Quantifying grooming in paired macaques

Quantitative tools are increasingly being used to study social interactions in natural, freely moving contexts. However the development of such tools for primate research has lagged behind that of other animals. To understand primate sociality, we need to measure how behavior varies as a function of the environment, social contexts, and bodily movement. Here, we present MacTrack (Macaque Tracking), a holistic behavioral tracking system which uses deep learning-based landmark tracking to predict the location of 134 landmarks (68 facial, 24 body, and 21 per hand) of interacting rhesus macaques. Using unsupervised learning, we find strong similarity between postures across behavioral tasks and social contexts and reveal novel, segregated behavioral motifs corresponding to previously indistinguishable behaviors (i.e., face groom vs. limb groom). Moreover, we explore how administration of oxytocin, a neuropeptide linked to social bonding and affiliation, impacts allo-grooming in opposite-sex pairings. We show that oxytocin dosages result in a greater number of grooming bouts and that, prior to grooming solicitation, monkeys in successful and unsuccessful grooming bouts display different behavioral phenotypes. By combining primatology and machine learning, this work advances our understanding of the macaque behavioral repertoire, with potential implications for understanding the building blocks of complex behaviors in humans.