Comparative analysis and integrative classification of NCI60 cell lines and primary tumors using gene expression profiling data

Cell lines derived from primary tumor tissues have provided a valuable tool for the understanding of cancer biology at the molecular level. Ever since the development of Hela, the first human cancer continuous cell line by George Gey, Margaret Gey and Mary Kubicek more than 50 years ago, cancer cell lines have been instrumental as in vitro model systems in cancer research [1]. Much of the knowledge that we have today on fundamental processes in cancer cells has largely depended on the use of cell lines. In addition, since cancer cell lines provide an unlimited source of malignant cells, they are widely used in screening for anti-cancer drugs. However, because cells cultured in vitro lack the overall tissue architecture including tumor microenvironment, the value of cancer cell lines depends on if and to what extent cancer cell lines represent primary tumors that they are derived from. Some cancer cell lines undergo phenotypic and genotypic changes due to genomic instability. Many factors such as cross-contamination can cause cell line misclassification [2]. A recent study of 500 leukemia cell lines determined that 15% of these cell lines had been misclassified [3].

Several approaches have been utilized to authenticate cancer cell lines. The ability to form tumors when cell lines were transplanted subcutaneously into nude mice allows a direct comparison of histopathology between tumors formed in nude mice and the human tumors of origin [4]. Efforts have been made to delineate morphological features of cell lines in comparison with archival tumor tissues that the cell lines are derived from [5, 6]. At molecular levels, expression of key proteins such as HER2/neu and p53 in breast and non-small cell lung cancer cell lines and their corresponding tumors have been assessed using immunohistochemistry [5, 6]. Widely used in forensic analysis, DNA finger printing has been a valuable technique in analyzing loss-of-heterozygosity and microsatellite alterations [7]. Through current finger printing technology, one can not only confirm the identity of established cell lines and identify new cell lines, but also evaluate the purity of a cell culture [3].

The advent of high-throughput technologies, together with the completion of human genome sequencing project has created a new paradigm of understanding biology by simultaneous measurement of tens of thousands of genes in each biological sample. Numerous studies have employed microarray technology for classification and characterization of cancers. Comparisons between breast [8] or lung tumors [9] and tumor tissue derived cell lines have been attempted by gene expression profiling. However, a two-way hierarchical clustering generated dendrograms with cell lines and primary tumors falling into two isolated groups [8]. Recently, a tissue similarity index was developed to compare cell line and primary tumor gene expression profiling data using expression of tissue specific genes [10]. However, this approach that depends solely on tissue specific gene expression is problematic since genes not selected in the analysis may represent key pathways in cancer development.

NCI60 represents the most commonly used cancer cell lines in cancer research and drug screening. NCI60 cell lines have been extensively characterized by karyotyping [11], gene expression profiling [12, 13] and protein arrays [14]. Their sensitivities to more than 100,000 compounds have been measured by the National Cancer Institute's Developmental Therapeutics Program (DTP) [15]. Here we report a direct comparison between NCI60 cell lines and 9 primary tumor types using publicly available gene expression profiling data generated from more than 500 primary tumor samples. We used Pearson's correlation coefficients to assess the similarities between cell lines and primary tumors. Correlation coefficients between a cell line and its presumed tumor of origin were tested against those between the cell line and other tumors to examine if the overall genome expression profiles derived from the cell line most closely resemble those derived from its presumed tumor origin. Furthermore, supervised learning methods were applied to classify cell lines into subtypes of cancer or into different tumor developmental stages for lung, CNS cancers and acute leukemia where clinical data of primary tumor samples were available. Our results suggest that the majority of the NCI60 cell lines are representative of their corresponding tumor types and thus provide suitable model systems for the study of cancer malignancies.

Cancer cell lines have served as the primary experimental system for exploring cancer molecular biology and pharmacology. Although the value of cell lines in cancer research and anti-tumor compound screening is much appreciated, there is continued skepticism that cell lines under-represent the features of the primary tumors that they were derived from. Previous studies to some extent addressed these concerns by applying experimental approaches such as DNA finger printing to validate and authenticate cancer cell lines. In the last several years, microarray technology has been used to generate gene expression data for hundreds of tumor samples and provided a new paradigm of molecular based cancer classification. Previous work using gene expression profiling to compare cell lines and primary tumors has only focused on individual cancer types. A systematic effort in this area has been lacking to investigate if NCI60, the panel of cancer cell lines most widely used in cancer research and drug screening, represent their tumors of origin. Here we describe a novel approach to compare cell lines and primary tumors by computational analysis of publicly available gene expression profiling data of NCI60 cell lines and more than 500 primary tumor specimens. We were able to not only provide evidence to determine if a cancer cell line is correctly labeled to represent its corresponding tumor origin, but also classify a cell line into tumor subtypes or stages that the cell line may be most appropriate as a model system.

