Fig. 1. Genes are stretches of DNA that give rise to proteins. Usually, one gene gives rise to one protein. Because of polymorphism in the population, genes that give rise to the same protein can have slight variations in the genetic code. Each variant is called an ‘allele’. Each gene is transcribed into mRNA, which is then translated into protein. Each protein has its own, specific mRNA. For simplicity, genes without introns (sequences within genes that are not translated into protein) are shown.

Fig. 2. (A) mRNA is synthesized with the help of RNA polymerase and with DNA serving as template. (B) Gene array experiments are performed on a substrate (array) to which oligonucleotides or cDNAs are tethered. The labeled target that is complementary to the oligonucleotide or cDNA will attach (hybridize) to it, whereas unbound sample will be washed off. The label is then used to visualize the intensity of the reaction. The intensity should be correlated with the amount of target in the sample.

Fig. 3. Data preparation for oligonucleotide arrays. After hybridization and staining is completed, the array is scanned and raw data are obtained. (A) Two hypothetical samples are hybridized to two separate arrays, with probe sets for four genes (w, x, y, z). Each gene is represented by eight probe pairs, with the upper cell of each pair as the ‘perfect match’ (PM), and the lower cell as the ‘mismatch’ (MM). After background is subtracted from all cells, both arrays are scaled to the same level of brightness (‘normalization’). Genes x, y, and z have the same expression level in both samples, whereas gene w has a higher expression level in sample 1. The values for mismatches are subtracted from the values for the perfect matches, and final expression levels of genes are calculated. Different programs apply different algorithms to calculate expression levels, and some programs do not incorporate MMs into the calculation. (B) Entire view of the Affymetrix human GeneChip® U133A array, which has more than 22,000 probe sets of 11 perfect match:11 mismatch pairs. An RNA sample from human lymphocytes was used for hybridization. A small area is magnified and shown with an overlaid grid that designates the borders for the cells. (C) The 11 probe pairs for GAD1 are shown, with PM cells above the MM cells. The histogram shows the intensity reading for each cell pair. Cell pairs are spread over the entire array, to minimize the impact of artifacts. Stars on the array mark the location of the cell pairs for GAD1. Notes: (a) The darker the cell, the lesser amount of RNA is bound. (b) Negative values can be obtained, such as in probe pair 2 of sample 2 of gene w (A), or probe pairs 5, 6, and 8 in GAD1 (C). Different algorithms might deal differently with these cell pairs, or not take MM into consideration at all.

Fig. 4. Example of three MAPPs drawn in GenMAPP, and of data visualization. Expression data were loaded into MAPPs representing serotonin synthesis (A), serotonin receptors (B), and GABA receptor subunits (C). GenMAPP color codes transcripts according to specifications set by the investigator. In the hypothetical sample shown, transcripts that are upregulated in the experiment group are labeled red (dark red if P level below 0.01, light red if P level between 0.05 and 0.01), transcripts that are downregulated are labeled blue (dark blue if P level below 0.01, light blue if P level between 0.05 and 0.01), transcripts that are below expression threshold (in this example set as ‘present’ in less than 20% of all samples) are labeled dark gray, transcripts that meet none of the criteria are labeled in light gray. A white box (no color) indicates that the transcript is not present in the data set. If the investigator has a large number of MAPPs, the hypergeometric distribution of a data set can be calculated to investigate which MAPP has more regulated genes than would be expected by chance. For example, the MAPP on GABA receptors (C) would likely be significant for downregulated transcripts. If a large number of these MAPPs are used, they can provide important biological information on gene array results. Many MAPPs are available at the GenMAPP website to which investigators can add their own MAPPs on their local computers, tailored to their biological system. Investigators are also encouraged to share their MAPPs via the GenMAPP website. In addition to tailored MAPPs, the GenMAPP program makes use of the Gene Ontology database. Numbers to the right of each box refer to fold regulation over control (A, C), or to P levels (B).

