An Informatics Approach Identifying Markers of Chemosensitivity in Human Cancer Cell Lines¹

The NCI-ACDS⁴panel (1, 2) consists of 60 cell lines selected from human tumors of different tissues of origin (breast, central nervous system,colon, bone marrow, lung, skin, ovary, prostate, and kidney). Nearly 100,000 chemical compounds have been tested for cytotoxicity in this panel using a 2-day assay to determine the GI₅₀in each cell line (1), with more compounds being tested continually. In addition to identification of potential antineoplastic drugs, these 60 cell lines are also being characterized on a molecular level to define alterations in genes that may contribute to carcinogenesis. Appropriate data mining of such databases may in turn aid in the development of compounds with specific cytotoxicity directed against cancer cells with particular molecular characteristics(3).

Although expression at the protein level (4) may be a more accurate predictor of activity in the cell, mRNA expression can be a very useful end point for comparison between cell lines. When mRNA levels are informative, their measurement would have the practical advantage of requiring a relatively small sample as well as being more quantitative, rapid, and sensitive. It is also possible to use RNA analysis for newly identified target genes in advance of antibody availability. Microarray technologies promise to yield quantitative measurement of thousands of mRNA levels at once (5, 6),which paves the way for such applications as molecular tumor profiling and precise tailoring of individual chemotherapy regimens.

Accurate measurements of relative basal mRNA levels in human cancer cell lines may provide insight into mechanisms of molecular regulation,interactions, and drug action when this information is considered in the context of the other data in the NCI-ACDS. For instance, our laboratory recently identified a subset of the p53 wild-type cell lines with reduced or absent induction of GADD45 afterγ-irradiation (7). When the NCI-ACDS database was searched for differential sensitivity of this subset of cell lines to cytotoxic drugs, topoisomerase II inhibitors were found to be significantly less toxic in the cell lines with defective GADD45 induction (P < 0.0001). This difference was confirmed by the demonstration that etoposide 16 toxicity and DNA-protein cross-links were decreased when GADD45 expression was blocked with an antisense vector and led to discovery of a role for Gadd45 in regulating chromatin accessibility (8). This example demonstrates how mechanistic insight can be gained by exploring hypotheses suggested from analysis of a large, complex, but well-controlled biological survey.

We have performed a careful quantitation of the relative basal levels of 10 transcripts chosen for their known roles in cellular damage responses or cancer biology and as potential modifiers of toxicity(O6AT, CIP1/WAF1, GADD34, GADD45, GADD153, cMYC, MDM2, BAX,BCL2, and BCL-X_L) in the 60 cell lines of the NCI-ACDS and looked for correlations between levels of these transcripts, p53 status, and cytotoxicity of the tested compounds in the NCI-ACDS database. Many statistically significant and interesting associations with existing molecular target profiles were observed. In this report, we highlight those we found most provocative,but other correlations can be explored.⁵ One of the most striking findings in the present study was the overall negative correlation between BCL-X levels and sensitivity to the 122 “standard” chemotherapy agents with the exception of one group of antimetabolites. This correlation was stronger than the positive correlation with cytotoxicity previously reported for p53 across diverse mechanisms of drug action(9) and suggests the importance of endogenous BCL-X levels in the cellular response to chemotherapeutic agents, independent of p53 status.

Cells from the NCI-ACDS cancer cell panel (9) were grown in RPMI 1640 supplemented with 10% fetal bovine serum and 100 units/ml penicillin (Sigma) and 100 μg/ml streptomycin (Sigma). All of the cultures were maintained at 37°C in a humidified 5%CO₂ atmosphere. Cells were lysed for RNA harvest at 40–60% confluence. Suspension cell lines were maintained between 1 and 10 × 10⁵ cells/ml and were harvested in log phase growth at 4–5 × 10⁵ cells/ml. All of the cultures were fed with fresh medium the day before harvest.

