Exposure‐Response Relationships for Axatilimab in Patients with Chronic Graft‐Versus‐Host Disease

Abstract

Axatilimab, a high‐affinity humanized monoclonal antibody that targets colony‐stimulating factor 1 receptor, is approved for the treatment of chronic graft‐versus‐host disease (cGVHD). Here, we describe the exposure‐response relationships for efficacy and safety in patients with cGVHD who received axatilimab. Exposure‐efficacy relationships were assessed in treated patients in the AGAVE‐201 study (n = 239); exposure‐safety relationships were assessed in treated patients in AGAVE‐201 (n = 239) and the phase 1/2 SNDX‐6352‐0503 study (n = 39). For binary or time‐to‐event endpoints, logistic or Cox regression analyses, respectively, were performed using axatilimab exposure metrics that were derived from a previously developed population pharmacokinetic/pharmacodynamic model. Overall response and ≥ 7‐point improvement in modified Lee Symptom Scale responses were associated with axatilimab exposure, with lower axatilimab exposure increasing the odds of a response. Duration of response was not associated with axatilimab exposure. Ten of 11 safety endpoints were associated with axatilimab exposure, with higher axatilimab exposure increasing the odds of adverse events. Among the 3 regimens evaluated in AGAVE‐201, the 0.3 mg/kg once every 2 weeks (Q2W) regimen had the highest predicted probability of response. Additionally, this dose group had the lowest predicted probability of event occurrence across all 10 safety endpoints associated with exposure among the evaluated regimens. Despite body weight influencing axatilimab exposure by > 20%, its effect on efficacy and safety endpoints remained minimal, with a maximum difference of ≤ 0.4% and ≤ 4.4% between the 1st and 4th quartiles of body weight, respectively. Taken together, these findings support the benefit–risk profile of axatilimab 0.3 mg/kg Q2W in patients with cGVHD.

Study Highlights.

• WHAT IS THE CURRENT KNOWLEDGE ON THE TOPIC?

Axatilimab, a monoclonal antibody targeting colony‐stimulating factor 1 receptor (CSF‐1R), was recently approved for chronic graft‐versus‐host disease (cGVHD) after ≥ 2 prior lines of systemic therapy. A population pharmacokinetic/pharmacodynamic model for axatilimab has recently been described.

• WHAT QUESTION DID THIS STUDY ADDRESS?

We aimed to characterize exposure–efficacy and exposure–safety relationships for axatilimab among patients with cGVHD.

• WHAT DOES THIS STUDY ADD TO OUR KNOWLEDGE?

In this analysis, lower axatilimab exposure was associated with higher clinical responses, and higher axatilimab exposure was associated with a higher frequency of adverse events. In patients with cGVHD, axatilimab 0.3 mg/kg every 2 weeks (Q2W) had the optimal benefit–risk profile. Model‐based predictions indicated that body weight–mediated changes in axatilimab exposure did not have a significant effect on efficacy and safety endpoints.

• HOW MIGHT THIS CHANGE CLINICAL PHARMACOLOGY OR TRANSLATIONAL SCIENCE?

This analysis provides further support for the dose selection of axatilimab 0.3 mg/kg Q2W in patients with cGVHD.

Chronic graft‐versus‐host disease (cGVHD) is an immune‐mediated, potentially life‐threatening complication that affects approximately 50% of patients who receive allogeneic hematopoietic stem cell transplantation (allo‐HSCT). ¹ , ² , ³ , ⁴ cGVHD is a leading cause of nonrelapse mortality following allo‐HSCT. ⁵ In addition, there is a considerable symptom burden with cGVHD (most often assessed using the modified Lee Symptom Scale [mLSS]). ⁶ Management of cGVHD is challenged by the need to balance immunosuppression, cumulative toxicity, and risk of infection from corticosteroids and other immunosuppressive therapies ⁷ , ⁸ ; infections are also a leading cause of nonrelapse mortality following allo‐HSCT. ⁵

Monocytes and tissue‐resident macrophages (including monocytes that have migrated out of the bloodstream) contribute to the multiorgan inflammation and fibrosis that are hallmarks of cGVHD pathogenesis. ⁹ Monocytes and macrophages are dependent on colony‐stimulating factor 1 (CSF‐1) or interleukin‐34 activation of the CSF‐1 receptor (CSF‐1R) for cell survival, migration, and proliferation. ¹⁰ Axatilimab is a high‐affinity anti–CSF‐1R humanized immunoglobulin G4 (kappa light chain) monoclonal antibody. ¹¹

Axatilimab was initially evaluated for the treatment of cGVHD in a phase 1/2 study (SNDX‐6352‐0503 [NCT03604692]). ¹² In phase 1, 17 patients received axatilimab at dose levels of 0.15 mg/kg once every 2 weeks (Q2W), 0.5 mg/kg Q2W, 1 mg/kg Q2W, 3 mg/kg Q2W, and 3 mg/kg every 4 weeks (Q4W). ¹² In phase 2, 23 patients received the 1 mg/kg Q2W dose level. ¹² Clinical axatilimab responses were accompanied by reductions in cGVHD symptom burden. ¹² The overall response rate (ORR) by Cycle 7 Day 1 was 67% (95% CI, 50%–81%) among all evaluable patients in the study. ¹²