Fig. 5. Hierarchical clustering. (A) Gene expression array analysis was performed with frontal cortex samples from rats treated for 26 days in single, daily injections with vehicle or the antipsychotic drug haloperidol (haldol; 0.5 mg/kg/day). Expression levels were calculated using RMAExpress in natural scale. The 100 transcripts with the most variability across all samples were used for hierarchical clustering (using a correlation distance metric and centroid linkage method). Because transcripts expressed around or below threshold have the highest variability but the least reliability, they were removed from the analysis. We chose to include only transcripts that were above detection threshold in more than 50% of all samples. Hierarchical clustering in panel A addresses the question if variability in gene expression levels can be explained by treatment. The results show that transcripts with high variability coincide with treatment, a desirable outcome that demonstrates that variability is not random. It also indicates which samples are most alike. This type of clustering can be used to control for artifacts that might be introduced if samples are worked up in batches. If the various batches are clustered together, it might indicate a batch effect. In addition, duplicate samples should cluster together. (B, C) In a second experiment, rats were treated for 26 days in single, daily injections with vehicle or the antipsychotic drug haloperidol (haldol; 0.2 mg/kg/day), and striata were subjected to gene expression microarray analysis. All transcripts that reached a P value ≤ 0.02, and that were above detection threshold in at least 50% of all samples were used for hierarchical clustering (n = 237). The FDR rate of these genes was below 1%. Fold difference was not used as a criterion because data were log-transformed (calculated with the RMAExpress algorithm). Clustering in this experiment was used to evaluate if the chosen genes clustered the samples. The low FDR rate is already an indication that the expression levels of these transcripts are quite specific for treatment, and clustering should reflect that fact. Indeed, the clustering of control and haloperidol samples was complete and almost insensitive to distance metric used. ‘Correlation’ distance metric, shown in panel B, rank correlation (not shown), and Euclidian distance (not shown) separated control and haloperidol samples. However, if absolute correlation distance metric was used (C), no significant separation was seen. The latter distance metric was therefore not suited for this experiment. The dendrogram above the samples shows the association of the samples, the dendrogram on the left side shows the association of the genes (gene names were removed). Gene dendrograms were removed in panels B and C for better clarity. The length of the branches represents the distance in expression levels of the genes and samples. Each row represents a gene and each column represents a sample. The red color represents expression levels above mean expression of a gene across all samples, the black color represents mean expression levels and the green color represents expression levels lower than the mean (see color scale). Since the expression level for each gene is standardized to have a mean of 0 and a standard deviation of 1, the standardized expression values fall predominantly within [−3, 3]. Significance was reached for the haloperidol and the vehicle clusters in panel A (P ≤ 0.024) and in panel B (P ≤ 0.0022). Significance was not reached in panel C. A white line was introduced between the two main sample clusters and the two main gene clusters in panels A and B for better visualization.

Fig. 6. LDA and PCA. Gene expression array analysis was performed with striata of rats treated for 26 days in single daily injections with vehicle, the antipsychotic drug haloperidol (0.2 mg/kg/day), or the antipsychotic drug clozapine (8 mg/kg/day). (A) LDA was performed with the same subset of genes shown in Fig. 5, using vehicle (blue boxes) and haloperidol (red boxes) samples as known classes. Clozapine was imputed as the unknown class (indicated by the gray border around the box). All clozapine samples grouped with haloperidol. (B) The same experimental setup was used as in panel A, but PCA was performed instead of LDA. Unknown samples are represented by black boxes. Although the results are somewhat similar, the clusters are much less obvious than with LDA. LDA and PCA classifications can be used to diagnose disorders or predict treatment response, and they can serve as a diagnostic tool in diseases such as cancer. In this particular example, the genes chosen were not just indicative for haloperidol treatment, but for clozapine treatment as well. Thus, the transcripts used for the classification could be important for the therapeutic properties of antipsychotic drugs. To evaluate if the transcripts correspond to particular biological functions, they were subjected to MAPPfinder analysis using over 400 individual MAPPs that we have assembled in the past. The most significant finding in these 237 genes was an accumulation of mitogen-activated protein kinase genes, which were upregulated with haloperidol in the MAPPfinder analysis (C), (z score > 5.8; P = 0.0). As hierarchical clustering (Fig. 5B) and LDA analysis might suggest, clozapine followed a similar trend (D). Legend: Red boxes denote upregulation of transcripts that have 50% or more present calls (dark red: P ≤ 0.02, light red: P ≤ 0.05, pink: trends, P ≤ 0.1), blue boxes denote downregulation of transcripts that have 50% or more present calls (no transcript met this criterion), light gray denotes transcripts that have 50% or more present calls and were not regulated, dark gray denotes transcripts that have less than 50% present calls, and unmarked boxes denote that transcript was not present on the array. P levels are shown to the right of the transcript boxes.