(636d) Regulatory T-Cell Differentiation Plasticity Model

Cells of the immune system undergo various responses that include differentiation into effector cells to combat a detected antigen. Particularly, T-lymphocytes are of interest because of their diverse range of possible immune responses. The differentiation of T-cells depends on the cytokine environment [1]. In vivo studies have established the plasticity of differentiation of T-cells to better allocate resources, reduce collateral tissue damage, and more effectively combat antigens [1-2]. T-cells can differentiate into many effector cells, but the focus of this work is upon Th1, Th2, Th17. Th1 cells produce IFN-γ and contribute to cellular immunity against microbes. Th2 cells produce IL-4, IL-5, and IL-13 where their immunity role is primarily humoral. Th17 produce IL-17 that is vital for combating extracellular bacteria and fungi. More importantly, Th17 differentiation is dependent on the induction of the orphan receptor (ROR)- γ which is stimulated by TGF-β, and IL-6 among other pro-inflammatory cytokines [1,3]. In summary, over-active Th1 and Th17 cells are responsible for organ-specific autoimmunity while Th2 cells are responsible for asthma attacks and allergies.

The immune system also regulates the strength of the response by causing a portion of the differentiated effector cells to revert to Regulatory T-cells (Treg). Studies have shown that cytokine as well as bacterial metabolite stimulation of a T-cell population has the potential of shifting the population towards Treg cells. FoxP3 is determined to be the key regulatory protein for Treg differentiation, as well as its functionality in immune suppression. This protein is utilized to  identify the differentiated Treg cells among the other effector cells. Being able to control the population composition of the immune system cells has important therapeutic applications that include control over infections and combating of autoimmune diseases through preventing the effector T-cells from being overly active.

This research proposes a differentiation model of T-cells upon simulation with different combinations of IL-2, IL-4, IL-6, TGF-β, and indole. The model essentially fits the change in the T-cell population composition in response to external cytokine stimulation. As the biological system is not well understood, the prediction of how the composition of the T-cells will shift in response to cytokine stimulation is a non-trivial problem. More importantly, it is even more challenging to predict the population after exposure to a combination of cytokines ̶ which are very likely to interact and influence the outcome of the population. Therefore, a data driven approach to this complex biological system is necessary to develop a quantitative understanding of the mechanisms at work that govern the differentiation plasticity of T-cells. By constructing an experimental dataset that transverses a variety of cytokine concentrations and combinations, the resulting model parameters are capable of predicting the response of the cell population to an unfamiliar cytokine stimulation more robustly.

The levels of FoxP3 and IL-17 within a cell population were measured. Since Treg cells predominantly express FoxP3, while Th17 cells express IL-17, the presence and detection of those proteins as well as their relative compositions is translated to reflect the T-cell population composition that results in the detected cytokine profile. A neural network model was used to capture the effect that exposure to the various stimulants has on composition of the population. Neural networks were chosen based on their flexibility for fitting data and their ability to accommodate a large degree of nonlinearity within the model [4-5]. However, it is important to note that an inherent challenge in neural networks is ensuring that the model does not over-fit the data through over-parametrization. The procedure used is that the developed model receives the concentration of the cytokines as input, and predicts the T-cell population compositions of Th17 and Treg. The network is trained using a Bayesian Regulation algorithm to generalize the model to all the training data. This method inherently embodies the task of avoiding over-parameterization of the model [6]. Different network architectures are constructed that include different transfer functions within the neurons of the network. The training data is applied to the different network architectures, and the output of the networks is analyzed. In addition, the network models are also challenged to predict the T-cell population upon input of cytokine concentration combinations that the model has not been trained with. The models are then compared and based on their overall performance with respect to the training dataset, the testing dataset, and their complexity in terms of number of parameters.

The identified model contains a total of 22 parameters, and 4 processing neurons–2 neurons within each hidden layer–that are able to capture the nonlinearity and correlate the inputs and outputs from the dataset. The model suggests an optimal, indole concentration that can be utilized to obtain the largest shift in the T-cell population composition to Treg cells. In addition, the model suggests a novel combination of cytokine compositions that display important dynamics in the metabolic response of the T-cells to the cytokine stimulations. This work is unique in its investigation of the effect of cytokines as well as bacterial metabolites on T-cell population and transcription factors as well as the modeling of the biological system.

Reference

[1]        Zhou L, Chong M.M.W, Littman D.R. Plasticity of CD4+ T cell lineage differentiation. Immunity 2009;30:646-655.

[2]        Josefowicz S.Z, Rudensky A. Control of regulatory T cell Lineage commitment and maintenance. Immunity2009;30:616-625.

[3]        Opal S.M, DePalo V.A. Impact of basic research on tomorrow's medicine: Anti-inflammatory cytokines. CHEST 2000;117:1162-72.

[4]        Qian N. Sejnowski T.J. Predicting the secondary structure of globular proteins using neural network models. Am J Mol Biol. 1988;202:865.

[5]        de Freitas J.F.G. Bayesian methods for neural networks. Trinity College University of Cambridge2003.

[6]        MacKay  D.J.C. Bayesian interpolation. Neural comput 1994;4:415-419.