PD‐1‐Cis IL‐2R Agonism Determines the Predicted Pharmacological Dose Range for the Immunocytokine Eciskafusp Alfa (PD1‐IL2v)

ABSTRACT

Binding in cis‐configuration to the PD‐1 receptor (PD‐1) and IL‐2$\mathit{\beta \gamma }$ receptor (IL‐2R$\mathit{\beta \gamma }$) has been shown to lead to differentiation of CD8 T cells to better effectors, which is anticipated to drive efficacy of the immune‐targeted cytokine eciskafusp alfa, or PD1‐IL2v. Here we present a geometrically driven mathematical formulation informed by in vitro and early clinical data which enables prediction of doses at which cis‐binding is at its highest, and explains observed differences in concentration‐time profiles for patients who had recent exposure to other anti‐PD‐1 molecules compared with those who had not. Furthermore, binding in cis‐configuration is expected to follow a “bell‐shaped” relationship with drug concentration such that high concentrations may lead to reduced benefit/risk ratio compared with concentrations around the peak of the bell‐shape. Model simulations identify patient cohorts for whom the upper limit of the pharmacological dosing range may be defined by either undesirable off‐tumor target engagement or a decrease in on‐tumor cis‐binding.

Study Highlights.

• What is the current knowledge on the topic?

  • ○
    Current knowledge indicates that stimulating cytotoxic CD8 T cells is key to cancer immunotherapy, and differentiating stem‐like CD8 T cells into better effectors can be achieved via IL‐2 receptor stimulation. However, stimulating the IL‐2R$\alpha$ chain on regulatory T cells may counteract the effect. Cis‐binding of PD1‐IL2v to the PD‐1 and IL‐2$\mathit{\beta \gamma }$ receptors has been shown to differentiate resource CD8 T cells while avoiding CD25 stimulation.
• What question did this study address?

  • ○
    This study addressed the question of how PD‐1 expression affects cis‐binding of PD1‐IL2v, the impact of prior anti‐PD‐1 CPI treatment on its pharmacokinetics and mode of action, and the importance of systemic IL‐2 binding for determining a pharmacologically active dose range.
• What does this study add to our knowledge?

  • ○
    This study adds a mechanistic model describing PD1‐IL2v cis‐binding, incorporating TMDD mechanisms, pharmacokinetics, and tumor uptake, using in vitro and early clinical data. Model simulations predict a bell‐shaped dose response for cis‐binding, non‐linear pharmacokinetics, and explain differences in concentration‐time profiles between anti‐PD‐1 experienced and naive patients.
• How might this change clinical pharmacology or translational science?

  • ○
    This may change clinical pharmacology by providing a framework to predict intra‐tumoral CD8 binding, inform dose optimization, and consider the effect of prior CPI treatment on PD1‐IL2v efficacy and safety. It also emphasizes the importance of considering cis‐binding in dose optimization of cis‐targeting compounds.

1. Introduction

Stimulating and expanding tumor‐specific cytotoxic CD8 T cells in the tumor microenvironment is central to cancer immunotherapy. In particular, differentiation of stem‐like CD8 T cells into “better effectors” is achieved via binding of all three of the interleukin‐2 receptor subunits (IL‐2R$\mathit{\alpha \beta \gamma }$) [1]. However, binding to the IL‐2R$\alpha$ chain (also known as CD25) which is also expressed on immunosuppressive T regulatory (Treg) cells and endothelial cells may reduce the overall tumor‐suppressive effect of CD8 effector cells and increase dose‐related toxicity. Although binding to only the $\beta$ and $\gamma$ subunits of the IL‐2 receptor is not sufficient to drive the “better effector” phenotype, binding in cis‐configuration to the PD‐1 receptor (PD‐1) and IL‐2$\mathit{\beta \gamma }$ receptor (IL‐2R$\mathit{\beta \gamma }$) has been shown to lead to differentiation of stem‐like CD8 T cells to better effectors while avoiding CD25 binding completely [2]. To this end, the antibody‐cytokine fusion protein, Eciskafusp alfa (PD1‐IL2v), has been developed such that the PD‐1 antibody delivers IL‐2R$\mathit{\beta \gamma }$ agonism to PD‐1 expressing T cells via the fused monomeric IL‐2 variant molecule, IL‐2v, providing cis‐binding that has been reported to generate the better effector population. Moreover, it is known that tumor‐specific CD8 T cells inside the tumor microenvironment express PD‐1 at higher frequencies and receptor densities than the CD8 cells in the blood [2], and the PD‐1 “docking” design of PD1‐IL2v exploits the disparity in PD‐1 receptor expression as well as the avidity effect of three binding sites. The theoretical consequence is that higher IL‐2R$\mathit{\beta \gamma }$ stimulation is delivered to the CD8 cells inside the tumor microenvironment than those in the blood and peripheral organs.

