Machine Learning Approaches for Behavioral Phenotyping of Humanized Knock-in Models of Alzheimer's Disease

点击翻译按钮获取中文摘要

PROJECT SUMMARY Preclinical efforts to develop treatments for cognitive impairment in Alzheimer’s disease have been hindered by two barriers: time-consuming behavioral assays that often lack sensitivity, and transgenic (TG) mouse models with APP overexpression that do not accurately recapitulate human pathogenesis. To overcome the second barrier, newly-developed humanized App knock-in (App-KI) mouse models express AD-related human genes at physiological levels. App-KI mice have prominent brain pathology, but inconsistent, milder or absent behavioral phenotypes in many traditional behavioral tests, including Morris Water maze (Saito et al., 2014). These limitations of standard behavioral testing, including lack of sensitivity, low throughput, and reproducibility represent key methodological barriers to proper development of therapeutics in newly developed App-KI mice. Without a solution to this problem, it is likely that translational and preclinical research will struggle to develop therapies in models of early pathogenesis or sporadic AD characterized by mild or subtle behavioral phenotype and lacking overt clinical disease manifestation. To overcome the limits of behavioral testing, we propose to implement machine-learning (ML) approaches that offer complete, unbiased, and robust behavioral characterization of even subtle behavioral phenotypes. Specifically, we propose to upgrade and refine our novel computer vision ML approach (Aim #1), to validate it in App-KI and TG AD mouse models (Aim #2), and to apply it to mice receiving a newly-developed anti-Aβ antibody to validate our approach in a preclinical setting (Aim #3). We recently published our first iteration of an ML package that developed the VAME neural network to identify behavioral motifs. VAME will be further developed to provide rapid testing of large cohorts, unbiased identification of disease-associated behavioral deficits, and reproducible phenotypes across experimental conditions and laboratories. Our proposal thereby addresses key limitations of standard behavioral testing and mouse modeling, vertically advances the methodology of behavioral neuroscience, launches innovative biotechnological development, and opens new horizons for dementia-related research, including the adaptation of the approaches to humans. Successful completion of the proposed studies will provide a new preclinical tool for diagnosis, assessment, and disease monitoring in mouse models of AD. In conclusion, will establish a novel machine- learning behavioral phenotyping platform with the power to non-invasively identify robust behavioral alterations in App-KI models of AD, removing a key methodological barrier to the translational study of MCI, and increasing the value of behavioral research broadly. This award will critically support the PI to undertake immersive entrepreneurial training experiences at local universities and at a startup company focused on developing novel therapeutics for neurodegenerative diseases.