Harnessing novel machine learning approaches for behavioral segmentation and brain signal integration in humanized models of Alzheimer's disease - SUMMARY Recent advances in computer vision and machine learning (ML), including pose estimation and behavioral segmentation, offer unprecedented improvements for acquisition and processing of behavioral data. These approaches enable comprehensive analysis of behavioral patterns frame-by-frame at a millisecond scale, allowing for the identification of dozens of behavioral states. These emerging methods greatly expand the depth, breadth, and sensitivity of behavioral quantifications and may provide novel insights into behavior, brain-behavior interactions, and disease pathogenesis. We found that these ML platforms, with improved sensitivity and breadth compared to conventional behavioral tests, are uniquely suited to assess subtle behavioral and cognitive changes in the recently-developed humanized Alzheimer’s disease (AD) knock-in (KI) models generated by the NIH (MODEL-AD), which express human proteins at physiological levels and avoid the pitfalls associated with transgenic overexpression. We seek to determine the co-pathogenic contributions of Aβ, apoE isoforms, and TAU to behavioral and cognitive functions using these novel humanized AD-KI mice. The proposal aims to further validate our ML platform for behavioral segmentation, called Variational Animal Motion Embedding (VAME), and to develop state-space computational models for comprehensive behavioral phenotyping of spontaneous behavior and cognitive functions in AD-KI mice (Aim 1). We will also perform validation studies of our ML approaches and computational analyses by introducing established therapeutic or mechanistic interventions, including reducing TAU expression (TAUKO) or blocking neuroinflammation (Fggγ390– 396A) in AD-KI mice (Aim 2). In Aim 3, we will exploit VAME’s capacity to integrate multi-modal inputs (brain oscillations and single-unit activity) alongside behavioral patterns to define meaningful relationship between brain activity patterns and behavioral states across disease progression. We will then use these multi-scale datasets to enrich our state-space models for assessing navigation in a novel labyrinth maze. As VAME developers, we will also improve the open-source VAME software and automate our ML pipeline for multi-scale integration to promote collaboration between computational scientists and translational researchers (Aim 4). We anticipate that the proposed development of robust and accessible ML and computational tools for behavioral and brain signal interrogation will drastically reduce some of the challenges associated with conventional behavioral testing, which often hinder progress in behavioral and translational neuroscience, by improving the sensitivity, depth, scalability, reproducibility, and statistical power of behavioral studies.
