Pembrolizumab Exposure–Response Assessments Challenged by Association of Cancer Cachexia and Catabolic Clearance

Exposure–response (E–R) assessments in oncology have played an increasingly important role toward understanding the impact of dose selection on patient outcomes (1). To date, several notable E–R analyses in different cancer types have considered a range of clinical endpoints, including overall survival (OS), progression-free survival (PFS; refs. 2–6), and overall response rate (ORR)/tumor kinetics (7–16). The registration of pembrolizumab for the treatment of melanoma and non–small cell lung cancer (NSCLC) at the dose of 2 mg/kg administered once every 3 weeks (Q3W) was supported by dose–response analyses and model-based E–R analyses of longitudinal tumor size (10, 17) using techniques adapted from earlier work (7, 18).

Pembrolizumab (KEYTRUDA; Merck) is a humanized IgG4 mAb that directly binds programmed death 1 (PD-1) expressed on T cells, blocking the interaction between PD-1 and its ligands, PD-L1 and PD-L2. The PD-1 pathway represents a major immune checkpoint that tumor cells exploit to evade antitumor T-cell activity. When pembrolizumab binds PD-1, this restores antitumor immunity and T-cell activity against cancerous cells. The randomized comparisons of pembrolizumab doses of 2 and 10 mg/kg have corroborated all E–R simulations to date showing similarity in efficacy outcomes for pembrolizumab-treated melanoma and NSCLC across this 5-fold dose range. In melanoma, three randomized comparisons demonstrated similar efficacy for pembrolizumab at 2 versus 10 mg/kg Q3W regimens (11, 12). In a phase II/III study of advanced NSCLC, pembrolizumab doses of 2 and 10 mg/kg Q3W provided superior OS compared with docetaxel, with similar ORR and PFS outcomes at each dose (19).

Outside of pembrolizumab development, relatively few E–R analyses involve data from mAb dose-ranging studies. Therefore, the present pembrolizumab evaluations from two large randomized trials of pembrolizumab at 2 and 10 mg/kg Q3W have provided an important opportunity to gain insight into the challenges and potential complications for E–R of large molecules in oncology. To our knowledge, this is the largest analysis of OS to date for pembrolizumab-treated patients and the first published exposure–survival analysis of pembrolizumab.

The primary analyses include data from two large randomized trials, KEYNOTE-002 in patients with melanoma (ClinicalTrials.gov, NCT01704287; ref. 20) and KEYNOTE-010 in patients with NSCLC (ClinicalTrials.gov, NCT01905657; ref. 21). Additional data from the pivotal study of pembrolizumab in first-line NSCLC (KEYNOTE-024; ClinicalTrials.gov, NCT02142738) served as a separate validation cohort (22). The KEYNOTE-002 and KEYNOTE-010 studies were expected to provide a robust, well-balanced data set for investigation of pembrolizumab exposure–survival relationships for each indication.

All studies were conducted in accordance with Good Clinical Practice guidelines and the Declaration of Helsinki. All protocols and subsequent amendments were approved by the appropriate institutional review boards or ethics committees at each participating institution. All patients provided voluntary written informed consent. Brief details on study treatment, enrollment criteria, and design are included in Supplementary Methods. Additional complete details regarding the designs for each trial have been published (19, 22, 23).

Clearance (CL) is the typical variable derived in pharmacokinetic (PK) analyses to reflect the kinetics of the elimination process. It relates concentration of drug measured in the body to a dose or amount administered. In this report, CL is specifically defined as the volume of serum from which pembrolizumab is completely removed per unit time. Results from a recently developed, time-dependent pharmacokinetic (TDPK) model provided post hoc CL estimates for E–R assessments in this report (sensitivity analyses conducted to confirm results were independent of the choice of model structure/exposure metric). A population PK approach was applied to determine typical pembrolizumab PK parameter values as well as associations of covariates and parameter values. In this type of analysis, data from every individual are considered simultaneously in a unified model. Further details on PK sampling and brief background information on the methods of this approach are provided in Supplementary Materials. Full details of PK model methods are published separately (24).

Serum exposure, i.e., area under the serum concentration–time curve (AUC), was calculated using the first-dose CL as Dose/CL₀. As in prior E–R analyses, AUC was normalized to a 6-week interval [AUC_6weeks,CL0 = Dose/CL₀ × (6/dosing frequency in weeks)] to provide an exposure value over an integer number of dosing intervals when considering every 2 weeks (Q2W) and Q3W.