Based on the results of SNDX‐6352‐0503, the efficacy and safety of axatilimab in patients with cGVHD were evaluated further in the pivotal phase 2 AGAVE‐201 study (NCT04710576). ¹³ The ORR by Cycle 7 Day 1 was 74%, 67%, and 50% in the 0.3 mg/kg Q2W, 1 mg/kg Q2W, and 3 mg/kg Q4W cohorts, respectively. ¹³ A clinically meaningful reduction in cGVHD symptom burden (defined as a ≥ 7‐point improvement in mLSS ¹⁴ ) in the first 6 cycles was reported by 55%, 54%, and 36% of patients in the 0.3 mg/kg Q2W, 1 mg/kg Q2W, and 3 mg/kg Q4W cohorts, respectively. ¹⁵ Adverse events (AEs) were mostly low grade and reversible with no unexpected safety concerns. ¹³ Drug discontinuation owing to treatment‐emergent adverse events (TEAEs) occurred in 6%, 22%, and 18% of patients with the 0.3 mg/kg Q2W, 1 mg/kg Q2W, and 3 mg/kg Q4W regimens, respectively. ¹³

A population pharmacokinetic/pharmacodynamic (PK/PD) model was previously developed based on pooled data from 4 clinical studies of axatilimab. ¹⁶ The structural model consists of a 2‐compartment axatilimab PK model with saturable clearance and turnover models for CSF‐1, nonclassical CD14+/CD16++ monocytic cells (NCMC), aspartate aminotransferase (AST), and creatine phosphokinase (CPK). ¹⁶ Three covariates were associated with PK parameters and included in the final model: body weight on volume of distribution, and baseline CSF‐1 and antidrug antibody (ADA) on the linear component of axatilimab clearance. Body weight and CSF‐1 covariates were adjusted for the population median and introduced via a power function, while ADA was incorporated as a time‐dependent categorical covariate. Among these 3 covariates, body weight was the only covariate that affected axatilimab steady‐state exposure by > 20%. ¹⁶ Here, we characterize axatilimab exposure–response relationships for efficacy and safety to further support the dose selection of axatilimab in patients with cGVHD.

METHODS

Study population and endpoints

Data for these analyses were derived from 2 clinical studies in patients with cGVHD, a phase 1/2 study (SNDX‐6352‐0503) ¹² and AGAVE‐201, a pivotal phase 2 study in patients with cGVHD, both of which are summarized in Table S1 . In SNDX‐6352‐0503, axatilimab doses of 0.15 mg/kg Q2W, 0.5 mg/kg Q2W, 1 mg/kg Q2W, 3 mg/kg Q2W, and 3 mg/kg Q4W were administered. In AGAVE‐201, 3 doses of axatilimab (0.3 mg/kg Q2W, 1 mg/kg Q2W, 3 mg/kg Q4W) were evaluated. Median treatment durations across doses were 7.1 months and 7.8 months in SNDX‐6352‐0503 and AGAVE‐201, respectively.

The exposure‐efficacy analysis dataset included efficacy‐evaluable patients in AGAVE‐201 who had received ≥ 1 dose of axatilimab and had ≥ 1 quantifiable postdose plasma concentration. Efficacy endpoints included 2 binary endpoints (overall response and ≥ 7‐point improvement in mLSS [a regulatory threshold for clinically meaningful cGVHD symptom improvement]). Duration of response (DOR), a time‐to‐event endpoint, was evaluated in patients who achieved an overall response and was defined as the interval from the date of the initial response until the first organ progression of cGVHD from nadir, the start of a new anti‐cGVHD therapy, or death from any cause, whichever occurred first.

The exposure‐safety analysis dataset included safety‐evaluable patients in SNDX‐6352‐0503 and AGAVE‐201 who had received ≥ 1 dose of axatilimab and had ≥ 1 quantifiable postdose concentration. Safety endpoints (all binary) included grade ≥ 3 TEAEs, TEAEs leading to dose modifications (including discontinuations, dose interruptions, and dose reductions), grade ≥ 3 adverse events of special interest (AESI), serious TEAEs, treatment‐related TEAEs, infections of unspecified etiology (grouped AE terms for infections that were not bacterial, viral, or fungal), amylase and lipase increases (grouped AE terms), CPK elevations (grouped AE terms), liver enzyme elevations (grouped AE terms), periorbital edema (grouped AE terms), and infusion‐related reactions (grouped AE terms including hypersensitivity reactions). AEs were treatment emergent regardless of causality. AESI included infusion‐related reactions, infections, hepatic enzyme elevations, CPK elevations, periorbital edema, and amylase and lipase elevations, which were determined based on the drug class and mechanisms of action of axatilimab.

PK exposure metrics

The PK exposure metrics of axatilimab were estimated for individual patients based on empirical Bayes estimates of PK parameters, taking into account individual covariate values (ie, body weight, CSF‐1, and ADA [time‐invariant covariate, where ADA is assigned a value of 1 if positive at any timepoint and 0 otherwise]), and patients' specific dosing history from a pooled population PK/PD analysis. ¹⁶ Body weight and CSF‐1 covariates were adjusted for the population median and introduced via a power function. A total of 6 exposure metrics of axatilimab and their log‐transformed counterparts were evaluated in exposure‐response analyses of both efficacy and safety. These exposure metrics included single‐dose maximum concentration (CMAXSD) and area under the curve (AUC) from time 0 to infinity after a single dose (AUCSDINF). The first 4‐week treatment cycle C _max (CMAX1) and AUC (AUC1) and steady‐state 4‐week treatment cycle C _max (CMAXSS) and AUC (AUCSS) were used for multiple dosage regimens. Two additional exposure parameters, C _max and AUC at the time of first event incidence (TOFI; CMAXSDTOFI and AUCSDTOFI, respectively), were also generated for exposure‐safety analyses. Because most measurements of trough concentrations (ie, C_min ) in the first and steady‐state cycles were below the limit of quantitation, C_min was not evaluated as an exposure metric in the exposure‐response analyses.

