase 3 dose.

Study Highlights.

• What is the current knowledge on the topic?

  • ○
    Favezelimab binds to the immune checkpoint receptor LAG‐3 and blocks the interaction with ligand MHC class II. Because of its proposed role on both effector T cells and Tregs, LAG‐3 is one of several immune checkpoint molecules for which blockade of both cell populations has the potential to enhance antitumor immunity.
• What question did this study address?

  • ○
    Data from 2 clinical trials of patients with advanced solid tumors (Phase 1) and metastatic CRC (Phase 3) were used to develop a popPK model for favezelimab and evaluate the effect of covariates on its PK. Favezelimab exposures were used to evaluate the E‐R relationship between favezelimab and ORR, Grade ≥ 3 drug‐related AE occurrence, and drug‐related AEOSI in gastric cancer patients in the Phase 1 study.
• What does this study add to our knowledge?

  • ○
    The popPK model for favezelimab and E‐R models for safety and efficacy endpoints adequately characterized the observed data in the 2 clinical trials. The PK was described by a 2‐compartment TMDD model with MM approximation with an effect of positive ADA status on the linear component of CL. Race, sex, body weight, and albumin were found to be clinically relevant covariates. The E‐R modeling showed a significant relationship between favezelimab exposure and ORR in gastric cancer patients, where higher favezelimab AUC was associated with a higher probability of objective response. No significant E‐R safety relationship was found between favezelimab exposure and the probability of Grade ≥ 3 drug‐related AEs or drug‐related AEOSIs.
• How might this change drug discovery, development, and/or therapeutics?

  • ○
    This study has improved scientific understanding of favezelimab PK and the determinants of drug exposure, safety, and efficacy.

1. Introduction

Favezelimab is an immunoglobulin G4 (IgG4) monoclonal antibody (mAb) that binds to the immune checkpoint receptor lymphocyte activation gene‐3 (LAG‐3) and blocks the interaction with its ligand, major histocompatibility complex (MHC) class II. LAG‐3 is an inhibitory immune modulatory receptor that regulates both effector T‐cell homeostasis, proliferation, and activation and plays a role in the suppressor activity of regulatory T‐cells (Tregs) [1, 2, 3]. Clinical trials investigating the efficacy of immunomodulatory antibodies, including favezelimab, show promising results across various cancer types in monotherapy and combination therapy [4, 5, 6, 7].

Favezelimab has been evaluated as monotherapy at doses ranging from 7 to 800 mg IV Q3W, as combination therapy with pembrolizumab (200 mg Q3W) at 800 mg Q3W, and as a co‐formulation (named MK‐4280A) of favezelimab (800 mg) with pembrolizumab (200 mg) administered Q3W [6, 7].

This analysis describes popPK model development using data from clinical studies P001 (NCT02720068 [Phase 1]) and KEYFORM‐007 (NCT05064059 [Phase 3]) to characterize the pharmacokinetics and quantify sources of variability for favezelimab. Logistic regression modeling was used to establish exposure‐response (E‐R) relationships for efficacy (ORR) and safety (Grade ≥ 3 drug‐related adverse event [AE] occurrence and drug‐related AE of special interest [AEOSI] occurrence) across two randomized doses in patients with gastric cancer in P001.

2. Methods

2.1. Clinical Trial Designs and Populations

Study P001 (NCT02720068) was a first‐in‐human Phase 1, multi‐site, multi‐cohort, open‐label study of favezelimab assessing both monotherapy and combination treatment with pembrolizumab with or without chemotherapy or lenvatinib in patients with advanced solid tumors that progressed on all available standard‐of‐care therapies or with select solid tumors (colorectal, head and neck, and gastric cancer). The co‐formulation of favezelimab with pembrolizumab was initially evaluated in this study [7].

Study KEYFORM‐007 (NCT05064059) was a Phase 3, randomized, active‐controlled, parallel‐group, multi‐site, open‐label, safety and efficacy study of co‐formulated favezelimab/pembrolizumab in patients with PD‐L1 positive metastatic colorectal adenocarcinoma and who had progressed on or could not tolerate previous treatment. The specific population of colorectal cancer (CRC) patients who were chosen for evaluation in Phase 3 (PDL‐1 positive patients, who had progressed on prior therapy) were not independently tested at lower doses in PN001. Thus, the Phase 3 data at a single Phase 3 dose could not be pooled with comparable data in PN001. This decision to not utilize efficacy data from the single‐dose level in the Phase 3 study was prespecified in a modeling analysis plan finalized before Phase 3 data readout.

