A latent variable model for quantifying social behavior in rodents

点击翻译按钮获取中文摘要

PROJECT SUMMARY Computational methods for quantifying mammalian natural behavior, including social interactions, are crucial for developing a sophisticated understanding of the neural basis of behavior. Yet a full description of behavior consists of much more than an animal’s actions. External cues (such as the actions of a social partner) drive our behavioral responses, and our responses to those cues can depend on context, our internal mental state, and prior experience. We may approach an individual when we feel safe, or attack that same individual when we feel threatened. The resulting complexity makes natural behavior — and social interactions in particular — challenging to study. To overcome this barrier, I propose to develop broadly applicable models to predict natural and social behavioral dynamics in mice based on changing external cues and internal states. These models will use unsupervised learning techniques to quantify and predict complex patterns of behavior in an interpretable manner while linking social behaviors to changes in neural activity across multiple timescales. Together, these models will provide an unprecedented view of how different neural populations encode the internal states that shape social behaviors as they unfold over time. The first aim is to fit a set of increasingly complex datasets with flexible latent-state models that describe how natural and social behaviors arise in response to factors such as external cues and time-varying internal states. In the second aim, I will apply this modeling framework to calcium recordings in dopaminergic projections to the Nucleus Accumbens and Tail of the Striatum as well as glutamatergic cell bodies in the Lateral Habenula — all neural populations shown to respond in social contexts. I will determine how these neural populations differentially encode sensory inputs, internal states, and behavioral outputs. I will also examine how the activity in each neural population correlates with transitions between different behaviors and internal states and how these representations change with experience. The proposed work will break new ground by applying novel computational tools and sophisticated, unsupervised behavioral quantification methods to discover the internal and external variables that shape natural behaviors as well as the underlying neural correlates of social interactions. Together, the new computational modeling techniques that I am proposing will advance several goals of the NIMH Theoretical and Computational Neuroscience Program: they (1) contain distinct levels of analysis, (2) link neuronal and behavioral processes, (3) enhance predictions of high-resolution behavioral data along with neural units of analysis, and (4) provide effective explanatory techniques and methods of interpretation for their results.