Albumin Levels Are Predictive of Cachexia‐Induced Time‐Dependent Clearance of Therapeutic Antibodies: A Physiologically Based Pharmacokinetic Model of Durvalumab

ABSTRACT

Cachexia is a metabolic condition that accelerates the clearance of monoclonal antibodies in cancer patients and is a known mechanism causing time‐dependent clearance. Successful anticancer treatment often ameliorates symptoms of cachexia, reducing the drug clearance over time especially in patients who respond. Serum albumin level is a common biomarker of cachexia that is frequently associated with antibody drug clearance. Physiologically based pharmacokinetic (PBPK) models of antibody drugs have incorporated albumin metabolism but have not been applied to describe time‐varying clearance due to improvement in cachexia. The objective of this analysis was to evaluate albumin levels as a biomarker that is predictive of changes in antibody clearance due to cachexia. A PBPK model that jointly describes metabolism of albumin and biologic drugs was fitted to longitudinal albumin data from cancer patients treated with durvalumab and was used to predict changes in durvalumab clearance over time. PBPK model predictions were compared to empirical population pharmacokinetic (PK) models of durvalumab and other checkpoint inhibitors fitted directly to clinical PK. The model fitted the observed albumin data in cancer patients closely, and the three fitted parameters showed low uncertainty (RSE < 10%). By accounting for longitudinal albumin data in patients, the PBPK model recapitulated the observed magnitude of the change in clearance of durvalumab without fitting to clinical PK data. The model simulation demonstrated that utilization of albumin levels as a marker of cachexia in PBPK models can be used to mechanistically predict time‐dependent clearance of monoclonal antibodies.

Study Highlights

• What is the current knowledge on the topic?

  • ○
    Cachexia accelerates monoclonal antibody clearance in cancer patients, and successful anticancer treatment can improve cachexia and reduce drug clearance over time. Existing PBPK models have not been applied to predict time‐dependent clearance arising from cachexia improvement.
• What question did this study address?

  • ○
    Can longitudinal albumin data, a known biomarker of cachexia, be incorporated into a PBPK model to mechanistically predict time‐dependent clearance of monoclonal antibodies in cancer patients?
• What does this study add to our knowledge?

  • ○
    This study demonstrates that albumin can serve as a mechanistic biomarker of cachexia and predict the associated time‐dependent clearance within a PBPK framework.
• How might this change drug discovery, development, and/or therapeutics?

  • ○
    This modeling strategy offers a mechanistic means to anticipate evolving clearance in oncology populations, with potential applications in dose optimization, PK prediction, and model‐informed drug development for monoclonal antibodies.

1. Introduction

Cachexia is an inflammatory hypermetabolic condition that occurs in advanced cancer and has been shown to have an important influence on pharmacokinetics of therapeutic antibodies [1]. Systemic inflammation triggered by cytokines released from the tumor results in anorexia, muscle wasting, weight loss, and declining health status [2]. Cachexia occurs more frequently in advanced cancer and is associated with worse outcomes and anticancer treatment resistance [3, 4]. Increased protein turnover caused by cachexia has been shown to decrease serum protein levels including albumin [5] and accelerate clearance of biologic drugs [6]. Consequently, albumin level is one of the most common patient factors associated with clearance of antibody drugs in population pharmacokinetic (popPK) analyses [7]. Patients with low serum albumin tend to have faster antibody clearance and lower drug exposure reflecting the accelerated consumption of protein throughout the body that is a consequence of cachexia.

Response to anticancer treatment often improves symptoms of cachexia, decreases the turnover of proteins over time, and results in time‐dependent pharmacokinetics of therapeutic antibodies [8]. As such, patients who respond to therapy demonstrate declining drug clearance over time and elevated drug exposure compared to patients who do not respond. This introduces a circular dependency in exposure‐response analysis where the drug exposure may influence the likelihood of a response, but the response to the drug may also influence the exposure [9]. A significant exposure‐response relationship typically indicates that higher exposure causes a better response and an increased dose may offer improved efficacy. However, an alternative interpretation is that patients who respond had declining clearance and increasing exposure over time, and the exposure had no influence on the response rate, and therefore an increased dose would have no benefit. Misleading exposure‐response analyses that require careful interrogation of confounding factors have been observed with several therapeutic antibodies in oncology [9]. Therefore, time‐dependent clearance of therapeutic antibodies due to cachexia has profound implications on dose optimization.