Both trials were conducted in accordance with the principles of the Declaration of Helsinki, the ethical guidelines of the Council for International Organizations of Medical Sciences, the Good Clinical Practice guidelines of the International Council of Harmonization, and all applicable laws and regulations. An institutional review board or independent ethics committee at each trial center approved the protocol, and all participants provided written informed consent.

2.2. Data Analyses and Procedures

PopPK modeling used data from all patients who received at least 1 dose and had at least 1 measurable favezelimab postdose concentration. Nonlinear mixed‐effects modeling using the stochastic approximation expectation–maximization (SAEM) method in Monolix Version 2024R1 was performed to characterize favezelimab PK. R Version 4.3.2 supported the exploratory analysis and post‐processing of model outputs. Logistic regression characterized E‐R relationships for efficacy (ORR) and safety (Grade ≥ 3 drug‐related AE occurrences and safety drug‐related AEOSI). AEOSIs were a predefined list of AEs of interest expected upon treatment with immune modulatory combinations (Table S10). All E‐R modeling was conducted using R Version 4.3.2.

The model development strategy for PK and E‐R included exploratory data analysis, base structural model development, evaluation of covariate effects, model refinement, and model evaluation.

During base model development, 1‐, 2‐, and 3‐compartment models were evaluated. Models with linear CL and both linear and nonlinear CL were also evaluated. Time‐dependent reductions in CL have been reported for pembrolizumab [8] and thus were also explored for favezelimab.

For each E‐R analysis, the relationships of popPK model‐derived favezelimab exposure measures (average area under the concentration‐time curve [AUC_avg] considering individual‐specific dosing history, average maximum concentration [C_max,avg], Cycle 1 AUC, and Cycle 1 C_max) on the E‐R endpoints were evaluated using linear and log‐linear functions of the logistic regression models. If an exposure metric did not demonstrate a statistically significant (p‐value < 0.05 for estimated exposure effect) E‐R relationship in the base model, no additional modeling was conducted for that endpoint.

Since ADA incidence occurs at different times for individual patients and can also occur transiently during treatment, no unique time to steady state was defined across patients. Consequently, steady‐state exposures were not separately evaluated as an exposure metric in E‐R analyses.

2.3. Assessing the Impact of Covariates

Following development of an appropriate PK base model, the influence of covariates (age, baseline body weight, sex, race, albumin, Eastern Cooperative Oncology Group [ECOG] performance status, and hepatic and renal impairment) was evaluated by developing a full model that included these covariates on linear CL and central volume (V_c). These covariates were prespecified to be of potential clinical significance prior to the analyses. Under this full model approach, all covariates were retained in the final model regardless of effect size or estimated precision. This allowed us to avoid making selections of covariate relationships (amongst correlated covariates) based on likelihood‐based p‐value assessments without multiplicity adjustments [9].

Continuous covariates were included using the power model and categorical covariates were incorporated using the exponential model. Missing values for continuous covariates were imputed with the median value. Because some categorical covariate groups had small patient numbers, grouping was performed for selected categories, as detailed in the footnote of Table S2.

Noncompartmental analyses of PK data from P001 have previously demonstrated that favezelimab exposures are similar in monotherapy and in combination with pembrolizumab. Similarly, pembrolizumab PK was shown to be unaffected by favezelimab. Since both favezelimab and pembrolizumab are cleared by protein catabolism pathways that are far from saturation at therapeutic doses of these antibodies, no PK interaction is expected, and pembrolizumab PK was not explored as a covariate on favezelimab PK.

No separate attempt was made to model dose dependence of ADA incidence or timing of onset. Observed ADA status was instead treated as a time dependent regressor. Simulations at the clinically relevant dose of 800 mg were performed using observed ADA incidence/timing data at this dose.

Based on the final model, individual empiric Bayesian PK parameter estimates for patients in the analysis population were used to simulate drug exposures at steady state (AUC across a 21‐day dosing period at steady state [AUC_{0_tau,ss}] and C_max at steady state [C_max,ss]) for the patients who received a dose regimen of favezelimab 800 mg IV every 21 days for 5 cycles. The effect of ADAs was included by carrying forward the last observation as a time‐dependent regressor on CL. Forest plots were created using the data from the 800 mg dose group to evaluate the clinical significance of the covariates included in the full model and to visually depict the relative difference in AUC_{0_tau,ss} and C_max,ss exposure (each expressed as a geometric mean ratio [GMR] and 90% confidence interval [CI]) due to the effect of each covariate.

For E‐R model development, continuous and categorical covariates (baseline tumor size [mm], sex, age [years], and race) were tested in models that showed a significant E‐R response. The choice of covariates evaluated was based on clinical relevance and initial data exploration. Forward selection (at p‐value of 0.01) and backward elimination (at p‐value of 0.001) procedures were used for this process.

