Identifying bimodality in data using a proximity-based null model with an application to classifying cell cycle phases using oxidative stress

Abstract

Detecting communities in large complex networks has found a wide range of applications in physical, biological, and social sciences by identifying mesoscopic groups based on the links between individual units. Moreover, community detection approaches have been generalized to various data analysis tasks by constructing networks whose links depend on individual units' measurements. However, identifying well-separated subpopulations in data sets, e.g., multimodality, still presents challenges as both community detection with existing null models and other partition methods either fail to give partitions that correspond to dips in the data or give partitions that do not correspond to dips in the data. Here we introduce a new spatially informed null model for this task that takes into account spatial structure but does not explicitly depend on distances between measurements. We find that community detection using this null model successfully identifies subpopulations in multimodal data and accurately does not for unimodal data. This new method represents a complement to statistical methods, as we treat data directly using a network science approach. We apply this new null model to first distinguish interphase and mitotic cell cycle phases and then S and G2 cell cycle phases in a group of Dictyostelium discoideum cells using measurements of oxidative stress, which have been shown to correlate strongly with cell-cycle behaviors.