Time‐dependent clearance of therapeutic antibodies is most commonly characterized using empirical PopPK models [8]. The drug clearance is described by a simple declining function of time, such as a sigmoidal or exponential decay function. Empirical models are descriptive and effectively summarize the observed clearance changes but offer no direct mechanistic interpretation from the fitted model. Time‐varying covariates offer a semi‐mechanistic approach where the effect over time of a clinically meaningful marker of disease status such as serum albumin or tumor size is evaluated as a covariate of drug clearance or other PK model parameters [10, 11, 12]. This approach enables interpretation of the potential contributing factors to the time‐dependent clearance but does not provide a straightforward summary of the magnitude and time scale of clearance changes and complicates simulation of covariate effects.

In contrast to empirical models, physiologically based pharmacokinetic (PBPK) models take a mechanistic approach where compartments correspond to specific tissues, and behavior mimics physiological processes. Several PBPK models of monoclonal antibodies have been developed [13, 14, 15], and typically have a whole‐body structure that divides each major organ into plasma, endosomal, and interstitial compartments with blood and lymph flow connected in physiological order. The metabolism of monoclonal antibodies is mechanistically described by pinocytosis of drug from the plasma to the endosome where it is proteolytically degraded or recycled to the cell surface by the neonatal Fc receptor (FcRn). As albumin follows the same clearance and distribution mechanisms, several PBPK models have jointly characterized the pharmacokinetics of albumin conjugates and albumin‐binding drugs alongside monoclonal antibodies [16, 17, 18]. These models offer a mechanistic link between the metabolism of albumin and clearance of antibody drugs, and can potentially enable prediction of the time course of therapeutic antibody clearance changes using serum albumin levels as an indicator of the severity of cachexia.

In this study, we describe the development of a mechanistic PBPK model of a therapeutic antibody exhibiting cachexia‐mediated time‐dependent clearance by extending a previously developed general PBPK model and using reported clinical data from an immune checkpoint inhibitor, durvalumab [19]. A model of changing protein turnover over time was fitted to longitudinal albumin data reported from durvalumab clinical trials without input from any clinical pharmacokinetic data, and was incorporated into the durvalumab PBPK model. Subsequently, the resultant PBPK model was used to simulate changes in durvalumab pharmacokinetics over time to assess its ability to predict cachexia‐mediated time‐dependent clearance. The PBPK model performance was evaluated against clinical PK outputs from empirical popPK models that were fitted directly to clinical PK data. Our hypothesis is that changes in albumin levels in cancer patients treated with durvalumab can be used to mechanistically predict the magnitude and change in durvalumab drug clearance observed in clinical studies.

2. Methods

2.1. Model Development

A previously developed physiologically based pharmacokinetic (PBPK) model (a translational two‐pore PBPK model that can characterize both endogenous and exogenous proteins) from Liu et al. [18] was extended to mechanistically characterize the time‐varying clearance frequently observed in the pharmacokinetics of immune checkpoint inhibitor antibody drugs. Briefly, the PBPK model characterizes PK of biologics of varying size including monoclonal antibodies, Fc‐fusion proteins, albumin, albumin conjugates, and albumin‐binding proteins using a whole‐body structure (Figure 1A). The model simultaneously describes endogenous albumin, endogenous immunoglobulin G (IgG), and exogenous therapeutic antibody concentrations. Protein catabolism was modeled by uptake and degradation in lysosomes within endothelial cells and recycling via the neonatal Fc receptor (FcRn) (Figure 1B). The rate of degradation of the unbound drug in the lysosome is described by the rate constant parameter k _deg, while drug that is bound to the FcRn receptor is protected from degradation and is recycled to the plasma. Endogenous and exogenous IgG compete for binding to FcRn, while albumin has a separate binding site on FcRn and does not compete with IgG. Tissue distribution was described by extravasation with a two‐pore model as a function of molecular size. Because this model jointly describes the PK of exogenous therapeutic antibodies and endogenous albumin through the same clearance pathways, it offers a link between the albumin level in the serum and catabolism of antibody drugs.

FIGURE 1.

