MIAME, we have a problem

Abstract

Microarrays have captured the imagination of geneticistsand molecular biologists like no other technology, with theexception of perhaps PCR. Descended from the humblenorthern blot, which semi-quantitatively measures theexpression level of one transcript at a time, through spotblots, which measures a few transcripts, microarraysclaim to measure several thousand transcripts simul-taneously – spot blots writ largescale! But are they anygood at doing what they claim? Two articles in this issue ofTrends in Genetics [1,2] suggest that confidence in theresults of the current generation of microarray experi-ments is misplaced.In the first article, Miron and Nadon [1] argue for theuse of ‘inferential literacy’ in microarray analysis. Inother words, it is important to have an understanding ofthe characteristics of high through-put data, whichneeds to be considered in the design and execution ofan experiment and not as an optional bolt-on at thedata-analysis stage. However, no amount of statisticalor algorithmic knowledge can compensate fordeficiencies in the technology itself. As Lord Rutherfordis supposed to have said ‘if your experiment needsstatistics, you ought to have done a better experiment.’But it seems that these deficiencies are whispered andnot discussed openly in polite society or in manyscientific journals.The second article by Draghici and colleagues [2] islonger than the normal TIG review, but I feel thespaceisjustifiedbytheimportanceofthemessage.Although microarrays are adequate for detecting thedirection of change in expression of genes expressedabove a certain threshold, a large proportion ofthe transcriptome is beyond the reach of currenttechnology [2]. This problem is glossed over by variousintensity filtering steps employed by researchers doingcomparative microarray studies to remove low inten-sity signals that are ‘unreliable’. Draghici et al. [2]point out that there are inconsistencies between thevarious microarray platforms (in situ synthesised shortoligos, longer oligos, spotted oligos and spottedcDNAs), making it almost impossible, for the moment,to compare results from different platforms. Althougha minimum requirement is that the same transcriptsare being detected by different platforms (amazingly asignificant number of probes intended to interrogatethe same transcript do not), it also seems importantfor cross-platform consistency that the probes ondifferent platforms correspond to the same part ofthe transcript. This is usually attributed to the need toassay the same splice variant, but could it be thatconsistency is improved because the same cross--hybridizing sequences are then detected by all plat-forms [3]?As if the problems associated with different platformswere not enough, a recent trio of articles [4–7] showednot only inconsistencies across platforms but alsoinconsistencies among laboratories that were using thesame platform, and even using the same RNA samples.Matters were improved by the use of common protocolsfor RNA work-up and also, and the importance of this isnot widely appreciated, common methods of datahandling and analysis. If scientists are to create geneexpression databases that incorporate results frommultiple laboratories, it is simply not good enough toadhere to the minimal information about microarrayexperiment (MIAME) guidelines, which only focus on thedocumentation of experimental details, while failing toaddress real problems with the technology and how itis used.Equally depressing is the rush to apply microarraysto obtain ‘gene signatures’ to aid disease diagnosis andprognosis. Again results from different groups studyingostensibly the same disease are frequently non-con-cordant [7,8]. The use of different microarray platformsis partly to blame for this. But perhaps most of theproblem comes from lack of ‘inferential literacy’ meetinglack of epidemiological savvy. The ToxicogenomicsResearch Consortium suggested that more-consistentresults would be achieved not with signatures fromindividual genes but by examining the gene ontology(GO) categories of the differentially expressed genes [6].Perhaps, but it is a sobering comment that when twoRNA samples were compared in different laboratories,on different platforms and analysed in the same way,gene-by-gene list comparisons varied. All that couldbe agreed on were the changes in different GO categories– representative of the tissue of origin of the samples [6].If scientists in different laboratories cannot agree on anordered list of gene-expression differences when pre-sented with the same two RNA samples, we really dohave a problem.So what is the solution? Obviously, putting theright probes on the array would be a start – interrogat-ing the same transcript or splice form is important.Consistent standards between laboratories would helpimprove the consistency of results – but consistency isnot enough – after all the results within a laboratorywere all consistent but the results can be consistentlywrong. What we need is a proper evaluation ofmicroarrays (including sample extraction and work-up,data handling and analysis) and an understanding ofwhat is important to achieve consistent, accurate andreproducible results across laboratories. But perhaps