Anti‐PD‐(L)1 Antibodies: Insights From QSP‐Based Meta‐Analysis

ABSTRACT

Checkpoint inhibitors that target PD‐1 or PD‐L1 have had a profound effect in a variety of cancers, both as a single therapy and in combinations. Meta‐analyses suggest that monoclonal antibodies (mAbs) targeting PD‐1 may yield better survival outcomes compared to anti‐PD‐L1 mAbs, however these conclusions are limited by a lack of direct clinical comparisons between the two classes. There is a shared hypothesis for the mechanism of action of these drugs: inhibition of the PD‐1:PD‐L1 signaling pathway through binding to either target. Using a Quantitative Systems Pharmacology (QSP) model‐based analysis, we test whether differential inhibition of PD‐1:PD‐L1 complex formation (a surrogate for inhibition of the signaling pathway) is sufficient to explain the efficacy difference between anti‐PD‐1 and anti‐PD‐L1 mAbs observed in clinical meta‐analyses. The model predicts that high levels of PD‐1:PD‐L1 complex inhibition are achieved by all the considered mAbs at their clinical dosing regimens, but it does not indicate that anti‐PD‐1 mAbs yield higher inhibition over anti‐PD‐L1s, in contrast to the meta‐analyses. Significant model parameter variability and a bootstrap sampling analysis mirroring the comparison from Duan et al. (2020) do not change this conclusion. This suggests that anti‐PD‐1 and anti‐PD‐L1 mAbs are not differentiable based on PD‐1:PD‐L1 complex inhibition alone, and that the hypothesized shared mechanism of action of the two classes of drugs is incomplete.

1. Introduction

Cancer immunotherapy has emerged as a promising strategy for harnessing the immune system to eliminate tumor cells [1, 2, 3]. Checkpoint blockade of the PD‐1: PDL‐1 interaction leads to tumor regression, and antibodies targeting PD‐1 and PD‐L1 have been approved for treating a variety of tumors. PD‐1 is primarily expressed on T cells and normally functions in immune‐mediated tolerance through interaction with PD‐L1, which is expressed in multiple tissues, and the PD‐1:PD‐L1 signaling pathway has been shown to negatively regulate effector T‐cell responses and protect against immune‐mediated tissue damage [4]. PD‐L1 is often highly expressed in tumors, where it exploits this pathway through binding to PD‐1 expressed on tumor‐infiltrating T cells and thereby prevents T cell activation. Blocking the PD‐1:PD‐L1 interaction using antibodies against either target releases inhibition of T cell signaling, leading to tumor cell killing [4, 5].

By binding to either target, both anti‐PD‐1 and anti‐PD‐L1 mAbs inhibit the formation of PD‐1:PD‐L1 complexes and thus the PD‐1/PD‐L1 signaling pathway, promoting T‐cell activation and antitumor activity. However, several studies have suggested that anti‐PD‐1 therapy may be more effective than anti‐PD‐L1 therapy. The meta‐analysis by Duan et al. [2] compared pembrolizumab, nivolumab, atezolizumab, and avelumab clinical studies using a mirror‐principle approach and found that anti–PD‐1 was more effective than anti‐PD‐L1 in terms of overall survival (HR, 0.75; 95% CI, 0.65–0.86; p < 0.001) and progression‐free survival (HR, 0.73; 95% CI, 0.56–0.96; p = 0.02). Other meta‐analyses have found similar conclusions, while others have found no significant differences between the two mAb classes. However, there are no direct clinical head‐to‐head studies to determine this definitively (see Zhao et al. [3] for a more thorough review).