Organ‐level structure of the whole body PBPK model (A), and internal compartments comprising each organ (B). (A) Boxes represent tissue compartments. Solid arrows represent the direction of blood flow between compartments. Dotted arrows represent the direction of lymph flow. (B) Ab, antibody; Q, blood flow; J, lymph flow; CL_up, pinocytosis uptake clearance; k _deg, degradation rate constant; k _on, association rate constant; k _off, dissociation rate constant. FcRn, neonatal Fc receptor; fr, fraction recycled to the plasma compartment; CL_TP,L, extravasation clearance via large pores; CL_TP,S, extravasation clearance via small pores.

In this study, we extended the PBPK model by Liu et al. to describe time‐dependent pharmacokinetics (TDPK) of monoclonal antibodies resulting from disease improvement and reduction in symptoms of cachexia. To describe declining serum protein catabolism over time associated with treatment response, the degradation rate constant of free proteins in the endosome k _deg(t) decreased over time (t) according to an exponential decay function (Equation 1). This function included three parameters: the degradation rate constant after the change in clearance has completed (k _deg,inf), the time‐dependent portion of the rate constant (k _deg,T), and the rate constant of the change in degradation (k _des).

$$k_{\mathrm{deg}}\left(t\right)=k_{\mathrm{deg},\inf }+k_{\mathrm{deg},\mathrm{T}}{e}^{-k_{\mathrm{des}}t}$$ (1)

The PBPK model was implemented as a system of ordinary differential equations (ODE) using the SimBiology package in MATLAB version 2023b. Model equations and parameters were adapted from Liu et al. 2024 [18], and are presented in detail in the Supporting Information. The compartment volumes and flow rates are shown in Table S1 and the remaining model parameters are described in Table S2. The three parameters of the k _deg model were fitted to longitudinal albumin data in patients with advanced cancer during treatment with durvalumab, an anti‐PD‐L1 antibody that demonstrated time‐dependent clearance. Albumin data were digitized from a popPK analysis of durvalumab [19] using WebPlotDigitizer version 4.0 (https://automeris.io). Data fitting was done in SimBiology using the fminsearch algorithm. The dependent variable was albumin concentration in the serum and only the three k _deg parameters were fitted, while all other PBPK model parameters were kept fixed. Adequacy of the fit was assessed by visual inspection of observed versus predicted data, and by statistical certainty of parameter estimates.

2.2. PBPK Model Simulations

The mechanistic durvalumab PBPK model prediction of changing clearance was compared to output from previously published empirical clearance models from population PK (popPK) analysis. Importantly, the time‐varying clearance of the PBPK model was governed only by a function describing clinical albumin data and was not fitted to any pharmacokinetic data for estimation of time dependence, while the empirical models were estimated directly from clinical PK data. As such, the purpose of this comparison is to evaluate the predictive capability of the PBPK model. Because the PBPK model was fitted to albumin data from durvalumab, the change from baseline clearance was first compared between the PBPK model and the durvalumab popPK model [19]. Durvalumab was simulated with a regimen of 10 mg/kg every 2 weeks for 60 weeks. Change from baseline clearance was compared between the popPK and the PBPK models every 10 weeks after the first dose. The concentration‐time profile of the PBPK model was simulated using the SimBiology package in MATLAB version 2023b. The apparent clearance of the PBPK model after each simulated dose over a 2‐week interval was calculated as $\mathit{CL}=\beta \times V_{\text{extrap}}$ from the simulated concentration‐time profile after repeated dosing, where the slope of the elimination phase β was calculated between two points on the concentration‐time curve in the elimination phase, and the extrapolated volume of distribution V _extrap was calculated at each dose time by log‐linearly extrapolating the concentration in the elimination phase backward to the time of dosing (C _extrap), and applying it to the equation: $V_{\text{extrap}}=\frac{\text{dose amount}}{C_{\text{extrap}}-C_{\text{trough}}}$ where C _trough is the trough concentration immediately prior to the dose.

The change from baseline CL of the popPK model over time was calculated according to Equation (2), where t is time, I _max is the logarithm of the maximal change in clearance, TI ₅₀ is the time of half‐maximal change, and γ is the Hill slope. The 95% confidence interval of the change in clearance was estimated by parametric bootstrapping of each model parameter from a normal distribution using the parameter estimate as mean and standard error of each estimate as standard deviation. Because standard errors (SE) were not reported in the popPK analysis of durvalumab, they were approximated from the reported non‐parametric 95% confidence interval according to $\mathrm{SE}\cong \frac{{\mathrm{CI}}_{\text{upper}}-{\mathrm{CI}}_{\text{lower}}}{2\times 1.96}$, where CI_lower and CI_upper are the lower and upper bounds of the reported confidence interval. The 95% confidence interval of the change in CL was calculated as the 2.5th and 97.5th quantiles across the bootstrapped samples.

