DeepOF: a Python package for supervised and unsupervised pattern recognition in mice motion tracking data

Abstract

DeepOF (Deep Open Field) is a Python package that provides a suite of tools for analyzing behavior in freely-moving rodents. Specifically, it focuses on postprocessing time-series data extracted from videos using DeepLabCut (Mathis et al., 2018). The software encompasses a diverse set of capabilities, such as: • Loading DeepLabCut data into custom objects and incorporating metadata related to experimental design. • Processing data, including smoothing, imputation, and feature extraction. • Annotating behavioral motifs in a supervised manner, such as recognizing huddling and climbing, and detecting fundamental social interactions between animals. • Embedding motion tracking data in an unsupervised manner using neural network models, which also facilitate end-to-end deep clustering. • Conducting post-hoc analysis of results and visualization to compare patterns across animals under different experimental conditions. The package is designed to work with various types of DeepLabCut input (single and multi-animal projects), includes comprehensive documentation, and offers interactive tutorials. Although many of its primary functionalities (particularly the supervised annotation pipeline) were developed with top-down mice videos in mind, tagged with a specific set of labels, most essential functions operate without constraints. As demonstrated in the accompanying scientific application paper (Bordes et al., 2022), DeepOF has the potential to enable systematic and thorough behavioral assessments in a wide range of preclinical research settings.