Exposure parameters associated with the single dose and the first treatment cycle were derived from the population PK/PD model simulations using each patient's nominal administered dose and regimen. The exposure parameters (CMAXSDTOFI and AUCSDTOFI) on the day of the first safety event were generated for each safety endpoint and individual patients using the single dose immediately prior to TOFI. To account for individual patient's dose changes that occurred throughout the course of the study, steady‐state PK parameters were simulated using cycle average dose for each efficacy and safety endpoint and individual patients using the following equation:

$$\text{Cycle average dose}=\frac{{\sum }_{i=1}^n\mathrm{Dos}{\mathrm{e}}_i}{t_n+\left\{\begin{array}{c}14,\text{for}\mathrm{Q}2\mathrm{W}\\ 28,\text{for}\mathrm{Q}4\mathrm{W}\end{array}\right.}*\left\{\begin{array}{c}14,\text{for}\mathrm{Q}2\mathrm{W}\text{regimen}\\ 28,\text{for}\mathrm{Q}4\mathrm{W}\text{regimen}\end{array}\right.,$$

where n is the number of dosing events before TOFI and t _n is the time of the nth dose.

A complex algorithm was developed to derive TOFI, which was then applied individually for each endpoint and patient. For DOR, TOFI matched the time of the response's end. For the ≥ 7‐point improvement in mLSS, overall response, and all safety outcomes, TOFI matched the time of the first response or safety event. If no event occurred, TOFI was defined as the end of treatment for safety and was defined as the end of treatment, Cycle 7 Day 1, or the start of a new anti‐cGVHD therapy, whichever happened first for efficacy.

Development of exposure‐response models

Exposure‐response analyses, including exploratory data analysis, model development, diagnostics, and forward simulations, were implemented in R software (v4.2.1) and followed a common model development workflow (Data S2 and S3 ). ¹⁷ , ¹⁸

First, exploratory plots were used to assess exposure‐response relationships. For binary endpoints, data were split by exposure quartiles, and the mean response probability was plotted vs. mean exposure with a fitted logistic regression line. For time‐to‐event variables (DOR), Kaplan–Meier curves were plotted by quartile together with a log‐rank test. Following graphical exploration, a linear logistic regression model was evaluated for binary endpoints, and a Cox proportional hazards regression model was used for DOR.

Once the exposure metric was selected into the model, a covariate analysis was conducted according to the standard stepwise forward‐addition and backward‐elimination processes. A summary of covariates for the efficacy and safety population is presented in Tables S2 and S3 , respectively. Covariates were tested on the intercept of the exposure‐response relationship (see Supplemental Methods for details).

Model‐based simulations

Based on developed exposure‐response models, dose responses were projected for binary endpoints with different nominal dosing regimens (0.3 mg/kg Q2W, 1 mg/kg Q2W, and 3 mg/kg Q4W for efficacy; 0.15 mg/kg Q2W, 0.3 mg/kg Q2W, 0.5 mg/kg Q2W, 1 mg/kg Q2W, 3 mg/kg Q2W, and 3 mg/kg Q4W for safety). Both uncertainty and variability were considered in the exposure predictions based on the variance–covariance matrix derived from the Fisher information matrix and random effects, respectively; each simulation set per dose level contained 100 populations with 500 participants each. Covariate values were sampled randomly with replacement (bootstrap sampling) from the AGAVE‐201 population for efficacy predictions and from both cGVHD studies for safety simulations. To calculate the median and CI for the probability of efficacy and safety outcomes, sets of events corresponding to the different exposure values were sampled 1,000 times from the Bernoulli distribution, and 95% CI was calculated as the 2.5% and 97.5% percentiles of the simulated values.

To evaluate the effect of body weight (a covariate identified as potentially clinically significant in the population PK/PD analysis) on efficacy and safety events, the probability of occurrence of responses and safety events was projected for the nominal dose of 0.3 mg/kg Q2W using the quartiles of observed body weight distribution of the populations for exposure‐efficacy analysis (N = 239; median [reference], 71.6 kg; range, 18.1–151 kg) and exposure‐safety analysis (N = 278; median [reference], 72.4 kg; range, 18.1–151 kg), respectively. To calculate the median and CI for the probability of efficacy and safety outcomes, sets of events corresponding to the different exposure values were sampled 1,000 times from the Bernoulli distribution, and the 95% CI was calculated as the 2.5% and 97.5% percentiles of the simulated values.

Ethics statement

All studies were conducted in accordance with applicable Good Clinical Practice guidelines and the provisions of the Declaration of Helsinki. Study protocols were approved by the relevant institutional review board or ethics committee at each study center. Written informed consent/assent was obtained from all patients (or their parents/legal guardians) before enrollment.

RESULTS

Exposure‐efficacy analysis

Among the 241 efficacy‐evaluable patients with cGVHD from the AGAVE‐201 study, 239 were included in this analysis; 2 patients who were not treated with axatilimab were excluded from the analysis. Among 239 patients, 79, 81, and 79 received doses of 0.3 mg/kg Q2W, 1 mg/kg Q2W, and 3 mg/kg Q4W, respectively. A summary of the observed responses for binary efficacy endpoints is shown in Table S4 .