Here we used QSP modeling to address whether disruption of PD‐1:PD‐L1 complexes is more efficient using anti‐PD‐1 antibodies compared to anti‐PD‐L1 antibodies, and whether this can explain the differences observed in clinical studies. Factors that could affect differential complex disruption by anti‐PD‐1 versus anti‐PD‐L1 include expression and turnover of both receptors (PD‐1 and PD‐L1), drug‐receptor affinities, and drug PK parameters. Included in this analysis are five approved mAbs targeting PD‐1: pembrolizumab (KEYTRUDA, Merck Sharp & Dohme LLC, Rahway, NJ), nivolumab (OPDIVO, Bristol‐Myers Squibb Company, Princeton, NJ), cemiplimab (LIBTAYO, Regeneron Pharmaceuticals Inc., Tarrytown, NY), dostarlimab (JEMPERLI, GlaxoSmithKline LLC, Philadelphia, PA), and retifanlimab (ZYNYZ Incyte Corporation, Wilmington, DE), and three approved mAbs targeting PD‐L1: atezolizumab (TECENTRIQ, Genentech Inc., South San Francisco, CA), avelumab (BAVENCIO, EMD Serono Inc., Boston, MA), and durvalumab (IMFINZI, AstraZeneca UK Limited, Cambridge, England).

2. Key Question

Despite the shared hypothesis for the mechanism of action of anti‐PD‐1 and anti‐PD‐L1 antibodies (inhibition of the PD‐1:PD‐L1 complex), better survival outcomes of anti‐PD‐1 treatment have been demonstrated by recent clinical meta‐analyses. Here, a QSP model‐based analysis is used to compare the level of PD‐1:PD‐L1 complex inhibition across approved anti‐PD‐1 and anti‐PD‐L1 antibodies. The key objective of this case report is to investigate whether the predicted level of PD‐1:PD‐L1 complex inhibition captures the efficacy differences seen in meta‐analyses.

3. Analysis Plan and Key Assumptions

Model simulations were run using the “Monospecific Anti‐Receptor with Membrane Ligand Competitor (4‐Compartment) with Avidity” model in Applied BioMath QSP Notebook (Certara Predictive Technologies) [6] shown in Figure 1. It is a mechanistic pharmacokinetic/pharmacodynamic (PK/PD) model of a mAb binding to a membrane‐bound receptor on the surface of one cell and blocking binding to a membrane‐bound ligand on the surface of another cell. In the model, the “receptor” is defined as the target of the mAb, and the “ligand” is the cognate binding protein of that target (i.e., for anti‐PD‐1 mAbs, the “receptor” is PD‐1 and the “ligand” is PD‐L1, and vice versa for anti‐PD‐L1 mAbs). There are three compartments in the model: central, peripheral, and disease (tumor). The key model output is the predicted level of PD‐1:PD‐L1 complex inhibition in the tumor. For more details about the model, see Applied BioMath 2025 [7], Head et al. [8], Marcantonio et al. [9].

FIGURE 1.

Model diagram. After administration in the Central compartment, the antibody can distribute to the Peripheral and Disease (Tumor) compartments. Elimination occurs in all compartments. The antibody can bind bivalently to membrane and soluble receptors. Synthesis, internalization, and binding of target receptor to ligand are also captured in every compartment. Figure created with BioRender.com.

The biological parameters of the model, including the PD‐1 and PD‐L1 target concentrations, receptor internalization rates, associated cell densities, and receptor‐ligand affinity, were informed by literature (see Table S1). A range of plausible tumor burdens with different PD‐1/PD‐L1 expression levels and cell densities were considered, giving six tumor parameterization scenarios (see Appendix S1). With the assumption that the number of tumor cells remained constant, PD‐L1 expression on tumor cells was varied from low (30%+) to high (90%+) and used to calculate the PD‐L1 concentrations and PD‐L1‐expressing tumor‐cell densities input into the model. The number of T cells was varied according to the desired effector‐to‐target‐ratio (E:T ratio) (from 1:10 to 1:100), and the number of PD‐1‐expressing T cells was determined using a nominal PD‐1 expression level on T cells from literature (see Appendix S1). The drug parameters of the model were set using published clinical and preclinical data for the different mAbs. Binding affinities for anti‐PD‐1 mAbs to PD‐1 and for anti‐PD‐L1 mAbs to PD‐L1 were set to values from corresponding meta‐analyses on PD‐1 or PD‐L1 binding [10, 11]. Drug‐specific PK parameters, specifically the measured linear elimination half‐lives, the peripheral distribution half‐lives, and the peripheral distribution partition coefficient, were set to literature‐derived values. Complete parameter tables for the model can be found in Table S1.

