A sparse genetic code underlies the neuroanatomical organization of the brainstem

A major challenge for neuroanatomy is the integration of classical characterizations of neuronal morphology, connectivity, and histochemistry, with recent characterizations of genetic expression. Large-scale genomic databases now provide a computational tool to understand brain organization both morphologically and genetically. Here, we combine neuroanatomical techniques with informatics and statistical learning to uncover principles of genetic organization. We focus on the mouse trigeminal brainstem because of its stereotyped connectivity, its highly conserved anatomy, and its primary importance to sensory inputs in all vertebrates.

While most studies of spatially registered genomic data use global expression to infer anatomical differences, we examine the inverse problem: what are the fewest genes necessary to correctly identify regions defined by cytoarchitecture and innervation? We generated a micron-resolution atlas of inputs using a sensitive neuronal tracer injected into each branch of the adult mouse trigeminal nerve. Brains were sectioned, stained, imaged, registered, and reconstructed to produce a volumetric, vectorized map of innervation and morphology. We coregistered these data with the Allen Gene Expression Atlas, a dataset of 20,012 usable in situ hybridization probes with expression intensities at 200um3 spacing throughout the brain (~10^6 voxels).

We next generated brain region classes according to their distinct anatomical attributes, and treated each gene expression set as a feature vector. Using supervised learning algorithms based on decision trees and L1-norm regularization, we find sets of gene pairs and triplets that uniquely specify classically described trigeminal nuclei. This uniqueness demonstrates that an extremely sparse representation from the large set of genes is sufficient to outline the anatomical substructure of the brainstem.

The combination of classical tract tracing and modern imaging with novel informatics and statistical learning techniques thus 1) exploits and builds on existing registration infrastructure, 2) extends the quantitative tools available for assessing spatially-registered genomic data, and 3) is essentially generalizable to any neural system. More fundamentally, it provides a framework for identifying putative novel genes for developmental studies, and a strategy for targeting individual brain regions for manipulation with optogenetics and for imaging with genetically encoded sensors.