2.4. Model Evaluation

The final popPK model was evaluated using prediction‐corrected visual predictive check (pcVPC) procedures [10]. Other diagnostic plots, including goodness‐of‐fit (GOF) plots and normalized prediction distribution errors (NPDEs) versus time and population predicted concentrations as well as parameter precision (relative standard error expressed as a percent [%RSE]) were also evaluated.

The final E‐R models were evaluated by comparing model predictions to the observed incidence of specified safety or efficacy endpoints in each exposure quartile to the fitted model predictions. Parameter precision (%RSE) was also used to assess model adequacy.

3. Results

3.1. Patient Characteristics

A total of 724 patients contributing 8980 concentrations from Studies P001 and KEYFORM‐007 were included in the popPK analysis. The analysis population had a median age of 59 years (range: 25 to 85 years), with a median body weight of 70.9 kg (range: 33.0 to 145.9 kg). Overall, the analysis population included more male patients (61.9%) than female patients (38.1%). This population was primarily Caucasian (68.2%) and Asian (27.1%), while the remaining racial categories (black/African American, American Indian/Alaska Native, Native Hawaiian/Pacific Islander, multiple, or unknown) accounted for a small percentage (4.6%) of the analysis population and, therefore, were later combined with the most common group (Caucasian) for the PK analysis of covariate effects.

A summary of study designs, patient demographics, and baseline covariates are shown in Table S1 and Table S2. In Table S3, the incidence of ADAs and timing of onset of ADA positivity by dose are summarized. Overall, the median time to ADA development was approximately 3 weeks. The frequency of ADA‐positive responses differed by dose and was higher in those patients who received favezelimab doses lower than 700 mg. The incidence of ADA development was 26.2% in patients who received favezelimab 800 mg Q3W, and median ADA onset occurred by 3 weeks after the first favezelimab dose.

Tables S4 and S5, and Figure S1 illustrate the correlation between covariates. It should be noted that in the presence of correlations among covariates, absolute values of covariate–effect estimates should be interpreted with caution. Instead, simulation‐based evaluation was used to assess the relevance of the studied covariates.

3.2. Population Pharmacokinetic Model Development

Figure S2 depicts mean concentration‐time profiles for favezelimab in Cycle 1 stratified by dose, and indicates that favezelimab PK exposure increased with increase in dose and exhibited nonlinear PK at lower doses. Concentration‐time profiles for favezelimab appear to decrease more rapidly as drug concentrations fell below approximately 1000 ng/mL. In contrast, at doses beyond 7 mg, the terminal decline was slower, and systemic exposure increased more proportionately. This is consistent with the expected PK behavior of an antibody exhibiting TMDD. Dose normalized AUC and Cmax values calculated from non‐compartmental analyses of PK data in PN001 are presented in Figures S3 and S4, respectively. Exploratory data analysis also showed that ADA‐positive patients had lower exposure of favezelimab compared to ADA‐negative patients at 70 and 200 mg (Figure S5). In addition, model fit improved when the effect of ADA on CL was included in the model. This behavior was less evident at the 800 mg dose level. In keeping with these exploratory plots, model fits improved when the effect of anti‐drug antibody (ADA) on CL (increase) was included in the model. Inclusion of both linear and nonlinear CL also improved the model‐corrected Bayesian information criterion.

The final base model was a 2‐compartment TMDD model with MM approximation with an effect of positive ADA status on the linear component of CL. The model was parameterized using CL (L/h), V_c (L), intercompartmental clearance (Q, L/h), peripheral volume (V_p, L), maximum velocity (V_max, μg/h), and Michaelis constant (K_m, μg/mL). Interindividual variability (IIV) was included on V_c and CL using the exponential model with a covariance parameter between CL and V_c, and residual variability (RV) was described using an additive plus proportional error model. During this stage, favezelimab concentrations associated with absolute population weighted residuals (|PWRES|) > 6 were identified and excluded from the base model.

The parameter estimates for the final popPK model, along with corresponding precision estimates (%RSE), are presented in Table 1. All fixed effect parameters were estimated precisely (%RSE < 15%), other than K_m (65.4% RSE), as well as all random effect parameters (%RSE < 5% and %RSE < 10% for IIV and RV parameters, respectively). The poor precision on K_m was likely due to the fact that K_m is not an easily identifiable parameter for individuals with data in a limited concentration range. Also, at the Phase 3 dose of 800 mg, trough concentrations were much higher than K_m, suggesting that TMDD was fully saturated by this level.

TABLE 1.

Final population pharmacokinetic parameters for Favezelimab.

