I am collaborating with Dr. Stephanie Miller and Katie Ly from the Palop Lab Machine Learning Team at the Gladstone Institutes at UCSF. Our project focuses on implementing machine learning tools, specifically DeepLabCut (DLC) and Variational Animal Motion Embedding (VAME), to evaluate the more subtle spontaneous behavioral alterations in AD mouse models. 

Within the field of neuroscience, and specifically neurodegenerative research, mice are commonly used to model certain diseases. To study AD, for example, mice are bred to have certain mutations that cause plaques to form, resulting in the onset of the disease. There are many ways that these mice are studied in experiments, commonly including behavioral tests. Most standardized behavioral tests utilized to study mice rely on discrete empirical measurements that have the potential to overlook the full scope of disease-related behavioral alterations. For instance, an Object Location Test (OLT) measures the amount of time a mouse spends exploring a novel object compared to an object it is already familiar with, and a Morris Water Maze test evaluates spatial learning by measuring the time a mouse takes to reach a designated location (a platform). 

While these tests can tell us important things about the mouse's working memory, they do not tell us how its overall behavior is altered. Thus, only using standardized tests allows a greater chance of overestimating the efficacy of novel treatments as we only observe very specific behaviors. For instance, using an OLT, we may conclude that treatment improved the condition of an AD mouse by rescuing its working memory, when in reality, the mouse's overall behavior, which may be influenced by AD-related anxiety, actually remains unaffected by the treatment. 

With the increasing capabilities of machine learning, we are now able to holistically evaluate alterations in mouse behavior between testing groups. This is the topic of my project. Here is an outline of the process for this method (outlined in the figure to the right): First, DLC is used to track specific body parts of a mouse in a video, utilizing a deep neural network algorithm built upon just a small amount of labeled training data (10 labeled frames per video). VAME is then used to analyze this tracking data by segmenting the data into a specified number of repeated actions (motifs). Typical examples of motif categories include running, walking, rearing, grooming, and more. 

We can then evaluate differences in the usage of these motifs across testing groups. How might this be useful? Let's say we have 4 test groups of mice: healthy/control, AD/sick, healthy+treatment, and AD+treatment. In the untreated AD mouse group (AD/sick), VAME may determine that they spent a significant percentage of time using the running motif. This could show that they 1) are constantly aimlessly wandering, demonstrating memory deficits, and 2) are hyperactive due to AD-induced anxiety. However, we see that the AD+treatment group uses the running motif much less, and when they do use that motif, their speed is slower on average. This shows that these AD-related symptoms in the untreated group are rescued in the treated group. 

Using these motifs, various other analyses can be conducted, ultimately providing a more comprehensive and holistic representation of altered behaviors in each mouse group.