The geometry and dimensionality of the PD1‐IL2v molecule play a key role in its binding to PD‐1 and IL‐2R$\mathit{\beta \gamma }$ and hence its efficacy via cis‐binding. The dependence of the mode of action on cis‐binding means that a bell‐shaped dose response is expected, similar to other bispecific molecules [3]. High concentrations of the molecule lead to saturation of PD‐1 and IL‐2R$\mathit{\beta \gamma }$ by monovalently bound molecules, and therefore a mathematical formulation is required to determine not only the minimum, but also the maximum expected pharmacologically active dose. By combining a mathematical, geometrically‐driven binding model representing receptor binding at all three of the molecule's binding sites (based on the model for bispecific antibody binding presented by Kaufman et al. [4]) with a target mediated drug disposition pharmacokinetic (TMDD‐PK) model and a tumor uptake model, a mathematical formulation is invoked to investigate the expected pharmacologically active dose range in patients. Furthermore, an NK‐sparing approach is considered wherein PD‐1 expressing CD8 cells are maximally stimulated whereas stimulation of non‐PD‐1 expressing natural killer (NK) cells in the blood and tissues is kept to a minimum. This approach considers a mode‐of‐action driven safety factor, which may complement traditional methods to safety characterization.

Checkpoint inhibitor (CPI) molecules are the standard of care for several solid tumor indications [5]. PD1‐IL2v entered phase 1 clinical trials in 2020 for patients with inflamed solid tumor types likely to respond to checkpoint inhibitor treatment and/or IL‐2 treatment (NCT04303858). Consequently, the patient population is comprised of patients who have previously been exposed to anti‐PD1 (aPD‐1) treatment and patients who have not. Since two of the most widely used CPI treatments (pembrolizumab and nivolumab) also occupy close or overlapping epitopes on PD‐1 receptors, there may be competition for binding to the PD‐1 receptor in these cases. A mechanistic model formulation is particularly useful to compare the expected PD‐1 and IL‐2R$\mathit{\beta \gamma }$ receptor occupancy in patients who have had recent exposure to aPD‐1 therapy with those who have not.

In this work, we present a mechanistic model to describe the cis‐binding of PD1‐IL2v to PD‐1 and IL‐2R$\mathit{\beta \gamma }$ that incorporates TMDD mechanisms, pharmacokinetics and tumor uptake. We estimate key model parameters using in vitro and early clinical data only, such that no animal data were used to train the model. We subsequently perform model simulations to investigate three key questions related to treatment with PD1‐IL2v:

1. How does PD‐1 expression affect mechanistic binding of PD1‐IL2v to its target cells?

2. How does prior aPD‐1 treatment affect the pharmacokinetics and mode of action of the molecule?

3. Which dose range is appropriate for maximizing the PD‐1 mediated IL‐2v delivery while reducing IL‐2R stimulation on non‐target immune cells?

2. Methods

2.1. Model Development

The PD1‐IL2v molecule is composed of two PD‐1 binders and one binder specific for the $\beta$‐ and $\gamma$‐subunits of the IL‐2 receptor, which we refer to as the IL‐2v binder. The mathematical formulation in the presented model specifically includes: the free antibody (Ab), unbound PD‐1 ($R_P$) and IL2 ($R_I$) receptors, and 5 possible configurations for antibody‐receptor complexes, as illustrated in Figure 1.

FIGURE 1.

Schematic diagram to illustrate the binding of PD1‐IL2v to PD‐1 and IL‐2 surface receptors and to introduce notation for the mathematical species to be modeled. $R_P$: Unbound PD‐1 receptors, $R_I$: Unbound IL‐2 receptors, $\mathit{Ab}$: Free antibody, $X_P$: Antibody‐PD1 complex, $X_I$: Antibody‐IL2R complex, $X_{P1I}$: Antibody‐PD1‐IL2R complex, $X_{P2}$: Antibody‐PD1‐PD1 complex, $X_{P2I}$: Antibody‐PD1‐PD1‐IL2R complex. Inset: Schematic illustration of the geometric considerations for secondary and tertiary binding events.