$$\%∆\mathit{CL}=100\%\left(1-\exp \left(\frac{I_{\max }{t}^{\gamma }}{{\mathit{TI}}_{50}^{\gamma }+t^{\gamma }}\right)\right)$$ (2)

The simulated concentration‐time profile of durvalumab 10 mg/kg every 2 weeks was compared between the PBPK model prediction and empirical popPK model of durvalumab. The PBPK simulation was generated by solving the system of ODEs using the SimBiology package in MATLAB version 2023b. The popPK simulation was generated using the RxODE package in R version 4.3.2 using a two compartment model with parallel linear and nonlinear clearance as described in the popPK manuscript [19]. Finally, the described PBPK model was used to simulate changes in endogenous IgG levels that occur with durvalumab administration.

As an additional simulation, the change from baseline clearance from our PBPK model was also compared to the empirical clearance model of 10 immune checkpoint inhibitors with time‐dependent clearance (durvalumab [19], pembrolizumab [20], nivolumab [10], dostarlimab [21], cemiplimab [12], avelumab [22], atezolizumab [23], ipilimumab [11], tremelimumab [24], and relatlimab [25]) according to Equation (2) as described above. The population PK model parameters for these 10 molecules are listed in Table S3. All model simulations were plotted using the ggplot2 package in R version 4.3.2.

3. Results

3.1. Characterization of Albumin Level Over Time

A PBPK model of diverse biologic drugs was extended to mechanistically characterize the time‐dependent clearance of monoclonal antibodies in cancer patients due to disease improvement over time. The model was altered such that the degradation rate constant of free proteins in the endosome (k _deg) declined over time, resulting in an increase in exposure. The three parameters that describe the decline in k _deg according to an exponential decay function (Equation 1) were fitted to longitudinal albumin data reported in the popPK analysis of durvalumab [19]. The model fit of albumin concentration in the plasma matches well to the mean albumin level observed in patients who received durvalumab up to 450 days after the first dose (Figure 2). The fitted albumin profile results in a roughly 12% increase in albumin over time. Similarly, levels of endogenous IgG, another serum protein, simulated using the resultant model increased by 14% during treatment (Figure S1).

FIGURE 2.

PBPK model‐simulated serum albumin level over time (black line) fitted to albumin in patients with advanced cancer treated with durvalumab (black circles).

Table 1 shows the estimates of the three parameters included in the extended PBPK model. The degradation rate constant is divided into a time‐invariant component (k _deg,inf) and a time‐varying component (k _deg,T). Initially, $k_{\mathrm{deg}}\left(0\right)=k_{\mathrm{deg},\inf }+k_{\mathrm{deg},\mathrm{T}}$, and over time the value of k _deg asymptotically approaches $k_{\mathrm{deg}}\left(\infty \right)=k_{\mathrm{deg},\inf }$. The value of the time‐invariant component was estimated to be 18.32 1/day, and the time‐varying component was estimated to be 9.48 1/day. The time scale of change in the degradation is estimated using an exponential function with a rate constant estimated as 0.010 1/day, which corresponds to a time to half‐maximal change of approximately 2 months. The relative standard error of all parameters was less than 10% indicating good numerical certainty, though the model was fitted to digitized summary data rather than individual data which eliminates variance in the fitted data.

TABLE 1.

PBPK model parameters fitted to longitudinal albumin data in patients with advanced cancer treated with durvalumab.

Name	Description	Units	Estimate	Standard error	RSE
k deg,inf	Time‐invariant component of endosomal degradation rate constant	1/day	18.32	0.292	1.6%
k deg,T	Time‐varying component of endosomal degradation rate constant	1/day	9.48	0.110	1.2%
k tv	Rate constant of decline in degradation rate constant	1/day	0.0100	0.000621	6.2%

Abbreviation: RSE, relative standard error.

The performance of the model with time‐varying k _deg was compared to a similar implementation with varying total FcRn concentration in the endosome over time (Figures S2 and S3), and the k _deg model resulted in lower root mean squared error (RMSE) for both albumin predictions and change in clearance of durvalumab (Table S4). Therefore, the k _deg implementation was accepted as the final model.

