Harnessing biological rhythms for a resilient social motif generator - Summary How does the brain enable social interactions? The study of social behavior in non-human animals has long relied on coarse behavioral metrics like time spent interacting with another animal or simply the numbers of interactions. Although this approach has informed major insights into neural circuits which have a role in sociability, we still do not know how these circuits orchestrate patterns of social behaviors, especially under different social contexts where interactions have nuanced differences. Our long-term goal is to identify the neural mechanisms supporting social behavior in affiliative vs. antagonistic social contexts. To close the knowledge gap towards this goal, in this R34 we will build artificial intelligence (AI) tools that are capable of integrating multivariate sources of behavior data to quantify spatiotemporal signatures or “motifs” of diverse repertoires of social behaviors. Behavioral motifs have the potential to be captured by means of examining concurrent autonomic rhythms, especially breathing and heart rate. Indeed, we have long known that changes in the frequency of these rhythms coincide with specific affective and behavioral contexts. However, spatiotemporal signatures of social behaviors have not been captured in prior studies which have considered either breathing or heart rate in isolation. Nor have prior studies unleashed the potential to identify novel social behavioral motifs by using these autonomic rhythms in combination with video measures. The research objective of this Brain Initiative proposal is to develop semi-supervised artificial intelligence methods that result in a hierarchical multi-timescale model of social behavioral motifs directly from video, breathing, heart rate, and movement data via a head-mounted accelerometer. To accomplish this, we will use partial labels of mouse social behaviors, as well as physiologic measurements, in order to elucidate the full range of social behavior motifs across affiliative vs. antagonistic contexts. In Aim 1, we will define low-dimensional social behavioral states while incorporating autonomic rhythms, while in Aim 2, we will elucidate a multi-timescale hierarchical representation of social behavior in affiliative vs. agonistic social contexts. For both aims, we will integrate computer vision techniques with high-dimensional video and physiological data from mice while varying their isolation levels and who they are interacting with. The end-product will be a validated toolkit enabling the sensitive and robust identification of behavioral motifs. The easy-to-use toolkit which we call the Social Motif generator (So-Mo) will enable future studies to probe neural circuits during complex mouse behaviors at unprecedented resolution.