2.1.1. Binding Model

The combinatorics of the three‐binder system result in 7 possible reversible reactions which can take place between the molecule and its receptors, as visualized in Figure 2A. The mathematical model for receptor binding takes into account all possible binding configurations of the molecule and its receptors and comprises eight ordinary differential equations (ODEs) which are given in full in the Data S1.

FIGURE 2.

(A) Schematic diagram summarizing the binding events included in the binding model. (B, C) IL‐2 receptor occupancy (B) and cis‐bound IL‐2 receptor occupancy (C) simulated using the binding model depicted in A for high PD‐1 (20,000 receptors per cell) and low PD‐1 (500 receptors per cell) expressing cells. (D, E) Observed IL‐2 receptor occupancy data (filled circles) from in vitro experiments using fresh (D) and 3 days activated (E) T cells, and PD1‐IL2v (red) and a non‐PD1 targeted molecule (black). Open circles represent simulated values after model fitting for corresponding experimental conditions. (F, G) Observed PD‐1 receptor occupancy data (filled circles) from in vitro experiments using fresh (F) and 3 days activated (G) T cells with PD1‐IL2v. Open circles represent simulated values after model fitting for corresponding experimental conditions.

We assume that, before the first binding event for a given molecule, the antibody and its receptors exist in a well‐mixed liquid, where the concentration of receptors available is given by the number of receptors per cell multiplied by the number of cells per liter. For subsequent binding events, the effective concentration of available receptors is determined by the number of receptors within reach of the cell‐surface bound antibody. As detailed in the work of Kaufman et al. [4], the effective number of receptors available for further binding events is modulated by the ratio of the average number of available receptors on the cell surface and the average number of receptors available in the hemispheric volume swept out by the bound antibody, as illustrated in the inset of Figure 1. This factor can be written as

$$\begin{array}{c}{A}_{\mathit{XY}}=\frac{3}{2\left[\text{cells}\mathrm{per}\mathrm{L}\right]\left[\text{cell surface area}\right]\left[\mathrm{X}‐\mathrm{Y}\text{inter}‐\text{binder distance}\right]}\hfill \end{array}$$ (1)

which we absorb into the on‐rates of cis‐binding events by considering, for example, in the case one PD‐1 binder is bound, the rate of binding of the IL2v binder is given by

$$k_{\mathit{\text{onPeffI}}}=A_v\times A_{\mathit{PI}}\times k_{\mathit{onI}}$$ (2)

where $A_v$ is a constant by which the geometrical factor may be modulated, $A_{\mathit{PI}}$ is the value of $A_{\mathit{XY}}$ from the formula above where the inter‐binder distance between the PD‐1 and IL‐2v binders is taken into account and $k_{\mathit{onI}}$ is the on‐rate for IL‐2v binding without avidity. Similar formulas apply for all other secondary and tertiary binding events. For full details of the mathematical model, the parameter values therein, and how the model is adapted to represent non‐PD‐1 expressing cells, refer to the Data S1.

Cis binding is defined as the total concentration of molecules occupying an IL‐2 receptor and at least one PD‐1 receptor at the same time, as given by the following expression:

$$X_{\mathit{cis}}=X_{P1I}+X_{P2I}$$ (3)

The receptor occupancy of cis bound molecules as a percentage is therefore expressed as

$${\mathit{RO}}_{\mathit{cis}}=100\times \frac{X_{\mathit{cis}}}{R_I+X_I+X_{\mathit{cis}}}$$ (4)

