Apoptotic Susceptibility of Cancer Cells Selected for Camptothecin Resistance: Gene Expression Profiling, Functional Analysis, and Molecular Interaction Mapping

Acquired resistance to multiple forms of treatment seriously limits the efficacy of cancer chemotherapy. One drug affected by acquired resistance is camptothecin, which poisons top1⁴ by reversibly stabilizing DNA-top1 cleavable complexes (1, 2, 3). Camptothecin analogues are used clinically against multiple forms of cancer, including lung, breast, ovarian, and colorectal, but one factor that limits their activity is drug resistance, either primary or else secondary to treatment. Top1 mutations can explain resistance in some experimental models (1, 4), but they have not been implicated in clinical resistance to camptothecins. Hence, it seems likely that molecular alterations downstream from the top1 cleavable complexes are involved. To investigate that possibility in the present study we compared the prostate cancer cell line DU145 with a selected resistant subline RC0.1 (5). It is worth noting that such a study could not have been done with transfected or knockout cells insofar as multiple gene changes are involved (6).

DU145 is an androgen-independent, mismatch repair-deficient (7), microsatellite-unstable (8) human cell line derived from metastatic disease (9, 10). In their late stages, clinically metastatic and androgen-independent prostatic cancers tend to be aggressive and are uniformly fatal. 9-Nitro-camptothecin has been shown to activate a CD95 (Fas)/CD95L-dependent apoptotic pathway in DU145 cells (11). DU145 cells are resistant to apoptosis induced by anti-Fas-IgM but become sensitized to that IgM in the presence of camptothecin (9). Furthermore, DU145 cells may be induced to apoptose by tumor necrosis factor-related apoptosis-inducing ligand (12). The RC0.1 cell subline is a 9NC-SN38-topotecan-resistant derivative of the DU145 cell line selected after continuous incubation with 0.1 μm 9-nitro-camptothecin (9, 11). The top1 in RC0.1 (but not DU145) cells contains an R364H mutation that leaves the enzyme catalytically active but resistant to inhibition by camptothecin analogues in biochemical assays. The normal top1 allele is not expressed in RC0.1 cells (5). However, identification of the R364H mutation does not rule out the possibility that additional molecular changes also contribute to the observed camptothecin resistance.

In this study, we demonstrate a difference in the profiles of mRNAs expressed in DU145 and RC0.1 cells using 1648-clone cDNA microarrays designed to focus on cancer-interesting genes. One hundred eighty one genes differed in expression level by a factor of ≥1.5 between the two cell types. When the genes in the array were classified by function, we identified several categories that contained a higher percentage (i.e., with P < 0.05) of the 181 changing genes than would have been expected by chance. The largest such class was that of the apoptotic genes (P = 0.04). Because camptothecin treatment (4), in addition to many other chemotherapies and irradiation, has been shown to induce apoptosis (4), we asked if a generalized functional shift in apoptotic response had occurred between DU145 and RC0.1 cells.

Cell culture medium RPMI 1640 was from BioWhittaker, (Walkersville, MD); fetal bovine serum from Atlantic Biologicals (Norcross, GA); l-glutamine from Life Technologies, Inc. (Rockville, MD); trypsin-EDTA solution from Quality Biologicals, Inc. (Gaithersburg, MD); camptothecin, staurosporine, cis-platinum (II)-diammine dichloride, DMSO, propidium iodide and RNase from Sigma-Aldrich (St. Louis, MO); the APO-BrdUrd kit from BioSource International, Inc. (Camarillo, CA); and Annexin V-FLUOS staining kit from Roche (Indianapolis, IN). The DU145 cell line was obtained from the American Type Culture Collection (Rockville, MD). The RC0.1 cell subline was derived from DU145 cells continuously exposed to 25 nm 9NC for 1 month, 50 nm 9NC for 3 months, and 0.1 μm 9NC for 3 months (13). In the present study, cells were cultured in RPMI 1640 containing 0.29 mg/ml l-glutamine and 10% fetal bovine serum, and harvested using trypsin-EDTA at 85% confluence. For studies of drug sensitivity, camptothecin, staurosporine, and cis-platinum were solubilized in DMSO and added to the cells for 48 h of continuous exposure before harvest. UVC irradiation at 254 nm was done with a UV Stratalinker 1800 (Stratagene, La Jolla, CA), and the cells were harvested after 48 h. γ-Irradiation was done with a 6000-Ci Cs137 Mark I irradiator, model 68 (J. L. Shepherd & Associates, San Fernando, CA), and the cells were harvested after 48 h. For serum starvation, cell cultures were divided into RPMI 1640 with l-glutamine, washed twice with PBS (Life Technologies, Inc.) the next day, incubated for 2 days in the medium, washed twice with PBS, incubated for 3 days in medium, and then harvested.