For the 2 binary efficacy endpoints (overall and ≥ 7‐point improvement in mLSS) included in this exposure‐response analysis, 7 overlapping statistically significant exposure metrics (CMAXSD, LOGCMAXSD, CMAX1, LOGCMAX1, AUCSDINF, LOGAUCSDINF, and AUC1) that were associated with a single dose or the first treatment cycle were identified, with comparable Akaike information criterion (AIC) values and P values. Since AUCSDINF is equivalent to the area under the curve from time 0 to the dosing interval (AUCτ) at steady state, AUCSDINF was selected as the exposure metric to be used in exposure‐response analyses for overall response and ≥ 7‐point improvement in mLSS. The final model parameters are shown in Table S5 . Both overall response and ≥ 7‐point improvement in mLSS had negative associations with axatilimab AUCSDINF (Figure 1 a,b ). The odds of a patient achieving an overall response and ≥ 7‐point improvement in mLSS decreased by 7.6% and 5.8%, respectively, for each 50 μg/mL*d increase in AUCSDINF. Although no significant covariate was identified for overall response, the odds of a patient achieving a ≥ 7‐point improvement in mLSS increased by 2.15‐fold if the patient had cGVHD with > 4 involved organs at baseline (Table 1 ), suggesting that patients with more extensive baseline disease may derive greater benefit from axatilimab treatment.

Figure 1.

Logistic regression analyses for exposure‐efficacy endpoints: (a) overall response and (b) ≥ 7‐point improvement in mLSS, and forest plots for exposure‐efficacy endpoints: (c) overall response and (d) ≥ 7‐point improvement in mLSS. AUCSDINF, area under the curve after single dose from time 0 to infinity; mLSS, modified Lee Symptom Scale; Q2W, once every 2 weeks; Q4W, once every 4 weeks. For panels a and b, red points indicate the binary status of response occurrence (yes = 1, no = 0). Numbers shown are n/N for each exposure quartile. Black points with vertical bars are medians and CIs for each exposure quartile. The solid black line with shaded region shows the prediction curve and 95% CI. Diamonds with horizontal bars indicate the 5th, 50th, and 95th percentiles of exposures for each dose.

Table 1.

Parameter estimates from the logistic regression models for exposure‐efficacy relationships

Efficacy endpoint	Predictor	OR Coefficient (95% CI)	P value	Median Predictor (5th, 95th Percentile)	5th Percentile of OR, Median (95% CI)	95th Percentile of OR, Median (95% CI)
Overall response	AUCSDINF, 50*μg/mL*d	0.924 (0.879–0.975)	0.003	64.0 (8.99, 714.525)	0.986 (0.977–0.996)	0.325 (0.157–0.7)
≥ 7‐point improvement in mLSS	AUCSDINF, 50*μg/mL*d	0.942 (0.89–0.995)	0.03	64.0 (8.99, 715)	0.989 (0.979–0.999)	0.427 (0.189–0.926)
	>4 involved organs	2.153 (1.23–3.887)	0.007	—	—	—

AUCSDINF, area under the curve after single dose from time 0 to infinity; mLSS, modified Lee Symptom Scale; OR, odds ratio.

The predicted overall response in patients with median AUCSDINF was 69%; the predicted overall response in patients with AUCSDINF at the 5th and 95th percentiles was 70.8% and 44.4%, respectively (Figure 1 c ). The predicted ≥ 7‐point improvement in mLSS in patients with median AUCSDINF values was 45.4%; the predicted ≥ 7‐point improvement in mLSS in patients with AUCSDINF at the 5th and 95th percentiles was 46.8% and 27.7%, respectively (Figure 1 d ). A visual predictive check of the final exposure‐efficacy model demonstrated that across all AUCSDINF quartiles of axatilimab exposure, there was agreement between the model‐predicted and observed values, with observed values overlapping the 90% prediction interval (Figure S1 ).

Axatilimab exposure was not a statistically significant predictor for DOR among the 153 patients who achieved an overall response in AGAVE‐201. A representative Kaplan–Meier plot showing DOR stratified by AUCSS quartiles is shown in Figure S2 (P = 0.85). Notably, 30 of 153 patients (19.6%) remained at risk beyond 6 months. Given that no statistically significant exposure‐response relationships were observed between all evaluated exposure metrics and DOR, no further covariate assessment was performed for this endpoint.

Exposure‐safety analysis

A total of 279 patients with cGVHD were evaluable for safety across 2 studies; 278 were included in this exposure‐safety analysis. One patient from the SNDX‐6352‐0503 study was excluded owing to missing PK samples. Overall, 79, 107, 85, 6, and 2 patients received doses of 0.3 mg/kg Q2W, 1 mg/kg Q2W, 3 mg/kg Q4W, 3 mg/kg Q2W, or lower doses (0.15 mg/kg Q2W, 0.5 mg/kg Q2W), respectively. A summary of the observed incidences of events for the safety endpoints is provided in Table S6 .

Two overlapping statistically significant exposure metrics, CMAXSS and AUCSS, were identified for 10 out of the 11 binary safety endpoints examined. AUCSS had the lowest P value and AIC; thus, AUCSS was selected as the exposure metric to be used for these binary safety endpoints. All 10 safety endpoints had positive associations with axatilimab AUCSS (Table S7 ). Logistic regression plots for safety endpoints are shown in Figure 2 and Figure S3 , and forest plots of safety event probabilities are shown in Figure 3 and Figure S4 . Visual predictive checks confirmed that there was good agreement between the observed data and the safety‐response simulations in the final model (Figure S5 ).