mRNA was isolated by the guanidine thiocyanate method of Chomczynski and Sacchi (10), followed by poly(A)purification using oligodeoxythymidylate cellulose as previously described (11). cDNA probes for GADD34, GADD45, and GADD153 were obtained by excising the insert from pHu34B (12), pHul45 (12), and pHul75 (13), respectively. Other cDNA plasmids were obtained from Oncor (c-MYC) or were generously provided by B. Vogelstein of The Johns Hopkins University, Baltimore, MD(MDM2), S. Korsmeyer of the Washington University School of Medicine, St. Louis, MO (BAX and BCL2), W. El-Deiry of the University of Pennsylvania School of Medicine,Philadelphia, PA (CIP1/WAF1), and L. Boise of the University of Miami School of Medicine, Miami, FL (BCL-X). The cDNA inserts were labeled with ³²P using the PrimeIt RT system (Stratagene) according to the directions of the supplier. Twofold serial dilutions of the mRNAs were fixed on nylon membranes, with six copies of each filter being made from the same RNA dilution at one time. For each transcript, hybridization to the labeled cDNA probe was carried out on a complete filter set (which included RNA samples from all of the cell lines in the screen) in the same hybridization mix. High-stringency hybridization and wash conditions were as previously described (11). Hybridization was quantitated on a phosphorimager (Molecular Dynamics). Relative signal levels, normalized to the poly(A) content of each sample, were determined using the RNA Analysis program as previously described(14). Relative protein levels of Bcl-X_L and Bax were measured using standard Western blotting techniques as previously described (15).

Growth inhibition patterns (IC₅₀ values for 60 human tumor cell lines) were those available from the NCI Developmental Therapeutics Program.⁵ Values were reexpressed as potency values by using the negative log of the molar concentration calculated in the NCI screen. The dependence of drug potency on gene expression levels was gauged using either the Spearman correlation coefficient for continuous value gene expression measurements, such as GADD45 expression, or the Wilcoxon Rank Sum test for binary gene expression measurements, such as Mer−/Mer+ or p53 wild-type/mutant. Positive correlations occurred when relatively high levels of gene expression were found in relatively sensitive cell lines. Negative correlations occurred when relatively high levels of gene expression were found in resistant cell lines.

Partial correlations were calculated using SAS Proc Corr (SAS Institute, Cary, NC); e.g., to find the residual correlation between chemosensitivity and BCL-X after “removing” any contributions attributable to a correlation between chemosensitivity and measured doubling time, Proc Corr was executed with drug potency and BCL-X levels listed as the variables to test correlation and with doubling time listed as the partial variable. In this way, the potential role of previously demonstrated sensitivity factors(“molecular targets”) as underlying factors responsible for the present observed correlations could be examined statistically.

To visualize a large number of correlation or Wilcoxon test results simultaneously, we first calculated the asymptotic Pfrom the test statistic for each test using normal approximation and then plotted the results for all of the tests in a histogram or color-coded table. The histogram reveals the number of drug sensitivity correlations that would have been considered statistically significant had each drug’s correlation with gene expression been evaluated alone. The same histogram can reveal positive or negative trends that might not have been detected by a simple count of individual results above a particular significance threshold. The colorized matrix (or in some cases a single column) of P also highlights the statistical significance of correlation test results as well as positive or negative trends but allows the display of results according to a logical (drug mechanism of action) or empirical (cluster) order.

The P reported here correspond to the single-tailed test, in which the alternative hypothesis is that the correlation is negative and the null hypothesis is that the correlation is zero. As such, a significant negative correlation (α = 0.05) will be reported on our one-sided P scale as 0.05, whereas a significant positive correlation (α = 0.05) will be reported as 0.95. The latter case can be understood as there being a 95% probability of finding a more negative correlation by chance,whereas more positive correlations happen because of chance alone only 5% of the time, indicating with reasonable certainty that the correlation is positive.