Adherent cells were removed after trypsinization and washed in PBS after gentle centrifugation at 610 × g for 5 min. The cell pellets were fixed by resuspending them in 0.5 ml of 70% ethanol for 30 min, centrifuging at 925 × g for 8 min, and washing twice with ice-cold PBS to remove residual ethanol. For cell cycle analysis, the pellets were resuspended in 0.5 ml of PBS containing 50 μg/ml of propidium iodide and 100 mg/ml of RNase, then incubated at 37°C for 30 min and studied using a FACScan flow cytometer (Becton Dickinson, San Jose, CA).

The RNeasy kit was from Qiagen (Valencia, CA); diethyl pyrocarbonate-treated water from Research Genetics (Huntsville, AL); Cy3-dUTP, Cy5-dUTP, deoxynucleotide triphosphates, oligo(dT)_12–18 and oligo(dA)_40–60 from Amersham Pharmacia Biotech (Piscataway, NJ); and reverse transcriptase Superscript II and hCOT-1 from Life Technologies, Inc.

Total RNA was isolated from cells using the RNeasy kit and diluted to a concentration of 5 μg/μl in diethyl pyrocarbonate-treated water. In cDNA labeling reactions, 90 μg of total RNA were used for Cy5 labeling, and 78 μg of total RNA were used for Cy3 labeling. Seventy-five μm Cy3-dUTP or Cy5-dUTP were used in each labeling. Pairs of Cy3- and Cy5-labeled cDNA samples (DU145 and RC0.1 for the experimental microarrays) were mixed and hybridized to a pin-spotted cDNA microarray (Oncochip) developed in the Microarray Facility, Advanced Technology Center, National Cancer Institute, NIH, and prepared by one of us (W. C. R.). This glass slide microarray contained 2208 cDNA spots corresponding to 1648 individual human cancer-interesting genes (array lot HS-OC-2p10–101899).⁵ Included were 780 spots representing 364 individual genes in replicate. Hybridization data were acquired using a GenePix 4000 fluorescence scanner (Axon Instruments, Inc., Union City, CA). Images were analyzed using software developed by Yidong Chen in the Laboratory of Cancer Genetics, National Human Genome Research Institute.⁶ Validation studies were done by real-time RT-PCR as described in supplemental material.⁷

Microarray data were preprocessed using the PreProc computer program developed by Lawrence H. Smith (National Cancer Institute, NIH, Bethesda, MD).⁷ Options used included Gaussian kernel fitting, no background subtraction, and whole-chip data normalization (14). Gene ratios for four experimental arrays (i.e., two color-reversed pairs) were reciprocally averaged.

Because we had four array measurements (arrays 72, 73, 74, and 75) for each gene with three categories of expression ratio levels, we performed a repeated measures analysis to test the hypothesis of no association between the four arrays and three response levels, adjusted for gene. Thus, we obtained 1648 independent two-way contingency tables such that each table had four rows, one for each of the four arrays and three columns corresponding to the expression ratio levels. Each entry n_hij (h = 1,… ,1648; i = 1,… ,4; j = 1,… ,3) in the table was given a value of 1 when gene h was classified in expression ratio level j for array i, and 0 otherwise. We then applied the extended Mantel-Haenszel mean score statistic

for i = 1, … ,4; where {a_hj} is the set of scores for response levels in the h^th stratum. Because the distribution of genes was not uniform over the three levels of expression ratio, we used rank scores standardized by the sample size in a nonparametric analysis. Because the across-strata sample sizes

i = 1, … ,4, were sufficiently large, the quantity

had approximately a χ² distribution with three degrees of freedom, where μ_∗ is the expected value of f_.1. with variance v_∗ under the null hypothesis of no association (see Ref. 15). An analogous formula was used to incorporate the adjustment of the reciprocal indicator variable with the corresponding number of degrees of freedom. This treatment assumes, as is often done, that the overall covariance of expression patterns among different genes is negligible in comparison with the total variance. All of the analyses were done using SAS software version 8.2.

Genes were assessed functionally based on molecular pathway information developed by Kohn (16).⁷ Additional assessments were made using GeneCards⁸ and MedMiner (17).⁷

Adherent cells were removed after trypsinization, collected in PBS, centrifuged at 610 × g for 5 min, fixed in 70% ethanol for 30 min, washed twice with cold PBS, resuspended in 0.5 ml of PBS with 50 μg/ml of propidium iodide and 100 mg/ml of RNase, incubated at 37°C for 30 min, and analyzed using a FACScan flow cytometer. The APO-BrdUrd assay was performed using the APO-BrdUrd kit following instructions of the manufacturer. In brief, DNA strand fragments in the cells were tagged with BrdUrd-dUTP at their 3′-OH ends using terminal deoxynucleotidyl transferase. The BrdUrd-dUTP was fluorescein-labeled with a monoclonal antibody to BrdUrd, and the cells were then analyzed in a FACScan flow cytometer. Phosphatidylserine translocation was measured using the Annexin V-FLUOS staining kit for fluorescence labeling and the FACScan for measurement of the resulting fluorescence.