Parameter	Population parameter estimate	Stochastic approximation				Conditional mode
		SE	%RSE	P2.5	P97.5	Shrinkage (%)
Fixed effects
Vc (L)	3.08	0.0665	2.16	2.95	3.21	24.4
ECOG effect on Vc	−0.0031	0.0186	600	−0.0395	0.0333
Sex effect on Vc	0.255	0.0208	8.14	0.214	0.296
Body weight effect on Vc	0.353	0.0433	12.3	0.268	0.438
Albumin effect on Vc	−0.116	0.0707	60.8	−0.255	0.0222
Asian race effect on Vc	−0.0974	0.0228	23.4	−0.142	−0.0528
Hepatic impairment effect on Vc	0.0123	0.0204	166	−0.0277	0.0524
Renal impairment effect on Vc	−0.0184	0.0184	100	−0.0545	0.0178
Vmax (μg/h)	13.6	1.64	12.1	10.7	17.1
Km (μg/L)	49.5	32.4	65.4	17.5	140
CL (L/h)	0.0195	0.000764	3.93	0.0180	0.0210	2.90
ECOG Effect on CL	0.0742	0.0330	44.5	0.00947	0.139
Sex effect on CL	0.200	0.0371	18.5	0.128	0.273
Age effect on CL	−0.115	0.0774	67.4	−0.267	0.0369
Body weight effect on CL	0.599	0.0785	13.1	0.445	0.753
Albumin effect on CL	−1.11	0.124	11.2	−1.35	−0.864
Asian race effect on CL	−0.245	0.0395	16.1	−0.322	−0.167
Hepatic impairment effect on CL	0.0461	0.0358	77.7	−0.0241	0.116
Renal impairment effect on CL	−0.0926	0.0337	36.4	−0.159	−0.0266
Q (L/h)	0.0322	0.00233	7.23	0.0280	0.0371
Vp (L)	1.22	0.0384	3.15	1.15	1.30
ADA Effect on CL	1.11	0.0123	1.10	1.09	1.13

	IIV magnitude expressed as SD	IIV magnitude expressed as %CV	Stochastic approximation
			SE	%RSE	P2.5	P97.5
Standard deviation of the random effects
omega_Vc	0.184	18.6	0.00802	4.35	0.169	0.201
omega_CL	0.409	42.6	0.0115	2.82	0.387	0.432
Correlations
corr_Vc_CL	0.242	24.6	0.0468	19.3	0.149	0.332

	RV magnitude expressed as SD	RV magnitude expressed as %CV	Stochastic approximation
			SE	%RSE	P2.5	P97.5
Error model parameters
Additive Error	30.3		2.98	9.83	25.0	36.7
Proportional Error	0.322	33.1	0.00298	0.924	0.316	0.328

Note: The equations to predict the typical values for clearance ($\widetilde{\mathit{CL}}$) and central volume ($\widetilde{V\mathrm{c}}$) for an individual, based on the parameter estimates from the final popPK model and statistically significant covariates, are provided in the equations below.

$$\begin{array}{c}{\widetilde{\mathit{CL}}}_{\mathit{ij}}=0.02\times {\left(\frac{{\mathit{\text{WTKG}}}_i}{71}\right)}^{0.6}\times {\left(\frac{{\mathit{Alb}}_i}{40}\right)}^{-1.11}\times \exp \left(0.074·\left[\text{ECOGBLN}=1\right]\right)\times \exp \left(0.2·\left[\mathrm{SEX}=\text{Male}\right]\right)\times \exp \left(-0.25·\left[\text{Race}=\text{Asian}\right]\right)\times \exp \left(-0.093·\left[\text{RFCAT}=\text{mild}/\text{moderate}/\text{severe}\right]\right)\times \exp \left({\mathrm{\eta CL}}_{\mathrm{i}}\right).\hfill \end{array}$$

$$\begin{array}{c}{\widetilde{\mathit{VC}}}_i=3.08\times {\left(\frac{{\mathit{\text{WTKG}}}_i}{71}\right)}^{0.35}\times \exp \left(0.26\times \mathrm{SEX}=\mathrm{\text{Male}}\right)\times \exp \left(-0.097\times \left[\text{RACE}=\text{Asian}\right]\right)\times \exp \left({\mathrm{\eta VC}}_{\mathrm{i}}\right).\hfill \end{array}$$

