A Novel Two-Part Mixture Model for the Incidence and Time Course of Cytokine Release Syndrome After Elranatamab Dosing in Multiple Myeloma Patients

Abstract

Cytokine release syndrome (CRS) is a common, acute adverse event associated with T-cell redirecting therapies such as bispecific antibodies (BsAbs). The nature of CRS events data makes it challenging to capture an unbiased exposure-response relationship with commonly used models. For example, simple logistic regression models cannot handle traditional time-varying exposure, and static exposure metrics chosen at early time points and with lower priming doses may underestimate the incidence of CRS. Therefore, more advanced modeling techniques are needed to adequately describe the time course of BsAb-induced CRS. Herein, we present a two-part mixture model that describes the population incidence and time course of CRS following various dose-priming regimens of elranatamab, a humanized BsAb that targets the B-cell maturation antigen on myeloma cells and CD3 on T cells, where the conditional time-evolution of CRS was described with a two-state (i.e., CRS-yes or no) Markov model. In the first part, increasing elranatamab exposure (maximum elranatamab concentration at first CRS event time (C_max,event)) was associated with an increased CRS incidence probability. Similarly, in the second part, increased early elranatamab exposure (C_max,D1) increased the predicted probability of CRS over time, whereas premedication including corticosteroids and IL-6 pathway inhibitors use demonstrated the opposite effect. This is the first reported application of a Markov model to describe the probability of CRS following BsAb therapy, and it successfully explained differences between different dose-priming regimens via clinically relevant covariates. This approach may be useful for the future clinical development of BsAbs.

Figure 1

Two‐part mixture model of CRS. The incidence of CRS is described by a logistic regression model where the probability that a patient will have at least one CRS event is dependent upon a baseline probability (or “intercept,” β₀) as well as covariates, including, but not limited to, drug exposure (or dose). In patients having at least one CRS event, a two‐state Markov model (i.e., CRS‐yes; 1, or no; 0) describes the evolution of CRS events over discrete‐time intervals where the probability to transition between states, or to stay within a state, is estimated from the observed transitions within each patient.

Figure 2

Population‐predicted incidence of CRS over the observed range of C _max,event. The open circles and whiskers represent the observed proportions and 95% CIs of grade ≥1 CRS at the median observed C _max,event values for each of the dose‐priming regimens. The solid blue line and ribbon represent the predicted probabilities and 95% CI of grade ≥1 CRS, respectively. The rug lines indicate the C _max,event values for those patients who did (top) and did not (bottom) experience grade ≥1 CRS. The dosing of these amounts occurred on Days 1, 4, and 8.

Figure 3

Population‐averaged joint‐predicted CRS probabilities over time. Open circles and whiskers represent the observed fraction and 95% CI of grade ≥1 CRS, respectively. The number of patients contributing to the observed summaries per dose‐priming regimen are given in the panel titles. Solid lines and shaded regions represent the median and 95% CI of the predictions of grade ≥1 CRS, respectively. The dosing of these amounts occurred on Days 1, 4, and 8.

Figure 4

Population‐averaged joint‐predicted CRS transition probabilities over time. Open circles and whiskers represent the observed fraction and 95% CI of each transition possibility for grade ≥1 CRS, respectively. The number of patients contributing to the observed summaries per dose‐priming regimen are given in the panel titles. Solid lines and shaded regions represent the median and 95% CI of the predictions for each transition of grade ≥1 CRS, respectively. The dosing of these amounts occurred on Days 1, 4, and 8.

Figure 5

Population‐predicted CRS incidence over a wide range of dose‐priming regimens. The CRS incidence probabilities were generated using all permutations of the following step‐up dose 1 (sequence from 4 to 44 mg by 4 mg) and step‐up dose 2 (sequence from 0 to 32 mg by 4 mg) proposals (99 data points total) on Days 1 and 4, respectively, with 76 mg on Day 8. All were produced with 100% premedication usage. The observed dose‐priming regimens (4/20/76, 12/32/76, and 44/0/76 mg) are plotted as green, red, and blue dots, respectively.