After 48 h of continuous exposure to 20 nm and 100 nm camptothecin, the parental DU145 cells showed dramatic arrest in late S phase and early S phase, respectively, and the emergence of presumably apoptotic cells with DNA contents lower than that of cells in G₁ phase (Fig. 1). After treatment with 100 nm camptothecin, there was a 74% reduction in DU145 cell number at 48 h in comparison with untreated DU145. In contrast, the resistant subline, RC0.1, showed no discernible treatment-related change in cell cycle distribution and only a 17% reduction in cell number. The percentages of cells with sub-G₁ DNA contents after treatment with 20 nm and 100 nm camptothecin were 17% and 24% for DU145 and 2% and 6% for RC0.1. As with many pharmacological experiments, a reduction in cell number (as seen here for DU145 cells) may have reflected a selection process for cells less affected by the drug. That selection process would be expected to decrease the differences observed between wild-type and resistant variant.

To determine which genes might have been involved in the differential apoptotic response, we used cDNA microarrays to compare gene expression in RC0.1 and DU145. Two arrays were hybridized with Cy5-labeled DU145 cDNA and Cy3-labeled RC0.1 cDNA. Two arrays were labeled in the reciprocal way with Cy3-labeled DU145 cDNA and Cy5-labeled RC0.1 cDNA. For a “split sample” control, RC0.1 RNA was the source for both Cy5- and Cy3-labeled cDNA. Overall, we observed that 181 of 1648 (11.0%) of the genes differed in expression level by a factor of ≥1.5 between DU145 and RC0.1. Table 1,A lists genes expressed at least 4-fold lower in RC0.1, and Table 1,B lists genes expressed at least 4-fold higher in RC0.1. Also included are genes belonging to functional groups enriched or depleted to a statistically significant extent by >1.5-fold (see Table 3).

Three hundred sixty four genes were replicated at least once on the array. Of those, 312 showed agreement across all of the replicates with regard to classification as unchanged or altered by 1.5-fold (either up or down). Genes that showed disagreement were excluded from consideration unless they had also been evaluated by real-time RT-PCR. The median SD across replicates for the “mean intensity ratios” (see Table 1, A and B) was 0.12 (i.e., 12%), indicating a high degree of reproducibility.

Replicate genes on the array that showed disagreement were deemed unreliable and excluded from the analysis, unless the disagreement was resolved by TaqMan real-time RT-PCR (data not shown). These RC0.1:DU145 ratios ranged from 0.13 to 13.1 and were consistent from one experimental array to the next. The median coefficient of variation for these 181 genes was 0.14 (i.e., 14%). For the split sample, 2.9% of the genes exceeded the 1.5-fold change cutoff. Use of the reciprocal-labeling experimental design and analysis eliminated red-green bias for any particular spot. Some genes appeared multiple times on the array, thus providing ≥8-fold replication of the results.

To test the reliability of the microarray data, we used TaqMan real-time RT-PCR assay because it is highly specific, and because its sources of error are very different from those of hybridization to a cDNA array. There was a general concordance between microarray and real-time RT PCR results (r = 0.89 for the 18 genes studied; see data in Supplementary Table 2).⁷ The overall Pearson correlation coefficient for the 18 gene ratios studied was 0.89 (data not shown). GSTP1, IL8, and NK4 differed most dramatically in expression between DU145 and RC0.1. The magnitude of the change was underrepresented by the microarrays, but the direction was correct. By real time RT-PCR, the RC0.1:DU145 ratios were 0.001, 0.002, and 0.03, respectively.

The gene expression ratios in DU145 and RC0.1 cells were compared across the four microarrays, 72, 73, 74, and 75. Microarrays 72 and 73 had ratios of expression intensity (reciprocal indicator = R/G) based on fluorescent dyes that were reversed with respect to those of the other two microarrays, 74 and 75 (reciprocal indicator = G/R). These ratios were compared for the 1648 individual genes present on the microarray.

The ratios of expression intensity were divided into three categories (′< 0.67′, ′[0.67, 1.5]′, ′> 1.5′). Table 2 shows the overall distribution of expression ratios. Nonparametric analysis was used to test the null hypothesis of no association. We applied the stratum-adjusted Kruskal-Wallis (extended Mantel-Haenszel) test, considering each gene to constitute a stratum. Because the three ordered levels for the expression ratio are not equally spaced, rank scores standardized by sample size were used for the mean score statistics.

After combining values from microarrays 72 and 73, and values from microarrays 74 and 75, we observed an association between the ratio levels and the reciprocal indicator (P = 0.0027), indicating a small but statistically significant “dye-effect.” After adjusting for the reciprocal indicator as well as the individual genes, we saw no association between the expression ratios and microarray number using the same nonparametric analysis (P = 0.73). We also saw no association between expression ratio and the microarray number within the two levels of the reciprocal indicator; that is, expression ratios had no association with microarray for arrays 72 and 73 (P = 0.26) or 74 and 75 (P = 0.58). This observation indicates that there was no significant array effect on the expression ratios.