where ${\widetilde{\mathit{CL}}}_{\mathit{ij}}$ is the typical value of clearance (L/h) in the ith patient at the jth time point; ${\mathit{\text{WTKG}}}_i$ is the baseline body weight (kg) in the ith patient; ${\mathit{ALB}}_i$ is the baseline albumin (g/dL) in the ith patient; $\mathit{\text{ECOGBLN}}$ is the baseline ECOG performance status of ith patient; RFCAT is the renal function category of the ith patient; and ${\widetilde{\mathit{VC}}}_i$ is the typical value of central volume (L) in the ith patient. ${\eta }_i^x$ is a random variable that represents the persistent difference between the “true” individualspecific estimate and the typical value of the X parameter in the ith subject; the ${\eta }_i^x$ are independent, identically distributed statistical errors with a mean of 0 and a variance equal to ${\omega }_x^2$. Continuous covariates were included in the model using the power model: $P^{*}={\theta }_x{\left(\frac{\mathit{\text{Covariate}}}{\mathit{\text{Median}}}\right)}^{{\theta }_y}$. where P* is the value of a model parameter P; ${\theta }_x$ is the typical value of this parameter in a patient whose covariate is at the median value; and ${\theta }_y$ is the fixed effect parameters quantifying the slope of the log (parameter) versus log (covariate) relationship. Categorical covariates were incorporated in the exponential models as follows: For the most frequent category: $P^{*}={\theta }_x$. For other categories: $P^{*}={\theta }_x\times {\mathrm{e}}^{\left({\theta }_y\times {\mathit{\text{Covariate}}}_{0,1}\right)}$. where ${\theta }_y$ quantifies the deviation of the parameter from its most frequent value, ${\theta }_x$, for this particular categorical variable state.

Abbreviations: %CV, coefficient of variation expressed as a percent; %RSE, relative standard error expressed as a percent; ADA, antidrug antibody; CL, clearance; ECOG, Eastern Cooperative Oncology Group; IIV, interindividual variability; K_m, Michaelis–Menten constant; P, percentile; popPK, population pharmacokinetics; Q, intercompartmental clearance; RV, residual variability; SD, standard deviation; SE, standard error; V_c, central volume of distribution; V_max, maximum elimination rate; V_p, peripheral volume of distribution.

Shrinkage in the distributions of the individual empiric Bayesian estimates was low for CL (2.9%) and moderate for V_c (24.4%). The magnitude of IIV in V_c and CL is moderate (18.6 coefficient of variation expressed as a percent [%CV] and 42.6% CV, respectively). After inclusion of covariates in the model, the variability in V_c reduced by 9.5% (base model 28.1% CV for V_c) in the final popPK model when compared to the variability in the base model, while the variability in CL decreased by 13.0% (base model 55.6% CV for CL). In the final popPK model, RV was moderate and was estimated as a combined additive plus proportional model (30.3 ng/mL standard deviation concentration units [additive] and 33.1% CV [proportional]). The unexplained variability remained similar between the base and the final popPK models.

While all covariates evaluated in the popPK analysis were included in the final model, only a subset of the covariate effect parameters had estimates which did not include 0 (corresponding to no effect) in the 95% CIs; these parameters were sex, body weight, and race effect on V_c, and ECOG, sex, body weight, albumin, race, and renal impairment on CL. The equations to predict the typical values for clearance ($\widetilde{\mathit{CL}}$) and central volume ($\widetilde{V\mathrm{c}}$) for an individual, based on the parameter estimates from the final popPK model and statistically significant covariates, are provided as a footnote in Table 1.

The forest plots illustrate the clinical significance of the covariates included in the final model on steady‐state exposures for favezelimab (Figure 1 [AUC_{0_tau,ss}] and Figure S6 [C_max,ss]). The point estimates of the GMRs for the covariate effects of hepatic impairment, renal impairment, and age on AUC_{0_tau,ss} were within the 0.8–1.25 bounds, indicating a lack of clinical relevance for these effects on favezelimab exposure.

FIGURE 1.

Forest plots of AUC_{0_tau,ss} after administering 800 mg dose at steady state. ADA, antidrug antibody; AUC_{0_tau,ss}, area under the concentration‐time curve across a 21‐day dosing period at steady state; ECOG, Eastern Cooperative Oncology Group.

The 800 mg dose was the highest dose evaluated in the favezelimab clinical program. Given the flat exposure‐safety relationship up to the 800 mg dose, it was unclear how much increased exposure beyond 800 mg would result in intolerable toxicity. Conservatively, we have assumed that a 25% increase in exposure will not result in a meaningful worsening of the safety profile compared to 800 mg. The 200 mg dose (4‐fold lower than the 800 mg Phase 3 dose) showed lower efficacy compared to 700 mg in a randomized dose finding cohort of patients with gastric cancer in P001, and a modest E‐R relationship was demonstrated in this range. While the exposure range of 800 mg in relationship to a true plateau in the exposure‐efficacy relationship remains unclear given that only 2 doses were evaluated during dose finding, conservatively we have assumed that a 20% reduction in exposure would not result in a meaningful reduction in efficacy. Thus, bioequivalence bounds of 0.8–1.25 were assumed to correspond to clinically relevant limits, and exposure changes within these limits are not expected to meaningfully alter the efficacy/safety profile.

