Unsupervised Quantification of Stereotyped Behavior in Flies and Mice

Abstract

Recent advances in the study of animal behavior have utilized tools from computer vision and machine learning in order to perform high-throughput behavior quantification. Here, we utilize a behavior analysis paradigm established by Berman et al. (2014). Using a principle component-based approach, raw behavior video data is reduced to a postural time series. A dimensionality reduction approach embeds the dynamics of this time series into a two-dimensional “behavioral map,” and discrete behaviors are identified as peaks in the map's probability density function. Here, we expand on the work of Berman et al. in two main ways. Firstly, we define a metric for quantifying the stereotypy of a given behavior, based on the variance in a behavior's postural trajectory over many instances of the behavior. We find that most identified behaviors in the fruit fly Drosophila exhibit a high degree of stereotypy, providing support for the view of behavior as a set of discrete states, at least to an approximation. Additionally, we apply this framework, which was originally developed for Drosophila, to the mouse. The mouse is a prominent model organism in neuroscience research, and we hope that these behavior quantification techniques can be combined with modern neural recording and perturbation techniques in order to understand the mechanisms through which the brain responds to external stimuli and generates complex behaviors. Finally, we utilize proof-of-concept experiments to suggest that our analyses can quantify the behavioral effect of neuropharmacological agents and neural perturbations with optogenetics.