The approach we took to finding putative in vivo binding sites for berenil was to consider the yeast genome as a bank of DNA sequences that, in the context of chromatin and the environment of the nucleus, have various affinities for the drug. Among these are sequences whose role in the regulation of transcription could be affected by berenil binding. Changes in gene expression at the level of transcription could therefore occur directly through mechanisms such as interfering with transcription factor binding or altering chromatin structure 217. Indirect effects on transcript levels may also occur through gene regulatory networks or general stress response 18. We sought to use microarray analysis to determine the set of yeast genes that are directly or indirectly affected at the level of transcription by berenil. We then addressed the significant challenge of distinguishing between these two categories during analysis of the upstream regions of affected genes.

Our experimental method was to culture yeast in the presence and absence of berenil, isolate total RNA, and conduct microarray hybridizations to measure changes in steady state transcript levels for all yeast genes. We gathered data suitable for analysis from five experiments, with two whole-genome microarrays in each experiment and two of them conducted as dye swaps of the other three. We analyzed microarray data using MicroArray Genome Imaging and Clustering (MAGIC) Tool 19. Once MAGIC Tool produced files of foreground and background intensities for all of the microarray experiments, we used Excel to analyze the data. Differences in the performance of the two dyes were accounted for by normalization 20. A series of filtering criteria were used to ensure that only reliable data were used in the production of an affected gene list. As described in the Methods section, the criteria for inclusion of the data for a given ORF were 1) the sum of the intensities for the two channels had to be greater than the minimum median of that sum for all the experiments, 2) the foreground for at least one channel had to be double the background, 3) data for a feature had to pass the first two filters in at least 5 of the 10 measurements of that feature, 4) data for a feature must pass the first two filters in both Cy3- and Cy5-labelled samples, and 5) the coefficient of variation of log transformed ratios across experiments had to be less than one. The culmination of our analysis was a list of ORFs for which we had reliable microarray data. This resulted in a final list of 52 genes whose mRNA levels decreased and two whose levels increased upon berenil treatment. We find it very interesting that the vast majority of the reliably affected genes were negatively affected; this observation is likely to be relevant to discovery of the mechanism of action of berenil and other MGBDs.

In order to provide validation of the microarray results from an independent method, we performed real time PCR measurements for selected genes. Quantitative reverse transcription real time PCR with SYBR Green reporting was used to generate the data presented in Table 1. The gene TUB1 was used as a standard as described 21 and was also unaffected in our microarray experiments. The genes STF2, HSP78, and SPI1 were found to have lower steady state levels of mRNA according to the microarray data and also had lower levels according to real time PCR. Although the genes appeared in the same order with regard to the magnitude of the effect, the real time data resulted in higher expression ratios, an effect that may be due to signal saturation in the microarray experimental approach.

The list of genes whose steady state transcript levels were shown by our microarray analysis to be reliably affected by berenil includes two genes whose levels increased. According to the Saccharomyces Genome Database (SGD), their functions are in phosphate metabolism and processing of 20S pre-RNA 22. Of the 52 genes whose mRNA levels decrease, 14 are involved in stress response, nine in carbohydrate metabolism, four in electron transport, two in meiosis and mitosis, one in regulation of redox homeostasis, one in regulation of proteolysis, one in salinity response, one in vacuole fusion, one in response to metals, one in phosphate metabolism, and one in DNA repair, according to the SGD. The remaining 16 genes have not had functions assigned to them by SGD. The list of affected genes is likely to include some that are directly affected by berenil. Genes in this category are expected to have upstream transcriptional control regions that include berenil binding sites. Genes whose transcript levels changed by indirect drug effects are also likely to be in the list. Such genes would not be expected to contain berenil binding sites in their upstream control regions. Although we cannot determine which genes may be indirectly affected through gene regulatory networks, we note that there are a number of genes that function in stress response. We chose to test the hypothesis that these 14 genes are indirectly affected by berenil and that the remaining genes are directly affected. Table 2 lists the 40 genes hypothesized to be directly affected while Table 3 lists those hypothesized to be indirectly affected, each with an expression ratio, function, and molecular process, if known.

