Van Eyk, Dunn - Proteomic and Genomic Analysis of Cardiovascular Disease - 2003 (522919), страница 40
Текст из файла (страница 40)
A moderately high quality digitalcamera may have three million pixels. Each pixel is capable of showing threechannels of color at one byte or 24 bits per channel, or about 16 ´ 1018,000,000 possible combinations in one digital microscope image. In contrast, if we split therange of expression of one gene into 8 different bins, we can calculate the maximum number of potential observable expression states. Thus, the complexity ofexpression genomics is 8 (different expression levels for each gene) raised to thepower of 35 ´ 103 (the rough number of genes) [7].
However, it appears that cellsrarely express more than 2 ´ 104 genes, and probably less than 103 genes actuallyare involved in any specific cell response. Thus, for any comparison between twocells, it is not likely that one has to think about more than 80,000 different possible phenotypes. This number, 80 ´ 103, while still impressive, is easier to imaginethan 16 ´ 1018,000,000.Tumor biologists have already begun to use arrays to classify specific neoplasms[8], or to define two functional states of a cell, e.g. comparing a metastatic cellwith its non metastatic parent [8].
A particularly impressive paper was publishedby Golub et al., comparing acute myelogenous leukemia with acute lymphatic leukemia [9]. Where the pathologist might make a decision based on subjective criteria only, the objective diagnoses by Golub et al. were made by looking at these7,000 transcription units, quantifying the amount of RNA made by all of these,doing this in a large number of cases, and showing that a pattern emerged. Fiftygenes were sufficient to define the two types of leukemia.8.2.4Methodological Biases in Array InterpretationOne common problem in array analysis is the use of biased arrays that focus onspecific sets of genes identified in such cellular processes as inflammation, development, replication, etc.
If, for example, one looks at an array limited to “inflammation”, we will inevitably find multiple members driven by NFjB. For example,McCaffrey et al. [10] used expression arrays focused on inflammation to show thatatherosclerotic plaques are enriched in transcripts driven by the EGR1 and EGR21311328 Skeptical Analysis of Arraystranscription factors downstream of TGF-b and PDGF, two growth factors believed to play a critical role in the formation of the plaque [10]. The use of a focused array with only 568 genes created the implicit assumption that the geneson the array are more important to the biology of interest than the other approximately 35,000 genes in the genome.8.2.5Usefulness of Clustered GenesThe use of some form of unbiased pattern recognition has become the most common approach to analysis of array data.
While the mathematics are beyond thescope of this review, it is useful to offer a somewhat general description of clustering and related analytical methods. Tab. 8.2 shows a hypothetical data set of thesort typically used in cluster analyses. With this kind of data set, one may clusterexperiments – e.g. identify experiments (or samples) in which genes behave similarly or identify genes that behave alike across experiments. Experiments 1 and“N” show identical values for all genes and hence these two experiments are moreclosely related to each other than they are to experiment 2. An alternativeapproach is to cluster genes; this is equivalent to looking for rows in Tab.
8.2 thatshow similar patterns. For example, genes 1 and 4 behaved similarly in these experiments and, we presume, have similar promoters.Clustering may be either biased or unbiased. Biased analyses ask about predetermined lists of genes, e.g.
genes belonging to specific functional classes (“showme genes that behave like these”). Unbiased analyses simply look for common expression patterns irrespective of previous knowledge. The former question mightalso be called “hypothesis-based research” and the latter, “systematic exploration”.Biased clustering analysis can be useful to help identify genes which behave similarly to specific genes of interest. This approach must be used with caution as hasalready been discussed above because of the “Kevin Bacon” effect. We also need tobe aware that a gene of interest may not be representative of its function or thefunction of genes under similar control.Tab. 8.2 Clustering by rows versus columns. An example of a hypothetical data set that may beused in a clustering analysis.
Rows represent genes, columns represent experiments or samples,and values in the table represent absolute or relative expression measurementsCluster by experimentClustery gy Gene" GeneGene" GeneGeneGene12345Expt. 1Expt. 2Expt. 3121–1.41.26161.01.0121–1.41.28.2 Applications of Array Analysis8.2.6False AssociationsA major concern with all approaches to cluster analysis is the lack of statistical criteria to decide which clusters are valid. The work of Golub et al.
cited above is instructive. Their ability to use the set of genes identified by a self-organizing mapto differentiate two kinds of leukemia was determined prospectively by demonstrating that this pattern could be used with known cases. It is important to realize, however, that no statistical method existed to give the investigators proof thatthis cluster was diagnostic. We do not know whether other patterns might havebeen even more successful. “SAM” (daisy.Stanford.edu/MicroArray/SMD/research.html), is an attempt to deal with this problem by scoring each gene on thebasis of expression relative to standard deviation of repeated measurements (in array data) [11].