There was a strong overall concordance in expression ratio level (κ = 0.67, where a value of 1 corresponds to perfect concordance and a value of 0 corresponds to that expected by chance, with 95% confidence interval, 0.64–0.71) between the gene ratios for microarrays 72 and 73, and between microarrays 74 and 75. Thus, the expression ratios of the four microarrays (after adjustment for the reciprocal indicator) were highly likely to fall into the same category (i.e., one of the three levels). Both the association and concordance findings indicate reproducibility of the ratios for the individual genes across the four microarrays.

Functional classification of the genes present on the microarray (16) yielded 179 functional categories. We asked which of the categories were statistically significantly enriched for genes that had a ≥1.5-fold level of change with respect to the null hypothesis that the expected 11% of genes (181 of 1648) in the category would differ in expression. Five of the 179 categories (Table 3) met that criterion: “apoptotic,” “fra-2-associated,” “MHC,” “mitochondrial,” and “NFκB-associated.” For the statistical analysis, we used the one-tailed exact binomial test without correction for multiple comparisons. The one-tailed test was used because only categories with proportions statistically >11.0% (181 of 1648) were a point of focus in our attempt to explain the observed shift in apoptotic response.

To test whether the statistical indications of alteration in the apoptotic pathway had functional implications, we assessed the apoptotic responses to various treatments by flow cytometry using the Annexin V fluorescence assay to detect early apoptosis events (i.e., phosphatidylserine translocation to the outer plasma membrane leaflet) and the APO-BrdUrd (TUNEL) assay for detection of late apoptosis events (i.e., chromosomal DNA fragmentation). Fig. 2,A indicates early apoptosis events in DU145 cells subjected to diverse treatments, including exposure to camptothecin, staurosporine, and cisplatin, as well as serum starvation, and UV and γ-irradiation. The most extensive response was observed in staurosporine-treated DU145 cells. In contrast, practically no late apoptosis was seen in RC0.1 cells exposed to camptothecin, cisplatin, serum starvation, or γ-irradiation (see histograms in Fig. 2,A). Staurosporine and UV irradiation induced the signature of late apoptosis in RC0.1 cells (Fig. 2,A). The results are summarized quantitatively in Table 4.

The PI3K molecule falls within the area of the molecular interaction map that changes in a direction consistent with the observed shift in apoptotic response. The apoptotic response to camptothecin treatment in the presence of wortmannin, an inhibitor of PI3K (18), was assessed by flow cytometry using the Annexin V and APO-BrdUrd assays. Wortmannin dramatically increased the apoptotic response of DU145 cells, particularly the late stage of apoptosis (Fig. 3, A and B). In contrast, wortmannin had no effect on apoptosis in RC0.1, with or without camptothecin. The apparent potentiation in DU145 cells suggests (but does not prove) PI3K involvement in the resistance of RC0.1 cells to camptothecin. In other experiments (see supplemental material⁷, Figs. 2A and B), wortmannin similarly potentiated cis-platinum-induced apoptosis in DU145 but not RC0.1, suggesting a defect in RC0.1 that relates directly to, or is downstream from, PI3K.

This study started with a comparison of gene expression profiles in the DU145 human prostate cancer cell line and a subline, RC0.1, selected for resistance to camptothecin (5). RC0.1 cells express top1 with a mutation that does not alter the catalytic activity of the enzyme but confers resistance to camptothecin analogues in biochemical assays (5). However, that mutation does not exclude the possibility that additional molecular changes also contribute to the observed resistance. To identify such changes, we assessed gene expression profiles in the cells using quadruplicate Oncochip cDNA microarrays. Expression level differences by a factor of ≥1.5 between the two cell lines were prominent for several categories of functional genes (Table 3): apoptosis-related (12 of 62 genes; P = 0.038), MHC (7 of 15 genes; P = 0.001), mitochondria-related (but not apoptosis-related; 4 of 7 genes; P = 0.004), NFκB-related (6 of 14 genes; P = 0.002), and fra-2 associated (2 of 2 genes; P = 0.012). The apoptosis-related group was particularly interesting because of its relevance to drug resistance and because it was the largest group of genes that showed statistically significant enrichment with genes that differed in expression by a factor ≥1.5 between the two cell lines. Given the unique status of the apoptosis category, no correction for multiple comparisons was applied. However, in interpreting these Ps, it is worth noting that the assumption of independence probably fails (to some unknown degree) for this calculation. Hence, the Ps should be considered as heuristic indices, not as the results of formal inference.