Patient demographics and laboratory values were explored to account for survival variation using the Cox model methods described in Supplementary Methods. Eastern Cooperative Oncology Group (ECOG) status (0 vs. 1) and region (East Asian vs. not East Asian) were two of the stratification variables for the primary KEYNOTE-010 efficacy analyses and therefore considered for the NSCLC modeling. Other potential baseline variables of interest included baseline sum of the longest diameter of the target lesions (BSLD), lactate dehydrogenase (LDH), serum albumin (ALB), platelet count (PLT), age, gender, cancer stage, presence of EGFR-sensitizing mutation, and histology (squamous vs. nonsquamous). For melanoma, potential covariates also included a variable indicating tumor PD-L1 expression positivity (PD-L1) and presence/absence of BRAF mutation. For evaluating the possible connection of on-study cancer cachexia progression and OS, postbaseline rates of body weight change (WTRATE) and serum albumin change (ALBRATE) were incorporated into the analyses as time-dependent covariates. These variables reflect an instantaneous rate of change (% change/day) from baseline and thus normalize for potential differences in follow-up. Using the time-varying values of WTRATE and ALBRATE, the Cox model compares risk of an event across updated values at each event/time interval when patient measurements are taken, re-evaluating which risk group each person belongs based only on values occurring up to and not beyond the considered time interval.

Separate Cox models for pembrolizumab-treated melanoma and NSCLC were implemented to determine if another factor or combination of factors unrelated to pembrolizumab exposure could explain a perceived survival gap between rapid and slow pembrolizumab elimination, i.e., CL₀ subgroups. Such a substitution, if possible, would facilitate an unbiased assessment of pembrolizumab exposure–related influence on patient outcome (as exposure is proportional to dose and inversely proportional to CL₀). Univariate OS models were established considering CL₀ alone. Final multivariate models were developed both with and without CL₀ considering all patient factors selected during model construction (Akaike information criterion used in selection). The ability of the multivariate models to account for the CL₀-associated survival gap was assessed by comparisons of the CL₀ HR with the univariate model and change in log-likelihood from adding CL₀ after entering other factors in each model.

To explore alternative E–R methodology proposed in earlier publications, a matched, case–control E–R analysis was also performed using the NSCLC KEYNOTE-010 data. Similar to methods from the prior analyses (4), patients of each pembrolizumab dose arm and respective exposure quartiles were included in case groups. Control patients receiving docetaxel alone were matched 1:1 by the Mahalanobis metric method to the case groups based on AIC-selected risk factors described above for the multivariate Cox models (25). The relative pembrolizumab treatment effect versus chemotherapy for the case–control matched datasets was summarized by HR. The intent of this method was to explain some or all of the confounded CL₀–OS association by alternative factors to permit unbiased E–R assessments.

Details regarding study centers and investigators have been previously reported (19, 22, 23). Only pembrolizumab-treated subjects with at least one PK measurement were included (n = 340 for melanoma, n = 652 for NSCLC). The subsets of these patients with no missing covariates of interest were included in multivariate Cox regressions and case–control analyses (complete-case dataset n = 211 melanoma; n = 537 NSCLC). The source of analysis datasets, number of subjects, and baseline characteristics are detailed in Supplementary Tables S1 and S2.

Supplementary Fig. S1 shows the OS curves for the two dose groups in melanoma and NSCLC (based only on subjects with available PK data from KEYNOTE-002/-010). The overlap in the confidence intervals and Cox HRs reflects comparability of OS across a 5-fold dose and exposure range from 2 to 10 mg/kg in both tumor types [melanoma Cox HR = 0.98; 95% confidence interval (CI), 0.94–1.02 and NSCLC Cox HR = 0.98; 95% CI, 0.95–1.01].

