Molecular Determinants of the Cytotoxicity of Platinum Compounds

INTRODUCTION

The discovery and development of platinum compounds has been one of the greatest achievements of cancer chemotherapy in the past three decades. Cisplatin was accidentally discovered by Rosenberg in 1968 as an antibiotic and very rapidly as a very powerful antiproliferative agent against tumor cells. It was rapidly introduced in clinical use, principally for the treatment of germ cell tumors of testis and ovary (1) . The spectrum of activity of cisplatin includes, in addition to germ cell tumors, all types of squamous cell carcinomas but only some adenocarcinomas such as gastric cancer. The clinical use of cisplatin has been considerably hampered by its renal toxicity, which may lead to definitive renal failure and which requires special procedures of administration (2) . Neurological toxicity is also a problem, because long-term cisplatin administration may lead to disabling peripheral neuropathy.

A considerable effort has been made by the pharmaceutical industry for the discovery of cisplatin analogs. The main goal of this research was the identification of compounds devoid of the limiting toxicity of cisplatin (3) . Also, the emergence of cisplatin-resistant relapsing tumors after initial sensitivity has stimulated the research of active compounds presenting no cross-resistance with the original compound. Several hundreds of cisplatin analogs have been synthesized and screened in various systems, but only very few of them have shown a clinical interest deserving their introduction in the anticancer armamentarium (4) .

Carboplatin (see structure in Fig. 1 ⇓ ), the first analog to be approved for cancer chemotherapy, is now widely used in ovarian and testicular cancer as well as in the other indications of cisplatin. In contrast, the toxicity profile of carboplatin is quite different from that of cisplatin, with myelosuppression as the major dose-limiting toxicity. In view of the similarity of their indications and the difference in toxicity targets, the combination of these two platinum compounds has been proposed as a means of minimizing the toxic symptoms while preserving the antitumor efficiency (5) .

Another analog, oxaliplatin (see Fig. 1 ⇓ ), is characterized by a completely different spectrum of activity: its main indication has been found in colorectal cancer, a malignancy that is quite refractory to the classical platinum compounds. In contrast, no indication of this analog has been found in germ cell tumors or squamous cell carcinomas. Oxaliplatin is now included in the classical protocols of treatment of advanced colorectal carcinoma in combination with fluorouracil (6) .

No other platinum compound is currently available in the clinical setting, although several have been tested in Phase I and II trials. Among them, tetraplatin (ormaplatin, NSC363812; Fig. 1 ⇓ ) presents original features but its anticancer spectrum and toxicity pattern have not yet been fully identified (7) .

These four platinum compounds have been extensively studied against in vitro and in vivo models, with a special emphasis on the 60-cell-line panel of the National Cancer Institute (NCI; Ref. 8 ). Rixe et al. (9) have established the comparative in vitro cytotoxicity profiles of these agents, and have shown a strong similarity between cisplatin and carboplatin on one hand and between oxaliplatin and tetraplatin on the other hand. Using the program COMPARE to assess similarities between cytotoxicity profiles, Rixe et al. found that the Pearson correlation coefficients were 0.807 between cisplatin and carboplatin and 0.756 between oxaliplatin and tetraplatin but only 0.30–0.34 between cisplatin and either oxaliplatin or tetraplatin, and <0.1 between carboplatin and either oxaliplatin or tetraplatin (9) .

This clear-cut distinction between the two groups of platinum compounds can be interpreted as due to the existence of different mechanisms of action of the compounds, or at least of mechanisms of resistance. This prompted us to explore the molecular determinants of the activity of these compounds by establishing relationships between gene expression profiles and sensitivity to platinum compounds of tumor models. The first step of this exploration can be the mining of available data in public databases, especially the NCI database. The NCI has recently explored the expression of a panel of 9400 genes in the 60 cell lines that had been used for 10 years for establishing drug cytotoxicity patterns (10) . Using this database, we have conducted an in silico analysis to identify, in the 60 cell lines of the panel and among the 1416 genes or markers that had given the strongest variation pattern, the genes the expression of which was positively or negatively correlated to the sensitivity to each of the platinum compounds of interest, cisplatin, carboplatin, oxaliplatin and tetraplatin. We present here the result of this research, which can be considered as a paradigm of the NCI public database mining.

MATERIALS AND METHODS

We have used for this study the databases of the NCI that are accessible through the Internet, 4 both for the cytotoxicity data of the four platinum compounds in relation to the 60 cell lines of the panel and for the gene expression data in the same cell lines. A comprehensive review on the access facilities and the generation of matrices has been published by the NCI (10) .

