Alex Lee

UC San Francisco

“A transformer model for brain region discovery in the mouse brain”

New spatial genomics technologies are revolutionizing our understanding of the organization and molecular composition of cells in organs and tissues, particularly in the mammalian brain. These technologies have facilitated the creation of comprehensive resources like the Allen Brain Cell dataset, which provides detailed information on millions of cells in the murine brain and their spatial organization. However, mapping these detailed cellular data to specific known brain regions is challenging with current methods. To address this, we’ve developed a novel method using deep learning that allows for highly accurate delineation of brain regions and their cellular makeup. Specifically we train a transformer model in a self-supervised manner, to identify cellular niches within the dataset itself, bypassing the need for pre-defined regional boundaries. This approach allows for precise mapping of brain regions’ cellular content directly from the data, overcoming previous limitations in spatial genomics. Using this resource we plan to construct a data-driven ontology of brain regions and their molecular composition to compare with the extensively human-curated ontologies from collaborators at the Allen Institute for Brain Science.


Neuroscientists often make inferences about the function of different areas of the brain based on their molecular or cellular characteristics. However, until recently, a detailed experimental cataloging of the various cells in the brain, even in a model organism such as a mouse, has been impossible. In my project we bring together exhaustive spatial transcriptomics measurements of the brain and novel deep learning algorithms to directly detect brain regions with novel cellular and molecular makeup, creating highly accurate maps of the composition of the brain.

7 + 2 =