In contrast to some perceptions, our analysis suggested that most cancer cell lines are representative of their original tumor types. Global gene expression profiles in 51 of 59 NCI60 cell lines are most similar to that of their corresponding tumor origins. Although 8 cell lines have gene expression profiles more similar to tumor types other than their presumed origins, they do have strong correlations to their corresponding tumor types with correlation coefficients in the range of 0.7 (Table 2). Extensive experimental follow-up studies are necessary to clarify our computational analysis. However, there are several possible explanations for some of the discrepancies between the labels of these 8 cell lines and their gene expression profiles. Both prostate cell lines PC-3 and DU-145 had expression profiles with lower correlation coefficients to prostate cancers than other tumor types (Figure 2; data not shown). The lack of maximal correlation could be due to that PC-3 and DU-145 are androgen independent but most primary prostate cancers are androgen dependent [21, 22]. The hormonal dependence affects cell growth and may cause significant changes of gene expression profiles [23]. Since the two cell lines were initiated from bone and brain metastases of prostate adenocarcinomas respectively [21, 22], an alternative explanation is that the progenitor cells that they are derived from had lost the gene expression patterns of differentiated cells from the prostate. Among the 7 breast cell lines, only NCI/ADR-RES did not have an expression profile most similar to breast cancer biopsies. NCI/ADR-RES, originally known as MCF7/adr, is adriamycin resistant and was established through selection of MCF7 cells that are resistant to stepwise increasing concentrations of adriamycin [24]. Therefore, it is possible that adriamycin resistant MCF7 cells had altered gene expression profiles that are characteristic of cells from breast tissue. Indeed, it has been shown that in drug resistant cell lines, expression of some genes are induced during the selection process [24–26]. Upon further inspection of our correlation analysis results, we discovered that NCI/ADR-RES gene expression profile is most similar to those of ovarian cancers (data not shown). This finding is consistent with a recent report that NCI/ADR-RES is strikingly similar to an ovarian cancer cell line OVCAR8 based on a karyotyping study [11] and therefore supports our approach in using global gene expression profiles to evaluate the similarity between cell lines and primary cancer tissues.

Because we were able to obtain clinical annotations of the tumor samples used in gene expression profiling studies [16, 18, 19], 7 non-small cell lung cancer, 4 CNS cancer and 3 leukemia cell lines were classified into tumor subtypes or stages using supervised learning methods. 6 of the 7 NSCLC lines were classified as adenocarcinomas. This result has clarified the confusion in the literature regarding classification of some NSCLC cell lines. For example, NCI-H226 has been annotated as an adenocarcinoma cell line in some studies [27] but as a squamous carcinoma cell line in others [28]. Our supervised learning based classification using more than 100 primary tumor samples as the training dataset strongly suggest NCI-H226 is of adenocarcinoma origin. We also predicted tumor stages for the 6 adenocarcinoma cell lines. A549 falls into the stage III/IV group and the other 5 were predicted to represent stage II tumors (Table 3). Since the 79 gene feature that we used to build the classification model exhibited a similar expression profile between the stage II and stage III/IV patients (Figure 4), the 5 cell lines classified as stage II tumors could be in the stage III/IV group as well. Nevertheless, our results are consistent with the fact that cancer cell lines are generally derived from late stage cancers that have accumulated necessary genetic mutations for unlimited growth in vitro.