To reduce the bias in the drug screen database toward the overrepresentation of some chemical structure and activity classes, a k-means, algorithm-based clustering process was developed.⁶A series of iterative clustering optimizations was performed on the NCI-ACDS IC₅₀ patterns to find 1200 clusters that were well populated but maximally diverse in pattern. Exemplar compounds, nonconfidential compounds closest to the center of each cluster, were taken to represent the entire database. In theory and as first demonstrated by early applications of the COMPARE program⁷and many examples since, highly correlated groups of compounds will share the same physicochemical properties and more importantly, often share mechanism of action properties. This phenomenon was named by the late Ken Paull the “COMPARE effect.” Appropriately, the sets of compounds represented by the database exemplars are termed compare effect clusters. When a molecular target pattern comprising some cell characteristic measured in the NCI-ACDS is used to search for correlating activity patterns, a search against the redundancy-reduced set of compare effect cluster exemplars is expected to provide a more concise and accurate survey of the molecular target’s role in measured cytotoxicity.

Many of the transcripts measured in this study previously have been shown to be induced by alteration in growth conditions, including confluence and starvation stress (16, 17, 18, 19, 20), so maintaining consistent growth conditions was crucial for meaningful and reproducible measurement of basal RNA levels. Culture conditions were especially important in lymphoid and myeloid lines, a point that is often overlooked as the concept of confluence does not apply to such cell lines. For instance, when the lymphoid cell line Molt4 was deliberately grown past exponential growth phase without the addition of fresh medium, GADD45, CIP1/WAF1, and MDM2 were all induced 10-fold or more over the levels present several days previously in the same culture while in exponential growth (data not shown). In the same experiment, however,no change was found in the levels of GADD34, GADD153, BAX,BCL-X_L, cMYC, or O6AT. To avoid the potential confounding effects of starvation, glucose deprivation,or other overgrowth-related stresses in various cell lines, all of the cultures were maintained in exponential growth, with suspension lines kept at densities of 2–10 × 10⁵cells/ml. All of the cell lines were fed with fresh medium the day before harvest of RNA. Attached cell lines were harvested at 40–60%confluence, whereas suspension lines were harvested at densities of 4–5 × 10⁵ cells/ml.

In previous studies with inducible genes (21, 22, 23)our laboratory has developed an accurate and rapid approach for comparative measurements of low-abundance transcripts (11)that is less labor intensive than RNase protection and that avoids the problems of normalization inherent in Northern blot analysis of weakly expressed transcripts. Serial dilutions of poly(A) RNA are hybridized to a radioactive probe and used to generate a standard calibration curve. This compensates for the nonlinearity (pseudo-first-order kinetics) of signal to mRNA content, which can be encountered in hybridizations, such as when the probe is not in excess(11, 24). Replicate membranes made with the same serially diluted RNA samples are hybridized to a polyuridylic acid probe to control for differences in the poly(A) RNA content between different samples or for any loading differences between lanes. The poly(A)content of different samples is generally within ∼25%, although greater variations did sometimes occur between different cell lines. Hybridization to a polyuridylic acid probe effectively corrects for such variations (14). This method is also much more reliable than normalization to so-called housekeeping genes, such as GAPDH or β-actin, which can show considerable variability between different cell lines. Overall, this technique has been shown to yield accurate determinations of low-abundance transcripts (on the order of 1/10⁵) and to reliably detect differences of 1.5-fold or greater (11, 14, 25) with results comparable with those of RNase protection (26). Such sensitivity and accuracy are critical for the meaningful comparison of uninduced basal levels of transcripts in different cell lines and exceed the degree of accuracy possible for measurements of relative protein levels or for many other methods of measuring relative mRNA levels, including most array applications. Using this approach, we determined the relative basal levels of the mRNAs for the genes GADD34, GADD45, GADD153, BAX, BCL2,BCL-X_L, MDM2, CIP1/WAF1, O6AT, and cMYC in the 60 human cancer cell lines of the NCI-ACDS. RNA samples from all of the 60 cell lines were hybridized together at high stringency in the same hybridization mix at the same time to avoid variations in hybridization conditions. The results of quantitative hybridization of these 10 transcripts are presented in Table 1 as basal levels relative to the levels in BT549 or in the case of BCL2, which was not detectable in this cell line, relative to the levels in MDA-N. Although some transcripts were more variable than others, in most cases variation was less than 10-fold across the NCI-ACDS.

