Unsupervised Classification of Behavior for Open Field Mouse Recordings

Advances in computer vision and deep learning have made it possible to explore animal behavior at fine temporal and spatial scales. In particular, new advances in automated pose detection make it possible to track fine-scale movements in mice, a model system for the study of many aspects of neural function, from locomotion and coordination to complex neurodevelopmental disorders such as autism. We combine the use of a deep-learning-based approach called LEAP (LEAP Estimates Animal Pose), which produces estimates of joint coordinates, with unsupervised classification in order to discover distinct behavioral bouts in 72 wild-type mice in an open field arena over five subsequent days. We use the resulting behavioral phenotypes to explore the evolution of behavior in individuals over time, as well as identify behavioral differences between wild-type individuals and genetic models of neurodegenerative disease. These differences elucidate complex phenotypes that arise from neuropathology, and capture greater complexity with fewer experimental constraints than simpler high-throughput behavioral assays.