Multi-scale correlation structure of gene expression in the brain

Abstract

The mammalian brain is best understood as a multi-scale hierarchical neural system, in the sense that connection and function occur on multiple scales from micro to macro. Modern genomic-scale expression profiling can provide insight into methodologies that elucidate this architecture. We present a methodology for understanding the relationship of gene expression and neuroanatomy based on correlation between gene expression profiles across tissue samples. A resulting tool, NeuroBlast, can identify networks of genes co-expressed within or across neuroanatomic structures. The method applies to any data modality that can be mapped with sufficient spatial resolution, and provides a computation technique to elucidate neuroanatomy via patterns of gene expression on spatial and temporal scales. In addition, from the perspective of spatial location, we discuss a complementary technique that identifies gene classes that contribute to defining anatomic patterns.

Introduction

An important question common to molecular neurobiology and neuroanatomy is whether regionalized gene expression patterns correspond to or define the basic architecture of the brain (Bota et al., 2005, Hawrylycz et al., 2010, Swanson, 2003, Thompson et al., 2008). A study based on scanning the Allen Brain Atlas (www.brain-map.org) of more than 20,000 genes in the C57Bl/6J mouse indicates that about 78.8% are expressed at some level in the adult murine brain (Lein et al., 2007). Although gene expression is a dynamic process, the spatial expression patterns at a specific time point may still reflect the anatomy and function of a region. This problem is complicated due to the wide range of expression patterns, ranging from more globally expressed transcripts to highly specialized regional markers. The combinatorial complexity of expression patterns and common activity of gene networks indicate that statistical approaches may be useful in analyzing gene expression data. However, asking which genes in general are expressed in a specific region of the brain may shed little light on the key genes that are most specific for the reticular nucleus nor indicate their relationship to genes relevant to the anatomy of that region of the brain. Developing a means to examine and compare expression patterns of genes and spatial domains could aid in relating gene function with neuroanatomy.

A key to this approach is to spatially map genomic-scale data, which is becoming more feasible on a large scale in a variety of model organisms as well as in humans (Sunkin & Hohmann, 2007). This mapping process effectively begins with voxelation of a standardized space (Lein et al., 2007, Petyuk et al., 2010) and represents a method for capturing the acquisition of three-dimensional (3D) gene expression patterns in the brain. It can be accomplished by direct tissue segmentation (Petyuk et al., 2010) for studying genomics or proteomics or virtually, by constructing a grid of in silico mapped voxels of image data (Lein et al., 2007, Ng et al., 2009). In either case, high-throughput analysis of spatially registered voxels produces maps of gene expression analogous to the anatomical images reconstructed in medical imaging systems. These maps can then be analyzed using clustering, dimension reduction, and other techniques to understand how the genome constructs the anatomy of the brain.

If many genes (or other modalities) are mapped at sufficient spatial resolution, then one can simultaneously consider two related concepts. For a given gene, it becomes possible to find all genes which are highly spatially correlated with that gene. Here, ‘correlation’ means any metric of similarity such as Pearson correlation, mutual information, or other measures. Similarity of gene expression across tissues is a common technique for examining expression profiles and studying gene expression in terms of “guilt by association” (Stuart, Segal, Koller, & Kim, 2003). This enables one to study variance in both gene expression values and patterns across an anatomic landscape. The notion is that genes sharing a common expression pattern may act together in a network and have a common role in building or influencing activity in the region where they are co-expressed.

An analogous and dual problem arises when the expression data are viewed from the spatial context. For a given spatial location, one can ask how gene expression varies in the spatial neighborhood of that location. While neuroanatomists have used gene expression data to guide their understanding of brain architecture for some time (Dong, 2008), only more recently have integrated studies over large classes of genes been possible (Dong et al., 2009, Ng et al., 2009, Thompson et al., 2008). If data are spatially mapped, these two fundamental approaches connecting gene expression and anatomy are dual to each other and one can develop powerful tools to navigate between sets of correlated genes and delineations of anatomic regions. We illustrate these concepts with two tools from the Allen Brain Atlas, NeuroBlast and the Anatomic Gene Expression Atlas, described below.