There's an interesting commmentary in Nature Cell Biology this month about the sources of variability in microarray studies. Briefly, microarrays (Wikipedia entry) are arrays of spots of short stretches of DNA (it can be other substances as well). Each spot is designed to be complementary to (match; but more importantly, to bind) the product of only a single gene. Using some care, you can put on a complex mixture, and the fluorescence seen at each spot will correspond to the level of expression of that gene. The fluorescence is read by a digitizer directly onto a hard drive for analysis.
The promise of this method is that, in principle, it vastly multiplies the genetic description associated with a particular condition, and specifically brings in the chance that something completely unexpected will fall out of the description of a known cell state. The difficulties have been information management issues and, more importantly, lack of agreement between studies, especially when different hardware is used.
The recent comment in NCB reviews some progress towards standardizing this method toward increasing agreement between different conditions. I think the results won't shock anyone- some labs are able to do the same work on different platforms with good reproducibilty, and others cannot. The articles cited in the comment have a set of procedures affecting both sample preparation and initial data analysis which seem to improve reliability. Although it's really nice that the hardware is not always to blame, it still suggests that microarrays are not going to be nearly as portable as PCR. Perhaps they need to be ceded to big centers just as cDNA cloning is effectively done now.
In my end of the field, people are using microarrays as a screen to identify new genes, which are then followed up one at a time. Given all of the work that needs to be done even once you've gotten a good read, I have heard some thoughtful people suggest you'd be better off picking an interesting sequence at random.
UPDATE 22 June: Ouch. A paper in this weeks' PNAS takes down a prominent Nature Genetics finding from 2003. Ramaswamy et al. compared tumor tissue to healthy tissue and concluded that the metastatic potential of the tumor is contained within the bulk of the cells, a conclusion with pretty big therapeutic implications. The PNAS work goes over the raw data and suggests that this conclusion is highly dependent on the exact threshold chosen while reading in the fluorescent chip. (The link to Ramaswamy et al is the PubMed entry, which shows the paper had attracted some flack already.)