Automated behavioral analysis of asleep fruit fly

Official URL: https://risc01.sabanciuniv.edu/record=b2758670_(Table of contents)

Sleep is a highly conserved behavior program across the animal kingdom, hinting at its essential value. In order to decipher the functions of sleep, careful characterization of the underlying changes in behavior and physiology is needed in powerful genetic model systems such as Drosophila Melanogaster. Recent advances in machine learning have enabled tracking of body parts and robust pose estimation in videos; however, automated quantification of behaviors requires mapping from a pair of spatial coordinates to behavioral categories. Detection and successful mapping of behaviors exhibited during sleep come with unique challenges. Existing methods and pipelines are developed by focusing on behaviors defined by macro postural changes ignoring subtle movements during sleep. On the other hand, our task of phenotyping sleep requires tackling unobtrusive changes that sparsely occur during long sleep cycles. To this end, we develop basty (Automated Behavioral Analysis of Asleep Fruit Fly), a novel, end-to-end pipeline made public as a configurable, open source, and easy-to-use software package. Our pipeline enables several analyses, such as computing meaningful behavioral representations, detecting activities in long sleep experiments (14-16 hours), and nearest neighbors analysis in a low-dimensional behavioral space. We evaluate our pipeline with a dataset of sleep deprivation and wild-type sleep experiments, focusing on five behavioral categories. Results show that our pipeline successfully maps behavioral categories, achieving an AUC score of 0.8, and it can also discover unobserved behaviors and differences in behavioral repertoires. Furthermore, we present an analysis of the behavioral repertoire exhibited during sleep by examining spatio-temporal characteristics of the behaviors and their temporal organization; in both wild-type sleep and sleep-deprivation experiments.