We sought to analyze the upstream regions of the hypothesized direct and indirect categories of genes for the occurrence of sequence features and to consider the results in light of in vitro berenil binding studies. In this way, we can address the validity of our categorization of the affected genes and extend knowledge of berenil binding preferences to the environment of cells. As reviewed above, studies conducted in vitro have shown that berenil binds 5–6 nucleotide A/T-rich tracts that tend toward heteropolymeric (alternating A and T) character 13141516. If these binding preferences extend to the cellular context, then the upstream sequences of the 40 yeast genes hypothesized to be directly affected by berenil and listed in Table 2 are expected to contain 5–6 nucleotide sequence elements that are A/T-rich and heteropolymeric. For the 14 genes hypothesized to be indirectly affected and listed in Table 3, the signal of berenil binding sequences should fall to the background level found in the upstream sequences of all yeast genes. Our challenge was to find ways to analyze the sequences in order to uncover any existing sequence patterns. We reasoned that if the effect of berenil on transcription levels is due to binding sites that they would be found upstream of directly affected genes as 5-mer and 6-mer sequences. In order to reduce the background noise, we limited our search to 200 nt upstream of the start site for translation and to the sense strand only. We developed two measures for determining whether the occurrence rate of a given sequence element is unusually high in the upstream regions of genes. The first was the difference between the percentages of the upstream regions of affected and unaffected genes containing a sequence. The second is the ratio of the number of occurrences of a sequence in the affected gene upstream regions to that of the unaffected gene upstream regions.

For use as a control group, we assembled a list of 56 genes that remained reliably unaffected by berenil treatment in the course of our microarray experiments. Sense strand sequences from the 200 bp upstream of these genes, the 40 directly affected genes, and the 14 indirectly affected genes were measured for the occurrence of all possible 5-mer and 6-mer sequences The percentage of genes with an. occurrence of each sequence was determined for each of the three categories and the difference in percentage was calculated between each of the two affected categories and the unaffected category.

Table 4 shows the top ten 5-mer and 6-mer sequences from the direct category according to this criterion. For example, AATAA occurred upstream of 71% of the directly affected genes, but in only 40% of the unaffected ones, for a difference of 31%. The sequences listed occur in an average of 61% of the direct gene upstream regions. All of them occur more frequently upstream of direct category genes than of the unaffected genes; the average occurrence is 23% higher. We also measured the number of occurrences of each sequence in the directly affected, indirectly affected, and unaffected gene upstream regions. The ratio of the number of occurrences in the affected category to that in the unaffected one was calculated and the sequences were ranked according to this ratio, with the top ten sequences listed in Table 5. For example, the sequence ATAAG occurs in 30 times in the 40 affected gene upstream regions, a rate that is 2.3 times higher than in the unaffected gene regions. The sequences listed occur an average of 28 times in the 40 directly affected genes, an average of 2.1 times the occurrence rate found in the 56 unaffected gene regions.

Table 6 shows the top ten sequences according to the difference criterion for the indirect gene category. For example, ACCTC occurs in 50% of the indirect gene regions but only 2% of the unaffected one, for a difference of 48%. The sequences listed occur in an average of 50% of the 14 genes hypothesized to be indirectly affected with an average difference of 35% between the affected and unaffected gene regions. Table 7 shows the results of ranking sequences in the upstream region of genes in the indirect category using the ratio criterion. For example, the sequence AATCT occurs nine times in the indirect sequence regions, a rate that is 3.3 times higher than that found for the unaffected genes. The sequences listed occur an average of 7.8 times in the upstream regions of the 14 indirectly affected gene set, an average of 4.4 times higher than the rate found for the 56 unaffected genes.

The difference and ratio criteria for choosing overrepresented sequences yield more similar results for the direct category than for the indirect category. Of the 20 sequences from the direct category listed in Table 4, 15 are found in Table 5, shown in bold. Only nine indirect category sequences are shared between Tables 6 and 7. The extent to which the difference and ratio lists share sequences can be attributed to three causes. First, the two measures are not unrelated. The number of occurrences of a given sequence affects the number of genes with which it is associated, and vice versa. Second, an average of one shared sequence is expected at random. Third, shared sequences may occur because they reflect sequence characteristics of the data set. Since the first two of these causes are not expected to be different for the direct and indirect categories, the increased level of shared sequences for the direct category is telling. It is likely to arise from characteristics of the upstream regions of the direct category genes. In order to investigate these characteristics, we formed sets of unique 5- and 6-mers for the direct and indirect categories that included each shared sequence only once and conducted several types of sequence analysis.

Several observations can be made regarding the sequences presented in Tables 4, 5, 6, 7 that can be used to address our hypotheses about genes that are directly or indirectly affected by berenil and to evaluate the extent to which the rules for drug binding in vitro can be extended to the cellular context. These observations relate to the A+T content and the extent of heteropolymeric character found in the sequences listed in Tables 4, 5, 6, 7 and are outlined in the following sections.