There was a modest decrease in AUC_{0_tau,ss} in ADA‐positive patients and in patients with ECOG performance status > 0. However, the point estimate of the GMRs for these covariates was very close to the lower bound of 0.8, suggesting a lack of clinical relevance for these effects on exposure.

In contrast, higher exposures were observed for females compared to males (GMR of 1.39 [90% CI: 1.29, 1.5]); Asians relative to patients in the combined Caucasian/Others/Missing race category (GMR of 1.41 [90% CI: 1.31, 1.53]); and patients in the lowest quartile of baseline body weight (GMR of 1.55 [90% CI: 1.41, 1.7]) compared to patients in the 2 higher weight quartiles (GMRs of 1.22 [90% CI: 1.1, 1.34] and 0.873 [90% CI: 0.788, 0.967]), respectively. The GMRs and 90% CIs for these covariate effects exceeded the 0.8–1.25 range, illustrating the clinical impact of these covariates on AUC_{0_tau,ss}.

Patients in the lowest quartile of albumin (≤ 37 g/L) demonstrated lower AUC_{0_tau,ss} (GMR of 0.629 [90% CI: 0.57, 0.7]) compared to patients in the quartiles of higher albumin levels (GMR of 0.873 [90% CI: 0.79, 0.96] and 1.05 [90% CI: 0.96, 1.15]). Since body weight was correlated with race and sex, and albumin levels were correlated with race (Table S4), the observed effects on exposure cannot be uniquely ascribed to particular covariates.

Similar trends were observed for C_max,ss, where female patients had higher C_max,ss than males, and lower body weight was associated with higher C_max,ss. Other evaluated covariates, including renal impairment and hepatic impairment, ADA, ECOG status, race, albumin, and age, did not substantially influence C_max,ss, demonstrating these effects were not clinically meaningful.

Based on the final model, the typical values of CL and V_c for a male (non‐Asian) weighing 71 kg, with an albumin level of 40 g/dL, an ECOG status of zero, and normal renal function were estimated to be 0.023 L/h and 3.94 L, respectively. The typical parameter values for V_p, Q, V_max, and K_m were 1.22 L, 0.032 L/h, 13.6 μg/h, and 49.5 μg/L, respectively. Model diagnostics, including GOF (Figure S7) and pcVPC (Figure 2) plots illustrate that the final model was able to describe the data adequately.

FIGURE 2.

Prediction‐corrected visual predictive check of the final model for favezelimab. CI, confidence interval; Conc, concentration.

The pcVPC was performed to evaluate the predictive performance of the final popPK model and results are shown in Figure 2. Good concordance is seen between the prediction percentiles (10th, median, and 90th) with 90% CIs of simulated favezelimab and the overlaid observed favezelimab concentration data and corresponding percentiles (10th, median, and 90th) plotted versus time since previous dose. Time since last dose was truncated to 700 h to reflect the model performance over a 4‐week time period, which is representative of the 3‐week dosing interval for favezelimab. The final model adequately characterized the central tendency and variability of the favezelimab observed concentrations over time.

3.3. Exposure‐Response Model Development

3.3.1. Objective Response Rate

Of the 80 gastric cancer patients randomized to 200 and 700 mg doses in Study P001, a total of nine patients (11.3%) in Study P001 achieved objective response during the treatment period (Table S6). The observed ORR generally increased as favezelimab exposure increased. The final E‐R efficacy model for ORR was described by a linear function of favezelimab average AUC.

The final model for ORR was estimated with reasonable precision (%RSE < 48%) for all model parameters (Table S7). Both linear and log‐linear forms of exposures were evaluated, and a linear form of AUC was found to have the most significant E‐R relationship with ORR. None of the evaluated covariates (baseline tumor size [mm], sex, age [years], and race) were found to be statistically significant predictors of ORR. Model predictions of ORR across exposure quartiles showed good concordance with observed data (Figure 3).

FIGURE 3.

Probability of objective response versus favezelimab AUC in the final E‐R model for ORR in gastric cancer patients. AUC, area under the concentration‐time curve; CI, confidence interval; E‐R, exposure‐response; ORR, objective response rate; p, probability.

3.3.2. Grade ≥ 3 Drug‐Related AEs

A total of 16 gastric cancer patients (20.0%) in Study P001 experienced a Grade ≥ 3 drug‐related AE during the treatment period (Table S8). The observed probability of a Grade ≥ 3 drug‐related AE generally remained constant as favezelimab exposure increased. In the modeling process, neither linear nor log‐linear logistic regression of favezelimab exposure measures showed a significant (α = 0.05) E‐R relationship with the probability of a Grade ≥ 3 drug‐related AE (Figure S8). The lack of an E‐R safety relationship is shown versus favezelimab average AUC in Figure 4A. Due to the lack of an observed E‐R relationship, no further modeling was done for Grade ≥ 3 drug‐related AEs.

