Computational neuroanatomy: mapping cell-type densities in the mouse brain, simulations from the Allen Brain Atlas

TL;DR

This paper uses the Allen Brain Atlas to computationally map and estimate the densities of different brain cell types in the mouse brain, accounting for variability in data through simulation.

Contribution

It introduces a Monte Carlo simulation approach to quantify uncertainty in cell-type density estimates from the Allen Brain Atlas data.

Findings

Estimated densities for 64 cell types with error bars

Demonstrated variability in cell-type localization

Provided a data-driven method for regional cell classification

Abstract

The Allen Brain Atlas (ABA) of the adult mouse consists of digitized expression profiles of thousands of genes in the mouse brain, co-registered to a common three-dimensional template (the Allen Reference Atlas). This brain-wide, genome-wide data set has triggered a renaissance in neuroanatomy. Its voxelized version (with cubic voxels of side 200 microns) can be analyzed on a desktop computer using MATLAB. On the other hand, brain cells exhibit a great phenotypic diversity (in terms of size, shape and electrophysiological activity), which has inspired the names of some well-studied cell types, such as granule cells and medium spiny neurons. However, no exhaustive taxonomy of brain cells is available. A genetic classification of brain cells is under way, and some cell types have been characterized by their transcriptome profiles. However, given a cell type characterized by its transcriptome, it is not clear where else in the brain similar cells can be found. The ABA can been used to solve this region-specificity problem in a data-driven way: rewriting the brain-wide expression profiles of all genes in the atlas as a sum of cell-type-specific transcriptome profiles is equivalent to solving a quadratic optimization problem at each voxel in the brain. However, the estimated brain-wide densities of 64 cell types published recently were based on one series of co-registered coronal in situ hybridization (ISH) images per gene, whereas the online ABA contains several image series per gene, including sagittal ones. In the presented work, we simulate the variability of cell-type densities in a Monte Carlo way by repeatedly drawing a random image series for each gene and solving optimization problems. This yields error bars on the region-specificity of cell types.