DeepLabCut Analysis of Social Novelty Preference

Abstract

The description and quantification of social behavior in laboratory rodents is central to basic and translational research. Conventional ethological approaches to social behavior are fraught with challenges including bias, significant human effort and temporal accuracy. Here we show proof of principle that machine learning can be applied to laboratory tests of social decision making. Rats underwent social novelty preference tests which were scored both by hand and again by a convolutional neural network generated in the DeepLabCut computer vision package of Mathis and colleagues. The CNN generated temporally (30Hz) and locally (<5pixels) accurate identification of rat nose, eye and ear positions which were then used to compute social interaction and topography heat maps. In sum, hand- and computer-scoring were strongly correlated, and each identified significant preferences to interact with novel conspecifics which sets the stage for applying DeepLabCut analysis to other types of social interaction in the future.