Searching for evidence of altered gene expression: a comment on statistical analysis of microarray data.

Hilsenbeck et al. (1) have presented an intriguing experimentthat raises important methodologic problems for approaching theassessment of the expression of specific genes. They comparedtumors from animals killed at various times to obtain estrogen-stimulated tumors before tamoxifen treatment, tamoxifen-sensitive tumors during tamoxifen treatment but before acquiredresistance, and tamoxifen-resistant tumors after resumption oftumor growth. They measured expression values for 588 genesfor each of the three tumor types. They argue that a significantdeviation in expression between tumor types for a particulargene represents evidence that the specific gene is altered duringthe development of tamoxifen resistance. In their report, Hilsen-beck et al. define a procedure for identifying “significant” de-viations, and they use this method to identify outlier genes. Infuture studies, they intend to study these outliers to identify“which of the outlier genes are most contributory to the tamoxi-fen-stimulated phenotype....”Before embarking on such a search for outlier genes, oneneeds to know whether different genes need different scales ofmeasurement. Hilsenbeck et al. (1) imply that one can indeedassume the same scale of measurement, i.e., the same variability,for each gene. We believe that replication is essential to deter-mine whether the 588 genes have the same natural scales ofmeasurement. If an individual gene is expressed differentially intamoxifen-resistant or tamoxifen-sensitive tumors, how do weknow that the differential expression reflects the true effect ofthe tumor rather than underlying differential variability? Forexample, erk-2 appears as an outlier in the ES/TS and ES/TRplots, where the variables ES, TS, and TR are estrogen-stimulated, tamoxifen-sensitive, and tamoxifen-resistant, respec-tively; the authors interpret the fact that the expression of erk-2differs in the estrogen-stimulated tumor as evidence that tamoxi-fen modifies erk-2 expression. The observation is, however, alsoconsistent with the quite different hypothesis that erk-2 has morevariability in expression than most of the other 587 genes. In thecase of erk-2, Hilsenbeck et al. have independent verificationthat the gene is involved in the pathway of interest. What aboutthe genes close to erk-2? How do we know that they are fruitfulcandidates for further study? Similarly, a gene that does notappear as an outlier and, hence, is not selected as a candidategene for study may be missed simply because its variability ofexpression is very low compared with that of other genes. Theauthors have some data relevant to assessment of within-genevariability, for they measured five tumors of each type; however,they do not describe whether they used those observations toconfirm their assumption that the same scale is appropriate forall genes.If indeed the same scale of measurement is appropriate for allgenes, the question then arises whether the authors have em-ployed a useful method for detecting outliers. They have chosen