2.1.2. Pharmacokinetics and Target Mediated Drug Disposition

The pharmacokinetic properties of PD1‐IL2v were assumed to be similar to the CEA‐targeted IL‐2v molecule, cergutuzumab amunaleukin, which have been reported in [6]. The pharmacokinetic model that was used to describe the target‐mediated drug disposition (TMDD) of this molecule includes binding to IL‐2R$\mathit{\beta \gamma }$, as well as internalization of this receptor and the molecule, upregulation of the receptor and natural turnover or receptors. Parameters that describe the IL‐2 receptor dynamics were assumed to be the same as those reported in [6]. Extending this model to include the PD‐1 binding kinetics, the receptor binding portion of the TMDD model was replaced with the full receptor occupancy model described above, as illustrated in Figure 3A (see Data S1 for the full list of equations). As shown in [2], complexes comprised of PD1‐IL2v, IL‐2R$\mathit{\beta \gamma }$ and PD‐1 are internalized. In order to capture this in the mathematical model, it was assumed that the internalization process of IL‐2 cis‐bound complexes occurs at the same rate as the internalization of bound IL‐2 receptors, and that PD‐1 receptors are recycled to the cell surface, but not upregulated.

FIGURE 3.

A: Schematic diagram illustrating the target mediated drug disposition extension to the binding model illustrated in Figure 2A. The model includes IL‐2R turnover, internalization of IL‐2R‐bound complexes $X_I,X_{P1I}\text{and}{X}_{P2I}$ and recycling of internalized IL‐2R and PD‐1. (B) Schematic diagram illustrating the pharmacokinetic and tumor uptake model extensions to the binding and TMDD models shown in Figures 2A and 3A. Doses of PD1‐IL2v are assumed to be delivered into the peripheral blood, accompanied by a first‐order clearance term. The free antibody is assumed to move between the peripheral blood and the tumor compartment. (C–F) Simulated pharmacokinetics in the blood (C), IL2 cis binding in blood (D), pharmacokinetics in the tumor (E) and IL2 cis binding in the tumor (F) of 3 one‐week treatment cycles for doses of 0.1, 1 and 10 mg using the full model and pharmacokinetic parameter values estimated from clinical data. Inter‐individual variability was incorporated by simulating 100 virtual patients. The medians are shown by solid lines and 50% and 90% confidence intervals are shown by the shaded areas.

2.1.3. Tumor Uptake

The model was extended to incorporate a tumor compartment in order to make predictions regarding the pharmacological activity. To this end, the model described in [7], which has also been used to describe the tumor uptake of cergutuzumab amunaleukin [6] was implemented. The receptor occupancy model was assumed to describe the binding in the tumor compartment. Since cellular dynamics (e.g., margination and expansion of T cells) are not within the scope of this model, we include only the uptake of the free antibody into the tumor environment, and not the antibody‐receptor complexes. The tumor uptake model is visualized in Figure 3B. The full model incorporating all of the aforementioned processes is summarized by the components depicted in the schematic diagrams in Figures 2A and 3A,B.

2.1.4. Simulations of CD8 and NK Cells

For simulations of PD1‐IL2v binding to both PD‐1 and non‐PD‐1 expressing cells, two cell types were represented in the full model. CD8 cells were assumed to express both PD‐1 and IL‐2R, whereas NK cells were assumed to express IL‐2R but not PD‐1. The binding kinetics and TMDD terms are expected to be shared among both cell types, and the absence of PD‐1 receptors on NK cells is explicitly taken into account for model simulations. Parameter values used to represent each cell type are provided in the Data S1.

2.1.5. Simulations of the Full Model

The pharmacologically active dose range for cis‐binding molecules such as PD1‐IL2v is bounded below by the minimum pharmacologically active dose (minPAD) and above by either the maximum pharmacologically active dose (maxPAD) or the maximum tolerated dose (MTD). To predict the optimal dose range, simulations of the full model were performed under the following assumptions:

1. Efficacy of PD1‐IL2v is likely driven by a combination of both IL‐2R agonism and checkpoint blockade. Our model predicts that PD‐1 receptor occupancy is maximal from 1 mg upwards (see Data S1). Consequently, we assume that target doses can be identified by maximizing and sustaining cis‐binding with no beneficial gain in PD‐1 blockade by dosing higher.

2. A proxy for efficacy is the area under the curve (AUC) of cis‐binding of IL‐2R on CD8 cells in the tumor microenvironment, illustrated in the shaded region in the left panel of Figure 5A. A demonstrative threshold of 75% of maximum cis‐binding AUC was used to determine the dose range for maximum efficacy, defining both the minPAD and maxPAD.

