(156d) CAR-Mediated T Cell Activation: Linking Molecularly-Detailed and Data-Driven Models | AIChE

(156d) CAR-Mediated T Cell Activation: Linking Molecularly-Detailed and Data-Driven Models


Finley, S. D. - Presenter, University of Southern California
Rohrs, J. A., University of Southern California
Cess, C., University of Southern California
Introduction: Chimeric antigen receptor (CAR) T cells are T cells that have been engineered to express a receptor that binds to a tumor-associated antigen so that the T cells can detect cancer cells better than the normal immune system. While this treatment is successful in many patients, there are also many patients where this treatment is unsuccessful. It has been shown that protein levels can vary between cells of the same type, and that this can lead to phenotypically different cells within the same genotype. Since reaction kinetics depend on many proteins, it is difficult to analyze ODE-based mechanistic models to the fullest extent due to their high dimensionality. Data-driven models, while lacking the biological detail of a mechanistic model, are able to provide a more generalized relationship between model inputs and outputs. Here, we use a data-driven method, in the form of partial least-squares (PLS), to generalize the relationship of protein expression and cell response in CAR T cells. We used the mechanistic model to generate the large amount of data needed for this approach. Thus, the data-driven model is being used as an analysis tool for the mechanistic model. The results of this study should provide better understanding for the improvement of CAR therapy.

Materials and Methods: Using a mechanistic model of CAR T cell signaling, a heterogeneous population of 10,000 cells was simulated. Cell response was determined using the relative phosphorylated ERK concentration. This response separated the population into two categories: responders and non-responders. To better understand the relationship between protein expression and cell response, a PLS model was developed using the initial protein concentrations as inputs and the response as the output. The data generated from the mechanistic model simulations was used to train and validate the accuracy of the PLS model. Examining the components and the weights of the PLS model allowed us to determine the general influence of each individual protein on the cell’s ability to respond. Since lower weighted inputs have less of an impact on the PLS outputs, proteins with lower weights were removed from the dataset and the PLS model was retrained to further determine which proteins are highly influential to the cell response.

Results and Discussion: From the population simulations, it was found that only about half of the cells in the population could be characterized as responders. A PLS model was developed that could classify this response with high accuracy. Out of the 19 proteins present in the network, only eight (LAT, Gads, SOS, Ras, RAF, MEK, RasGAP, and RasGRP) were found to be highly influential based upon the weights of their initial concentrations. The influence of these proteins on the system was found to match literature descriptions of their biological roles. Interestingly, it was found that the upstream proteins that interact with the CAR had very little influence on whether or not the cell would respond. All of the influential proteins identified are related to the activation of the MAPK pathway and its signal transduction. This approach of using a data-driven model as a way to analyze a mechanistic model allowed for a generalization of model parameters and investigation of their effects on the system as a whole. Additionally, we have used the data-driven model as a means of easing the computational burden of simulating the full mechanistic model. This enables a multi-scale framework to study immunotherapy, drug resistance, and the effects of tumor cell mutations.