Figure 2.

Logistic regression analyses for representative exposure‐response safety endpoints: (a) grade ≥ 3 TEAEs, (b) TEAEs leading to dose modification, (c) grade ≥ 3 AESI, (d) treatment‐related TEAEs. AE, adverse event; AESI, adverse events of special interest; AUCSS, area under the curve for steady‐state 4‐week treatment cycle; Q2W, once every 2 weeks; Q4W, once every 4 weeks; TEAE, treatment‐emergent adverse event. Red points indicate the binary status of AE occurrence (yes = 1, no = 0). Numbers shown are n/N for each exposure quartile. Black points with vertical bars are medians and CIs for each exposure quartile. The solid black line with shaded region shows the prediction curve and 95% CI. Diamonds with horizontal bars indicate the 5th, 50th, and 95th percentiles of exposures for each dose.

Figure 3.

Forest plots for exposure‐response safety endpoints: (a) grade ≥ 3 TEAEs; (b) TEAEs leading to dose modification; (c) grade ≥ 3 AESI; (d) treatment‐related TEAEs. AESI, adverse events of special interest; AUCSS, area under the curve for steady‐state 4‐week treatment cycle; TEAE, treatment‐emergent adverse event.

For all safety endpoints included in the final logistic regression safety models, an increase of 50 μg/mL*d in AUCSS increased the odds of TEAE occurrences; the lowest increase was for serious TEAEs (2.8%), and the highest increase was for CPK elevations (23.8%; Table 2 ). A few statistically significant covariates were identified for these safety endpoints. Although the odds of serious TEAEs increased by 2‐fold for patients with a baseline Karnofsky/Lansky performance score ≤ 70% (ie, unable to work), an increase in plasma albumin by 1 g/L decreased the odds of serious TEAEs by 9.1%. The odds of an event of amylase and lipase increase occurring in a patient increased by 1.4% for each 1 U/L increase in baseline lipase. The odds of an event of infusion‐related reactions occurring in a patient decreased by 1.7% with each 1 U/L increase of baseline alanine aminotransaminase. The odds of an event of CPK elevation occurring in a patient increased by 1% with each 1 U/L increase in baseline CPK.

Table 2.

Parameter estimates from the logistic regression models for exposure‐safety relationships

Safety endpoint	Predictor	OR Coefficient (95% CI)	P value	Median Predictor (5th, 95th Percentile)	5th Percentile of OR, Median (95% CI)	95th Percentile of OR, Median (95% CI)
Grade ≥ 3 TEAEs	AUCSS, 50 μg/mL*d	1.086 (1.043–1.131)	< 0.001	128.506 (18.876, 1442.416)	1.031 (1.014–1.047)	10.023 (2.811–34.257)
TEAEs leading to dose modifications	AUCSS, 50 μg/mL*d	1.055 (1.026–1.087)	< 0.001	128.711 (19.413, 1315.382)	1.022 (1.009–1.034)	4.277 (1.796–9.73)
Grade ≥ 3 AESI	AUCSS, 50 μg/mL*d	1.08 (1.043–1.124)	< 0.001	128.276 (18.197, 1333.464)	1.029 (1.016–1.042)	8.015 (3.13–19.795)
Serious TEAEs	AUCSS, 50 μg/mL*d	1.028 (0.999–1.055)	0.048	120.329 (18.197, 1112.353)	1.01 (1–1.02)	1.817 (0.995–3.393)
	ALB, U/L	0.909 (0.859–0.964)	0.001	40 (32.085, 47)	2.144 (1.353–3.38)	0.509 (0.341–0.765)
	Unable to worka	2.002 (1.183–3.403)	0.009
Treatment‐related TEAEs	AUCSS, 50 μg/mL*d	1.112 (1.038–1.194)	0.003	132.849 (19.254, 1403.762)	1.042 (1.015–1.068)	19.645 (2.863–121.591)
Amylase and lipase increasesb	AUCSS, 50 μg/mL*d	1.157 (1.103–1.21)	<0.001	117.732 (18.197, 1403.762)	1.055 (1.038–1.073)	61.692 (17.662–224.974)
	Lipase, U/L	1.014 (1.006–1.023)	0.003	33 (11.85, 107.9)	0.748 (0.616–0.895)	2.799 (1.478–5.571)
Periorbital edemab	AUCSS, 50 μg/mL*d	1.125 (1.081–1.17)	< 0.001	114.966 (18.197, 1233.407)	1.044 (1.03–1.059)	18.541 (7.355–49.993)
Infusion‐related reactions b , c	AUCSS, 50 μg/mL*d	1.088 (1.051–1.129)	< 0.001	126.017 (18.272, 1339.036)	1.032 (1.019–1.045)	9.749 (3.861–25.442)
	ALT, U/L	0.983 (0.97–0.996)	0.012	30 (11.85, 82.3)	1.368 (1.077–1.759)	0.405 (0.196–0.808)
CPK elevationsb	AUCSS, 50 μg/mL*d	1.235 (1.164–1.313)	< 0.001	120.651 (18.01, 1426.255)	1.079 (1.055–1.102)	401.082 (70.097–2266.125)
	CPK, U/L	1.01 (1.005–1.016)	< 0.001	64 (25.85, 224.45)	0.675 (0.552–0.815)	5.21 (2.361–12.168)
Liver enzyme elevationsb	AUCSS, 50 μg/mL*d	1.109 (1.066–1.154)	< 0.001	130.325 (18.609, 1341.576)	1.04 (1.025–1.055)	16.826 (6.071–46.579)

AESI, adverse events of special interest; ALB, albumin; ALT, alanine aminotransaminase; AUCSS, area under the curve for steady‐state 4‐week treatment cycle; CPK, creatine phosphokinase; OR, odds ratio; TEAE, treatment‐emergent adverse event.