3. Similarly, a proxy for safety is the AUC of IL‐2R binding on non‐PD‐1 expressing cells systemically (illustrated in the shaded region in the right panel of Figure 5A), such as NK cells. A demonstrative threshold of 75% of maximum systemic IL‐2R binding to non‐PD‐1 expressing cells was used to determine the lower limit for safety concerns or MTD.

FIGURE 5.

(A) Plots illustrating the calculation of the metrics related to the AUC of cis‐bound IL‐2 receptor occupancy on CD8 cells in the tumor (green shaded area) and the AUC of IL‐2 receptor occupancy on NK cells in the blood (orange shaded area) after a single dose of PD1‐IL2v. (B, C) Simulated pharmacologically active dose ranges for aPD‐1 CPI naïve (B) and aPD‐1 CPI experienced (C) patient cohorts. The AUC of cis‐bound IL‐2 receptor occupancy on CD8 cells in the tumor (normalized relative to the maximum value) is shown by the solid green curve, and the normalized AUC of IL‐2 receptor occupancy on NK cells in the blood is shown by the dashed orange curve. The green and orange dotted vertical lines mark the concentrations at which the efficacy and safety relevant thresholds are reached, respectively, which are both set to 75% of maximum. The range of doses that fall within the derived pharmacologically active range is shaded in yellow. (D) The simulated AUC of cis‐bound IL‐2R on CD8 cells in tumor for aPD‐1 CPI naïve (black) and aPD‐1 CPI experienced (red) patients. (E) The simulated AUC of cis‐bound IL‐2 receptors on high PD‐1 expressing CD8 cells in the tumor (orange) and low PD‐1 expressing CD8 cells in the blood (purple).

2.1.6. Nivolumab Model

To simulate PD1‐IL2v administration after recent Nivolumab treatment, the published clinical PK model [8] was combined with a mechanistic binding model taking into account reported binding affinities [9]. For more details on all model structures and parameter values, refer to the Data S1. Unless specified otherwise, simulations of aPD‐1 CPI experienced patients consider the scenario where patients receive 12 cycles of 240 mg of nivolumab on a Q2W schedule, where their last dose of aPD1 CPI is 2 weeks prior to their first dose of PD1‐IL2v.

2.2. Clinical Data

Pharmacokinetic data from 30 patients included in Part One (single agent dose escalation) of the phase 1 clinical trial (NCT04303858) were used to estimate the model parameters.

2.3. Technical Specifications

Model simulations were performed in Matlab 2021b using the Simbiology toolbox. Monolix version 2023R1 was used to estimate the volume of distribution and clearance rate for the extended model.

2.3.1. Parameter Estimation

Parameters $k_{\mathit{onI}}$ and $k_{\mathit{onP}}$, representing the forward binding rate for the IL‐2v and PD‐1 binders respectively, as well as the avidity constant $A_v$, were estimated using in vitro data for which experimental conditions, including number of cells per microliter and receptors per cell, were known or quantified as described in the Data S1. The built‐in lsqnonlin function in Matlab was used to estimate the parameter values.

Parameter estimation for clinical data was performed in Monolix 2023R1 using the SAEM algorithm and a population approach. See Data S1 for further information.

3. Results

3.1. The Mechanistic Binding Model Predicts Preferential Cis‐Binding on High PD‐1 Expressing Cells

The interactions of PD1‐IL2v with its receptors, taking into account valency and all possible binding combinations, are represented by the binding model, as illustrated conceptually in Figure 1 and by the diagram in Figure 2A.

Since PD‐1/IL‐2R cis‐binding is not directly measured in receptor occupancy assays, the mathematical model is used to infer the proportion of occupied IL‐2 receptors and those engaged in cis‐binding, as depicted in Figure 2B,C. The model predicts that more cis‐binding occurs at lower drug concentrations on high PD‐1 expressing cells compared to lower PD‐1 expressing cells (Figure 2B). Moreover, the overall proportion of IL‐2 receptors involved in cis‐binding is significantly higher for high PD‐1 expressing cells (Figure 2C).

Simulations indicate that the proportion of IL‐2 receptors engaged in cis‐binding increases with PD1‐IL2v concentration, reaching a maximum before decreasing at higher concentrations, forming a bell‐shaped profile. At higher concentrations, although IL‐2 receptor occupancy may be high (Figure 2B), the cis‐binding effect diminishes since most molecules are monovalently bound to either PD‐1 or IL‐2R. Thus, the PD‐1 anchoring effect is lost at high drug concentrations. Consequently, the mechanism of action (MoA) of PD1‐IL2v may necessitate not only a minimum pharmacologically active dose (minPAD) but also a maximum pharmacologically active dose (maxPAD).