Because changes in expression were prominent for genes associated with apoptosis, we next compared the apoptotic responsiveness of DU145 and RC0.1 cells after exposure to various drugs or other stressors (Fig. 2, A and B). Apoptotic response was markedly reduced in RC0.1, not only in response to camptothecin (top1 poisoning), but also in response to cisplatin (DNA cross-linking), staurosporine (protein kinase inhibition), serum starvation (growth factor deprivation), UV radiation (DNA cross-linking and major strand breakage), and ionizing radiation (DNA strand breakage). This pervasive functional response would not have been expected on the basis of the known top1 mutation.

The relatively large number of apoptosis-related genes that differed in expression between DU145 and RC0.1 cells was puzzling for two reasons. First, it was not clear why expression changes in so many genes would be required to achieve resistance to apoptosis. Second, many of the expression differences seemed to be in the direction contrary to what would be expected given the resistance of RC0.1 cells to apoptosis. To address these findings systematically, we constructed a molecular interaction map of the relevant apoptosis-related pathways, based on evidence cited in recent literature (Fig. 4,A). The map, which can be found in a navigable electronic form,⁷ was prepared using the notation of Kohn (16). Definitions of the symbols used are summarized in Fig. 4,B. The molecular species associated with significant gene expression differences can be located via the map coordinates listed in Fig. 4,C. In the map, the observed RC0.1:DU145 (R/D) gene expression ratios are in brackets within the molecular species symbols. The major findings, summarized in Fig. 4 C, reflect the complexity of the biological system. It will be necessary to discuss those complexities here in some detail.

For these gene families, which affect apoptosis most directly, our microarray data included the following (R/D = ratio of mRNA levels in RC0.1:DU145). Proapoptotic genes: A1/Bfl-1 (R/D = 0.76), Bad (R/D = 0.50, 0.45), Bak (R/D = 0.86), Bik (R/D = 1.09), caspase-1 (R/D = 1.09), caspase-3 (R/D = 0.74), caspase-4 (R/D = 0.89), caspase-5 (R/D = 0.95), caspase-6 (R/D = 0.58), caspase-7 (R/D = 0.95), caspase-8 (R/D = 0.90), caspase-9 (R/D = 0.90), and caspase-10 (R/D = 0.84, 0.88). Antiapoptotic gene: Bcl-2 (R/D = 1.48, 1.32). Pro- and antiapoptotic isoforms: Bcl-x (R/D = 1.10).

We see that 2 of these 14 genes exhibited ≥1.5-fold expression differences: BAD and caspase-6 (in the right lower quadrant of Fig. 4,A). Both are proapoptotic genes that show reduced expression (R/D <0.67) in the apoptosis-resistant RC0.1 cells. These differences are therefore in the expected direction. The reduced expression of BAD in RC0.1 cells was confirmed by real-time RT-PCR, which gave a ratio R/D = 0.12 (see supplemental material).⁷ The difference for antiapoptotic Bcl-2 was also in the expected direction (R/D = 1.32, 1.48), although the magnitude of the observed difference did not quite meet our threshold criterion of 1.5. However, the significance of this change is supported by the similar increase in mRNA level of c-Myb (R/D = 1.28, 1.48; Fig. 4 A, coordinates 8D), a transcriptional activator of Bcl-2.

BAD is induced by staurosporine to migrate to mitochondria in prostate cells (19, 20), and it is up-regulated by serum starvation, leading to apoptosis (21). Several genes exhibited coherent expression differences stemming from reduced BAD mRNA levels in RC0.1 cells. (By “coherent” we refer to multiple pathways leading to the same effect; thus, multiple expression differences that have a common final molecular impact would be considered coherent.) Two Bcl-2-binding proteins, Bag1 and 53BP2, exhibited significantly different mRNA levels in RC0.1 and DU145 cells. Bag1, which binds Bcl-2 and inhibits apoptosis (22), was expressed at a higher level in RC0.1 for two different clones on the array (R/D = 1.65, 1.57). In contrast, 53BP2, which binds Bcl-2 and enhances doxorubicin-induced apoptosis (23), was expressed at a lower level in RC0.1 cells (R/D = 0.58). Therefore, these differences were in the expected direction (Fig. 4,A, lower right quadrant). Coherent with the reduced expression of BAD in RC0.1 cells, we observed expression differences in 3 genes that could suppress the proapoptotic effect of BAD by way of Akt/PKB, a kinase that inhibits BAD by phosphorylation of critical sites. These 3 genes are phosphoinositide-3-kinase (PI3K catalytic γ-polypeptide; R/D = 1.60), integrin-β1 (R/D = 2.78), and receptor tyrosine kinase RON (R/D = 2.53). These genes appear to suppress BAD function by the mechanism indicated in the upper right quadrant of Fig. 4 A. PI3K phosphorylates phosphatidylinositols at the 3′ position, thereby generating binding and activation sites for Akt/PKB. Although the role of integrins in apoptosis is complex (e.g., see Ref. 24), the increased expression of integrin-β1 and RON in RC0.1 cells is noteworthy because these two receptors can cooperate to inhibit BAD by way of the Akt/PKB pathway (25, 26). Therefore, the expression differences of these 3 genes were in the expected direction and could additionally reduce the proapoptotic activity of already lowered levels of BAD. Thus, the reduced susceptibility of RC0.1 cells to apoptosis may have been attributable in part to reduced activity of BAD.

