Title: Multimodal Probabilistic Deconvolution and Semi-Supervised Latent Factor Discovery: A Set of Software Tools and Analytical Framework for Resolving Regional Cellular Heterogeneity with Specific Application to Human Alzheimer’s Disease Brain

Alzheimer’s disease begins years before symptoms appear, and some brain regions are affected earlier than others. Two especially vulnerable areas are the entorhinal cortex and the amygdala. Yet we still do not understand why certain cells in these regions are damaged first. This project will develop a new computational tool, called pyroNMF, to better interpret a powerful new kind of brain data. New spatial gene-mapping technologies can measure which genes are active in tiny locations within tissue while preserving the tissue’s structure. However, each measured spot contains a mixture of cell types, so current analysis methods often miss rare cells or fail to detect disease-related changes that are not already represented in reference datasets. pyroNMF is designed to address this problem by identifying the contributions of known cell types while also discovering previously unrecognized disease-associated patterns. The method will be applied to donated human brain tissue from people with and without Alzheimer’s disease, combining gene activity maps with measures of toxic tau protein buildup and MRI-based maps of brain anatomy. The result will be a detailed picture of where and how specific cell populations change in the earliest stages of disease. This work could reveal why particular neurons are selectively vulnerable, identify molecular signals linked to tau pathology, and provide a foundation for future biomarkers and treatments.