The profile over time of k _deg(t) that results from the fitted parameters is shown in Figure 3. The value of k _deg declines by 34% relative to the time of first dose according to the exponential decay function fitted to the longitudinal albumin data of durvalumab. According to the fitted function of k _deg(t), 80% of the change in clearance occurs by Week 23, and 90% of the change occurs by Week 33.

FIGURE 3.

Model predicted change in degradation rate constant of free proteins in the endosome fitted to longitudinal albumin data from durvalumab.

3.2. Evaluation of Time‐Dependent Clearance Prediction

Figure 4 shows the concentration‐time profile predicted by the PBPK model compared to the simulated profile of the durvalumab popPK model with intravenous administration of 10 mg/kg every 2 weeks for 40 weeks. The predicted concentration of the two models is highly similar with a close match between maximum and trough concentrations over time. Both models show a typical biphasic profile typical of monoclonal antibodies. The PBPK model predicts faster tissue distribution resulting in a shorter alpha phase and lower concentration particularly in the first week after dosing compared to the popPK model. The trough concentrations of the two models match well over time, though the PBPK model shows lower troughs at late time points likely due to slight underprediction of the change in clearance.

FIGURE 4.

Concentration‐time profile of 10 mg/kg every 2 weeks for 40 weeks of the time‐varying PBPK model versus the previously published popPK model of durvalumab [19].

Figure 5 compares the change from baseline clearance predicted by the PBPK model to the change in clearance empirically estimated from clinical PK data in the popPK analysis of durvalumab [19]. Overall, the trend over time is similar between the two models, and the PBPK prediction falls within the 95% confidence interval of the change in clearance estimated by the popPK model. This indicates that the PBPK prediction falls within the statistical uncertainty estimated directly to the observed clinical PK data. Notably, the 95% confidence intervals of the change in clearance according to the popPK model are wide, indicating a substantial uncertainty in the magnitude of changes in clearance estimated from the clinical PK. After 60 weeks, the mechanistic PBPK model predicts a 12% decline in CL, and the empirical model popPK predicts a 14% decline; therefore, the magnitude of the effect is similar between the two models.

FIGURE 5.

PBPK model‐predicted change from baseline clearance (CL) over time (red line) versus empirical popPK model of durvalumab with 95% confidence intervals (black points and bars).

In order to assess the predictive performance of our PBPK model for other checkpoint inhibitors, we compared the predicted change in clearance from our model to the output from empirical popPK models of 10 checkpoint inhibitors. As the included checkpoint inhibitors were all humanized or fully human IgG antibodies, the similarity in pharmacokinetics along with an assumed comparable improvement in disease would imply that we could predict the overall time‐dependent changes in clearance of these additional 10 therapeutic antibodies. Figure 6 shows the change from baseline clearance over time predicted by the PBPK model compared to the change in clearance of 10 immune checkpoint inhibitor monoclonal antibody drugs that exhibited time‐dependent clearance in cancer patients. A high degree of variability is observed in the popPK model‐estimated magnitude and time scale of the change in clearance between the 10 antibodies. Absolute clearance predicted by the final PBPK model versus the comparator antibodies is shown in Figure S4. While the PBPK model was fitted only to albumin data from durvalumab as longitudinal albumin was not available for other molecules, overall the PBPK model prediction is representative of a typical profile.

FIGURE 6.

PBPK‐model predicted change from baseline clearance over time (black line) versus empirical popPK models of 10 immune checkpoint inhibitor monoclonal antibody drugs (points).

4. Discussion

Time‐dependent clearance due to cachexia has been observed with many therapeutic antibodies [8], and often confounds interpretation of exposure‐response analyses used for dose optimization, particularly among immune checkpoint inhibitors [26]. In this study we developed a PBPK model describing the changes in clearance over time of durvalumab, an anti‐PDL1 antibody, by incorporating reported longitudinal serum albumin level in patients with advanced cancer. To our knowledge, this is the first PBPK model to mechanistically predict time‐dependent clearance of a therapeutic antibody using longitudinal albumin levels, while previous analyses have fit directly to observed clinical PK. Additionally, we examined the generalizability of this model to other immune checkpoint inhibitors, including antibodies targeting PD1, PDL1, CTLA4, and LAG3. As of January 2024, 11 checkpoint inhibitor antibodies have been approved by the US Food and Drug Administration for treatment of cancer [27], and all of these have demonstrated time‐dependent pharmacokinetics with varying magnitudes. Longitudinal albumin data was publicly available only for durvalumab, and so we compared the prediction from our durvalumab PBPK model against reported time‐dependent CL models for other checkpoint inhibitors to evaluate whether it is representative.

