MUDflow: Combining Neural Networks, UMAP and DBM Clustering to Identify Cell Populations Accurately, Quickly and Easily in Mass and Fluorescence Cytometry

Abstract

Identifying cell populations in flow cytometry data is mostly done via a "manual gating" method that often lacks verifiability and reproducibility, even in the hands of experienced investigators. Recently developed automatic gating methods have been shown to have good performance in cell population identification, but may require fine-tuned setup from experts or struggle to identify small populations. Here, we introduce an easily trainable multilayer perceptron neural network for automatic gating (MLPgater). Compared to the three popular automatic gating methods LDA, FlowSOM and PhenoGraph on three mass and six fluorescence cytometry datasets, MLPgater is most accurate by a substantial margin, tied with LDA as the fastest and uniquely able to replace manual gating except for training purposes. Furthermore, we show that combining MLPgater with UMAPs guided dimensionality reduction feature and DBMs clustering (MUDflow) effectively detects new populations that did not exist or were not identified in the training set.