Our prediction of leukemia subtypes for 4 of the 6 acute leukemia lines is not in agreement with its description provided by American Type Culture Collection (ATCC). Specifically, SR and RPMI-8226 are classified into the AML category. However, they are lymphoma and myeloma cell lines, respectively, and therefore we recognize it would be inappropriate to attempt to classify them into either AML or ALL since lymphoma and myeloma are pathologically different than leukemia. Moreover, K-562 is also classified into the AML class, but it is derived from a patient with chronic myeloid leukemia (CML). Similarly, it would be inappropriate to classify it into one of the acute leukemia subtypes, AML or ALL. MOLT-4 is documented to be derived from a patient with ALL. However, we classified it into the myelogenous origin because it exhibits gene expression patterns more similar to AML than ALL (Table 3; data not shown). Characterization using immunological, cytogenetic and molecular biology approaches has clearly confirmed the identify of MOLT-4 as an ALL cell line [29]. One possible explanation for our misclassification is that we built classifiers based on gene expression profiles of 20 ALL and 28 AML patients from a single cohort. Therefore, even we achieved an error rate of zero in leave-one-out cross validation, it is possible that our classifier is not generally applicable. When microarray data on more ALL and AML patients from different cohorts become available in the future, we might be able to more accurately classify leukemia cell lines by building an improved classifier.

Attempts have been made in the past for comparison and integrative classification of primary tumors and tumor derived cell lines using gene expression profiling [8, 9]. However, hierarchical clustering analysis generated dendrograms with cell lines and primary tumors falling into two distinct branches [8]. We also observed such separations when array datasets for NCI60 and 9 primary tumor types were subjected to hierarchical clustering (data not shown). One possible explanation is that gene expression differences between cell lines of different tumor origins are overshadowed by more significant differences between cell lines and primary tumors. Here we developed a simple analysis of correlation coefficients as a metric to measure the similarity between each cell line and different tumor types. In our approach, the general differences between cell lines and primary tissue samples would not interfere with comparisons between a cell line and multiple tumor types. Unlike previous studies that artificially selected genes differentially expressed between cell lines and primary tumors [9], our testing is robust and unbiased as we only removed genes that are deemed unexpressed in more than 80% of the array datasets. More significantly, to our knowledge, this is the first attempt to predict tumor subtypes or stages that the cell lines are suitable as model systems by applying supervised learning methods that have error rates close to zero in cross validations.

We also recognize the limitations in our study. First, gene expression profiles in cell culture in vitro may not reflect gene expressions evaluated when cells are grown in vivo. A recent study has shown that although two glioblastoma cell lines (U251 and U87) have disparate gene expression profiles when grown in monolayer cell cultures, they had similar gene expression patterns when grown as intracerebral xenografts in nude mice [30]. Therefore, a more insightful approach would be comparing gene expression profiles between primary human tumors and cell lines grown in xenograft models when such data become available. Second, correlation based analysis only provides hints on if a cell line is more representative of its presumed tumor origin than other tumor types, but does not provide a quantative measure on how close the cell line represents the corresponding cancer. Third, subclassification of cancer cell lines is not in fact available. Therefore, it is not possible to experimentally validate our prediction. Our classification based on supervised learning approach only provides suggestions on which cancer subtype that a cell line most closely represents. Fourth, overall genomic gene expression profiles were used in our statistical testing of gene expression correlations between cell lines and tumor samples. However, specific genes and pathways may be involved in development of different types of cancer. When this manuscript was in preparation, a study published by Minn et al. has identified 54 genes differentially expressed between one of the NCI60 breast cell lines, MDA-MB-231 and its single-cell-derived sub-line that has significantly higher lung metastatic activity [31]. Expressions of this lung metastasis signature not only mediate lung metastasis of MDA-MB-231 cells in mice, but are also significantly associated with breast cancer metastasis to lung in human [31], suggesting that the same set of genes plays a critical role in lung metastasis both in cell line models and in humans. Therefore, future analysis using expression of selected genes in known cancer development pathways would certainly shed light on the activity of these pathways in cancer cell lines in comparison with human tumors.

We developed and applied a novel approach for systematic comparative analysis and integrative classification of NCI60 cell lines and primary tumors. Comparative analysis suggested that 51 of 59 NCI60 cell lines we analyzed represent their presumed tumors of origin. Some of the 51 cell lines with the correct origins were further classified into different subtypes of cancer or different stages in cancer development based on supervised learning methods. Our results could provide guidance to the selection of appropriate cell lines for cancer research and pharmaceutical compound screenings. Furthermore, this gene expression profile based approach can be generally applied to evaluate experimental model systems such as cell lines and animal models for human diseases.