OS stratified by exposure and dose is presented in Supplementary Figs. S2 and S3, and subjects in the 1st quartile of exposure within the 2 mg/kg arms of KEYNOTE-002/-010 have similar OS compared with the 1st quartile within the 10 mg/kg arms, despite a 5-fold dose and exposure range across these subgroups. However, subjects in the 4th quartile of exposure for 2 mg/kg demonstrate substantially better OS than the 1st quartile exposure of 10 mg/kg, despite having a lower exposure [4th quartile of exposure for 2 mg/kg (∼2,000 μg/mL x day); 1st quartile exposure of 10 mg/kg (∼4,000 μg/mL x day)]. This unusual pattern of improved survival in subjects with higher exposure within each dose is incongruent with the similarity in OS across the 5-fold dose/exposure range, suggesting a confounding of PK and OS independent of direct pharmacologic effects on patient outcome. Thus, E–R trends were further explored by comparing the relationships of OS and CL₀ both within and across 2 and 10 mg/kg (Fig. 1 focusing on the outer quartiles). These data reveal a considerable difference in median OS for pembrolizumab-treated NSCLC between subjects with rapid (4th quartile) and slow (1st quartile) CL₀, i.e., 8.4 months (95% CI, 6.4–11.0) versus >∼23 months (lower 95% CI not reached in subjects with slow CL). Additionally, Supplementary Fig. S4 demonstrates 2nd and 3rd quartiles showing a pattern of graded response between the 1st and 4th quartiles. These same CL₀–OS trends are observed in melanoma (Fig. 1; Supplementary Fig. S4) and previously untreated (first-line) NSCLC (Supplementary Fig. S5; KEYNOTE-024, n = 152). Taken together, the dose–response analyses reinforce a lack of exposure-dependency in outcome, with CL₀–OS trends highlighting an underlying correlation of OS with pembrolizumab elimination.

Table 1 describes distributions of demographic and other baseline/on-study characteristics for subjects treated in KEYNOTE-002 and -010, stratified per CL₀ quartile within each indication; Fig. 2 shows results of the multivariate Cox analyses for melanoma and NSCLC. The multivariate Cox model in pembrolizumab-treated melanoma indicated BSLD, PD-L1, PLT, WTRATE, ALB, BRAF mutation status, and ECOG score were independently (P < 0.05) associated with OS in advanced melanoma (Fig. 2A). ALBRATE, ALB, WTRATE, baseline LDH, histology, gender, BSLD, and ECOG status were associated with OS in pembrolizumab-treated NSCLC (Fig. 2B). Notably, the cachexia-related factor associated with change in body weight (WTRATE) was found to account for a portion of survival variability in both populations. Similar OS trends for rate of on-study weight change were also observed in docetaxel-treated NSCLC (Supplementary Fig. S6), suggesting a disease-level involvement of weight loss and OS. Generally, the 4th quartile of CL₀ in both melanoma and NSCLC corresponds to a point estimate for 9-week weight loss exceeding the consensus 5% cutoff commonly understood as the diagnostic criterion for cachexia (Table 1). This binary cutoff was initially considered in the Cox OS models, but ultimately it was not found to be significant as it did not afford the level of granularity that a continuous measure of weight loss provides. This is an oft-cited critique of dichotomizing continuous variables in regression analyses (26, 27).

Table 2 summarizes the relationship of CL₀ and OS for melanoma and NSCLC, both in the unadjusted, univariate Cox models and the multivariate models, where the relative risk is adjusted for potential confounders such as known clinical risk factors and other derived variables which serve as a proxy for on-study progression of cancer cachexia (WTRATE/ALBRATE; defined above). Overall, the univariate, unadjusted CL₀ HR for KEYNOTE-010 was estimated to be 2.64 (95% CI, 1.94–3.57, P ≤ 0.001), representing the incremental risk of death per one unit increase of log-transformed CL. After entering the other baseline patient factors and on-study cachexia-related factors, WTRATE and ALBRATE, the adjusted point estimate of CL₀ HR was 1.53, but the 95% CI overlapped unity (0.97–2.41). The version of this model only adjusting for risk factors known at the start of treatment, i.e., excluding WTRATE/ALBRATE factors, showed CL₀ HR to be 1.88 (95% CI, 1.22–2.89). The resultant −2 × log likelihood increased by 4.9 units in the version of this baseline-risk-factor model without CL₀, confirming an inability to fully explain a prominent CL₀–OS association without adjusting for the postbaseline cachectic markers, again despite the presence of 2 to 10 mg/kg dose–OS similarity. Similarly, the presence of a hidden E–R confounder is also demonstrated in the pembrolizumab-treated melanoma population (KEYNOTE-002; Table 2), where the univariate (unadjusted model) CL₀ HR was 2.56 (95% CI, 1.72–3.80), and the multivariate-adjusted (Fig. 2) CL₀ HR was found to be 1.60 (95% CI, 1.04–2.47), and in the baseline-factor-only version of this same multivariate model (without WTRATE), the CL₀ HR was 1.64 (95% CI, 1.06–2.52).

Figure 3, illustrating NSCLC case–control analyses, further substantiates these findings. The trends for slope of HR versus exposure appear markedly steeper within-dose than the overall across-dose exposure comparisons, including the case–control analyses in which pembrolizumab exposure subgroups and the docetaxel control arm were matched on similar AIC-selected risk factors (methods as described above). E–R analyses using early/late exposure metrics from the TDPK model or from the previously described, simpler, static, two-compartment model yield similar conclusions (Supplementary Table S3).