Because expression of PI3K was greater in RC0.1 than in DU145, we exposed the cells to wortmannin, a PI3K inhibitor, in the presence and absence of camptothecin. Wortmannin potentiated apoptosis in DU145 cells (as assessed by the TUNEL assay), as predicted. That is, wortmannin appeared to inhibit a survival signal that is transmitted by or through PI3K to Akt/PKB (Fig. 4 A, top right quadrant). Similarly, wortmannin potentiated cisplatin-induced apoptosis (see Supplemental Material, Fig. 6, A and B),⁷ indicating that the suggested PI3K involvement is not limited to top1-active agents.

The gene expression differences considered thus far were consistent with the hypothesis that the apoptosis resistance of RC0.1 cells is largely because of reduced expression and/or enhanced post-transcriptional suppression of BAD. We now focus on the remaining apoptosis-related genes that exhibited differences. Unexpectedly, most of those genes exhibited differences in the contrary direction. Most are located on the left half of the map (Fig. 4 A, blue-shaded genes). Many of them fall into two coherent groups governed by the NFκB and TGF-β transcription pathways.

In the NFκB pathway (Fig. 4 A, left side), we see reduced expression in RC0.1 cells of three components of the NFκB transcription factors, namely p65RelA, c-Rel, and their transcriptional partner p100/p50. Although there was reduced expression of IκBα, a change that would tend to increase NFκB activity, the mRNA levels of NFκB transcription products were reduced (see below), suggesting that IκBα level may not have been limiting in this case.

The NFκB-stimulated transcription products for which we have data include A20, IAP2/MIHB, Bfl/A1, and c-FLIP. These genes have been reported as antiapoptotic. In addition, we have data for the NFκB-inhibited transcription product, Gadd153/CHOP, which has proapoptotic effects. A20, IAP2/MIHB, and Gadd153/CHOP showed expression differences ≥1.5-fold.

A20 and IAP2/MIHB inhibit apoptosis, and their transcription is positively controlled by NFκB (27, 28). A20 is a cytoplasmic zinc finger protein that inhibits tumor necrosis factor-stimulated apoptosis (29). IAP2/MIHB is a general caspase inhibitor. Gadd153/CHOP enhances apoptosis, and its transcription is negatively controlled by NFκB (30, 31). Gadd153/CHOP down-regulates Bcl-2 and sensitizes cells to apoptosis induced by stress at the level of the endoplasmic reticulum (31).

All of these differences are in a direction consistent with reduced NFκB transcriptional activity in RC0.1 cells. It is noteworthy that the mRNA level of IAP2/MIHB, a gene stimulated by NFκB, was diminished, whereas the mRNA level of Gadd153/CHOP, a gene inhibited by NFκB, was elevated. The decreased activity of the NFκB pathway in RC0.1 cells was unexpected because NFκB activity has mainly antiapoptotic effects (see Ref. 32). Another factor that may have contributed to down-regulation of NFκB-dependent transcription was the observed increase in the mRNA level of Egr-1 (R/D = 5.22). Egr-1 has been reported to bind p65RelA and thereby inhibit NFκB function (33). Egr-1 will be additionally considered in the context of the TGF-β pathway. In summary of this section, the NFκB pathway operates in the contrary direction with respect to the apoptosis resistance of RC0.1 cells.

The TGFβ-Smad pathway is noted for its proapoptotic effects in epithelial cells (34). In this pathway, we found increased expression of TGFβ2, Egr-1, Smad7, and c-Fos. The latter 3 genes are controlled by the Smad2/3:Smad4 transcription factors in response to TGFβ (Fig. 4 A).

Egr-1 has a radiation-inducible promoter, and it enhances apoptosis in response to ionizing radiation in human PC-3 prostate cancer cells (35). Egr-1 may enhance apoptosis via at least two routes shown in Fig. 4 A: (a) it stimulates transcription of PTEN, an indirect inhibitor of Akt/PKB (36); and (b) it binds and inhibits the RelA subclass of NFκB (33). Thus, the difference in Egr-1 expression (R/D = 5.22, one of the largest differences in our data set) was in the contrary direction.

Smad7, a feedback inhibitor of the TGFβ pathway, was up-regulated in RC0.1 (37), suggesting that increased transcription in these cells dominates over the feedback inhibition. Smad7 has proapoptotic actions via unknown connections to the JNK pathway (38) or possibly by reducing the antiapoptotic effect of the NFκB pathway (39). Either way, the increased Smad7 mRNA level in RC0.1 cells was in the contrary direction.