To determine whether the mAbs would achieve similar levels of PD‐1:PD‐L1 complex inhibition at their clinical doses, the antibodies of interest were simulated using each of the six different target parameterizations. Each of the antibodies was dosed at its clinical dosing regimen for 48 weeks to ensure PK steady state and Q2W/Q3W/Q4W cycle alignment (see Figure S1), and estimates of PD‐1:PD‐L1 complex inhibition were compared in the tumor compartment at the end of week 48.

To determine the impact of variability and sampling on the simulation conclusions, a virtual population (Vpop) was used to perform both sensitivity analysis and bootstrap analysis simulations. For Vpop details, see the Appendix S1.

4. Results

The model simulations for the eight mAbs at their respective clinical doses are shown in Figure 2. Figure 2a shows the predicted plasma PK profiles for the eight mAbs; the choice of tumor target parameterization has no effect on plasma PK. Figure 2b shows the site‐of‐action (tumor) PK profiles for the eight mAbs using each of the six tumor target burden parameterizations. At the highest PD‐1 target burden (E:T of 1:10), the model predicts a slight localized target‐mediated drug‐disposition (TMDD) effect for retifanlimab, where internalization of the drug:PD‐1 complex clears the antibody faster than it can be replenished from plasma distribution. This is due to the tighter affinity of retifanlimab for PD‐1 relative to the other anti‐PD‐1 mAbs, the relatively short half‐life of the PD‐1 receptor, and the high PD‐1 levels in the tumor at E:T of 1:10.

FIGURE 2.

Model simulations of the eight mAbs at their respective clinical doses. (a) Steady state plasma PK profiles for the eight mAbs. (b) Steady state tumor PK profiles; (c) Steady state PD‐1 or PD‐L1 target engagement in the tumor (% total receptor); and (d) Steady state PD‐1:PD‐L1 complex inhibition in the tumor (% baseline) for the eight mAbs in each of the six parameterizations.

Figure 2c shows the percent target engagement for each mAb, and in each target burden parameterization the model predicts > 97% target engagement to either PD‐1 or PD‐L1 receptors. This is consistent with previous clinical predictions summarized in Kasichayanula et al. [12], in which most of the mAbs simulated here were expected to achieve at least 90% receptor occupancy/target engagement at clinically active doses. To directly compare PD‐1 target engagement to PD‐L1, the PD‐1:PD‐L1 complex inhibition was calculated. In Figure 2d, the number of PD‐1:PD‐L1 complexes remaining following the addition of the antibody therapy (as % complex out of drug‐free baseline) was used as the metric for complex inhibition. For all the different mAbs (anti‐PD1 and anti‐PDL1) and considered target burden parameterizations, the model predicts > 96% complex inhibition (< 4% complexes remaining) at each clinical dose. Durvalumab, at its clinical dose of 1500 mg Q3W, sustains the highest level of PD‐1:PD‐L1 complex inhibition (~0.01% complexes remaining). The other two anti‐PD‐L1 mAbs, atezolizumab and avelumab, have the next highest inhibition (between 0.1% and 1% PD‐1:PD‐L1 complexes remaining), but not substantially more than most of the anti‐PD‐1 mAbs. Thus, simulations suggest that at clinical doses for each antibody, the anti‐PD‐L1 antibodies exhibit equal or more inhibition than the anti‐PD‐1 antibodies, despite their reported overall lower efficacy.

A sensitivity analysis was also performed on the model across the tumor parameterization scenarios using a Vpop approach. The Vpop model simulations of the eight mAbs at their respective clinical doses are shown in Figure 3. The anti‐PD‐L1 mAbs again sustain higher complex inhibition and show less variability than the anti‐PD‐1 mAbs, but all sustain relatively high levels of complex inhibition. The conclusions of this analysis are consistent with the results in Figure 2: even when accounting for parameter variability, the anti‐PD‐L1 mAbs exhibit higher PD‐1:PD‐L1 complex inhibition than the anti‐PD‐1 mAbs.

FIGURE 3.