The durvalumab PBPK model was developed by extending a previously published model that jointly characterized the pharmacokinetics of albumin, monoclonal antibodies, and other biologics [18]. Longitudinal albumin data from cancer patients receiving durvalumab [19] was fitted and the resultant function was used to estimate the changing rate of protein degradation throughout treatment as a result of disease improvement. No clinical PK data was used for the fitting, and the change in clearance over time predicted by the PBPK model was compared to the change in clearance estimated directly from clinical PK data by popPK analysis of durvalumab and nine other immune checkpoint inhibitors. The PBPK model matched well to the observed albumin data over time and showed low standard error of the estimated parameters indicating good precision. The fitted value of k _deg in our model characterizing patients with advanced cancer ranges from 27.8 to 18.3 1/day over time, while it is 15.3 1/day in the model described by Liu et al. [18] fitted to data primarily from healthy volunteers and non‐cancer patients, reflecting a similar but elevated rate of degradation in advanced cancer. Overall, the magnitude and time scale of time‐varying clearance is predicted by the PBPK model within the uncertainty of the empirical popPK model of durvalumab and captures a representative profile across 10 checkpoint inhibitor antibodies.

The rate constant of degradation of free proteins in the endosome was varied over time to fit to the clinical albumin data in the durvalumab PBPK model, reflecting an accelerated rate of protein degradation in cachectic patients prior to treatment that decreases following therapy and disease improvement. This is consistent with increased rates of protein turnover observed in cachexia [5], and suggests increased degradation is likely the cause of hypoalbuminemia. While lower albumin could potentially be explained by a lower rate of albumin synthesis, clinical evidence suggests that synthesis rates are unchanged in cachexia [28]. Altered expression of FcRn in cachexia is a potential alternative mechanism. Predictions of longitudinal changes in albumin and CL from a PBPK model implementation using varying FcRn expression over time was compared to the model using time‐varying degradation rate and resulted in greater error when compared to observations (Table S4), suggesting that predictions from the k _deg model are more consistent with the observed data. In line with this, altered drug clearance of proteins that lack FcRn interactions has been observed in cachexia mouse models, suggesting that changes in FcRn expression over time cannot fully explain the phenomenon [29]. Protein losing enteropathy, where serum proteins are lost via the gastrointestinal tract, has been proposed as a potential mechanism of elevated antibody clearance, and while it occurs frequently in gastrointestinal cancers, it is uncommon across other indications, and therefore is unlikely a driving mechanism [30]. Interactions with Fc gamma receptors (FcgR) expressed on immune cells sometimes result in elevated antibody clearance, particularly following formation of multimeric complexes of antibody bound to soluble antigen. However, monomeric immune complexes (an unbound antibody or a single antibody bound to a soluble antigen) are not cleared efficiently by this route [31, 32]. Because immune checkpoint inhibitors target membrane‐bound proteins with a generally low concentration of shed monomeric soluble target, formation of multimeric immune complexes is unlikely.

The decline in serum protein clearance in the model resulted in a predicted 14% increase in endogenous IgG level. Similar differences in serum IgG have been observed between patients with malignancy and healthy controls in clinical studies [33, 34, 35]. Increases in IgG have also been observed in patients following tumor resection surgery [33]. Furthermore, low serum IgG has been associated with cancer risk [36], and increases in IgG during treatment with checkpoint inhibitors have been associated with improved survival [37], both suggesting a potential link with cachexia and disease severity. Overall, this suggests that the model prediction is consistent with clinical data and is supportive of the mechanism described by our model.

The change in albumin level observed in clinical studies of durvalumab was sufficient to explain the magnitude in time‐varying clearance estimated by population PK analysis. Clinical PK data was not used to fit the model, and the purpose of the model was to explore whether albumin alone could mechanistically recapitulate the observed time course and magnitude of the changes in PK. The match between the PBPK prediction and observed trends substantiates the hypothesis that time‐varying clearance of immune checkpoint inhibitors is driven by altered rates of non‐specific protein turnover due to cachexia. Furthermore, accounting for changes in protein catabolism during anticancer treatment will improve predictive performance of PBPK models and enable simulation of time‐varying clearance, a phenomenon that has had significant influence on dose optimization of monoclonal antibodies [38].

