Mitigating the risk of cytokine release syndrome in a Phase I trial of CD20/CD3 bispecific antibody mosunetuzumab in NHL: impact of translational system modeling

Mosunetuzumab, a T-cell dependent bispecific antibody that binds CD3 and CD20 to drive T-cell mediated B-cell killing, is currently being tested in non-Hodgkin lymphoma. However, potent immune stimulation with T-cell directed therapies poses the risk of cytokine release syndrome, potentially limiting dose and utility. To understand mechanisms behind safety and efficacy and explore safety mitigation strategies, we developed a novel mechanistic model of immune and antitumor responses to the T-cell bispecifics (mosunetuzumab and blinatumomab), including the dynamics of B- and T-lymphocytes in circulation, lymphoid tissues, and tumor. The model was developed and validated using mosunetuzumab nonclinical and blinatumomab clinical data. Simulations delineated mechanisms contributing to observed cell and cytokine (IL6) dynamics and predicted that initial step-fractionated dosing limits systemic T-cell activation and cytokine release without compromising tumor response. These results supported a change to a step-fractionated treatment schedule of mosunetuzumab in the ongoing Phase I clinical trial, enabling safer administration of higher doses.

. e$PnoV M . 5 %& , and is assumed to result in 1 activated and 1 post-activated daughter cell, to allow for some serial activity. Thus, each T-cell proliferation event maintains the activated population (no change) and increases the post-activated T-cell number. For traffic, and interstate conversion, each arrow shows whether the corresponding reaction is a unidirectional or bidirectional reaction, e.g. for 5: %& , deactivation to $ %& is unidirectional whereas activation to  In the absence of the drug, CD8+ T-cells in all the tissues are at equilibrium. When the drug or vehicle (placebo) is injected, a very rapid decline in the number of T-cells in the peripheral blood is observed followed by a gradual recovery. This phenomenon of T-cell margination is presumably due to endothelial adhesion/transmigration of PB T-cells that are transiently stressed or partially-activated by injection. This post-treatment margination, adhesion, and efflux effect has been observed with other therapies such as IL2 and IL15 1-3 . In the model, this is represented semi-mechanistically by transient post-injection augmentation of cell trafficking from PB to tissue with the terms ovw . ovw and y$Nz . y$Nz , which respectively represent injection-and drug-induced margination signals that peak immediately upon injection to relative magnitudes ovw and y$Nz ( ovw ≪ y$Nz ) and decline exponentially thereafter ( ovw = y$Nz = exp (− )).

Traffic of resting T-cells from peripheral blood to spleen:
All parameters are tuned to match observed placebo and drug-induced T-cell margination (Error! Reference source not found.A-C and Supplementary Figure 2A; see Supplementary in which the parameter represents the maximal activation signal.  in which the parameter represents the maximal killing rate.

Pharmacokinetics (PK) of mosunetuzumab and blinatumomab
A standard two-compartment PK model with non-specific clearance (captured by a linear term) was used to describe the measured concentration of mosunetuzumab in serum in cynomolgus monkeys as shown below: compartments. The units for volume and clearance terms are mL/kg and mL/day/kg, respectively.
‚ denotes the drug concentration in the central compartment; the concentration of mosunetuzumab in other tissues is assumed to be ‚ multipled by a tissue-specific partition coefficient with a value lower than one 5 . Note that we used the actual PK measurements as the forcing function when we performed model calibration and generated virtual cohort of cynos. The PK model with allometric scaling (exponent for CL = 0.85) was only used to project the human PK profiles and predict the results in patients.
For the PK of blinatumomab, we used the model in 6 , where blinatumomab is administered as continuous infusion to patients and is rapidly cleared from circulation: . 1000 ‚ (Nz) and ‚ denote the systemic amount (ug) and concentration (ng/mL) of the drug, respectively, and we used the same tissue-specific partition coefficients that we used for mosunetuzumab to estimate the tissue concentration of blinatumomab.

