Maitreyee Bose, PhD: No financial relationships to disclose
Background: T-cell engagers (TCEs) represent a transformative approach in cancer therapy, designed to redirect the immune system to selectively target cancer cells. Despite their promising efficacy, the activation of immune system poses the risk of rapid release of cytokines, leading to occurrence of Cytokine Release Syndrome (CRS), potentially leading to severe concerns around patient safety. Several semi-mechanistic modeling approaches have been employed to understand cytokines in context of the drug action [1,2], however extending these to have a predictive model of clinical CRS risk for individual patients remains elusive. Heterogeneity in patients responses and clinical read outs and incidence of CRS events for the same dosing remains the key challenge in linking such mechanism driven models to clinical observations. Data driven models can be useful to regress the observed CRS incidence as a function of the patient attributes and the observed dynamics of contributing factors, like elevation in pro- and anti-inflammatory cytokines. Nonetheless, the sparsity of the data on the different components of CRS and the large inter-individual variability can limit the utility of models in predicting CRS risk for future dosing regimens. The primary objective of this work is to link mechanism-driven models of cytokines to CRS events using a hybrid machine learning (ML)-based classifier approach that can overcome the aforementioned limitations to predict CRS risk for specific patient cohorts.
Methods: In this work, we introduce a two-step predictive framework for predicting CRS in TCE therapy. The core of the framework is a mechanistic model capturing the homeostatic dynamics and the interplay between the immune system, TCE dosing, and cancer cells that can predict the dynamics of cytokine release. The output of this model is then used as predictor to a ML classifier that can take in several subject specific attributes, drug exposure metrics, baseline factors, and cytokine peaks for prediction of incidence of clinical CRS. We explore the utility of our approach using simulated data across various scenarios corresponding to different TCEs and cancer indications.
Results: Our results indicate that using the mechanism-based approach can provide additional information to the traditional ML based classifier providing a more robust prediction of the clinical endpoint of interest, i.e. the CRS incidence. Using simulated examples corresponding to various tumor indications and targets, we show that using our hybrid approach can improve the accuracy of the classifier, potentially enabling optimization of drug dosing. We illustrate the utility of our approach by exploring the effectiveness of a step dosing/dose fractionating design to mitigate CRS risk.
Conclusions: Recognizing CRS as a limiting factor for TCE development, this work highlights a practical tool for strategic model-based in silico optimization of TCE dosing and designing safer clinical trials.
Citations: [1] Hosseini, I., Gadkar, K., Stefanich, E. et al. 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. npj Syst Biol Appl 6, 28 (2020). https://doi.org/10.1038/s41540-020-00145-7
[2] Chen X, Kamperschroer C, Wong G, Xuan D. A Modeling Framework to Characterize Cytokine Release upon T-Cell-Engaging Bispecific Antibody Treatment: Methodology and Opportunities. Clin Transl Sci. 2019 Nov;12(6):600-608. doi: 10.1111/cts.12662. Epub 2019 Jul 26. PMID: 31268236; PMCID: PMC6853151.
