Whole-genome expression analysis: challenges beyond clustering

Abstract

Measuring the expression of most or all of the genes in a biological system raises major analytic challenges. A wealth of recent reports uses microarray expression data to examine diverse biological phenomena — from basic processes in model organisms to complex aspects of human disease. After an initial flurry of methods for clustering the data on the basis of similarity, the field has recognized some longer-term challenges. Firstly, there are efforts to understand the sources of noise and variation in microarray experiments in order to increase the biological signal. Secondly, there are efforts to combine expression data with other sources of information to improve the range and quality of conclusions that can be drawn. Finally, techniques are now emerging to reconstruct networks of genetic interactions in order to create integrated and systematic models of biological systems.

Introduction

The enthusiasm about microarray expression data analysis in the bioinformatics community has been remarkable. The peer-reviewed conference proceedings in the field have often provided the initial presentation of new methods, including the early application of clustering [1], linear decomposition [2] and algorithms to discern genetic networks 3•., 4•., 5•.. (All references to the Pacific Symposium on Biocomputing can be found at http://www.smi.stanford.edu/projects/helix/psb-online/) The public release of expression data sets 6., 7., 8. created a de facto set of benchmarks for analysis by the bioinformatics community. There remains a risk, however, that the community has tuned these algorithms to perform well on this small set of training examples and that the algorithms will not perform well on entirely new data sets. Thus, the continued release of data from different groups using different detailed methods, and even measurements from redundant experiments, will be critical [9].

In a typical array experiment, many genes (frequently all known) in an organism are assayed under multiple conditions. The data can be represented as a matrix in which the rows are genes and the columns are conditions. These conditions may be different time points during a biological process, such as the yeast cell cycle 7., 8. and Drosophila development [10], or they can be different tissue samples with some common phenotype, such as tissue type or malignancy. Although the amount of data generated in an expression experiment is tremendous, this is not yet a data-rich analytical task by statistical standards. The complexity of genomic systems, with N genes and thus N² potential pairwise interactions (not to mention higher order interactions), is even larger than the expression data sets and thus the ratio of data to unknown variables is still small. The major initial efforts at clustering and linear decomposition (such as principal components analysis) not only assist humans in understanding the data, but also demonstrate that the amount of independent new information may be much smaller than the number of raw data points suggests 2., 11.. (Some microarray analysis tools are available at http://classify.stanford.edu/)

Whole-genome expression data affect structural biology by providing valuable functional information about when and where a protein is expressed, when it is degraded and with which other proteins it may interact. Early work has surveyed the ability of expression data to yield clues about common sequential/structural motifs for regulatory elements (as reviewed below). It also addresses issues such as protein localization or the justifiability of predicting function using ‘guilt-by-association’ techniques, whereby similar expression may be a component of the association (as reviewed below). Jansen and Gerstein [12] have analyzed the sequential and structural features of highly expressed genes and found biases (more alanine/glycine, less asparagine, shorter sequences, more TIM barrels) in a group of highly expressed proteins.

Although not the main focus of this review, there has been a satisfying focus on maximizing the reproducibility and analyzability of microarray experiments 9., 13., 14., 15.. The ‘fold difference’ is widely used as a quantitative measure of the differential expression. The fold difference is the ratio of the expression in cells of interest versus the control cells. Genes expressed at low levels require higher fold differences in order to rise above the noise 16•., 17. and duplicate measurements of identical experiments can be very valuable for reducing noise and simplifying subsequent analysis [9]. There are also emerging methods for assigning confidence to differentially expressed genes [18]. The noise in expression data can confound analyses and rank data are often more robust than absolute measured values because of the variation in methods for subtracting out background noise and quantifying expression levels 19., 20.. Methods have emerged for imputing missing data in incomplete data sets (O Troyanskaya et al., unpublished data; see Now in press [78]). Finally, aneuploidy (and therefore the number of copies of a gene in the effective genome) has been shown to affect the expression level of a gene, either confounding the analysis or providing insight into the mechanisms of abnormal biology 21., 22..

The remainder of our review is organized around the result of a hierarchical clustering of the literature, in which the word counts are the features of articles used to cluster them, as shown in Fig. 1.