This report describes the first E–R assessment of survival for the PD-1–targeted mAb, pembrolizumab. Consistent with prior dose–response assessments and E–R results showing a flat relationship by tumor size response (19, 23, 28), similar survival outcomes were observed here at both 2 and 10 mg/kg. A lack of clinically relevant exposure-dependency in OS is demonstrated both by exploratory Kaplan–Meier plots, stratified per AUC quartiles, and in the Cox model–derived HRs (Fig. 3), reflecting similar values comparing the same quartile of AUC for 2 versus 10 mg/kg (∼5-fold exposure differences). Apparent within-dose differences in response are therefore driven by a relationship between CL and survival rather than reflecting a true causal exposure effect on survival. These results confirm 2 mg/kg (or a comparable fixed dose) as appropriate for treatment of melanoma and NSCLC, providing near-maximal efficacy such that dosing 5-fold higher does not meaningfully increase OS (29, 30). Of note, recent pembrolizumab trials in NSCLC and melanoma have begun to evaluate the 200 mg Q3W regimen in lieu of weight-based dosing, based on doses of 200 mg and 2 mg/kg, providing similar exposure distributions with no advantage to either dosing approach with respect to controlling PK variability (30).

It is noteworthy that despite the confirmation of flat dose/E–R relationship, the analyses described here also reveal a prominent association of pembrolizumab CL₀ at baseline and OS, whereby subjects with slower CL₀ have a more than doubled life expectancy. Pembrolizumab CL₀ combined with dosing regimen determines exposure (e.g., AUC = Dose/CL); thus, the overall lack of influence of pembrolizumab dose and considerable within-dose CL₀/exposure–OS trends signify latent confounding between pembrolizumab elimination and disease status. Pembrolizumab has demonstrated target-mediated drug disposition (TMDD) at doses up to 0.1 mg/kg—a 20-fold lower dose compared with 2 mg/kg. Therefore, the impact of TMDD on total pembrolizumab CL is considered negligible at clinically relevant doses and is not thought to be a driver of the CL₀–OS associations observed here. A more plausible hypothesis for this relationship is due to an altered catabolic state among subjects with end-stage disease, independent of any causal relationship of exposure and OS.

Cancer anorexia-cachexia, with an estimated prevalence of 15% to 40% in the general cancer population (31), is known to portend poorer prognosis in advanced malignancy (32, 33). Affected individuals can suffer dramatic loss of body weight and muscle strength (34–36), and those same catabolic drivers accompanying skeletal muscle loss also constitute a primary elimination pathway of the humanized IgG4/κ mAb, pembrolizumab (and similar biologics). In recently published findings, Flint and colleagues reported patients receiving immunotherapy may be particularly susceptible to cancer-associated cachexia (37). Their results revealed that cancer cachexia, triggered by tumor-induced changes in liver metabolism, produces an upregulation of stress hormone production which can ultimately lead to systemic immunosuppression and a loss of immunotherapy efficacy in some patients. A correlation of rapid mAb catabolism (the mechanism of pembrolizumab CL) with poorer OS is demonstrated in the present analyses. Moreover, numerous recent analyses suggest these patterns may not be unique to pembrolizumab, but rather, a broader phenomenon generalizable to other antineoplastic mAbs with similar catabolic CL mechanisms. For example, the initial 2010 biologic license application for the IgG1 anti–CTLA-4 mAb, ipilimumab, contained E–R results describing a pronounced exposure–OS relationship in melanoma (38). This prompted a postmarketing commitment for a large comparative trial, prospectively evaluating OS with randomized doses of 3 and 10 mg/kg (clinicaltrials.gov: NCT01515189); yet, a recent integrated analysis of ipilimumab phase II and III data showed little OS difference across this dose range (39). The early trastuzumab (anti-HER2/neu IgG1 mAb) population PK and E–R analyses similarly suggested a strong PK (C_min) association with OS that was later acknowledged as capturing, at least partially, imbalances in general disease risk factors (4, 40). More recent examples of this same phenomenon have been described for nivolumab (anti–PD-1 IgG4 mAb) and atezolizumab (anti–PD-L1 IgG1 mAb; ref. 41). These repeated findings across biologics involving differing therapeutic targets imply a likely common source of confounding which has been postulated in other specific instances to involve a correlation of cachexia and increased mAb catabolism secondary to generalized protein turnover. (42, 43) Their shared metabolic pathways and the well-established link of anorexia/cachexia-related metabolic wasting and patient outcome thus present a credible explanation that should be further explored.

