[18] Interpreting Experimental Results Using Gene Ontologies

Abstract

High‐throughput experimental techniques, such as microarrays, produce large amounts of data and knowledge about gene expression levels. However, interpretation of these data and turning it into biologically meaningful knowledge can be challenging. Frequently the output of such an analysis is a list of significant genes or a ranked list of genes. In the case of DNA microarray studies, data analysis often leads to lists of hundreds of differentially expressed genes. Also, clustering of gene expression data may lead to clusters of tens to hundreds of genes. These data are of little use if one is not able to interpret the results in a biological context. The Gene Ontology Consortium provides a controlled vocabulary to annotate the biological knowledge we have or that is predicted for a given gene. The Gene Ontologies (GOs) are organized as a hierarchy of annotation terms that facilitate an analysis and interpretation at different levels. The top‐level ontologies are molecular function, biological process, and cellular component. Several annotation databases for genes of different organisms exist. This chapter describes how to use GO in order to help biologically interpret the lists of genes resulting from high‐throughput experiments. It describes some statistical methods to find significantly over‐ or underrepresented GO terms within a list of genes and describes some tools and how to use them in order to do such an analysis. This chapter focuses primarily on the tool GOstat (http://gostat.wehi.edu.au). Other tools exist that enable similar analyses, but are not described in detail here.

Introduction

DNA microarrays are a high‐throughput experimental technique used for gene expression analysis (Lockhart 1996, Schena 1995). Such high‐throughput techniques have greatly facilitated the discovery of new biological knowledge. However, this kind of knowledge is often difficult to grasp, and turning raw microarray data into biological understanding is by no means a simple task. Classical analysis of microarray data is based critically on comparing gene expression at the level of single genes and determining the significantly differentially expressed genes (Ayroles 2006, Beissbarth 2000, Downey 2006, Saeed 2006, Smyth 2004) or on clustering techniques used to determine genes with a similar expression pattern in a series of experiments (Eisen et al., 2000; Gollub 2006, Saeed 2006). In other popular techniques, such as serial analysis of gene expression (SAGE), the focus is often just to find the transcripts expressed in a given tissue (Beissbarth et al., 2004). In each case, results can be expressed as a list of genes or as a sorted list of all genes on the array ranked by a score (see Fig. 1).

In the more recent literature, many groups came up with methods that use prior biological knowledge in order to help interpret these lists of genes (e.g., Al‐Shahrour 2004, Beissbarth 2004, Dennis 2003, Doniger 2003, Yue 2005). Together with the application of clustering techniques to microarray data the term “guilt by association” has been used, indicating that by using prior biological knowledge it may be possible to infer the function of coexpressed genes (Clare 2002, Quackenbush 2003). Furthermore, with a single experiment producing a list of hundreds of differentially expressed genes, automatic annotation and grouping of genes are necessary merely to make interpretation of the results possible (Beissbarth 2003, Boon 2004). In addition, the grouping of genes and subsequent testing for the differential expression of these groups of genes may help find significant biological processes even in cases where single gene analysis fails to produce significant results.

The main problem is how to incorporate prior biological knowledge and where to get it from. The type of knowledge that comes to mind first that would be useful to incorporate is that of biological pathways. However, the information on this as stored in pathway databases such as the Kyoto Encyclopedia of Genes and Genomes (KEGG) (Kanehisa 2000, Mao 2005) is quite sparse. Other sources of information could be SWISS‐Prot key words, presence of transcription factor‐binding sites, protein domains, etc. Further, it is important to organize this information in a structured way. Ontologies are a commonly used concept in computer science (Draghici et al., 2003). They are used to define terms in a controlled vocabulary as well as relations between those terms. They can also be viewed as a directed acyclic graph, which defines a hierarchy of terms. The Gene Ontology Consortium (http://www.geneontology.org) is a worldwide consortium that defines an ontology that can be used to annotate genes (Ashburner et al., 2000). Furthermore, there are a number of Gene Ontology (GO) databases, where annotation for genes of a certain organism or several organisms are stored (e.g., Blake 2003, Camon 2004). The hierarchical structure of GO annotations is illustrated in Fig. 2.

A biological annotation can be expressed as a gene set. Gene ontology annotations can be cut at different levels, allowing the definition of gene sets for each GO term, leading to a hierarchy of gene sets. The subsequent statistical test for whether a certain biological function (or gene set) is associated with the experimental outcome of a microarray is often referred to as gene set enrichment analysis (Lamb 2003, Mootha 2003), and analysis with GO groups has become widely used for the analysis of microarray data (also see Gollub 2006, Hennetin 2006, Whetzel 2006). This analysis allows testing for gene sets (or functional groups) that are represented significantly more frequently in a list of genes that appears interesting in a microarray experiment than in a comparable list of genes that would be selected randomly from all genes on the microarray. This then allows the association of functions to a list of genes.