The basal transcript levels in a subset of the lines were rechecked by comparing the RNA samples from the original isolation side by side with newly isolated RNA from freshly grown cells. Cell lines representing both the median and the most extreme expression levels were chosen for each probe. The values of the repeated measurements agreed well with the original values, showing only slight variations(Fig. 1).

For a number of the transcripts measured in this study, previous measurements of relative protein levels were available in the NCI-ACDS database. Comparison of our quantitation of basal mRNA levels with these protein data sets (Table 1) gives an additional indication of the robust nature of our measurements. As might be expected, there is a strong positive correlation between these two measurements where relative protein levels are available for comparison (BAX: r = 0.41, P = 0.001; BCL-X_L: r = 0.55, P = 0.000006). These mRNA and protein level measurements are in good agreement, but there is also substantial variation in the protein level that is not explained by the mRNA measurements. Some of this unexplained variation may be attributable to posttranslational protein regulation, whereas some may reflect the different sources of error in the very different quantitation techniques necessarily used and the inherent difficulties in normalizing relative protein content. Although mRNA levels may not be a direct substitute for protein measurements, they can nonetheless be of predictive value.

We first looked for correlations between the basal mRNA levels of the 10 genes measured in the cell lines of the NCI-ACDS. The strongest correlation between these targets is for CIP1/WAF1 and BAX (Fig. 2; r = 0.687; P = 0.0001). CIP1/WAF1 basal levels also correlated well with the basal levels of GADD45 (r = 0.471; P = 0.0001), a gene similarly regulated during cell growth arrest and stress responses(27). This may suggest that when both of these genes are expressed in a cell, they tend to maintain a relatively constant ratio with respect to each other. Previous experiments had suggested a compensatory mechanism whereby under stress conditions, a cell could offset a lack of GADD45 induction with an unusually prolonged and large magnitude CIP1/WAF1 response(7). Because the basal levels of these two genes correlate positively with each other rather than negatively, such compensation would appear to be engaged only during stress response, however, not during normal growth. The basal levels of the three major GADD genes as measured here also correlate with each other(GADD45 and GADD153: r = 0.36; P = 0.004; GADD45 and GADD34: r = 0.42; P = 0.0007; GADD153 and GADD34: r = 0.52; P < 0.00005), suggesting coordinate regulation of these genes under normal growth conditions and during stress response (12).

Despite reports that both basal (28) and stress-induced levels (29) of the GADD genes are down-regulated by overexpression of c-Myc, basal levels of the GADD genes do not correlate well with endogenous c-MYC mRNA levels (r = 0.13; P > 0.25). This may indicate that for Myc,mRNA is not a good predictor of protein levels or that additional regulators of the GADD genes may be more important in determining basal levels of these genes during exponential cell growth. It should also be noted that the previously reported down-regulation of the gadd genes by c-Myc occurred only in the presence of Myc overexpression, so it is possible that the effect of c-Myc levels on gadd gene regulation is not significant within the normal endogenous range of c-MYC expression.