Vpop simulations of the eight mAbs at their respective clinical doses. Ribbons indicate the lower and upper quartile profiles, with the central solid/dashed line as the median profile. (a) Steady state plasma PK profiles for the eight mAbs. (b) Steady state tumor PK profiles for the eight mAbs. (c) Steady state PD‐1:PD‐L1 complex inhibition in the tumor (% baseline) for the eight mAbs. (d) Boxplot of steady state trough PD‐1:PD‐L1 complex inhibition in the tumor (% baseline) for the eight mAbs.

The bootstrap analysis was performed using the model simulations for the pair of drugs with the closest PD‐1:PD‐L1 complex inhibition, avelumab 10 mg/kg Q2W (anti‐PD‐L1) and pembrolizumab 10 mg/kg Q3W (anti‐PD‐1). This corresponds to the Mirror 2 scenario (KEYNOTE‐010 vs. JAVELIN Lung 200) from the meta‐analysis by Duan et al. [2], which showed that patients treated with pembrolizumab had a 27% lower risk of death over time compared to avelumab (OS HR, 0.73; 95% CI, 0.57–0.95). Using the model Vpop simulations of the two drugs at these doses, 100 Vpatients were sampled out of the 500 total Vpatients (without replacement) 10 k times, and only 3.92% of samples showed higher median PD‐1:PD‐L1 complex inhibition for pembrolizumab compared to avelumab (Figure 4). Additional bootstrap analysis results are presented in the Appendix S1. Vpatient sampling supports the conclusions of the model that, based on the mechanism of PD‐1:PD‐L1 complex inhibition, anti‐PD‐L1 mAbs are predicted to perform better than anti‐PD‐1 mAbs.

FIGURE 4.

Bootstrap analysis of avelumab 10 mg/kg Q2W vs. pembrolizumab 10 mg/kg Q3W. (a) Vpop simulation of PD‐1:PD‐L1 complex inhibition in the tumor (% baseline) for the two mAbs. (b) Boxplot of trough PD‐1:PD‐L1 complex inhibition from the Vpop simulation. (c) Histogram of median PD‐1:PD‐L1 inhibition for each of the two mAbs in each bootstrap sample. (d) Histogram of the median differences in PD‐1:PD‐L1 inhibition between the two mAbs in each bootstrap sample. Only 3.92% of bootstrap samples show a median difference greater than zero (greater inhibition for pembrolizumab).

5. Impact Assessment

This modeling work is certainly not the first in this space; other QSP models and meta‐analyses have also compared PD‐1 and PD‐L1 binding antibodies [2, 3, 13]. However, in the QSP model of Bazzazi and Shahraz [13], tumor growth rates are fit to relevant patient data for each drug, so the difference in model‐predicted efficacy cannot be attributed to a difference in PD‐1:PD‐L1 complex inhibition. That QSP model's approach is top‐down rather than bottom‐up, with the model implicitly starting with the assumption that PD‐1 occupancy offers an advantage over PD‐L1 occupancy in tumor growth inhibition. On the other hand, in the meta‐analyses of Duan et al. [2] and Zhao et al. [3] the goal is to estimate the efficacy difference in the aggregated clinical data, not to draw conclusions about where those differences are derived at a mechanistic level.

This modeling work is unique in that no priors are inputted into the model for either class of antibody to perform better than the other. Instead, this QSP model tests whether there is a mechanistic explanation for the observed difference in anti‐PD‐1 vs. anti‐PD‐L1 efficacy that is based on PD‐1:PD‐L1 binding. Through integrating each drug's PK and binding parameters in a single QSP model with PD‐1 and PD‐L1 receptor dynamics, the model shows that anti‐PD‐1 and anti‐PD‐L1 mAbs are not differentiable based on PD‐1:PD‐L1 complex inhibition alone.

Other pathways or mechanisms not included in this model could differentiate the two classes of therapeutic antibodies. For example, the PD‐1:PD‐L2 axis is regulated by PD‐1 binding (Liu et al. 2021) [4] but not PD‐L1 binding, and recent research has suggested that blocking PD‐1:PD‐L2 interactions may be an integral part of anti‐PD‐1 treatment efficacy [14, 15]. Similarly, the PD‐L1:B7.2 axis is affected by PD‐L1 binding [4], but presumably not by PD‐1 binding. Additional resistance mechanisms to PD‐1/PD‐L1 inhibition (e.g., upregulation of PD‐L1 or tumor neoantigen expression [5]) could also have differential impacts on clinical outcomes. As part of a future study, these additional pathways and mechanisms could be added to the model, and this shared QSP model framework could be used to explore the relative strengths of the new mechanisms with respect to PD‐1:PD‐L1 complex inhibition.