Data from in vitro receptor occupancy assays (see Data S1 for details) were used to estimate model parameters, including the binding rates of PD1‐IL2v to PD‐1 ($k_{\mathit{onP}}$) and IL‐2R ($k_{\mathit{onI}}$), as well as the avidity constant ($A_v$). Figure 2D–G show a close fit between the observed and simulated data for both IL‐2R and PD‐1 receptor occupancy.

The model predicts, in alignment with experimental observations, that PD‐1 docking of PD1‐IL2v leads to preferential IL‐2R binding compared to a non‐PD‐1 binding molecule, FAP‐IL2v, in both freshly isolated and 3‐day activated CD8 T cells (Figure 2D,E). Additionally, IL‐2R binding of PD1‐IL2v occurs with a lower EC50 in 3‐day activated CD8 cells, which express high levels of PD‐1 receptors, compared to freshly isolated CD8 cells.

3.2. The Extended Model Predicts Nonlinear Pharmacokinetics and Non‐Monotonic Increasing Association of IL‐2R Cis‐Binding With Dose

To translate the binding model to a clinically relevant scenario, the model was extended to include receptor internalization and turnover, pharmacokinetics, and tumor uptake, as shown in Figure 3A,B. Using a population approach informed by drug concentration data from Phase I study patients, we estimated the volume of distribution ($V$) to be 2.39 L and the elimination rate ($k_{\mathit{el}}$) of PD1‐IL2v to be 0.0197 $h^{-1}$. These values align with similar molecules; detailed parameter estimation can be found in the Data S1.

Simulations of PD1‐IL2v doses from 0.1 to 10 mg (within the range explored in Phase I study NCT04303858) indicate that the pharmacokinetics of the unbound molecule in blood (Figure 3C) and tumor (Figure 3E) follow the expected pattern, with TMDD effects most pronounced at the lowest simulated dose. The mechanistic model was used to predict IL‐2R cis‐binding at corresponding doses. A positive feedback loop is accounted for in the TMDD model where the treatment leads to upregulation of IL‐2R on target cells. The simulations indicate that IL‐2R cis‐binding in the blood is higher during the first cycle (one dose per cycle and cycle length 7 days) at a dose of 1 mg compared to 10 mg (Figure 3D). During the second and third cycles, IL‐2R cis‐binding is similar for doses of 0.1 mg and 1 mg, and comparatively less cis‐binding is expected at 10 mg. In the tumor (Figure 3F), at the end of the third cycle, more cis‐binding is predicted at the lowest dose of 0.1 mg compared to 1 and 10 mg.

The bell‐shaped relationship between drug concentration and IL‐2R cis‐binding (Figure 2C) explains these predictions. At high local concentrations, the molecule predominantly binds monovalently to IL‐2R and PD‐1. Thus, while IL‐2 receptor occupancy may be high at such concentrations, the proportion of IL‐2R engaged in a cis‐bound complex decreases.

3.3. Model Simulations Explain Differences in Concentration‐Time Profiles Between aPD‐1 CPI Experienced and aPD‐1 CPI Naïve Patients

The following simulation results concern patients who have current exposure to aPD‐1 CPI such that molecules are still present in the blood and tissues and those who have either had no exposure to aPD‐1 CPI or no exposure recently enough to consider detectable levels of the molecule in the blood or tissues (referred to hereafter as aPD1 CPI experienced/naïve).

For the lowest doses administered in the Phase I study NCT04303858, the effect of TMDD on the blood concentration of PD1‐IL2v was much more pronounced in pharmacokinetic profiles for patients who were aPD‐1 CPI naïve compared to those who were aPD‐1 CPI experienced. Cycle 1 concentration‐time profiles for three example aPD‐1 CPI naïve patients and one aPD‐1 CPI experienced patient, who had received their last dose of aPD‐1 treatment 33 days prior to the start of PD1‐IL2v treatment, are shown in Figure 4A. The aPD‐1 CPI naïve patients had undetectable concentrations of PD1‐IL2v in the blood within 2 days of the first 0.1 mg dose. The aPD‐1 CPI experienced patient, who also received a 0.1 mg dose of PD1‐IL2v, still had detectable levels of PD1‐IL2v in the blood 1 week after dosing. A similar trend is observed for patients receiving the 0.3 mg dose (Figure 4B).