The most significant relationship between p53 status and any of the basal mRNA levels measured here was for CIP1/WAF1(P = 0.002; Fig. 3,A). Induction of CIP1/WAF1 by ionizing radiation previously has been shown to be dependent on wild-type p53 function(30, 31). p53 may also play an important role in the basal regulation of CIP1/WAF1, because among p53 mutant cell lines, basal expression was extremely low in most cases. Basal levels of BAX, another gene with p53-regulated stress response(32, 33), also tended to be higher among p53 wild-type cell lines (P = 0.003; Fig. 3 B). Interestingly, no significant correlation was found between p53 status and basal levels of GADD45, despite the dependence of ionizing radiation induction of this gene on wild-type p53 function(34, 35) and the positive correlation of basal levels of GADD45 with those of CIP1/WAF1. Induction of GADD45 by other DNA-damaging stresses, such as UV radiation or alkylating agents, is regulated by both p53-dependent and-independent pathways (34, 36). The present results perhaps suggest a greater role for p53-independent mechanisms in maintenance of basal GADD45 levels.

The role of genes such as GADD45, CIP1/WAF1, and MDM2 in stress response has been well characterized both in our laboratory and others (18, 20, 37, 38), and the ionizing radiation response of all three genes has been found to be dependent on wild-type p53. Although in this study the basal level of only BAX and CIP1/WAF1 correlated with p53 status, we were interested to see if there was a relationship between basal levels and ionizing radiation induction of these genes or if induction raised the absolute expression to similar levels in all of the cell lines, irrespective of basal levels. The relative induction of mRNA by γ-rays was measured previously for these three genes in the 60 cell lines of the NCI-ACDS (9). Comparing the basal levels of these genes with their absolute induced levels (the product of the basal level and the relative fold-induction in each line) there is a trend of increased absolute induced levels with increasing basal expression both within the p53 wild-type subset of cell lines (Fig. 4) and across the entire NCI-ACDS panel (GADD45: r = 0.71; CIP1/WAF1: r = 0.63; MDM2: r = 0.94; all P < 0.00005), despite the relative lack of induction in the p53 mutant cell lines. This indicates that the stress response of these genes does not compensate for their variations in basal expression. Rather, the relative basal expression levels of these genes are an important determinant of absolute stress-induced expression levels, regardless of the fold-induction in a particular cell line. The finding that the relative relationship of gene expression levels among the cell lines is similar before and after genotoxic stress, irrespective of p53 status,supports the use of basal expression level data as an informative marker for predicting genotoxic response.

We next compared basal expression levels of the 10 genes quantitated here with sensitivity to a set of 122 standard chemotherapy agents in the NCI-ACDS (Table 2). Loss of detectable expression of O6AT at either the mRNA or protein level is estimated to occur in ∼15–25% of primary human tumors (39), resulting in the Mer− phenotype. A slightly higher fraction, 18/60 (30%), of the cell lines in the NCI-ACDS panel expressed undetectable levels of O6AT in our study (Table 1). To test the ability of our gene expression measurements to identify mechanisms of drug action, we first examined the correlation of Mer+/Mer−phenotype with activity of O⁶-alkylating agents(r = −0.44; P < 0.05). This contrasted with the lack of correlation between the Mer phenotype and toxicity of other alkylators (r = −0.04). This would be expected because the O6AT DNA repair protein is specific for the reversal of the O⁶lesions generated by these agents but is not active against other lesions or other mechanisms of toxicity (Fig. 5,A and B). As a specific example, a comparison of the GI₅₀ values for PCNU, one of the agents in this mechanistic class, is illustrated in Fig. 5 C for cell lines expressing (Mer+) or not expressing (Mer−) O6AT. Expression of O6AT has previously been shown to be a good predictor of sensitivity of a cell line to nitrosourea compounds such as this(39). These results provide a positive control illustrating the identification of drugs by mechanism of action through informatic analysis of RNA expression data and provide a validation of the methodology used here.