The relationship of c-Fos to apoptosis is complex. c-Fos binds c-Jun to form the AP-1 transcription factor, and the activity of c-Jun requires that it be phosphorylated by JNK (40). AP-1 can have pro- or antiapoptotic effects under different circumstances (41, 42). JNK generally has proapoptotic effects, some of which may be mediated via AP-1 (41, 43). c-Fos mRNA is controlled transcriptionally and also by mRNA stabilization in response to DNA damage (44). Thus, the increased level of c-Fos mRNA in RC0.1 cells cannot simply be assigned to a pro- or antiapoptotic category.

A Smad-independent pathway from TGFβ to apoptosis may go by way of Daxx and JNK, perhaps through a phosphorylation sequence involving ASK1 and MKK4 (Fig. 4 A; Ref. 45). Daxx, which binds to the cytoplasmic domain of TGFβRII, can mediate TGF-β-induced apoptosis by stimulating JNK via ASK1 (see Ref. 46 and references cited therein). JNK is mainly proapoptotic (42). It activates c-Jun by phosphorylation of Ser-63 and Ser-73, and may thereby exert some of its proapoptotic effects (41, 42). The activity of the c-Fos:c-Jun heterodimer at AP-1 promoter elements may have been additionally increased in the RC0.1 cells by the reduced expression of Fra-1, a close relative of c-Fos. Like c-Fos, Fra-1 forms heterodimers with c-Jun, and it binds the same AP-1 promoter sites. Fra-1 has a transactivation domain that requires phosphorylation for activity. It may competitively inhibit AP-1 promoters (47). However, when its transactivation domain is phosphorylated, it may activate some promoters, such as that of GSTP1 (discussed below).

E2F-1 and c-Myc are prominent stimulators of apoptosis. Both had mRNA levels higher in RC0.1 than in DU145 cells, the contrary direction. c-Myc is a transcription factor that may contribute to up-regulation of E2F1 (Ref. 48; Fig. 4 A). The c-myc promoter contains an inhibitory Smad-binding element and, thus, may be down-regulated by TGF-β (49). However, despite the elevated activity in the TGF-β-Smad pathway, c-myc mRNA levels were elevated in RC0.1 cells. This increase may have been because of post-transcriptional regulation, because c-myc mRNA and protein both are short-lived, and regulation by mRNA stabilization has been reported (44).

The increased E2F1 mRNA level in RC0.1 cells was consistent with the increased c-myc mRNA level, because c-Myc has been reported to increase E2F1 mRNA levels (48). Moreover, Myc-induced apoptosis requires E2F1 (48). c-Myc and E2F1 may have stimulated apoptosis by suppression of Bcl-2 mRNA and protein (50, 51).

The contrary elevation of E2F1 and c-myc mRNAs in RC0.1 are the most striking of the seemingly paradoxical changes that we observed. However, we consider next an almost equally unexpected contrary reduction of GSTP1.

The glutathione S-transferase gene, GSTP1, was markedly suppressed in RC0.1 cells (R/D = 0.12 by microarray; ∼0.001 by real-time RT-PCR). GSTP1 protects cells against apoptosis induced by oxidative stress or DNA-damaging drugs (52). Therefore, the markedly reduced GSTP1 mRNA level in RC0.1 cells is in the contrary direction. Protection against apoptosis by GSTP1 may be attributable, in part, to inhibition of JNK. Inhibition of JNK by GSTP1 is abrogated by oxidative stress, which causes GSTP1 to form covalent dimers that cannot interact with JNK (53, 54). An AP-1 site in the GSTP1 promoter can be activated by Fra-1:c-Jun heterodimer (55). Thus, the reduced GSTP1 mRNA level may have been attributable in part to lower Fra-1 mRNA. Fra-1 has a transactivation domain that can be activated through phosphorylation by ERK (56). GSTP1, which is expressed in DU145 cells, can be silenced by hypermethylation of CpG islands, particularly in prostate cancer (57, 58, 59). We are investigating whether the contrary decrease in GSTP1 mRNA in RC0.1 cells was because of hypermethylation, an epigenetic event that has a relatively high probability of spontaneous occurrence.

Dad1 mRNA was increased in RC0.1 cells (R/D = 2.29, 2.12). Dad1, a M_r 12,000 membrane protein, is a subunit of the oligosaccharide transferase complex of the endoplasmic reticulum (60). Although deletion of Dad1 has been found to sensitize cells to apoptosis (61), there is no evidence that overexpression of Dad1 would protect against apoptosis. Because Dad1 is only one of several subunits constituting the antiapoptotic oligosaccharide transferase, the activity of this enzyme complex may not normally be limited by Dad1 availability. Therefore, we are uncertain of the relationship (if any) between the difference in DAD1 mRNA level and the resistance of RC0.1 cells to apoptosis.