FIGURE 4.

(A, B) Observed concentration‐time profiles from the first cycle of treatment of PD1‐IL2v from patients receiving 0.1 mg (A) or 0.3 mg (B) doses of PD1‐IL2v shown by open circles for patients who had not received prior anti‐PD1 CPI treatment (black) and patients who had received prior anti‐PD1 CPI treatment (red). Superimposed with thick solid lines are the corresponding simulated concentration‐time curves using the full model with inter‐individual variability generated by 100 virtual patients (90% confidence intervals shown). (C, D) Full model simulations of IL‐2 receptor occupancy (solid lines) and IL‐2 cis‐binding (dashed lines) after a single dose of PD1‐IL2v at 0.1 mg (C) or 0.3 mg (D) for patients who had not received prior aPD‐1 treatment (black) and patients who had received a course of Nivolumab where the last dose was 4 weeks prior to the simulated PD1‐IL2v dose (red). Inter‐individual variability is included as in A and B.

By superimposing simulations of the mechanistic model, including inter‐individual variability, with these concentration‐time profiles, it is clear that this behavior aligns with the expected model predictions. aPD‐1 CPI experienced patients have fewer PD‐1 receptors available to form cis‐bound complexes, leading to lower IL‐2R binding (Figure 4C,D) and therefore less TMDD, which is primarily driven by IL‐2R internalization. Thus, although the fast dynamics of PD1‐IL2v in the blood are observed, the cis‐ and total IL‐2R binding is predicted to be much higher for aPD‐1 CPI naïve patients compared to aPD‐1 CPI experienced patients at the lower dose levels. The model predicts sustained IL‐2 cis‐binding for the entire dosing interval for both the 0.1 and 0.3 mg doses in aPD‐1 CPI naïve patients, although this has not been experimentally verified.

3.4. Model Simulations Highlight Dosing Ranges That May Optimize Intra‐Tumor Cis‐Binding and Minimize Systemic Monovalent IL‐2R Binding

From our simulations, it is possible to derive summary metrics to represent efficacy and safety. As depicted in Figure 5A, we consider the AUC of cis‐binding in the tumor as a marker of efficacy and the AUC of total IL‐2 receptor occupancy as a marker of potential safety concerns such as systemic NK cell activation. By deriving these metrics for a broad range of doses, a pharmacologically active dose range can be identified, delimited by the minPAD on the lower end and either the maxPAD or the MTD on the upper end.

Figure 5B,C illustrate that the pharmacologically active dosing range differs between aPD‐1 CPI experienced and naïve patients. For aPD‐1 CPI naïve patients, the dose range is constrained by the maxPAD and includes the lowest dose simulated (0.1 mg), while for aPD‐1 CPI experienced patients, the range is narrower, includes higher doses, and is limited by the MTD. A sensitivity analysis of how the dose range depends on model parameter values is presented in Figure S5.

Since it is anticipated that the amount of cis‐bound IL‐2R drives efficacy, the absolute AUC of cis‐binding is shown in Figure 5D. Cis‐binding is predicted to be higher in aPD‐1 CPI naïve patients than in aPD‐1 CPI experienced patients. This aligns with the PD‐1 targeting of the molecule and the results presented in the previous section, as cis‐binding is limited when PD‐1 is occupied by another molecule.

Finally, as shown in Figure 5E, the model predicts a tumor targeting effect driven by higher PD‐1 expression on CD8 cells in the tumor for the entire dose range simulated. The differential in cis‐binding amount between the tumor and the blood is smaller for higher doses. These results indicate that higher doses may not lead to increased efficacy, and the PD‐1 targeting effect that drives differentiation of CD8 cells to better effectors in the tumor may be best exploited at intermediate dose levels. Furthermore, appropriate identification of the pharmacologically active dose range allows the anti‐tumor effect to be maximized while minimizing peripheral immune cell activation.