One of the most striking findings in this analysis is a broad negative correlation (r = −0.3) between basal BCL-X_L levels and overall drug sensitivity to all of the categories of drugs tested, with the exception of one subgroup, which includes antimetabolites such as thioguanine,thiopurine, and 5-azacytidine (Fig. 6,A). This pattern is almost an exact mirror image of the pattern of drug sensitivities correlated with p53 status (Fig. 6 B; an overall positive correlation of ∼0.23; Ref.9). Bcl-X_L is a member of the Bcl2 family of apoptosis-regulating proteins, which has been shown to protect against p53-mediated apoptosis (40, 41). An alternately spliced form of the same transcript codes for Bcl-X_S, which antagonizes the activity of the major Bcl-X_L protein to promote apoptosis(42), but this message is expressed at very low levels in most cell lines examined. Overexpression of BCL-X in murine lymphoid cells has been reported to have a protective effect against the toxicity of diverse agents including cyclosporin A, rapamycin,FK-506 (43), bleomycin, cisplatin, etoposide,vincristine, hygromycin B, mycophenolic acid (44),vinblastine, teniposide, methotrexate, fluorouracil, hydroxyurea, andγ-irradiation (45). Conversely, suppression of Bcl-X_L levels in a human cancer cell line by use of either BCL-X antisense or overexpression of its antagonist Bak was found to sensitize the cells to apoptosis after treatment with 5-fluorouracil or cisplatin (46).

Although overexpression of Bcl2, another major antiapoptotic protein,has also been shown to protect against killing by many different agents(47, 48), the basal levels of BCL2 mRNA in untreated cells show no correlation with overall sensitivity to the chemotherapy agents in this study (Fig. 6,C), possibly indicating Bcl-X_L as a more important regulator of cell death. In addition, levels of BAX, which codes for an apoptosis-promoting protein, do not show a comprehensive correlation with drug toxicity across the NCI-ACDS (Fig. 6 D) as might have been expected. This may suggest that Bax-independent mechanisms of Bcl-X_L action may play a predominant role in determining the survival of cancer cells exposed to chemotherapeutic agents. On the other hand, because only basal expression levels were measured here, we cannot exclude the possibility that drug-inducible levels of BAX or BCL-2 may correlate with tumor cell line responses to anticancer compounds. The pattern of correlations between BCL-X basal expression and cytotoxicity does appear to be specific to expression of this gene, however, and not merely a general hallmark of apoptosis-regulating genes.

Because BCL-X previously had been shown to be induced by ionizing radiation in a small subset of cancer cell lines(49), we measured γ-ray induction of this gene in an additional 31 of the 60 cell lines (data not shown). We found a relationship between basal and absolute induced levels similar to(r = 0.79; P < 0.0001) that found for the previously measured genes GADD45,CIP1/WAF1, and MDM2 discussed above, indicating that the relative relationship of BCL-X levels are also very similar before and after treatment. The relative inductions of BCL-X were furthermore not predictive of cytotoxicity in either the Standard set of 122 agents or the larger Unidrug set.⁸

Previous measurements of doubling time for the cell lines of the NCI-ACDS (9) also show a significant negative correlation across all of the drug classes examined (Fig. 6,E). This indicates, as might be expected, general increased chemoresistance the more slowly a cell divides. This pattern was similar to that found for BCL-X, so it was possible that BCL-X expression levels were acting as a marker for cellular growth rate rather than chemosensitivity. The correlation between doubling time and BCL-X expression is not significant, however(P > 0.05). Furthermore, examining the partial correlation of BCL-X levels and drug sensitivity by“removing” the statistical effect of doubling time still demonstrates a highly significant effect for BCL-X (Fig. 6,F), indicating an effect independent of doubling time. Similarly, the protective effect of BCL-X also appears to be independent of p53 status because there was no difference in distribution of BCL-X levels among the p53 wild-type and p53 mutant cell lines in the NCI-ACDS. The significance of correlations between drug sensitivities and BCL-X expression was also unaffected by p53 status of the cell lines, as demonstrated by either partial correlation (Fig. 6,G) or by the exclusion of the p53 wild-type cell lines in determining the drug sensitivity correlations(Fig. 6 H).