Unfortunately, SAM and similar methods provide only an arbitraryway of setting a threshold. In effect, one can increase or decrease the number ofacceptable genes but we still lack any rigorous test that would allow us to statethat the possibility of a specific set of genes being differentially expressed is at aspecific p value [12].8.2.7Hypotheses Based on Expression PhenotypesThe task of “hypothesis-based” research then becomes identification of the mechanisms controlling an expression pattern. For example, Clark et al. used expressionphenotypes to define melanoma cells selected for their metastatic capacity by serial harvesting of cells that metastasize after implantation of the tumor in mice [8].Besides defining the metastatic phenotype, they were able to identify a specificmember of the Rho small GTPase family as being differentially expressed.
Theywere also able to show that RhoC alone was necessary and sufficient to supportmetastasis and that the closely related RhoA or B were not differentially expressedand were not needed to support metastasis. It would have been very interesting ifthey had gone on to tell us whether or not the rest of the metastatic expressionphenotype was RhoC-dependent.Another example of the hypothesis-based approach to analysis of array wastaken by Svaren et al. in their study of the role of EGR1 in prostate cancer [13].These authors began by using adenovirus to transfect prostatic cells with EGR1,then analyzed the effects with an array containing 5,600 randomly chosen genes.Genes identified as EGR1-dependent were subsequently confirmed by quantitativePCR in cancer tissue.
Intriguingly, many of these mapped to the 11p15 chromosomal locus, an imprinted region implicated in other cancers. One such gene,IGF-II has been implicated in prostate cancer [13].1331348 Skeptical Analysis of Arrays8.2.8Multiple Venn Diagrams: Decreasing the Size of the CatalogA review of the array literature in preparation for this paper suggests that themost successful way to reduce the size of the data set produced by an array studyto a more manageable and often more meaningful list is to use multiple comparisons between strategically chosen data sets, including data independent of the array itself.
As more comparisons are done, fewer genes will achieve significance.For example, a recent paper in the diabetes literature compared the expressionpatterns of adipose tissue in several strains of obese rats with the expression patterns of the same rats made hyperglycemic by streptozoticin (Fig. 8.2) [14]. Theuse of multiple strains of obese animals, even though each strain was studiedonly once, led to identification of only 92 genes that marked the diabetic-obese animals. It is likely that these numbers would be further reduced by a study thattakes more account of animal-to-animal variability, or even more powerful byusing additional Venn diagrams to focus attention on those genes that are regulated by diabetes alone and then comparing that group with the two publishedstudies.As another example, positional cloning within the areas identified by geneticstudies [15] is difficult unless there are obvious candidate genes.
Fig. 8.3 suggestsone possible resolution to this problem. Any over- or under-expressed gene foundin a locus associated with the disease phenotype would become an obvious candidate, as would any transcription factor located in such a locus and known to control genes showing a clustered expression pattern in the array.
Mirnics and collaborators compared expression of mRNA from brains of hospitalized patients withand without schizophrenia with that of patients with an unrelated psychiatric disorder [16]. Monkeys treated with antipsychotic drugs provided a further control.Only one gene, RGS4, showed consistent change in expression.
RGS4 was theonly gene in their study to map to the major schizophrenia susceptibility locus1q21–22.Venn Diagram. Theuse of multiple types of data isa powerful tool to filter expression arrays.Fig. 8.28.3 Analysis of Measurement ErrorsVenn diagram of transcriptome versus genome. Use of transcription patterns tomap multiple genomic loci. These Venn diagrams show comparisons of genes (on theright) with transcripts (on the left). The firstset of intersects include those genes that aredifferent in sequence (right) versus expression (left), producing the “subsets of interest”.
The third level set shows the intersec-Fig. 8.3tion of the genomic subset with the expression subset. Differences in transcription thatmap to differences in the genome are suspectas sites where mutation has affected a promoter, mRNA stability, or message function.At a more sophisticated level, differences intranscription may soon be mapped to geneticdifference marking functional mutations intranscription factors, receptors, or cytokines.8.3Analysis of Measurement ErrorsNo review of microarray technology would be complete without a discussion ofappropriate criteria for quality control. Microarray technology is essentially ahighly parallel form of a dot blot.
Correlations with assays by other methods, e.g.Northern blot, are surprisingly good [17, 18]. Unfortunately, even when mostgenes are accurately measured, data for individual genes may be discrepant. Inone study, four of forty-six differences substantiated by Northern analysis werenot demonstrable in the array [19]. This error may reflect the inability of a dot onan array to distinguish between related genes, splice forms, or alternative polyadenylation sites. Another problem is the error due to “non specific” cross hybridization. In one study in our lab we found differential hybridization at a spot thatturned out to carry an ALU-rich sequence.
![](https://s.studizba.com/z.php?f=/templates/si/i/noavatar.png&w=100&h=100&t=1"){kind=avatar}