FIGURE 4.

Probability of safety AE occurrence versus favezelimab AUC in gastric cancer patients. AE, adverse event; AUC, area under the concentration‐time curve; CI, confidence interval; p, probability.

3.3.3. Drug‐Related AEOSI

A total of 23 gastric cancer patients (28.8%) in Study P001 experienced a drug‐related AEOSI during the treatment period (Table S9). The specific safety events that were characterized as AEOSI for these 23 patients are summarized in Table S10. The observed probability of a drug‐related AEOSI generally remained flat as favezelimab exposure increased. In the modeling process, neither linear nor log‐linear logistic regressions of favezelimab exposure measures showed a significant (α = 0.05) E‐R relationship with the probability of a drug‐related AEOSI (Figure S9). The lack of an E‐R safety relationship is shown versus favezelimab average AUC in Figure 4B. Due to the lack of an observed E‐R relationship, no further modeling was done for drug‐related AEOSI.

4. Discussion

Pharmacokinetic data showed that serum favezelimab exposures increased in a dose‐dependent manner. Concentration‐time profiles suggested that the TMDD exhibited by favezelimab was saturated at the Phase 3 dose of 800 mg. Because favezelimab exhibited nonlinear PK consistent with TMDD, a 2‐compartment model incorporating MM approximation for the nonlinear clearance was employed at the base model stage [11]. The final popPK model adequately described the data as illustrated by the pcVPC and GOF plots and good precision in the PK parameter estimates.

The model also included the effect of time‐varying ADA on linear CL to account for immunogenicity. In ADA‐positive individuals, an increase in CL was estimated, likely representing enhanced nonspecific elimination due to immune‐complex formation and accelerated CL. Addition of time‐varying ADA on CL was found to be a significant effect, improved the model fit, and was consistent with the observed decrease in favezelimab concentrations in ADA‐positive patients. Of note, ADA incidence was not explicitly modeled but considered as a time‐dependent regressor. Therefore, no attempt was made to describe dose‐dependent changes in ADA incidence. Beyond the effect of ADA, the CL of favezelimab was not time dependent, unlike pembrolizumab, which demonstrates time‐varying CL.

Female and Asian patients exhibited higher exposures compared to males and non‐Asian patients. These results are similar to previously reported demographic effects for therapeutic antibodies [12, 13]. In addition, lower body weight was associated with higher exposure, reflecting the impact of body size on the distribution and CL of favezelimab. This observation is in line with the well‐established influence of body weight on mAb PK, and may also, in part, contribute to the lower exposure for female and Asian patients.

Serum albumin emerged as a relevant covariate, with hypoalbuminemia (≤ 35 g/L) associated with reduced drug exposure. This finding is consistent with the known mechanistic link between albumin and IgG recycling through the neonatal Fc receptor (FcRn) pathway, where low albumin may serve as a surrogate marker of impaired FcRn function or increased catabolism [14, 15]. In oncology, hypoalbuminemia frequently reflects cachexia, systemic inflammation, or poor nutritional status, all of which can accelerate mAb CL [12, 16]. Similarly, impaired ECOG performance status reflects greater disease burden, which has been associated with elevated catabolic activity and reduced antibody exposure in multiple studies [17, 18]. Together, the association of low albumin and higher ECOG scores with reduced exposure underscores the importance of disease‐driven catabolism in driving variability of mAb PK in cancer patients. Furthermore, other evaluated covariates including age and renal and hepatic impairment did not meaningfully influence favezelimab exposure, consistent with the catabolic elimination pathway of IgG4 antibodies that did not involve renal elimination and hepatic metabolism [19].

These findings confirm that certain intrinsic factors and renal and hepatic impairment are unlikely to necessitate dose adjustments and highlight the interplay of immunogenicity, systemic catabolism, and FcRn‐mediated salvage in determining mAb PK in oncology. Similar to prior studies, reduced exposure in patients with cachexia or hypoalbuminemia may serve as a surrogate marker of aggressive disease biology and poorer outcomes. These observations reinforce the need to consider disease‐related covariates when interpreting PK data in cancer populations.