The seemingly paradoxical contrary differences between DU145 and RC0.1 mRNA may be explainable by a two-stage model (Fig. 5) for the development of drug resistance. In this model, the first stage alters the expression or activity of one or more gene products (Fig. 5,box A) that provide direct resistance to apoptosis. Once this change is in place, genes in upstream pathways (Fig. 5, box B) are relieved of constraints imposed by the proapoptotic effects of their over- or underexpression. That is, expression changes that would normally favor selection if they did not also induce apoptosis would become possible without death of the cell. For example, E2F1 and c-Myc stimulate cell cycle progression as well as apoptosis. According to the above hypothesis, once relieved of the proapoptotic impact of these genes, the cells would be free to express the genes at higher levels and thereby achieve a selective advantage in the presence of drug.

In the current study, Fig. 5, box A, includes caspases and the Bcl-2 protein family, particularly the proapoptotic member BAD. It also includes some proteins that affect the activity of BAD post-transcriptionally, particularly by way of Akt/PKB. In the current study, Fig. 5, box B, includes transcription products governed by the NFκB and Smad transcription factors. Fig. 5, box A, corresponds approximately to the right half of Fig. 4,A; Fig. 5, box B, corresponds approximately to the left half.

Fig. 5, box B, proteins may enhance the ability of cells to proceed through the cell cycle despite various types of stress and/or to continue proliferating despite relatively high cell densities. Either mechanism could provide a selective advantage in the presence of drug. However, according to the model, this route to increased drug resistance is prevented as long as changes in the expression of Fig. 5, box B, proteins would also induce apoptosis. This model for a route to drug resistance is analogous to a possible mechanism for the origin and progression of malignant tumors recently discussed by Hickman (62); malignancy is deterred as long as unfettered proliferation remains linked to apoptosis. When the route to apoptosis is blocked, proliferative mechanisms are free to wreak havoc.

In the present study, we have used cDNA microarrays in tandem with pharmacological experiments to detect widespread molecular changes in cells and relate them to functional consequences. The overall protocol of the study included: (a) Annexin V and APO-BrdUrd flow cytometric assays to measure early and late stages of apoptosis, respectively; (b) hybridization of mRNA to cDNA microarrays for gene expression profiling; (c) real-time RT-PCR for validation of the microarray data (data not shown); (d) formulation of an annotated molecular interaction map for visualization and interpretation of the molecular changes and their consequences; (e) statistical analysis of the categories and relationships; (f) a first, prospective pharmacological test (using wortmannin) of PI3K involvement in the resistance of RC0.1 to apoptosis; and (g) formulation of a two-stage model for the development of resistance.

The study began with the hypothesis that DU145-RC0.1 differences other than the point mutation reported previously in the top1 of RC0.1 were influential in the development of resistance to camptothecin. To explore this hypothesis, we did microarray studies to identify the range of molecules involved and to assess their patterns of change. We first observed that the Akt-PI3K, caspase, and Bcl-2 family portions of the apoptotic machinery showed expression changes in the direction expected for increased resistance to apoptosis in RC0.1. However, the microarray study also yielded the puzzling finding that other sets of molecules associated with apoptosis, principally in the NFκB and TGFβ pathways, seemed to be changing in the contrary direction. Analysis of a molecular interaction map of the relevant genes led us to a two-stage “permissive apoptosis-resistance” hypothesis and model to explain the overall results.

Several obvious limitations of this study should be borne in mind. First, at this point in the development of the technology, mRNA expression profiling is attended by uncertainties, particularly with respect to the distinction among close gene family members. Second, not all of the genes potentially involved in drug resistance were present on the Oncochip array. Absent, for example, were the ATP-binding cassettes, subfamily G (WHITE), member 2 (bcrp), and subfamily C (mrp). Genes for multidrug resistance (63), mdr1 and 3 genes, were present but not significantly changed (R/D = 1.02 and 1.22, respectively), consistent with prior studies demonstrating that camptothecins are not mdr1 substrates (64). Third, differences in transcript expression may not reflect differences in protein expression. Fourth, functional differences may not be reflected directly at the transcript level. Single nucleotide polymorphisms, chromosomal aberrations, and post-translational modifications may be influential. As in any feasible biological study, we were able to investigate only a subset of the possible influences and secondary effects that may have been functionally important. The incompleteness of even this extensive, multifaceted study indicates the need for high-throughput, quite comprehensive approaches to biological systems, to synergize with work done on a gene-by-gene, factor-by-factor basis (65, 66, 67).

With respect to differences at the chromosomal level, we do have some contributory evidence, which will be presented in extenso elsewhere.⁹ Very briefly: (a) in bacterial artificial chromosome array-based comparative genomic hybridization studies done in a collaboration with the laboratory of Joe Gray (UCSF, San Francisco, CA)^10,11 we have observed a statistically highly significant correlation between DNA copy number and the expression levels of genes important in the results of the present study; and (b) routine cytogenetics plus spectral karyotyping showed greatly increased genomic instability in RC0.1 and a ploidy-relative amplification of the antiapoptotic gene BAD.

We thank Althea Jackson for editorial assistance in preparation of the manuscript, and Dr. Kimberly Bussey for advice and consultation on cytogenetic issues.