^a Defined as a baseline Karnofsky/Lansky performance score ≤ 70%.

^b Grouped adverse event terms.

^c Including hypersensitivity reactions.

Infections of unspecified etiology did not have a statistically significant association with axatilimab exposure (Figure S6 ); thus, no further covariate assessment was performed for this endpoint.

Model‐based predictions

Because the sign of β coefficients was negative in logistic regression models where axatilimab exposure was identified as a significant predictor of efficacy response (β _AUCSDINF of −0.0016 [95% CI, −0.0026, −0.0005] and − 0.0012 [95% CI, −0.0023, −0.0001] for overall response and ≥ 7‐point improvement in mLSS response, respectively), the predicted probability of response declined as the axatilimab dose increased. As a result, the highest probability of response was observed with the 0.3 mg/kg Q2W regimen (Figure 4 a ), with a median overall response probability of 70.8% (95% CI, 65.9%–75.7%) and a median ≥ 7‐point improvement in mLSS probability of 53.2% (95% CI, 47.7%–58.4%), which is consistent with observed values for these endpoints (74.7% and 54.4%, respectively). The maximum change in probability between 0.3 mg/kg Q2W and 3 mg/kg Q4W for overall response was −14.1%, and −9.9% for ≥ 7‐point improvement in mLSS.

Figure 4.

Model‐predicted event rates of evaluated dose levels for (a) efficacy endpoints, (b) general TEAE endpoints, and (c) additional safety endpoints. AESI, adverse events of special interest; CPK, creatine phosphokinase; mLSS, modified Lee Symptom Scale; Q2W, once every 2 weeks; Q4W, once every 4 weeks; TEAE, treatment‐emergent adverse event. *Grouped adverse event terms. ^†Including hypersensitivity reactions.

In contrast, the probability of all safety events for which axatilimab exposure (AUCSS) was a statistically significant predictor increased with axatilimab exposure (Figure 4 b,c ). Predicted probabilities of event occurrence were consistent with observed values for the 0.3 mg/kg Q2W, 1 mg/kg Q2W, and 3 mg/kg Q4W treatment regimens. The largest increase in the probability of safety events was observed for CPK elevations with mean probabilities of 13.9% (95% CI, 9.1%–18.8%) and 81.7% (95% CI, 77.2%–86.1%) for the 0.15 mg/kg Q2W and 3 mg/kg Q2W regimens, respectively (Figure 4 c ).

To evaluate the effect of body weight on efficacy and safety, drug exposures in efficacy‐evaluable patients in AGAVE‐201 and safety‐evaluable patients in SNDX‐6352‐0503 and AGAVE‐201 were simulated using individual PK model parameters at a dose of 0.3 mg/kg Q2W. The exposure metrics, AUCSDINF and AUCSS —predictors of efficacy and safety, respectively—increased with body weight (Figure S7 ). The median AUCSDINF in the exposure‐efficacy analysis population increased from 9.98 μg/mL*d in the first body weight quartile (median, 53.6 kg) to 15.2 μg/mL*d in the fourth quartile (median, 94.8 kg). Similarly, the median AUCSS in the exposure‐safety analysis population increased from 18.3 μg/mL*d in the first quartile (median, 53.8 kg) to 29.8 μg/mL*d in the fourth quartile (median, 97.6 kg). The median AUCSDINF and AUCSS values for each quartile of body weight were used to predict the probability of response and occurrence of safety events, respectively. The exposure difference between the first quartile and fourth quartile of body weights resulted in minimal relative changes both in efficacy (< 0.4% for overall and ≥ 7‐point improvement in mLSS response probabilities) and safety (< 4.4% for all safety endpoints; Figure 5 ).

Figure 5.

Model‐predicted event rates for 0.3 mg/kg stratified by body weight quartiles for (a) binary efficacy and (b) safety endpoints. AESI, adverse events of special interest; AUCSDINF, area under the curve after single dose from time 0 to infinity; AUCSS, area under the curve for steady‐state 4‐week treatment cycle; CPK, creatine phosphokinase; mLSS, modified Lee Symptom Scale; Q, quartile; TEAE, treatment‐emergent adverse event. *Median (range) body weight Q1, Q2, Q3, and Q4 for efficacy endpoints were 53.6 (18.1–60.3), 67 (60.8–71.6), 76.2 (71.9–84), and 94.8 (84.1–151) kg, respectively. ^†Median (range) body weight Q1, Q2, Q3, and Q4 for safety endpoints were 53.8 (18.1–62), 67.9 (62.2–72.3), 79.2 (72.5–85.7), and 97.6 (85.8–151) kg, respectively. ^‡Grouped adverse event terms. ^§Including hypersensitivity reactions.