Of the assessed relationships, GMR (90% CIs) were contained within 0.8–1.25 for all covariates except weight, race, sex, and albumin. However, given the flatness of all of the E‐R and exposure‐safety relationships, none of the covariates assessed were expected to meaningfully alter efficacy or safety. The E‐R efficacy model for ORR characterized the observed data well from the Study P001 gastric cancer patients. The modeling showed a positive E‐R relationship, where higher favezelimab exposure was shown to be associated with a higher ORR. The model‐based summary plots of ORR versus favezelimab exposure showed that the E‐R model adequately characterized the observed data. The E‐R safety analyses of Grade ≥ 3 drug‐related AEs and of drug‐related AEOSI for Study P001 found no significant E‐R relationship between favezelimab exposure and occurrence of either safety endpoint.

Since a positive E‐R trend was observed, while exposure‐safety relationships were flat, choosing the highest feasible dose does not worsen the safety profile compared to lower doses, optimizes the overall risk/benefit. Overall, these PK and E‐R analyses helped advance the pharmacometric characterization of favezelimab and further the understanding of the determinants of drug exposure, safety, and efficacy.

A limitation of this work is that Phase 3 data were not included in the E‐R analyses, and no attempt was made to account for potential confounding or to correct potential bias resulting from the evaluation of a single dose in Phase 3. Consequently, the ability to extrapolate benefit–risk determinations based on exposure‐response and safety relationships established in a randomized Phase 1 gastric cancer cohort to the Phase 3 CRC patient population is limited.

Another limitation is the empirical assessment of the effect of ADA on PK. Since observed ADA status was treated as a time dependent predictor of clearance, no attempt was made to explicitly model the dose dependence of immunogenicity. Moreover, ADA status (positive/negative) was treated as a simple binary modifier of clearance. This simplification ignores the fact that the true effect of ADA may be a continuous function of the relative strength of the immune response (titer).

Author Contributions

B.K. and M.S.C. designed the research. M.G., H.B., and K.F.M. performed the research, and M.G., H.B., K.F.M., B.K., and M.S.C. analyzed the data. All authors wrote the manuscript.

Funding

These studies were funded by Merck & Co. Inc., Rahway, NJ.

Conflicts of Interest

M.G., H.B., and K.F.M. are employees of Simulations Plus Inc., and may hold stock options in Simulation Plus Inc. B.K. and M.S.C. are employees of Merck Sharp & Dohme LLC, a subsidiary of Merck & Co. Inc., Rahway, NJ, USA and may hold stock options in Merck & Co. Inc., Rahway, NJ, USA.

Supporting information

Table S1: Summary of Favezelimab Studies Used in the Population Pharmacokinetic Analysis.

Table S2: Demographic and Baseline Covariates Table.

Table S3: Summary of Onset and Incidence of Anti‐drug Antibodies, by Dose.

Table S4: Statistical Test of Correlation Between Continuous and Categorical Covariates.

Table S5: Statistical Tests of Correlation Between Categorical Covariates.

Table S6: Summary of Dose–Response Objective Response Rate, by Treatment.

Table S7: Final Exposure‐Response Objective Response Rate Model Parameters.

Table S8: Summary of Dose–Response Grade ≥ 3 Drug‐Related Adverse Event Occurrence, by Treatment.

Table S9: Summary of Dose–Response Drug‐Related Adverse Event of Special Interest Occurrence, by Treatment.

Table S10: Summary of Drug‐Related Adverse Events of Special Interest Occurrence in Patients With Gastric Cancer of Study P001.

Figure S1: Covariate Correlations.

Figure S2: Mean Favezelimab Concentrations Versus Time Following Cycle 1, Stratified by Dose.

Figure S3: Dose‐normalized Exposure (dnAUC0‐21 days) vs. Dose Comparison of Favezelimab (MK‐4280) in Cycle 1 when Administered Alone as Monotherapy or in Combination with Pembrolizumab (MK‐3475) with or without Chemotherapy or with Lenvatinib or as a Coformulation Therapy with Pembrolizumab in Participants with Advanced Solid Tumors.

Figure S4: Dose‐normalized Exposure (dnCmax) vs. Dose Comparison of Favezelimab (MK‐4280) in Cycle 1 when Administered Alone as Monotherapy or in Combination with Pembrolizumab (MK‐3475) with or without Chemotherapy or with Lenvatinib or as a Coformulation Therapy with Pembrolizumab in Participants with Advanced Solid Tumors.

Figure S5: Mean Favezelimab Concentration Versus Time, by Dose and Anti‐drug Antibody Effect

Figure S6: Forest Plots of Cmax_ss After Administering 800 mg Dose at Steady State.

Figure S7: Goodness‐of‐Fit Plots for the Final Population Pharmacokinetic Parameters for Favezelimab.

Figure S8: Probability of Grade ≥ 3 Drug‐Related AE Occurrence Versus Favezelimab Exposures in Gastric Cancer Patients.

Figure S9: Probability of Drug‐Related AEOSI Occurrence Versus Favezelimab Exposures in Gastric Cancer Patients.