In a first part of this work, the cytotoxicity data of the four platinum compounds in relation to the 60 cell lines were extracted and converted into Microsoft Excel format (A-matrix). We have selected −(log GI₅₀) as the parameter representative of the cytotoxicity of the compounds. Then, the gene expression data of the 60 cell lines were extracted and also converted into Microsoft Excel format (T-matrix). As specified by Scherf et al. (10) , these gene expression data had been obtained at the NCI using glass microarrays. They correspond, for each gene, to the log of the ratio of the signal provided by a cell line to the one provided by a pool of 12 cell lines chosen as a reference. As in the study of Scherf et al. (10) , the analysis was limited to a subset of 1376 genes that showed strong patterns of variation among the cell lines and had <5 of 60 values excluded on the basis of visual quality control or low signal. Besides the available gene markers, 40 other markers corresponding to protein quantification by Western blotting or functional assays were included in the database as carried out by the NCI (10) . Because they could be relevant for tumor cell characterization, they have been treated as gene expression data.

For each molecular marker, we calculated the Pearson coefficients of correlation r between the level of its expression in the 60 cell lines and the degree of cytotoxicity of each of the four compounds. These coefficients of correlation have, therefore, 58 degrees of freedom and are representative of the association between the activity of a drug and the expression of a gene. A significantly positive r value allows one to consider a gene as associated to drug efficacy, whereas a significantly negative r value allows one to consider it as associated to drug resistance. Finally, we calculated the Pearson coefficients of correlation between the r values obtained for two different drugs, to evaluate the degree of similarity of the two drugs in terms of the genes associated to drug efficacy or inactivity. These coefficients of correlation have, in this case, 1414 degrees of freedom. In all of the analyses, the Pearson coefficients of correlation have been considered as significant only for P < 0.001.

For the genes identified as significantly correlated to the activity of at least one of the four platinum compounds, we completed the information provided by the NCI database by searching all of the information available in the various public databases (UniGene, LocusLink). 5 Gene functions were identified by mining the database of the Gene Ontology Consortium, 6 which gathers in a convenient manner all of the information available concerning molecular and cellular functions of the genes products.

In a second part of this work, we have explored the literature dealing with the identification of individual genes involved in cell sensitivity or resistance to platinum compounds. This identification originates mainly from studies on cisplatin- or oxaliplatin-resistant cell lines. Using this approach, we identified 18 genes or protein markers; the expression of these individual markers was searched in the whole NCI database, covering not only the original subset of 1376 genes and 40 individual molecular targets but also the complete set of 9400 genes and 255 molecular targets. The coefficients of correlation between the cytotoxicity of the four platinum compounds toward the 60 cell lines and the level of expression of these 18 markers were then calculated as for the exploratory step of 1416 markers.

Coefficients of correlation should be used with caution in such studies: a highly significant correlation is in no way indicative of a causal relationship; in addition, a coefficient of correlation can be highly significant without showing a gene to be a strong determinant of drug activity. With 58 degrees of freedom, a coefficient of correlation is significant at the P < 0.001 level from 0.394 and above; this also means that only 16% (0.394²) of the variation is explained by the correlation. In contrast, a coefficient of correlation of 0.85 indicates that 72% of the variation observed can be explained by the correlation. Finally, because a significant coefficient of correlation does not indicate the existence of a linear relationship between two groups of values, and because one or two extreme values can drive a coefficient of correlation to significance, we have first examined the graphs visually as often as possible, and we have also systematically recalculated the coefficients of correlation after elimination of the extreme values on both sides. No variations higher than 10% of the r values was found in these conditions.

RESULTS

Data Mining in the NCI Database.

We first compared the cytotoxicity of platinum compounds to gene expression in the 1416-gene data set. The Pearson coefficients of correlation r between drug cytotoxicity and gene expression ranged between −0.55 and 0.55. In view of the high number of degrees of freedom, a r value of 0.394 or higher is significant at the P < 0.001 level. In these conditions, 204 genes had their expression level significantly correlated, negatively or positively, with the cytotoxicity of at least one of the four drugs (Table 1) ⇓ . In comparison with cisplatin or carboplatin, high numbers of genes had an expression significantly correlated with the cytotoxicities of oxaliplatin and tetraplatin. For comparison, the same analysis was conducted for drugs of other classes; the numbers of genes having their expression correlated with cytotoxicity was of the same order of magnitude for drugs like doxorubicin, vincristine, paclitaxel, fluorouracil, and melphalan (Table 1) ⇓ . When considering two platinum-containing drugs together, it appeared that the expression of a large number of markers was significantly correlated to the cytotoxicities of both oxaliplatin and tetraplatin, whereas a smaller number of markers were correlated to the cytotoxicities of any other group of two drugs (Table 2) ⇓ . For only two markers was the level of expression correlated to the degree of cytotoxicity of three drugs together.