By explicitly considering the implications of the first‐pass assumption—that PD‐1:PD‐L1 complex inhibition drives the efficacy differences between anti‐PD‐1 and anti‐PD‐L1 antibodies—this modeling work lays the foundation for future research efforts. There is an observable difference in clinical outcomes across the two antibody classes, but this model suggests that this difference cannot be attributed solely to complex inhibition. Research that isolates any of the additional mechanisms discussed above (or others) could further delineate the source of the observed efficacy difference.

Future research should also continue to integrate QSP approaches such as this with statistical model‐based meta‐analysis (MBMA) methods, in line with International Society of Pharmacometrics (ISoP) cross‐special interest group (SIG) goals [16]. The construction of this QSP model with literature‐informed hypothesis and parameterization can inform the search for relevant covariates in future MBMA. Beyond PK‐ or exposure‐related factors, which are largely nonmechanistic, variability in target expression and signaling context (e.g., PD‐1 and PD‐L1 expression levels or T‐cell infiltration) may be the most informative for differentiating efficacy across antibody classes. A systematic approach could therefore involve incorporating available data on PD‐1/PD‐L1 pathway components or tumor immune context. We suggest a focus on other components of the PD‐1/PD‐L1 signaling axis since our differential is the signaling level—we assume comparable tumor and immune microenvironments across molecules and scenarios, which is also the basic assumption of any MBMA.

6. Conclusion

According to the model, high levels of PD‐1:PD‐L1 complex inhibition are achieved by the approved molecules at their current dosing regimens. The model predicts that anti‐PD‐L1 mAbs achieve slightly higher complex inhibition than anti‐PD‐1 mAbs, though whether this effect would lead to differential clinical outcomes is unclear. The model predictions of high inhibition may reflect that clinical doses are selected to achieve high inhibition regardless of patient variability in PD‐L1 expression and E:T ratio, though other considerations for dose selection are also likely (e.g., overcoming ADAs, safety, or population variability). Even with significant parameter variability, the model does not indicate that anti‐PD‐1 mAbs at the clinical doses yield higher complex inhibition, and hence higher efficacy, over anti‐PD‐L1s, in contrast to the meta‐analyses of Duan et al. [2] and Zhao et al. [3]. Bootstrap sampling analysis mirroring the comparison from Duan et al. [2] does not change this conclusion. This analysis suggests that PD‐1:PD‐L1 complex inhibition may not be a predictive surrogate for efficacy, since it does not account for the differences in reported efficacy for anti‐PD‐1 vs. anti‐PD‐L1 antibodies. Overall, this modeling work suggests that anti‐PD‐1 and anti‐PD‐L1 mAbs are not differentiable based on PD‐1:PD‐L1 complex inhibition alone, and that the hypothesized shared mechanism of action of the drug classes is incomplete.

Author Contributions

C.L.J. wrote the manuscript. C.L.J. and G.I.K. designed the research. C.L.J., S.A.H., D.A.F., D.H.M., and D.F. performed the research. C.L.J. analyzed the data. D.F., A.M., J.M.B., and J.F.A. contributed new analytical tools.

Funding

The authors have nothing to report.

Conflicts of Interest

C.L.J., D.A.F., S.A.H., D.F., A.M., D.H.M., J.F.A., G.I.K. are all employees of Certara Predictive Technologies and may hold shares in the company. John Burke was an employee at Certara at the time of this work.

Supporting information

Table S1: psp470195‐sup‐0001‐TableS1.xlsx.

Appendix S1: psp470195‐sup‐0002‐AppendixS1.docx.

Johnson C. L., Flusberg D. A., Head S. A., et al., “Anti‐PD‐(L)1 Antibodies: Insights From QSP‐Based Meta‐Analysis,” CPT: Pharmacometrics & Systems Pharmacology 15, no. 2 (2026): e70195, 10.1002/psp4.70195.