Wide and Deep Learning for Automatic Cell Type Identification

Abstract

Cell type classification is an important problem in cancer research, especially with the advent of single cell technologies. Correctly identifying cells within the tumor microenvironment can provide oncologists with a snapshot of how a patients immune system is reacting to the tumor. Wide deep learning (WDL) is an approach to construct a cell-classification prediction model that can learn patterns within high-dimensional data (deep) and ensure that biologically relevant features (wide) remain in the final model. In this paper, we demonstrate that the use of regularization can prevent overfitting and adding a wide component to a neural network can result in a model with better predictive performance. In particular, we observed that a combination of dropout and{ell} 2 regularization can lead to a validation loss function that does not depend on the number of training iterations and does not experience a significant decrease in prediction accuracy compared to models with{ell} 1, dropout, or no regularization. Additionally, we show WDL can have superior classification accuracy when the training and testing of a model is completed data on that arise from the same cancer type, but from different platforms. More specifically, WDL compared to traditional deep learning models can substantially increase the overall cell type prediction accuracy (41 to 90%) and T-cell sub-types (CD4: 0 to 76%, and CD8: 61 to 96%) when the models were trained using melanoma data obtained from the 10X platform and tested on basal cell carcinoma data obtained using SMART-seq.