Full details of the MajesTEC-1 study have been described previously [13, 14]. Briefly, eligible patients had RRMM after three or more prior lines of therapy, including triple-class exposure to an immunomodulatory drug, a proteasome inhibitor, and an anti-CD38 monoclonal antibody, and had an Eastern Cooperative Oncology Group performance status score of 0 or 1. A total of 165 patients received subcutaneous teclistamab at the RP2D of 1.5 mg/kg QW preceded by SUDs of 0.06 mg/kg (SUD 1) and 0.3 mg/kg (SUD 2), with 2–4 days permitted between each SUD and the first treatment dose. Patients had the option to switch to 1.5 mg/kg Q2W if they experienced a partial response or better (≥ PR) after ≥ 4 treatment cycles in phase 1 of the study (cycle duration 21 days) or a ≥ CR for ≥ 6 months in phase 2 [7].

Prolonged dosing intervals occurred due to changes in dosing frequency (in phase 1), logistical or patient issues, or dose modifications due to adverse events. Repeat SUD was not mandated by protocol following prolonged delays for patients in phase 1 and, as such, use of repeat SUD in this phase varied. Guidance for repeat SUD dosing was added in subsequent protocol amendments for patients in phase 2.

The population PK model has been previously described and comprised a two‑compartment model with first-order absorption and parallel time‑independent and time-dependent elimination pathways. The time-independent clearance component is thought to reflect the endogenous catabolic processes of IgG degradation, whereas the time-dependent clearance component corresponds to the decrease in drug clearance as disease status improves over time, which may be related to tumor burden or target amount [10]. This analysis used serum concentration data from MajesTEC-1. The population PK model and a virtual population of 1000 patients generated from individual baseline demographics and disease-related characteristics representative of the MajesTEC-1 population were used to project concentration–time profiles at steady state following two dose delay scenarios. These dose delays were simulated to occur after the thirteenth 1.5 mg/kg QW dose and after the seventh 1.5 mg/kg Q2W dosing to represent steady state. These serum concentrations were then compared with the estimated median trough concentrations (Ctrough) associated with SUDs, which informed time windows for the retrospective analysis of clinical data related to CRS.

Considering that BCMA re‑expression and immune recovery following treatment interruption could restore target‑mediated drug disposition and increase teclistamab clearance, a sensitivity analysis of the population PK model predictions following prolonged dose interruptions was conducted. Model-based simulated PK profiles were generated for each participant treated at the RP2D who experienced prolonged dosing intervals (> 28 days; N = 21), using individual PK parameter estimates (post hoc) derived from the population PK model and each participant’s teclistamab dosing history. Subsequently, the individual model-simulated concentrations were graphically compared with observed PK data.

A QSP model based on the mechanism of action for teclistamab and calibrated by clinical data was used to simulate cytokine dynamics. This model has been used to support teclistamab QW to Q2W dose switching in patients who reached a sustained CR [12]. It contains a cytokine release module that tracks immune cells in three states: inactive cells that do not secrete cytokines but can be activated to become cytokine-secreting cells that actively secrete cytokines, including interleukin-6 (IL-6) and interleukin-10 (IL-10); and refractory cells that are post-activated or dysfunctional cells that cannot secrete cytokines. This structure was consistent with the core structure from previously published approaches [15, 16]. Following teclistamab administration and immune synapse formation, immune cells are activated to secrete cytokines; however, cytokine secreting cells are short-lived and transition into a post-activated state. Therefore, there is a reduction in the pool of inactive immune cells and a transient increase in cytokine secreting cells. This new state is less sensitive to further teclistamab stimulation because the number of cytokine secreting cells is limited, and there are fewer inactive immune cells to be activated. When teclistamab administration is delayed for a long time, there is no longer a driving force for immune activation; therefore, cytokine secreting cells are reduced but inactive immune cells are replenished. Consequently, the system is restored for its sensitivity to teclistamab-induced cytokines release.

The hypothetical “immune cells” in the model represent a collective of cells implicated in cytokine secretion, including T cells and other myeloid-derived cells, such as monocytes, macrophages, and neutrophils, in the blood or bone marrow, and their activation is expected to be driven directly or indirectly by the formation of teclistamab-mediated immune synapse. The site of action for MM (i.e., bone marrow) is highly permeable for cytokines [16, 17]. Therefore, the IL-6 and IL-10 profiles in the blood compartment, especially their peak concentrations, were characterized and used as the indicators of the risk of CRS. CRS risk was modeled to have a positive association with peak cytokine concentrations in line with observed clinical data. Time to CRS onset was not included in the model.

The key data used to inform the QSP model included longitudinal serum concentrations of IL-6 and IL-10, along with the occurrence and grade of CRS, and dosing records of teclistamab and tocilizumab, an approved IL-6 receptor antagonist used to treat CRS [18,19,20]. Data from patients in the phase 1 RP2D cohorts in MajesTEC-1 with intense cytokine sampling (n = 40) within the first 21 days of teclistamab treatment were used for calibration. Intersubject variability (± tenfold from the population mean value) was included for the baseline cytokine concentrations, as observed in the cytokine data. Sparse cytokine data from patients treated at the RP2D in phase 2 (n = 125) were used for model validation. Data from phase 1 MajesTEC-1 patients who received tocilizumab prophylactically (n = 18) were used to quantify the impact of tocilizumab on total IL-6 concentrations. The QSP model estimates the “active” (i.e., uninhibited) IL-6 concentration by adjusting the teclistamab-induced IL-6 concentration to account for both the impact of tocilizumab-induced increases in IL-6 and inhibition of IL-6 by tocilizumab in patients who received tocilizumab (n = 16 in phase 1 and n = 52 in phase 2).

The virtual population parameters calibrated with the observed clinical responses of teclistamab in patients with RRMM were maintained as previously described [12]. Additional parameters from the cytokine module were calibrated using a Markov Chain Monte Carlo parameter estimation method with the observed longitudinal cytokine data to minimize the differences (i.e., sum of squared error) between observed and model estimated values.

Following model calibration, the QSP model was employed to simulate IL-6 and IL-10 concentrations in virtual patients who achieved and sustained a ≥ PR at steady state with either 1.5 mg/kg QW or 1.5 mg/kg Q2W dosing. In the simulated scenarios, dosing delays occurred at steady state with either QW or Q2W dosing, and teclistamab dosing was restarted either at SUD 2 following a 111-day delay or at 1.5 mg/kg after a 62-day delay. For each dose delay–restart scenario, virtual patients were stratified into two subgroups based on whether disease progression occurred during the dosing interruption.

Population PK predictions informed the time windows applied to a retrospective clinical analysis of MajesTEC-1 (data cut-off: August 22, 2023) to evaluate repeat SUD and incidence of CRS events in the setting of prolonged dosing intervals (> 28–62 days, ≥ 63–111 days, and ≥ 112 days). CRS events were graded according to the American Society for Transplantation and Cellular Therapy criteria [21].

The population PK analysis was performed using NONMEM (version 7.4.3, ICON Development Solutions, LLC, Ellicott City, MD, USA). Post-processing of the simulated outputs was performed using R software (version 3.4.1, Comprehensive R Network, http://cran.r-project.org/) [22]. The QSP model was implemented in MATLAB using the Simbiology toolbox (version 6.0, MATLAB version 2022b, Mathworks, Inc., Natick, MA, USA).