Mapping Mouse Behavior with an Unsupervised Spatio-temporal Sequence Decomposition Framework

Objective quantification of animal behavior is crucial to understanding the relationship between brain activity and behavior. For rodents, this has remained a challenge due to the high-dimensionality and large temporal variability of their behavioral features. Inspired by the natural structure of animal behavior, the present study uses a parallel, multi-stage approach to decompose motion features and generate an objective metric for mapping rodent behavior into the animals feature space. Incorporating a three-dimensional (3D) motion-capture system and unsupervised clustering into this approach, we developed a framework that can automatically identify animal behavioral phenotypes from experimental monitoring. We demonstrate the efficacy of our framework by generating an "autistic-like behavior space" that can robustly characterize a transgenic mouse disease model based on motor activity without human supervision. Our results suggest that our framework features a broad range of applications, including animal disease model phenotyping and the modeling of relationships between neural circuits and behavior.