We then calculated the Pearson coefficients of correlation between the coefficients of correlation, relating cytotoxicity to gene expression for each of the four drugs (Table 2) ⇓ . As expected, we obtained very high values for the correlations between oxaliplatin and tetraplatin on one hand, and carboplatin and cisplatin on the other hand. Much lower figures were obtained when comparing oxaliplatin and either carboplatin or cisplatin and when comparing tetraplatin and either carboplatin or cisplatin. Thus, the molecular markers of drug activity are very similar for oxaliplatin and tetraplatin, and for cisplatin and carboplatin.

Among the 204 markers significantly correlated with the cytotoxicity of at least one of the four platinum compounds, 4 of them have been detected from individual protein assays. Among the 200 remaining probes, 33 were not correctly identified in the NCI database, with the UniGene accession numbers corresponding to another gene than the description provided. Among the remaining 167 nucleic acid probes, a large amount of expressed sequence tags were used on the NCI glass microarrays; 46 of them are not described, whereas 37 received an official gene symbol and name by the HUGO Gene Nomenclature Committee. 7 In some cases, different probes corresponded to the same gene, with very similar expression patterns, which, finally, limited to 116 the total number of identified genes for which expression was significantly correlated to the activity of any of the platinum drugs. A complete list of these 116 identified genes markers and 4 individual protein markers is given on Table 3 ⇓ . A total of 191 different functions have been attributed to these 116 probes by the Gene Ontology Consortium. Among the functions or groups of functions, the most represented are the following: cytoskeleton-related protein activities (actin, myosin, kinesin, cell motility, and so forth; 22 genes); signal transduction-related proteins (16 genes); cell adhesion-related proteins (16 genes); cell proliferation control (15 genes); extracellular matrix proteins (14 genes); and transcription factors and related proteins (12 genes).

Some features apparent in Table 3 ⇓ are especially worth the mentioning. First, the activity of all platinum compounds was positively correlated with the γ-ray inducibility of the MDM2 protein or of G₁ arrest, two p53-dependent events requiring integrity of the p53 pathway (11) . Second, the activity of oxaliplatin was significantly higher in cell lines with a mutation in one of the RAS genes, whereas there was no correlation between RAS mutations and the activity of the other platinum compounds. This is in contrast to the observation that transfection with mutated Ha-RAS gene induced cisplatin resistance in NIH 3T3 cells (12) . It is worth remembering that cisplatin is particularly effective against ovarian and testicular tumors, which have a low frequency of mutated RAS alleles, and that oxaliplatin is rather active against colon cancers, which frequently exhibit a RAS mutation. This may justify the experimentation of this drug in pancreatic cancer, in which RAS mutations are especially frequent. And third, the activity of oxaliplatin and that of tetraplatin were positively correlated to the expression of the MYC oncogene: there again, this may suggest that the tumors that overexpress this oncogene could be particularly sensitive to these platinum analogs. The activity of tetraplatin was highly correlated with the level of topoisomerase II; this result can be interpreted as a mere relationship between cytotoxicity and cell proliferation, and it may also signify that tetraplatin-induced DNA damage preferentially occurs at the level of DNA–topoisomerase II cleavable complexes. We present in Fig. 2 ⇓ the diagrams showing some significant relationships between molecular markers and the cytotoxicity of selected platinum compounds.

When considering the genes associated with the activity of drugs of other classes (doxorubicin, vincristine, paclitaxel, fluorouracil, and melphalan), we observed that numerous genes were shared by drugs having no mechanistic relationship: 4 identified genes were significantly (P < 0.001) related to both doxorubicin and cisplatin or carboplatin activity, and 13 to both doxorubicin and oxaliplatin or tetraplatin activity. The corresponding figures were 0 and 11 for paclitaxel, 0 and 7 for fluorouracil, 6 and 33 for paclitaxel. No gene was shared by vincristine and any platinum compound.

Literature Mining in Relation to the NCI Database.