Cytokine dynamics
In the model, cytokine production is assumed to be the consequence of T-cell activation. In the presence of blinatumomab or mosunetuzumab, cytokines are produced by (or induced by) locally activated T-cells that receive an instantaneous local cytokine production signal, which is determined by drug concentration and decay in the number of target cells. For example, the PB IL6 levels is governed by following equation. where ro´2 F³ represents contribution of tissue IL6 to the measured IL6 levels.

Calibration of a Reference Virtual Cynomolgus Monkey
We used a scatter search methodology to calibrate adjustable model parameters to match the preclinical measurements in cynos. We defined the objective function as the sum of the difference between simulation results and the preclinical data for circulating CD8+ T-cell numbers, B-cell numbers, and percentage of CD8+CD69+ T-cells and limited B:T ratios in spleen and lymph nodes (only one data point after the last dose) across all animals in the three dosing groups as well as the placebo group. The outcome of the calibration is a single parameter set that best described the observed data, and the model structure together with this parameter set is termed the "reference virtual cyno". The search space for parameter values was restricted to biologically relevant ranges and consistent with the in vitro data and previously published measurements (see Supplementary Table 2 for list of parameters, values and references). In our optimization, we used the pre-treatment levels of B-cells and CD8+ T-cells for each individual animal to set the initial conditions. Due to large variability in the data and additional impact of ADA on PK profiles, we used the actual PK data (Supplementary Figure 1) for each animal as a forcing function to calibrate the model. For ADA-positive data points, we used the measured concentration if it was higher than the level of quantification (LOQ) for the PK assay; for data points below LOQ, the concentration was estimated using the estimated slope of the PK profile.

Generation of a Virtual Human ALL Patients
The model was adjusted to represent patients with acute lymphoblastic leukemia. We first replaced cyno physiological parameters with the appropriate human physiological volumes and T and B-cell numbers for peripheral blood and lymphoid tissues, preserving the model structure and other parameter estimates (See Supplementary Table 2 for human numbers). We then added representations of the pharmacokinetics of blinatumomab and its mechanistic effects including activation of CD8+ T lymphocytes, and subsequent killing of target CD19+ B lymphocytes. For blinatumomab PD, published data provided an estimate of concentration-dependent potency and suggested a similar quantitative cell-ratio dependence as was seen for mosunetuzumab [7][8][9][10] .

Generation of a Virtual Cohort of Cynomolgus Monkeys
To capture the biological variability observed in preclinical data, we generated a cohort of virtual cynos using the method introduced in 11 , with three main steps: (1) subject generation: to explore uncertainty by generating parameter sets locally randomized around the initial fit (reference virtual cyno); (2) subject selection: to select those virtual cynos whose simulated profiles lie within the range of preclinical data for circulating and tissue measurements at all dose levels.
Additionally, we selected the virtual cynos independent of the initial values for circulating B and T-cell levels and used the average PK concentration for a given dose level, so that for predictions, virtual cynos can be used for different dose levels and regimens (e.g. PK profiles) and a wide range of initial conditions for circulating B and T-cells; and (3) subject validation: to validate the virtual cohort of cynos against another preclinical dataset, not previously used for training.

Generation of a Virtual Population of NHL Patients
To show mean +/-standard deviation. In our optimization, we used the measured PK for each animal as a forcing function to calibrate the model. This was done by a linear interpolation of log-transformed PK data points measured for each animal. For ADA-positive data points, we used the measured concentration if it was higher than the level of quantification (LOQ) for the PK assay; for data points below LOQ, the concentration was estimated using the estimated slope of the PK profile. Red curves are the average of reconstructed PK profiles for each dose group.

Multiple Dose Study of Control Vehicle in Cynomolgus Monkeys (Preclinical Study 1)
The reference virtual cyno model replicates T and B-cell dynamics in the peripheral blood for the