Gene expression patterns can be used to evaluate the consequences of, and predict response to, chemotherapy. (A) Gene expression changes after short-term in vivo drug exposure can potentially reveal molecular mechanisms of action. RNA isolated from purified plasma cells before and after 48 hours in vivo treatment with PS 341 (VEL, n = 15), CC-5013 (REV, n = 13), dexamethasone (DEX, n = 20), or thalidomide (THAL, n = 18) was applied to U95Av2 microarrays (Affymetrix, Santa Clara, CA). Significant induction or suppression of gene expression after drug treatment was calculated on the basis of the percentage of change from baseline ([follow-up – baseline]/baseline). The top 15 genes in each drug group were selected according to rank P value and plotted on the basis of the percentage of change (red = increase in expression; green = decrease in expression). Drug treatment samples are grouped in columns, and genes are in rows. Note that genes up-regulated by dexamethasone (first 15 genes from the top) tend to be unique to that drug. Also note that most of the genes up-regulated by CC-5013 are also up-regulated by thalidomide, thus indicating possible common molecular targets for these 2 drug analogs. (B) Gene cluster of normalized expression values of a group of genes demonstrating differences between myeloma cases exhibiting response or no response to PS 341 proteasome inhibitor therapy. Plasma cells were purified from bone marrow aspirates from 40 patients before the initiation of therapy and after informed consent. Total RNA was isolated, and gene expression levels of approximately 12 000 genes were analyzed using the Affymetrix U95Av2 GeneChip microarray. After sufficient follow-up, 21 responders (exhibiting at least a 50% reduction in serum M protein) and 19 nonresponders (exhibiting progression, stabilization, or no > 25% reduction in M protein) were identified. With the use of a combination of chi square, Wilcoxon rank, and discriminant analysis, 40 genes were identified as being expressed higher in the nonresponder group and 42 higher in the responder group. Genes are in rows, and patient samples are in columns. The nonresponders are grouped under the green bar and responders under the red bar. The table below panel B shows the correlation of results from a gene expression–based prediction model of response and actual response to PS 341 therapy by using the same 40 patients analyzed in panel B. The results presented in the matrix were derived by first assessing the mean difference for all 12 625 probe sets between responders and nonresponders, and then obtaining P values and adjusted P values based on permutation tests (20 000 samples). The 40 samples were then randomly grouped into 5 strata of n = 8. Using 30 of the most significantly differentially expressed genes, a prediction model was developed by using stepwise logistic regression on data from 4 of the 5 strata (80% of total data, n = 32). The model developed on 80% of the data was applied to the 20% (n = 8) left out as a validation group. With 5 strata we developed 5 separate models and validated them on 5 separate validation samples. We averaged over the 5 validation samples to get an estimate of 88% (35 of 40 correctly classified).