While the fitted model describes the time profile of durvalumab drug clearance well, there is a significant variability in the magnitude of time‐varying clearance observed among immune checkpoint inhibitors. Because all of the checkpoint inhibitors included in this analysis are humanized or fully human IgG1, IgG2, or IgG4 antibodies (Table S3), the nonspecific clearance was expected to be comparable [1]. Consistent with this, baseline linear clearance values reported in popPK analyses are similar between these molecules aside from avelumab, which shows faster clearance due to suspected increased tendency of associating with membranes (Table S1). However, the maximal decrease in clearance over time ranges from 6% to 25% with a mean of 17% across the 10 checkpoint inhibitors in this analysis. The durvalumab PBPK model predicts a 12% decrease in clearance, which is below the mean across these 10 drugs, but nonetheless represents a typical profile of time‐varying clearance. Longitudinal albumin data in patients was published only for durvalumab, so the model could not be individually fitted to each drug. The predictive performance of the model likely would benefit from drug‐specific albumin data, as each compound was tested in different patient populations, diverse disease types, with varying combination treatments, and demonstrated a range of response rates. All these factors could influence the degree to which cachexia improves throughout treatment, result in a different albumin profile, and alter the magnitude of time‐varying clearance. However, the many factors that could influence the magnitude of change in clearance are not well understood quantitatively.

The PBPK model presented here was fitted to summary‐level albumin data for durvalumab, as individual‐level data was not publicly available. However, the response to therapy and improvement in health status differ significantly between patients, and this would be reflected in the albumin and drug clearance profiles over time. With individual‐level albumin data from clinical study of checkpoint inhibitors, interpatient variability could be incorporated into the PBPK model, facilitating exploration of factors that may influence time‐varying clearance. For example, patients who respond well to therapy often have a greater decline in clearance compared to those who do not. Drug clearance and albumin level have both demonstrated prognostic value in predicting the response to immune checkpoint inhibitors [39]. Fitting a mechanistic model to individual albumin profiles could enable meaningful simulation and prediction of drug clearance and health outcomes given albumin measurements, a routine laboratory test collected in cancer treatment.

PBPK models of biologic drugs have been developed to mechanistically characterize various molecular properties, including electrostatic charge [40], molecular size [18], target binding [15, 41, 42], and pH sensitivity [43]. PBPK models of antibodies against lymphocyte targets have been applied to describe the time‐varying clearance caused by drastic changes in target expression due to rapid lymphocyte depletion. However, time‐dependent clearance due to declining nonspecific protein catabolism during cancer treatment has not been previously investigated using a PBPK approach. In this study, we mechanistically demonstrated that changes in albumin level, a commonly used biomarker of cachexia, can explain the altered rate of protein degradation, and that its inclusion allows for the prediction of time‐dependent changes in clearance due to cachexia. Time‐dependent clearance due to cachexia has been shown to have a profound impact on the interpretation of exposure‐response analyses that mislead accurate dose optimization [8].

This study focused on the influence of albumin level on antibody drug clearance, but incorporation of other factors may further improve model predictions. Tumor burden is frequently associated with clearance of antibody drugs [22, 23, 44]. Cancer cells have been shown to consume extracellular protein as a source of nutrients [45], so high tumor burden could potentially contribute to non‐specific drug clearance. Furthermore, a high tumor burden results in more target‐expressing cells, which could increase drug clearance via target‐mediated drug disposition. Addition of tumor size to the PBPK model could interrogate the relative contribution of cachexia versus tumor‐mediated mechanisms in time‐dependent clearance under different scenarios. Albumin level is predictive of the likelihood of response to checkpoint inhibitors [46], and therefore tumor burden and albumin level are likely to be highly interrelated throughout treatment, and longitudinal modeling may reveal their relative contributions to time‐dependent clearance.

Author Contributions

J.R.P. and H.W. wrote the manuscript and designed the research. J.R.P. performed the research and analyzed data. Both authors reviewed and approved the final manuscript.

Funding

We gratefully acknowledge The University of British Columbia and The Killam Doctoral Scholarships for providing financial support in the form of a graduate stipend to J.R.P. throughout this research.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Appendix S1: psp470185‐sup‐0001‐AppendixS1.docx.

Appendix S2: psp470185‐sup‐0002‐AppendixS2.docx.

Appendix S3: psp470185‐sup‐0003‐AppendixS3.xlsx.