4. Discussion

The complex mode of action of PD1‐IL2v motivates the development of a mathematical model in order to predict the effects of the molecule which may not easily be measured, such as intra‐tumoral CD8 binding, and to inform clinical study design. In particular, it is anticipated that molecules which rely on cross‐linking to two or more targets for efficacy may exhibit a bell‐shaped dose response, which should be considered for dose optimization (and may contradict the traditional pharmacological dose escalation approach). The final model enables simulation of 3 key aspects of treatment with PD1‐IL2v:

1. Mechanistic binding of PD1‐IL2v to its target cells: high PD‐1 expressing CD8 cells in the tumor.

2. The effect of prior aPD‐1 CPI treatment on the pharmacokinetics and mode of action of the molecule.

3. Dose ranges to maximize PD‐1 mediated IL‐2v delivery while reducing IL‐2R stimulation on non‐target immune cells.

Simulations of IL‐2R and PD‐1 receptor occupancy in vitro support the hypotheses that PD‐1 docking enhances IL‐2v delivery to target cells and also achieves a higher potency on cells with high PD‐1 expression. Model simulations of cis‐bound IL‐2 receptor occupancy, the moiety responsible for driving efficacy, predict a bell‐shaped dose response with optimal formation at an intermediate concentration.

The model provides a mechanistic interpretation for the differences in clinical pharmacokinetics of PD1‐IL2v between patients who had and had not received prior aPD‐1 CPI treatment. Furthermore, the model predicts a stronger PD‐1 targeting effect in aPD‐1 naïve patients.

The model is used to explore dosing ranges which may simultaneously deliver optimal cis‐bound IL‐2 receptor occupancy, while constraining unwanted IL‐2R trans binding on non‐target cells such as NK cells. Due to the bell‐shaped nature of the relationship between concentration and cis‐binding, and in contrast to traditional approaches, the upper bound for the dosing range may be imposed either by reaching a maximum tolerated systemic IL‐2R binding (safety driven), or by a loss of cis‐bound IL‐2R in the tumor (efficacy driven). For given safety and efficacy thresholds, the determination of this upper bound may even be predicted to differ between patient cohorts or indications, depending on previous treatment, potentially resulting in a narrower pharmacologically active dose range in the case of prior aPD‐1 CPI treatment. Thus, the model may bring together all of the elements previously described to provide a holistic in silico approach to predicting clinically relevant dose ranges for relevant patient cohorts and indications, for example indications with particularly high PD‐1 expressing tumors. It is plausible that baseline PD‐1 expression may vary between CPI‐experienced and CPI‐naïve patients, potentially due to PD‐1 downregulation or inherently lower expression in CPI‐resistant (non‐responding) patients. This aspect was not explicitly accounted for in the analysis related to Figure 4. Further investigation would be a valuable extension of this analysis especially if PD‐1 expression data was available.

The translational approach was optimized using in vitro and early clinical data, and no animal data were used to train the model. To develop this in silico approach further, the model should be continuously refined by integrating more clinical and experimental data as they become available for precise parameter estimation. Substantial pharmacokinetic and pharmacodynamic inter‐individual variability would be expected, and this could be characterized using a mixed‐effects modeling approach. Pharmacometric extensions to the model may incorporate other important components such as anti‐drug antibodies or traditional individual covariates. Additional data would also enable determination of metrics (e.g., AUC, Cmax) and thresholds regarding drivers of efficacy and safety in collaboration with subject matter experts. As clinical data become available, the hypothesis that cis‐binding drives efficacy could be tested using a model‐informed approach. The model provides an integrated approach to therapeutic window prediction by incorporating biophysical binding processes. As such, a similar approach could be used in molecular design to optimize pharmacological properties and enhance the therapeutic window. This methodology may be applied to other cis‐binding molecules.

Author Contributions

L.G.H. and T.D.L. wrote the manuscript; L.G.H., T.D.L., L.L., M.M.‐F., M.M., V.T., and L.C.D. designed the research; L.G.H., T.D.L., L.L., M.M.‐F., M.M., V.T., and L.C.D. performed the research; L.G.H., T.D.L., L.L., and L.C.D. analyzed the data.

Conflicts of Interest

All authors are employees of Hoffman‐La Roche. L.C.D. and L.L. are named inventors on the patent application 15/943,237 which has relevance to this work. L.G.H., T.D.L., L.L., M.M.F., M.M., V.T., and L.C.D. declare ownership of Roche stock.

Supporting information

Data S1: psp470112‐sup‐0001‐Supinfo.pdf.