The NCI-ACDS screen for drug activity, the associated molecular target databases, and correlation analysis tools make a potent combination to identify markers associated with chemoresistance and other parameters relevant to cancer treatment. The novel factor analysis described here, which uses partial correlations to remove potential confounding effects of previously described sensitivity factors, increases the utility of this approach. We have used these analyses to find correlations between the parameters in the NCI-ACDS database and the basal expression of 10 genes involved in the cellular response to stress. Although many interesting relationships were uncovered, the most striking was the overall protective effect found for expression of BCL-X_L.

It is interesting that the protective effect of Bcl-X_L extends across all of the cancer cell types in the NCI-ACDS panel and is not restricted to lymphoid and myeloid cell lines, which are especially prone to undergo apoptosis. BCL-X_L induction by ionizing radiation previously has been shown to occur primarily in cell lines with both wild-type p53 function and a propensity to undergo rapid radiation-induced apoptosis (49). In contrast to the specificity of cell type required for BCL-X induction, the protective effect of high basal levels of BCL-X extends across the NCI-ACDS, depending neither on wild-type p53 function, nor on cell type. In a smaller set of cell lines, relatively low basal levels of BCL-X_L have been shown to correlate with a greater degree of γ-ray-induced apoptosis(49), suggesting the protective effect of high basal levels may extend also to ionizing radiation. Although no quantitative measurement of radiation-induced apoptosis has been made for the 60 cell lines of the NCI-ACDS, a comparison of the basal BCL-Xlevels in the lymphoid/myeloid panel (lines very prone to apoptosis;mean, 0.34 ± 0.10) with levels in all of the other cell lines (mean, 0.93 ± 0.08) indicates a similar trend. This implies that basal levels of BCL-X can play a major role in the determination of cellular response to a wide variety of chemotherapeutic agents, perhaps through modulation of apoptosis.

Unlike the sensitizing effect on cell killing previously seen for wild-type p53 (9), the protective effect of BCL-X extends beyond the more classical cancer therapeutics. This is illustrated with the “Unidrug 1247,” a set of 1247 agents from the NCI-ACDS database having maximally diverse activity patterns,and in theory, 1247 distinguishable biochemical mechanisms of cell killing.⁶ Although the dependence of toxicity on p53 status is no longer evident in this larger survey of agents, the(inverse) association between toxicity and BCL-X expression(Fig. 7, A and B) remains essentially unchanged from that seen across the smaller set of drug mechanisms. A similar pattern of correlations is also observed between drug activity and measurements of Bcl-X_L protein expression (Fig. 7, C and D). A χ² comparison of the drug activity patterns in the Unidrug 1247 for Bcl-X_Lmeasured as mRNA versus protein had a P < 0.0001, again indicating good agreement between the two measurements. In contrast, comparison between BAX expression at either the mRNA or protein level and sensitivity to this larger drug set shows no pattern of overall correlation (data not shown),consistent with the results in the initial set of 122 agents (Fig. 6,C). Unlike the case with BCL-X, when the bias introduced by multiple drugs sharing a small number of cytotoxic mechanisms is removed by considering the diverse mechanisms included in the Unidrug set, the strongly significant dependence of drug toxicity on wild-type p53 status (as seen in Fig. 6,B) is no longer strikingly apparent (Fig. 7 E), indicating that the protective effect of BCL-X expression may be much broader than the sensitization by wild-type p53.

The broad nature of the association between endogenous BCL-Xexpression and protection from chemotoxicity implicates this gene as an extremely important general determinant of cell death and suggests a p53-independent mechanism of Bcl-X_L action as a major component of chemoresistance in cancer cells. In addition, the lack of dependence on wild-type p53 function increases the attractiveness of Bcl-X_L as a potential target of cancer therapy, because the majority of human cancers have lost p53 function (50, 51). Such p53-mutant tumors are generally more refractory to chemotherapy, as reflected in the increased chemoresistance of cell lines with aberrant p53 (9). A therapeutic approach targeting Bcl-X_L may be able to circumvent the protective effect conferred by loss of normal p53 function in many tumors.

We thank Edward A. Sausville for his contributions to the NCI-ACDS database.