The second part of this work was based on literature analysis. From numerous studies performed with cell lines selected for resistance to cisplatin, several genes have already been associated with the acquired resistance to this compound (13, 14) : this is the case for mismatch repair proteins MLH-1 and MSH-2 and for high-mobility group proteins, which have been found to be deficient in cisplatin-resistant cells (15) . This is also the case for the detoxification systems involving glutathione or metallothionein, the levels of which have been found to be increased in cisplatin-resistant cell lines (16, 17, 18) . Consequently, the glutathione conjugate export pump, MRP1, appears overexpressed in cisplatin-resistant cell lines (19) . Some early-response genes encoding transcription factors such as c-MYC, c-FOS, or c-JUN are also overexpressed in cell lines selected for resistance to cisplatin (20, 21, 22) . Finally, a number of genes involved in the cellular response to DNA damage, such as TP53, BAX, BCL-2, and BCL-X_L, have been shown to interfere with cisplatin sensitivity (23, 24, 25) . For instance, transfection of the antiapoptotic gene BCL-2 confers resistance to cisplatin in the ovarian cell line A2780 (23) . Concerning oxaliplatin, some determinants of its activity, either in cell lines or in human tumors, have been described previously (26, 27, 28, 29, 30) . In particular, the DNA nucleotide excision repair (NER) has been shown to be involved in the repair of oxaliplatin-induced lesions, and the expression of the ERCC1 gene negatively correlated with oxaliplatin activity (29 , 30) .

We have thus identified 18 markers among those available in the complete NCI database (9400 genes and 255 molecular markers) as potentially related to cisplatin or oxaliplatin sensitivity or resistance from the data available in the literature. We have indicated in Table 4 ⇓ the coefficients of correlation that were obtained between the cytotoxicity of the four platinum compounds and the level of expression of these markers. No significant relationship was found between the cytotoxicity of any platinum compound and the level of p53 protein, evaluated by Western blotting; however, there was a relationship between the presence of a TP53 mutation in a cell line and its sensitivity to most platinum compounds, although with a relatively low significance level (0.01<P < 0.05). This explains the important role of the p53-dependent events such as the radiation inducibility of MDM2 and G₁ arrest, which were found to be correlated with the cytotoxicity of most platinum analogs in the previous step. The ERB-B2 oncogene, which had been mentioned as overexpressed in cisplatin-resistant cell lines (31) , was found negatively correlated to the cytotoxicity of tetraplatin and oxaliplatin (P < 0.001) and, to a lesser extent, to that of cisplatin and carboplatin (P < 0.05). In the BCL-2 family, the expression of the antiapoptotic gene BCL-X_L was the only one to be (negatively) correlated with the cytotoxicity of the platinum compounds, especially cisplatin and carboplatin. Concerning the DNA mismatch repair proteins MLH1 and MSH2, which have been found to be deficient in cisplatin-resistant cell lines (15) , we obtained positive coefficients of correlation between the level of their expression and the cytotoxicity of tetraplatin and oxaliplatin, but much below the level of significance (P > 0.05). Unfortunately, the nucleotide excision repair proteins ERRC1 and XPA were not present in the NCI database, which prevented the establishment of relationships between the level of their expression and oxaliplatin resistance. Finally, the platinum detoxification systems involving glutathione or metallothionein were weakly correlated to the cytotoxicity of platinum compounds: only the mRNA expression of the ATP-binding cassette pump MRP1 seemed correlated to the resistance to tetraplatin (P < 0.001) and to oxaliplatin (P < 0.05). In addition, metallothionein seemed paradoxically correlated to the sensitivity to tetraplatin.

DISCUSSION

In this study, we have explored the public databases of the NCI to identify the molecular markers associated with cell sensitivity and resistance to four platinum compounds of clinical interest. A total of 204 marker genes have been found, which are associated with a high level of significance to drug activity, but only 120 of them could be identified with certainty. In a first approach, we can consider these molecular markers as involved in the cellular response to the drugs. A large number of the markers that were identified as associated with the activity of the drugs correspond to genes involved in cell proliferation, in cell adhesion, in the cytoskeleton, and in transcriptional control. If one considers that chemotherapy should be prescribed in the future as a function of the molecular characteristics of the tumor, and not only as a function of its localization, the identification of the molecular determinants of drug action is a crucial enterprise. In this respect, the availability of large databases relating drug activity to gene expression profiles constitutes a powerful tool to identify the tumor characteristics that will allow the optimal choice of the anticancer drug on an individual basis. The NCI has implemented such a tool that allows a multitude of possible minings such as the one we have performed. Other databases have recently been constituted and made available (32 , 33) with the same purpose. Starting from the clues given by in silico research, one can validate in clinical studies the set of markers that appear, simultaneously, the most significantly predictive of drug action and the easiest to detect in routine pathology laboratories.

