Microarray Data Analysis: Separating the Curd from the Whey

For biologists, DNA microarrays present at once unprecedented opportunities and monumental challenges. In the opportunities column, microarrays produce genome-wide gene expression snapshots, facilitating a migration from gene-by-gene hypothesis-driven research to a relatively unbiased "discovery mode." The challenges broadly include data quality, analysis, and interpretation--that is, reaching an accurate and useful biological conclusion from the correlations identified within the data. Progress on these three fronts will yield substantial dividends both in medicine and in understanding the complex networks of signal pathways that cannot be elucidated through studies of small gene groups in isolation, according to Almut Schulze, a signalling specialist at the Imperial Cancer Research Fund in London. "It would be possible [after solving the data analysis problems] to measure the expression profile of a cell in a certain situation and to be able to deduce information about the activation state of signalling pathways," he says. Until recently the problem of data noise seemed the most intractable. The challenge is to filter out erroneous data through regression and normalization, but this task can be impeded by uncertainty over whether some apparently outlying data values in an expression set really are invalid. In some cases it will be the outliers rather than the main body of the data that are of greatest interest. "We can ask why are we getting dirty data and what can we learn even with the dirty data," says Georges Grinstein, professor and director of the Center for Biomolecular and Medical Informatics at the University of Massachusetts, Lowell. By definition such outlying data points will be small in number, and virtually impossible to detect by generic statistical analysis methods that fail to consider the underlying physical and chemical behavior of the whole microarray process, according to Gustavo Stolovitzky, manager of functional genomics and systems biology at IBM. The only hope of resolving such noise lies in dissembling the different contributing components to the errors, and determining their contribution. Then by various means, such as repeating experiments and subtracting systematic errors, the genuine data representing expression levels can be derived. NOISE CONTROL Several steps in the microarray process contribute to the total data noise. First, because microarrays require a significant amount of RNA, typically 20 to 200 µg per test,1 the RNA samples used to probe the chips must often be amplified. This is often the case for nucleic acids obtained from tissue (as opposed to cell culture) samples. The array itself is also a significant source of error, largely because of the tendency for cross-hybridization, leading to false assessment of expression levels. Research by IBM in conjunction with Columbia University succeeded in separating the contributions of the initial sample preparation, including amplification, from the hybridization, and found that the latter made a far greater contribution to the overall data noise or error.2 Furthermore they found the sample preparation noise was fairly constant and independent of the expression levels of the genes, suggesting that it could readily be filtered out of the data. But cross-hybridization noise has a totally different profile that is much harder to handle, being strongly dependent on expression levels, and varying in a non-linear fashion. This makes it almost impossible to filter out the effects of cross-hybridization through traditional generic statistical methods that don't take the underlying processes involved into account. For this reason IBM, this time working internally at its Thomas J. Watson Research Center in New York, decided to investigate the detailed working of cross-hybridization at the molecular level, in order to predict its effect given knowledge of the sequences involved. IBM concluded that a correlation exists between observed levels of cross-hybridization and the calculated free hybridization energy levels of the associated oligonucleotide (typically between 20 and 200 bases long).3 So given knowledge of the sequences it is possible to calculate the contribution to the total error made by cross-hybridization. The paper describes an algorithm IBM developed to calculate the effects of cross-hybridization, and identify outlying data points, which can then be removed by an additional technique. But there is a caveat: IBM admits that the theoretical levels of cross-hybridization tend to be greater than the levels measured by the experiment. This is partly because the calculations assume that the DNA probe and target strands of RNA are either completely hybridized or not at all, while in practice some are only partially hybridized. Although the joint IBM/Columbia University study found that errors resulting from RNA amplification in sample preparation are relatively small, they can still be significant in some cases. The amplification error for a gene is obtained by measuring the ratio of its RNA expression level to the total amount of RNA in the system both before and after amplification. The extent to