DISCUSSION

Axatilimab, a monoclonal antibody that blocks ligand binding to CSF‐1R and downregulates the development and differentiation of the pathogenic monocyte‐derived macrophages, is a promising immunotherapeutic agent for cGVHD. The goal of this analysis was to characterize exposure‐efficacy and exposure‐safety relationships between multiple preselected efficacy and safety endpoints as well as axatilimab exposure, and to deconvolute underlying variability through a covariate search.

Exposure‐safety analyses were based on data from the phase 1/2 SNDX‐6352‐0503 and phase 2 AGAVE‐201 studies, comprising 278 patients with cGVHD, whereas exposure‐efficacy modeling was performed using the data only from AGAVE‐201 (239 patients). To account for dose adjustments in the studies and the nonlinear PK of axatilimab with tri‐phase elimination, a previously developed semimechanistic population PK/PD model with target‐mediated axatilimab clearance was used to generate exposure metrics, including AUC and C _max. ¹⁶ Moreover, these metrics were explored in 3 scenarios: at steady‐state, at the first 4‐week treatment cycle, and after a single dose. For the 2 latter options, nominal dose and regimen were used. In addition, the single dose–related AUC and C _max exposure metrics with dose prior to TOFI were tested for the exposure‐safety analysis. The rationale for using single dose and first cycle to generate exposure was that these metrics are less susceptible to the confounding effect of target‐mediated drug disposition and ADA on clearance in the beginning of treatment compared with their steady‐state counterpart. To account for the interpatient heterogeneity in axatilimab dose due to dose modifications, steady‐state exposure was generated using individual average cycle doses calculated for each endpoint as daily average doses from Day 1 to TOFI multiplied by the number of days in 2‐week or 4‐week treatment cycles (for Q2W and Q4W regimens, respectively) within this period. If no event occurred, the time cut‐off was the beginning of Cycle 7 or the end of treatment for efficacy and safety analyses, respectively. Finally, each axatilimab exposure metric was evaluated in untransformed and log‐transformed form. In total, 12 and 16 types of axatilimab PK parameters were derived and used in model‐based analyses of efficacy and safety, respectively, to ensure the breadth of possible associations in the analysis.

A single optimal exposure‐based predictor was selected for each class of outcomes, separately for efficacy and safety analyses, acting as a base model for subsequent stepwise covariate search. For efficacy metrics, axatilimab peak concentration or AUC, either at the first treatment cycle or after a single dose, was considered. Given that the complex target‐driven PK of axatilimab is also susceptible to ADA effects, AUC was chosen over C _max as a metric more reflective of the patient‐level properties of axatilimab PK profiles. The probability of safety events is associated predominantly with steady‐state PK parameters.

Based on the results of logistic regression modeling, a statistically significant association was observed between axatilimab exposure and efficacy endpoints, including an overall response and a ≥ 7‐point improvement in mLSS response. The coefficients for axatilimab exposure were negative for overall response and ≥ 7‐point improvement in mLSS response, indicating a decrease in pharmacologic effect is observed with increases in drug concentration. Consistent with these findings, the mean probability of response was predicted to decline by 14.1% and 9.9% for overall response and ≥ 7‐point improvement in mLSS outcomes, respectively, between the 0.3 mg/kg Q2W and 3 mg/kg Q4W dose levels. The negative association between several axatilimab exposure metrics and the incidence of overall response or ≥ 7‐point mLSS improvements might be attributable to multiple factors. All exposure metrics that were identified as statistically significant predictors of axatilimab efficacy were associated with a single dose or the first treatment cycle. These parameters represent naive exposure achieved with a nominal dose, which is susceptible to dose adjustments, dose interruption, and ADA, especially for the higher doses. Steady‐state PK metrics partially account for these factors; however, they do not correlate with axatilimab efficacy. Fewer patients in the 0.3 mg/kg Q2W dose group (6%) discontinued treatment due to an AE than those in the 1 mg/kg Q2W or 3 mg/kg Q4W dose groups (22% and 18%, respectively). ¹³ Consequently, patients in the 0.3 mg/kg Q2W dose group had the longest median duration of treatment (9 cycles) compared with those in 1 mg/kg Q2W (8 cycles) and 3 mg/kg Q4W dose groups (6 cycles). ¹³ As a result, a greater proportion of patients receiving 0.3 mg/kg Q2W, compared with those on 1 mg/kg Q2W and 3 mg/kg Q4W, reached Cycle 7 Day 1, at which point efficacy was assessed. Thus, it could be hypothesized that the exposure‐efficacy relationship plateaus with the 0.3 mg/kg Q2W dosing regimen, and further dose increases do not provide substantial increases in treatment benefit, and the negative association between exposure and efficacy may be attributed to tolerability issues at higher dose levels. A dose higher or lower than 0.3 mg/kg Q2W would, at best, provide similar efficacy to 0.3 mg/kg Q2W.

Overall response was not associated with any other covariates. For the probability of ≥ 7‐point improvement in mLSS, an additional predictor (number of organs involved at baseline) was identified. Axatilimab therapy increased the odds of achieving a ≥ 7‐point improvement in mLSS by 2.15 times for patients with > 4 organs involved at baseline, suggestive of systemic improvement occurring with axatilimab treatment. Among the 64% of patients in the AGAVE‐201 study population who achieved an overall response, DOR was not affected by the differences in axatilimab exposure.

