Spectral transformations of facial behavior, associated neural activity patterns, and novel tools for the longitudinal tracking of neurons

Abstract

Understanding how patterns of neural activity give rise to behavior requires advanced tools to both quantify complex actions and track the underlying neural substrates over time. This dissertation presents two complementary contributions to address these challenges. First, this work introduces face-rhythm, an unsupervised computational pipeline that decomposes facial movements into interpretable behavioral components. This is accomplished using markerless point tracking, spectral analysis, and tensor component analysis. Applying this method to mice engaged in learning tasks reveals that the extracted uninstructed behaviors are predictive of internal states, such as reward expectation and goal-oriented intent. We also use this framework to investigate a core principle of facial motor control—are rhythmic facial movements encoded in the motor cortex by rhythmic control signals? We find that M1 employs a multiplexed coding strategy, representing both the precise, moment-to-moment position of the face (a phase-variant code) as well as the slower, overall energy of a movement (a phase-invariant spectral envelope). Notably, this work demonstrates that the majority of neural activity associated with facial movements greater than ~0.5 Hz is represented via a phase-invariant envelope. This result generalizes existing work showing that specific circuits and behaviors exhibit this behavior, and points towards generalizable models of hierarchical motor control where the cortex parametrically drives downstream central pattern generators. Second, to enable accurate and scalable longitudinal tracking of neural activity, this dissertation introduces ROICaT (Region of Interest Classification and Tracking), an analysis pipeline for tracking neurons and other ROIs in optical imaging data over months. The core of this system is ROInet, a self-supervised vision foundation model trained via contrastive learning on a diverse dataset of three million region of interest (ROI) images. ROInet learns a generalizable representation of cellular morphology, allowing for highly accurate zero- or few-shot classification of ROIs. In addition, by combining this morphological "fingerprint" with spatial information, as well as several other novel algorithms, ROICaT significantly outperforms state-of-the-art ROI tracking methods, particularly in challenging datasets with high cell densities or large inter-session tissue drift. Together, these works provide both biological insights and the enabling technologies to pursue them further.