To further investigate the CL₀–OS association and determine which clinical factors could possibly be linked to hypercatabolism and contribute to PK–OS convolution, a multivariate Cox regression model was explored. Similar to techniques implemented by the FDA in a prior case–control E–R assessment for trastuzumab, this survival model sought a potential replacement for CL₀ as a correlate of survival among various relevant prognostic factors (4). Although these analyses demonstrated that some survival variation linked to pembrolizumab CL₀ can alternately be ascribed to other risk factors, some correlation remained unexplained. The change in log-likelihood from the multivariate survival model of KEYNOTE-010 with and without CL₀ further confirms this finding. This highlights a recognized limitation of the E–R case–control approach, i.e., that one cannot discern whether any remaining unexplained CL₀–OS association is attributable to other hidden confounders or to a true E–R relationship (4, 44, 45). This investigation suggests that simply matching case controls on standard baseline disease factors may be inadequate to distinguish on-treatment cause and effect and delineate the influence of biologic drug exposure in this causal sequence. One could speculate that the association between pembrolizumab CL₀ and survival is difficult to displace by traditional risk factors such as ECOG, cancer stage, etc. because the measurement of CL₀ here provides a more precise estimate of catabolic rate, reflecting overall health and degree of cachexia.

In summary, a lack of clinically relevant exposure-dependency in OS with pembrolizumab across the dose range of 2 to 10 mg/kg was demonstrated for both melanoma and NSCLC. Consistent with prior randomized comparisons, these data support that increasing exposures above those attained at 2 mg/kg do not meaningfully improve response. In addition, we have shown rapid CL was strongly linked to decreased OS likely due to it being a proxy of disease severity and overall patient health. Given that the confounded association of longitudinal disease burden and PK has been observed across a class of oncology therapies, caution is warranted in interpreting E–R relationships, especially in the context of oncology trials evaluating a single dose level of biologic/mAb. Though challenging, randomized dose-ranging studies appear to be the only viable approach for decoupling PK of mAb and other latent confounders to delineate the role of exposure, and thus the impact of dose selection, on patient outcome. This ultimately underscores a broader need for better predictive clinical biomarkers of cachexia to understand the role of confounders of CL and survival in patients with cancer (46). Validation of a clear and standardized biomarker capturing the spectrum of cachexia in cancer could significantly aid in disease staging, refining trial inclusion/exclusion criteria, and, as evidenced here, E–R deconvolution.

A.G. Robinson reports receiving speakers bureau honoraria from Merck & Co., Inc., and is a consultant/advisory board member for AstraZeneca, Bristol-Myers Squibb, and Merck & Co., Inc. E.B. Garon reports receiving other commercial research support from AstraZeneca, Bristol-Myers Squibb, and Merck & Co., Inc. S.W. Ebbinghaus and J.A. Stone hold ownership interest (including patents) in Merck & Co., Inc. No potential conflicts of interest were disclosed by the other authors.

Conception and design: A.G. Kondic, K.M. Anderson, K. Mayawala, S.W. Ebbinghaus, J.A. Stone

Development of methodology: D.C. Turner, A.G. Kondic, K.M. Anderson, K. Mayawala, J.A. Stone

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): A.G. Kondic, A.G. Robinson, E.B. Garon, J.W. Riess

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): D.C. Turner, A.G. Kondic, K.M. Anderson, A.G. Robinson, L. Jain, K. Mayawala, J. Kang, S.W. Ebbinghaus, D.P. de Alwis, J.A. Stone

Writing, review, and/or revision of the manuscript: D.C. Turner, A.G. Kondic, K.M. Anderson, A.G. Robinson, E.B. Garon, J.W. Riess, L. Jain, K. Mayawala, J. Kang, S.W. Ebbinghaus, V. Sinha, D.P. de Alwis, J.A. Stone

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): A.G. Kondic, K.M. Anderson, J. Kang

Study supervision: A.G. Kondic

The authors thank all of the patients and their families, nurses and clinical and laboratory personnel, study coordinators, and operations staff who participated in these studies. Editorial assistance in the preparation of this report was provided by Christine McCrary Sisk and Sheila Erespe (both from Merck & Co., Inc.).

Funding for this research was provided by Merck Sharp & Dohme Corp., a subsidiary of Merck & Co., Inc.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.