Quantitative systems pharmacology modeling sheds light into the dose response relationship of a trispecific T cell engager in multiple myeloma

Abstract

In relapsed and refractory multiple myeloma (RRMM), there are few treatment options once patients progress from the established standard of care. Several bispecific T-cell engagers (TCE) are in clinical development for multiple myeloma (MM), designed to promote T-cell activation and tumor killing by binding a T-cell receptor and a myeloma target. In this study we employ both computational and experimental tools to investigate how a novel trispecific TCE improves activation, proliferation, and cytolytic activity of T-cells against MM cells. In addition to binding CD3 on T-cells and CD38 on tumor cells, the trispecific binds CD28, which serves as both co-stimulation for T-cell activation and an additional tumor target. We have established a robust rule-based quantitative systems pharmacology (QSP) model trained against T-cell activation, cytotoxicity, and cytokine data, and used it to gain insight into the complex dose response of this drug. We predict that CD3-CD28-CD38 killing capacity increases rapidly in low dose levels, and with higher doses, killing plateaus rather than following the bell-shaped curve typical of bispecific TCEs. We further predict that dose-response curves are driven by the ability of tumor cells to form synapses with activated T-cells. When competition between cells limits tumor engagement with active T-cells, response to therapy may be diminished. We finally suggest a metric related to drug efficacy in our analysis-"effective" receptor occupancy, or the proportion of receptors engaged in synapses. Overall, this study predicts that the CD28 arm on the trispecific antibody improves efficacy, and identifies metrics to inform potency of novel TCEs.

Figure 1

Our trispecific T-cell engager model was designed through a rigorous development process. Here we depict the four key steps in our model development process. (A) We first formulated a model diagram showing all the cells and interactions to be included in the model. We included eight key cells—naïve, effector memory, and active T-cells of CD4 or CD8 lineage, multiple myeloma cells, and CD38 + PBMCs. Interactions and processes included in the model are shown with arrows and labeled. (B) Assumptions made to determine how interactions should be mathematically formulated in the model are shown. Assumptions were made about the process of activation, synapse formation, killing, and resistance to killing based on literature research and internal discussions (Table S1). (C) The model code was generating by encapsulating all cells, synapses, receptors, and interactions in a rule-based model generation code. An example of the template used for the synapse formation term is shown. This term is added to the ODE for the synapse and subtracted from ODEs of the cells joining the synapse. (D) We verified that the assumptions and ODEs generated were properly executed by ensuring that flux balances were conserved in the model. The rule-based code produced an ODE for the net flux in each cell and receptor type (i.e. amount added/subtracted over time), which was then added to the initial cell/receptor number (blue lines). We compared this to the total number of cells and receptors across the simulation (black lines) to ensure that all species are conserved. In each panel, results are shown for two doses (8.4e−4 nM and 0.672 nM).

Figure 2

In vitro model training and qualification process ensured accurate prediction of both T-cell activation and tumor cell killing. (A) Diagram depicts the overall process of our model calibration, which involved calibration to three separate in vitro datasets, then compilation of all optimized parameters into one QSP model used for analysis and prediction. (B) Two model formulations were used for model calibration, designed to replicate the setup of the in vitro experiments performed to generate the data. A model of pre-activated CD8 T-cells and tumor cells was used to train the cytotoxicity parameters to data. A model of PBMCs was used to train T-cell activation and cytotoxicity data. (C) Parameters optimized and final model simulation compared to data on cytotoxicity and T-cell activation. (D) Parameters optimized to MIMIC assay cytotoxicity measurements. Blue line shows mean + /− SD of population generated from optimization in comparison to range measured experimentally (gray shaded bars). Boxes show optimization results from a traditional distance-based fitting approach to the maximum (red) or minimum (light blue) values of the MIMIC data. (E) Model diagram and output generated from tumor proliferation qualification. (F) Cytotoxicity model diagram used for qualification of model killing predictions for RPMI-8226 (left) and KMS-11 (right) cell lines.

Figure 3

Efficacy at low doses increases with increasing synapse number while high dose level efficacy is driven by decreasing the number of ineffective synapses through better access to active T-cells. One set of values from the final population was selected and run for three days for different doses of interest. The results for tumor killing (A), Total active T-cells (B), Free active T-cells (C), and ineffective T-MM synapses (D). The sensitivity of these outputs to parameters is plotted as an AUC heatmap, where dark red = positive correlation, dark blue = negative correlation (E–H). Sensitivity is shown at two doses, indicated on plots. Parameters shown in each heatmap have 80% or more correlation to least one evaluation metric (such as AUC) derived from the corresponding output. Variability of each output across doses in the sensitivity analysis population (I–L).

Figure 4

T-cell engager efficacy is driven by several factors leading to highly superior efficacy of trispecific antibodies at lower doses but comparable performance as dose increases. Comparison of trispecific simulated over three days to comparable bispecific with no CD28 binding. MM cell receptor occupancy and “effective” receptor occupancy for bispecific and trispecific, respectively (A, B). (C) Killing was predicted across the in vitro population for varying doses of the bispecific compared to trispecific. MM cell distribution and T-cell distribution are shown at final timepoint (D, E) for bispecific and trispecific, respectively. (F) Schematic of the prevalent forces driving dose–response curve. Blue line and boxes describe trispecific dose–response and levels of ineffective synapses, receptor occupancy, total T-cells, and active T-cells. Green line and boxes show bispecific simulation.