In addition to grade ≥ 3 TEAEs, serious TEAEs, TEAEs associated with dose modification (discontinuation/interruption/reduction), and treatment‐related TEAEs, increases in amylase, lipase, CPK, and liver enzymes were also observed in SNDX‐6352‐0503 and AGAVE‐201. For these safety endpoints, steady‐state axatilimab exposure (AUCSS) was identified as a statistically significant predictor. For 10 of 11 evaluated safety outcomes, the $\beta$ coefficient by exposure was positive with increases in the odds of event occurrence ranging from 2.8% for serious TEAEs to 23.5% for CPK elevations per 50 μg/mL*d of AUC. The occurrence of infections of unspecified etiology was not affected by differences in axatilimab exposure. Five clinically relevant covariates were found to be associated with 4 of 10 safety endpoints for which axatilimab exposure was a significant predictor. The incidence rate of CPK elevations was positively correlated with the baseline levels of CPK. The corresponding OR for CPK increase at the 95th percentile of baseline CPK (224 U/L) was 5.19 relative to a patient with a reference CPK level of 64 U/L. The incidence rate of lipase and amylase elevations was positively correlated with baseline lipase levels. At the 95th percentile (108 U/L) for lipase, the OR for lipase and amylase elevation was 2.8‐fold higher relative to a patient with a reference lipase level of 33 U/L. Low baseline albumin and decreased functional status (defined as a baseline Karnofsky/Lansky performance score ≤ 70%) were shown to increase the risks of serious TEAEs. At the 95th percentile (47 g/L) for albumin, the OR for serious TEAEs was 0.51 relative to a patient with a reference albumin level of 40 g/L. Decreased functional status at baseline corresponded to an increased incidence rate of serious TEAEs (OR, 2.0).

To further investigate the previous finding that baseline body weight affected axatilimab exposure by > 20%, ¹⁶ median probability of response for efficacy and incidence for safety with quartiles of observed body weight distribution in the cGVHD population was predicted for efficacy and safety outcomes using the final models with the 0.3 mg/kg Q2W regimen. Comparing with the predicted median exposure for the patients in the first quartile of body weight distribution, the AUCSDINF and AUCSS increased 1.52‐fold and 1.63‐fold, respectively. However, the relative difference in mean probability of response between the percentiles of body weight was limited to 0.4% for overall response and ≥ 7‐point improvement in mLSS. Among the 10 safety endpoints associated with axatilimab exposure, the most notable effect was observed with grouped CPK elevation, showing a 4.4% increase in the probability of response (from 0.113 in the first body weight quartile to 0.118 in the fourth quartile). Although body weight at 0.3 mg/kg had relatively strong effects on exposure in the population PK/PD analysis, ¹⁶ body weight at 0.3 mg/kg had minor effects on the probability of efficacy and safety events, so it was not considered a clinically meaningful covariate. Therefore, body weight–based dosing (mg/kg) is appropriate for the body weight range (18.1–151 kg) of patients in the AGAVE‐201 study.

Based on the developed exposure‐efficacy models and forward simulations, 0.3 mg/kg Q2W provided the highest probability of response among the 3 dosing regimens tested in the AGAVE‐201 study. At 0.3 mg/kg Q2W, the mean probability of response was 70.8% for overall response and 53.2% for ≥ 7‐point improvement in mLSS, which is consistent with observed values (74.7% and 54.4%, respectively). The 0.3 mg/kg dose group also had the lowest predicted probability of event occurrence for all 10 safety endpoints among the 3 dose regimens evaluated in AGAVE‐201. At 0.3 mg/kg Q2W, 6.3% of patients had TEAEs leading to dose reductions, 38.0% had TEAEs leading to dose interruptions, and 6.3% had TEAEs leading to discontinuation. These results demonstrated that the safety profile of axatilimab administered at 0.3 mg/kg Q2W is acceptable with manageable toxicities. In addition, because the predicted probability of safety event occurrence for the 0.3 mg/kg Q2W dose was similar to that of the 0.15 mg/kg Q2W dose across all evaluated safety endpoints, a lower dose was not expected to provide a meaningful improvement in safety.

In summary, within the dose range tested in the pivotal AGAVE‐201 study, efficacy decreased as axatilimab exposure increased, while the opposite trend was observed for safety endpoints. Among the 3 dosing regimens evaluated in AGAVE‐201, 0.3 mg/kg Q2W demonstrated the most favorable efficacy and safety profiles. A dose higher or lower than 0.3 mg/kg Q2W would be unlikely to provide a better efficacy or safety profile. Although body weight was identified as a significant covariate for axatilimab exposure, its effect on efficacy and safety was minimal and therefore not considered clinically meaningful.

In conclusion, 0.3 mg/kg Q2W achieved the optimal benefit–risk profile and is the recommended dosing regimen for treating patients with cGVHD.

FUNDING

This study was funded by Incyte Corporation (Wilmington, Delaware) and Syndax Pharmaceuticals, Inc (New York, New York).

CONFLICTS OF INTEREST

Y.Y. is an employee and shareholder of Incyte Corporation. X.L., J.S., and X.C. were employees and shareholders of Incyte Corporation at the time the work was completed. A.V., V.S., C.L., and Y.K. are employees of M&S Decisions FZ LLC, a modeling consultancy contracted by Incyte Corporation.

AUTHOR CONTRIBUTIONS

Y.Y., A.V., V.S., X.L., C.L., Y.K., J.S., and X.C. wrote the manuscript. Y.Y., A.V., V.S., and X.C. designed the research. Y.Y., A.V., and V.S. performed the research. Y.Y., A.V., V.S., X.L., C.L., Y.K., and J.S. analyzed the data.

Supporting information

Data S1

Data S2

Data S3