which this ratio has changed is a measure of the amplification error. In the absence of any error, this ratio would be unchanged by amplification. But in a quantitative study carried out in Norway,1 the expression ratios of 10% out of 4,331 genes in an experimental system were found to be significantly altered by the amplification process. RNA can be amplified in several ways, and as the Norwegian study admits, more work needs to be done to determine the impact of amplification for all of them. Also, work needs to be done to identify the genes most susceptible to amplification error, given that in the Norwegian study, expression ratios were virtually unaltered for 90% of the genes. DATA INTERPRETATION Even if the huge task of cleansing the data can be performed satisfactorily, there remains the almost equally daunting challenge of making sense of it. For large global investigations potentially spanning an organism's whole genome, the first task may be to identify clusters of genes whose expression patterns are closely correlated, either over a period of time (for example while the system is under the influence of a drug), or spatially across multiple tissues. A variety of statistical analysis packages are available for the immediate task of identifying correlated clusters, but the harder problem is then to identify what the common biological properties are. The best hope of reaching a speedy conclusion lies in the existing literature--either reports of other (possibly related) micro-array experiments, or studies that elucidated similar correlations in some different way. Researchers have had some success in this arena, using methods based either on keyword analysis4 or the frequency of commonly occurring terms within the text.5 Such techniques have the advantage of being unbiased and requiring no prior knowledge of the gene functions. But they cannot shed any light on gene expression clusters for which no published literature exists or can be found. In many cases the first port of call in the quest for biological meaning will be one of the fast-expanding compendiums of microarray data. Of these, one of the largest and best established is the Stanford Microarray Database, which now holds data on 37,000 microarrays, totalling about 1 billion spots (each representing expression levels of an individual gene in the experiments). Clearly such databases are far more useful if they provide direct links to publications and also allow searches for specific types of experiments based on straightforward keywords. "Somebody should be able to ask for liver, and find all experiments associated with liver," says Gavin Sherlock, director of microarray informatics at Stanford. "That's not a trivial task, because something might be labelled as pavo carcinoma and we need to know that's liver cancer." Another database issue is performance. Currently the Stanford database takes five hours to retrieve data from 1,000 micro-arrays, which is admittedly a larger than average requirement at present. "If that could be reduced to half an hour, that would have a huge impact on research," Sherlock admits. For this reason, Stanford hopes to upgrade its database storage hardware to achieve the desired order of magnitude increase in search and retrieval speed. Such developments will help usher in the era of full genome-scale microarray experiments. But it is not certain yet that large-scale microarrays will prove sufficiently reliable for medical applications such as cancer diagnosis, where the main requisites are accuracy and speed, often over a limited number of genes. For such applications, the data quality issues of microarrays may yet prove insurmountable, and emerging protein arrays may win the day. Philip Hunter (phunter@philiphunter.com) is a freelance writer in London. References 1. V. Nygaard et al., "Effects of mRNA amplification on gene expression ratios in cDNA experiments estimated by analysis of variance," BMC Genomics, 4:11, March 23, 2003. 2. Y. Tu et al., "Quantitative noise analysis for gene expression microarray experiments," Proc Natl Acad Sci, 99:14031-6, 2002. 3. G.A. Held et al., "Modeling of DNA microarray data by using physical properties of hybridization," Proc Natl Acad Sci, 100:7575-80, June 24, 2003. 4. D.R. Masys et al., "Use of keyword hierarchies to interpret gene expression patterns," Bioinformatics, 17:319-26, 2001. 5. D. Chaussabel, A. Sher, "Mining microarray expression data by literature profiling," Genome Biol, 3:research0055.1-research0055.16, 2002. function sendData() { document.frm.pathName.value = location.pathname; result = false if (document.frm.score[0].checked) result = true; if (document.frm.score[1].checked) result = true; if (document.frm.score[2].checked) result = true; if (document.frm.score[3].checked) result = true; if (document.frm.score[4].checked) result = true; if (!result) alert("Please select") return result; } .myradio{background-color : #94bee5} Please indicate on a 1 - 5 scale how strongly you would recommend this article to your colleagues? Not recommended 1 2 3 4 5 Highly recommended Please register your vote

Microarray Data Analysis: Separating the Curd from the Whey

Interested in reading more?

Become a Member of