Spatial Cell-type Enrichment Predicts Mouse Brain Connectivity

UCSF researchers Harry (Shenghuan) Sun, Justin Torok, Daren Ma, and Ashish Raj, PhD, with Christopher Mezias of Cold Spring Harbor Laboratory, published “Spatial cell-type enrichment predicts mouse brain connectivity” in Cell Reports. Co-first author Sun is a PhD candidate; co-first author Torok and Ma are specialists in Raj’s Brain Networks Laboratory at UCSF. The study found that machine learning methods can accurately predict whole-brain connectivity from neural cell types, and in particular oligodendrocytes, which produce and maintain myelin around neuronal axons. Using feature importance analysis, the researchers also identified key contributors to this connectivity prediction by cell types within the mouse brain.Cell Type Density Inference with Spatial Deconvolution

How the structural connectome – the web of connections within the brain – is related to the distributions of neural cell types is an important but unresolved question in neuroscience. Previous methods of analysis were mainly correlative and only utilized gene expression, not cell types. Exploring specifically how cell-type composition of brain regions relates to connectivity can deepen our understanding of how brain circuits mature during central nervous system development and how neurodegenerative diseases disrupt these circuits.

To create predictive cell-type-based models of structural connectivity, the researchers utilized recently published, whole-brain distributions of neuronal and non-neuronal cell types obtained using an in-house spatial deconvolution algorithm. They then implemented a random-forest-based machine learning approach to create predictive models of connectivity, supplying these distributions as features. Using random forest models also allowed the researchers to determine the individual contributions of each cell type and rank them in importance.

Using this approach, the researchers were able to predict both the presence or absence of brain connections and how strong the connections were with a surprisingly high level of accuracy, even though the construction of fiber connectivity is a highly complex process not strictly initiated by the constituent cell types. These results were replicated across two independently mapped sets of neural cell types, demonstrating that they were robust and not a function of the dataset used. They also showed that long-range and short-range connections could be predicted with cell types with a similar level of accuracy as the whole connectome.

Surprisingly, non-neuronal cells, and in particular oligodendrocytes, were more important than neuronal cells in predicting overall brain connectivity, even though connections are physically made up of neuronal projections. The researchers also found that cell-type importance was distance-dependent: non-neuronal types such as astrocytes and vascular cells were most important for predicting short-range connections, while medium spiny neurons and forebrain glutamatergic neurons were most important for predicting long-range connections. This suggests that the formation and maintenance of different kinds of connections may be mediated by different cell types. Overall, this link between cell types and brain connectivity provides a roadmap to expand the examination of this relationship in other species, including humans.