Cell-type–based model explaining coexpression patterns of genes in the brain

Abstract

Spatial patterns of gene expression in the vertebrate brain are not independent, as pairs of genes can exhibit complex patterns of coexpression. Two genes may be similarly expressed in one region, but differentially expressed in other regions. These correlations have been studied quantitatively, particularly for the Allen Atlas of the adult mouse brain, but their biological meaning remains obscure. We propose a simple model of the coexpression patterns in terms of spatial distributions of underlying cell types and establish its plausibility using independently measured cell-typespecific transcriptomes. The model allows us to predict the spatial distribution of cell types in the mouse brain.neuroscience | bioinformatics | neuroanatomy B rain-wide and genome-wide maps of gene expression are now available (1, 2), due to the development of high-throughput neuroanatomical methods (3-6). This has enabled analysis of the spatial correlation structure of gene expression (7-12). In the Allen Brain Atlas (ABA) of the adult mouse, the brain is divided up into cubic voxels of size 200 μm. The expression energies of up to 20,000 genes in the adult C57BL/6J mouse are given by automatic processing of in situ hybridized (ISH) brain sections, coregistered to the Allen Reference Atlas (ARA) (13). Coexpression of two genes in a voxel may arise from two sources: ðiÞ both genes are expressed within the same cell type or ðiiÞ the two genes are expressed in two different cell types, both present in the voxel. These two possibilities cannot be disentangled solely on the basis of the ABA. Ideally, gene expression profiles should be experimentally obtained for each cell type in the brain, and indeed such transcriptome profiles are now available (14-21). This cell-based approach to the study of gene expression is complementary to the gene-based approach of the ABA. In ref. 22, the data of ref. 19 were used to extract neuron-specific genes, astrocyte-specific genes, and oligodendrocyte-specific genes, which resulted in three combinations of brain-wide maps from the ABA, whose clustering showed anatomical signatures of major brain subdivisions (see also ref. 23 for estimates of neuron-specific and oligodendrocyte-specific expression patterns, both in the mouse and in the human brain). The present paper goes beyond the broad classification of cell types into three classes and attempts to estimate density profiles of every cell type known by its transcriptome profile. To study the genes collectively we use a voxel-by-gene data matrix E, corresponding to V = 49;742 cubic voxels, and 3,041 genes, as in refs. 24-26. The entry Eðv; gÞ is the expression energy of the gene labeled g in the voxel labeled v [a measure representing the level of mRNA in situ hybridization (1, 10)]. We combine the ABA with cell-type-specific transcriptome profiles, to gain biological understanding of the coexpression patterns of the genes. Our model is based on G = 2;131 genes found in all these datasets and in the coronal ABA. The model proposed to estimate the brain-w...

Bayesian Sparse Regression Analysis Documents the Diversity of Spinal Inhibitory Interneurons

“…This amounts to a non-negative constrained least squares (NNCLS) minimization problem (see Experimental Procedures; Wang et al , 2006; Abbas et al , 2009; Gong et al , 2011; Grange, et al , 2014). But the NNCLS approach fails in this case because, despite the constraint of non-negativity, it generates an infinite number of equally valid solutions.…”

“…These can be divided into two major methodologies. Regression approaches can be applied when the expression profile of cell types of interest is known a priori (Wang et al , 2006; Abbas et al , 2009; Gong et al , 2011; Zuk et al , 2013; Grange, et al , 2014). In contrast, matrix-factorization approaches become relevant when cell type expression profiles are not known (Repsilber et al , 2010; Erkkilä et al , 2010; Bazot et al , 2013; Zhong et al , 2013; Liebner et al , 2014).…”

Reconstruction and Simulation of Neocortical Microcircuitry

“…Although the distribution of protein and genetic markers for different neurons (Grange et al, 2014;Hendry et al, 1989;Kawaguchi and Kubota, 1997;Meyer et al, 2002;Toledo-Rodriguez et al, 2004) and the relative proportions of some morphologically and electrically classified neurons (Beaulieu and Colonnier, 1983;Cauli et al, 1997;Hendry et al, 1984;Meyer et al, 2010a;Rudy et al, 2011) have been described, we lack a comprehensive view of the number of each type of neuron in each layer. Since the advent of paired recording techniques, several studies have characterized the anatomical and physiological properties of synaptic connections between some types of neurons (Cobb et al, 1997;Feldmeyer et al, 1999;Frick et al, 2008;Gupta et al, 2000;Mason et al, 1991;Reyes et al, 1998;Thomson et al, 1993), but a large proportion have yet to be studied.…”

Brain-mapping projects using the common marmoset

“…An important aspect of mesoscopic structural mapping is geneexpression mapping at the brain-wide level (Lein et al, 2006). The spatial co-expression of genes reflects the underlying spatial distribution of cell types (Grange et al, 2014) and provides a complementary data set to combine with a mesoscale connectivity map. As a first step in investigating the spatial co-expression of genes in the marmoset brain, we examined the expression of 26 genes that are known to play an important role in cortical development by performing in situ hybridization in the marmoset brain (Mashiko et al, 2012) using anatomical terminology of the common marmoset brain (Tokuno et al, 2009a,b).…”