There are, however, several limitations to the type of study that we have developed and the reader should be aware of them:

1. The evaluation of gene expression was performed on a subset of 1416 genes and molecular markers, which represents a relatively small fraction of the total transcriptome, because ∼30,000 genes are present in the human genome, 10,000 of which are thought to be expressed in a single cell. The original glass arrays concerned a total of 9400 genes (nearly one-third of the genome), but the results had not been obtained on a sufficient number of cell lines of the panel to be included in the database (10) . As a consequence, not included in the database are many genes that may play an important role in determining cell sensitivity to the drugs. The existence of other databases including the expression levels of larger sets of genes in the NCI cell line panel will allow the extension of this type of in silico research to the complete genome (34) .

2. The level of expression of the 1416 molecular markers was determined with a technique that was still under development and not fully validated. Nothing is known on the accuracy of the gene expression data provided in the database. In particular, the lack of reproducibility of the crucial reverse-transcription step is a relatively important cause of error. A recent study comparing the gene expression data obtained for the 60 cell lines of the NCI panel by three different laboratories evidenced not more than 36% of statistically significant results for those markers present in the three data sets (35) . In the NCI data set that we have used, the level of expression of the 1376 genes in the 60 cell lines is normalized to the mean expression of these genes in a subset of 12 cell lines (10) . These levels of expression are, therefore, only relative to an arbitrary standard and this may introduce an important bias, especially for the genes that are expressed at a low level.

3. The criterion for drug cytotoxicity that has been retained by the NCI is the 50% growth inhibitory concentration (GI₅₀) and not cell kill. There has been much controversy about the relevance of growth inhibition to evaluate drug cytotoxicity (36) . For instance, growth inhibition by any drug is clearly related to the p53 status of a cell line, whereas clonogenic survival is not (37) . Therefore, one can wonder whether the GI₅₀ used by the NCI is the ultimate test to evaluate drug activity. Indeed, an observation made in the NCI database raises the problem of the relevance of this parameter: when considering not only the four platinum compounds, but the 121 active drugs of the NCI core database, the coefficients of correlation between the r values relating drugs and gene expression are in majority highly significant; when comparing each of the 121 drugs to the 120 other ones, the correlation is significantly positive at the 0.001 level in 80% of the possible cases. In other words, the levels of gene expression are correlated to a general drug sensitivity phenotype, independently of the drug. These markers may, therefore, be involved in cell proliferation rather than in drug activity and may reflect the properties of the cell lines rather than the determinants of drug activity. The fact that numerous genes are shared by drugs with quite different mechanisms of action argues in favor of this interpretation.

Several markers that we have identified as significantly correlated to the cytotoxicity of platinum compounds were already known as determinants of drug activity: this is the case for the functionality of the p53 pathway or for the expression of ERB-B2 or BCL-X_L proteins. Numerous other markers were also identified as involved in platinum activity, with a marked difference between the group constituted of cisplatin and carboplatin and the one constituted of tetraplatin and oxaliplatin. For instance, the significant link that we observed between oxaliplatin sensitivity and c-MYC expression was rather unexpected, although platinum compounds have been shown to activate the c-MYC gene promoter (38) . So, too, was the highly significant correlation between tetraplatin sensitivity and topoisomerase II expression.

Concerning the relationship with the presence of RAS mutations, it had previously been shown on the NCI human tumor cell line panel that there was a strong correlation between the sensitivity to cytarabine and, to a lesser extent, to topoisomerase II inhibitors, and the presence of activating RAS mutations (39) . This suggests that a common pathway of drug-induced cell death may exist for oxaliplatin, cytarabine, and some topoisomerase II inhibitors, and that it involves a dependency on the signal transduction pathway mediated by RAS. The RAS gene could be considered, therefore, as a molecular target for several apparently unrelated drugs. The concept of “oncogene addiction,” recently proposed by Weinstein (40) , may help understand how the targeting of an activated oncogene results in successful chemotherapy. In relation to this concept, the finding that several classical drugs may be active through oncogene targeting is of uppermost interest for drug development.