Neural network models for classifying immune cell subsets

The immune system is composed of heterogeneous cell populations and it includes several hundreds of distinct cells types such as neutrohils, eosinophils, basophils, macrophages, dendritic cells, CD4+ and CD8+ T cells, γδ T cells, mast cells, and B cells and each main cell type can be further differentiated into subsets with unique and overlapping functions. For example CD4+ T cells can be differentiated into T helper (Th)1, Th2, Thl7, and regulatory T cell (Treg) subsets. To study molecular mechanisms of cell differentiation, Systems Biology Markup Language (SBML) based Ordinary Differential Equation (ODE) models can be used for representing such processes. These intracellular signaling models often require many equations to accurately represent intracellular pathways and biochemical reactions. Another challenge in studying the immune system and immune responses is the need for integration of complex processes that occur at different time and space scales (i.e., populations, whole organism, tissue level, cellular and molecular) through multi-scale modeling. This study presents two novel neural network models for modeling CD4+ T cell differentiation and immune cell subset classification. The first model reduces the complex ODE intracellular model by focusing on the input and output cytokines and the second model establishes an automated subset classification based on molecular patterns expressed in immune cells. After training, the first model achieves small prediction errors of cytokine concentrations and the second model achieves 98% prediction rate for subset classification. Neural network algorithm and models have been widely used for many data mining tasks such as classification and pattern recognition. However, to the best of our knowledge this study is the first one applying the neural network algorithm for immune cell differentiation and subset classification. In addition, these novel neural network models can be computationally efficiently integrated into multi-scale models with limited computational costs.