The power and promise of microarrays are vast. Offering the ability to run tens of thousands of experiments in parallel on a small glass slide, gene chips have transformed functional genomics, whether through expression analysis or genotyping, from fantasy to reality with dizzying speed.

And yet, longstanding doubts persist as to the worth of the torrent of data that microarrays produce. Experiments have proved difficult to reproduce, and the lists of genes found in similar studies often have only limited overlap. Myriad protocols and platforms, and the proliferation of homemade arrays, means that there is no one standardized set of procedures to follow when designing or replicating an experiment. Some analyses have found cross-platform reproducibility to be poor.1

But, microarray technology isn't as flawed as it was feared to be. At least, those are the findings of a trio of papers published in the May issue of Nature...


The recent studies confirmed three persistent criticisms: That the bewildering array of platforms and research protocols available can make results from different studies hard to compare; that, in the hands of less-experiened labs, homemade arrays are less dependable than commercial chips; and that different labs doing the same study can often get very different results.

But the studies also had some surprisingly good news for the microarray community. First, they found that the various platforms, both commercial and homemade, could all deliver good results in experienced hands. They found also that standardized research protocols go a long way toward increasing reproducibility. And perhaps most importantly, they determined that the vast bulk of microarray data is driven by underlying biology, even in cases where platforms differ over a particular gene's expression.

Despite efforts in the three papers and other studies to crown a victor, no one microarray platform or set of experimental protocols has yet emerged as the best. Still, there are lessons to be learned.


If you're just getting into the microarray research game, and you don't have access to a core facility, consider using a commercial chip. The experimental protocols will be clearer and more standardized. Labs can achieve good results with in-house spotted arrays, but the recent studies indicate that success with homemade arrays depends largely on expertise and experience. And price is no longer a deal-breaker: Costs have come down in recent years to the point where paying more for a commercial chip might appear competitive with the cost of equipping a lab to produce spotted arrays.

"If you're good at tinkering and making something work and you have a lot of time to do that, you might want to go with one of these homegrown platforms. You'll save money and maybe get better results. But if you just want to stick something in some hole and press a button, you're probably better off with an industrial product," says Rafael Irizarry, an author on one of the Nature Methods papers3 and a biostatistician at the Johns Hopkins Bloomberg School of Public Health.



Courtesy of Affymetrix

This image of raw Affymetrix GeneChip array data highlights the problem microarray users face as they try to make sense of their gene expression figures: How likely is it that another researcher, in another lab, using (possibly) a different microarray, will see the same expression patterns as this array yields?

"If you want to collaborate with another researcher, you want to consider using the same platform," says Brenda Weis, research coordinator for the Toxicogenomics Research Consortium that produced one of the three papers.2 "That's a good way to make sure at the end of the day that your data can be compiled." Also, stick to identical experimental protocols and data-normalization procedures: The more you can standardize your methodology, the less you will have to contend with the lab effect when comparing your data.


Don't expect the magnitudes of individual gene expressions to be comparable across platforms; they're usually not. You can make more valid comparisons by looking at relative, rather than absolute, gene expression.5

For a more biologically meaningful and statistically robust approach to data analysis, look at the level of the biological process (i.e., the Gene Ontology (GO) annotation) rather than the gene. "Our paper2 showed that the biological representations of the genes are more important than the genes themselves. Even if on a gene-by-gene basis we found inconsistency, when we looked at the processes themselves, those conserved processes are there," says Weis.

Looking at the process level allows researchers to make sense of inherent variation, says Aviv Regev, a research fellow at Harvard's Bauer Center for Genomics Research and coauthor of a recent Nature Genetics paper that proposed a process-level method of looking at microarray data.6 "When you analyze things at the level of gene sets, there is noise at the level of individual sets. This might not only be a result of methodological problems with your microarrays, they might be inherent biological variability," she says.



© 2005 Nature Publishing Group

More than half of the variability in gene expression measurements between labs and platforms is attributable to the microarray platform itself. Contributions of different sources of variability were estimated using an analysis of variance (ANOVA) mixed model. (From Bammler et al., Nat Methods, 2(5):351–6, May 2005.)

One of the major barriers to the reproducibility of microarray data is a basic one: Given the plethora of platforms and protocols, it's often hard to precisely reconstruct the experimental methods used in a study from the published paper.

Several years ago, MGED developed its standard for data reporting, called Minimal Information about a Microarray Experiment (MIAME), to cut down on the widespread methodological confusion. The guidelines require researchers to go above and beyond what would ordinarily be required in a methods section, including detailed descriptions of the protocols used in RNA extraction, labeling, and hybridization; the data normalization algorithms used in preprocessing; and the design of the arrays themselves. MGED has also called for researchers to submit their raw data to public databases such as the Gene Expression Omnibus and ArrayExpress.

Although many specialized journals have yet to adopt MIAME, the standard is increasingly becoming the law of the land.7Nature, Cell, and The Lancet adopted MIAME compliance requirements for their authors in 2002, and many other journals have followed suit. Several commercial array manufacturers and software developers have written protocols and software to help their users follow MIAME guidelines. (See, for instance, http://www.mged.org/Workgroups/MIAME/miame_software.html for a list) "Most of the requirements are difficult to comply with if you don't have the tools," says Regev.


But while standards for data reporting are becoming well established, the same is not true for data collection, says Chris Stoeckert, associate professor at the University of Pennsylvania's Center for Bioinformatics and a member of MGED's advisory board. As they seek to develop best practices for experimentation, MGED will be looking at studies like the ones recently published in Nature Methods to shed light on what remains a murky issue.

"We all want to be able to look at a microarray experiment and assess whether it was done correctly or not. To a large degree we are unable to do that, because there are no standards yet as to how to assess it," Stoeckert says. "We need standards on the informatics side – how do you evaluate the data that's there – and also standards in terms of how to experimentally control for quality."

The intense spotlight currently being cast on microarray research may shed light on other research as well. All researchers can learn a few lessons on the importance of the lab effect, says Irizarry. "It's been known forever, not just in microarrays, that there is a lab effect. There's a paper I cite where the speed of light is plotted from different labs across time, and the difference between labs is statistically significant," he says.8 "I think microarrays are getting a bum rap because of one of their strengths, which is that they produce a lot of data. I think if other technology produced this much data, we would start seeing similar problems,"

Irizarry sees the beginnings of hopeful trends toward standardization in the research community. For instance, many universities are moving towards using core facilities rather than processing microarrays in individual labs. "In academia, you are seeing a movement towards standardization," he says. "Given what I've seen in terms of the lab effect, this is a move in the right direction."

Interested in reading more?

Magaizne Cover

Become a Member of

Receive full access to digital editions of The Scientist, as well as TS Digest, feature stories, more than 35 years of archives, and much more!
Already a member?