The development of a successful radiation sensitivity predictive assay has been a major goal of radiation biology for several decades. We have developed a radiation classifier that predicts the inherent radiosensitivity of tumor cell lines as measured by survival fraction at 2 Gy (SF2), based on gene expression profiles obtained from the literature. Our classifier correctly predicts the SF2 value in 22 of 35 cell lines from the National Cancer Institute panel of 60, a result significantly different from chance (P = 0.0002). In our approach, we treat radiation sensitivity as a continuous variable, significance analysis of microarrays is used for gene selection, and a multivariate linear regression model is used for radiosensitivity prediction. The gene selection step identified three novel genes (RbAp48, RGS19, and R5PIA) of which expression values are correlated with radiation sensitivity. Gene expression was confirmed by quantitative real-time PCR. To biologically validate our classifier, we transfected RbAp48 into three cancer cell lines (HS-578T, MALME-3M, and MDA-MB-231). RbAp48 overexpression induced radiosensitization (1.5- to 2-fold) when compared with mock-transfected cell lines. Furthermore, we show that HS-578T-RbAp48 overexpressors have a higher proportion of cells in G2-M (27% versus 5%), the radiosensitive phase of the cell cycle. Finally, RbAp48 overexpression is correlated with dephosphorylation of Akt, suggesting that RbAp48 may be exerting its effect by antagonizing the Ras pathway. The implications of our findings are significant. We establish that radiation sensitivity can be predicted based on gene expression profiles and we introduce a genomic approach to the identification of novel molecular markers of radiation sensitivity.
- Radiation Sensitivity
- Genomic Expression
- Leave-one-out cross-validation
Radiation therapy has played a major role in cancer therapeutics since its discovery more than 100 years ago. It is estimated that in the United States and Europe, more than 1 million individuals receive radiation therapy every year as part of their cancer treatment ( 1). Radiation therapy can be curative, particularly in prostate cancer, head and neck cancer, and cervical cancer, in which cure rates with definitive radiation therapy are comparable with those achieved with radical surgery. It also plays an important role in multimodality organ conservation protocols as in breast cancer, rectal cancer, soft tissue sarcoma, and laryngeal cancer, in which cure rates have been equaled to those achieved with radical surgery with the advantage of organ function preservation.
The development of a radiation sensitivity predictive assay has been a central goal of radiation biology for several decades ( 2). The clinical value of a radiation sensitivity predictive assay is significant, as it could potentially lead to better selection of patients for radiotherapy protocols, an improved ability to predict prognosis, and perhaps a decrease in therapy-related side effects. Although a number of approaches have been developed (tumor hypoxic fraction prediction, tumor proliferative potential determination, and determination of tumor inherent radiosensitivity), none have become routine in the clinic ( 3). There are several reasons for failure; however, it could be argued that a major reason is our limited knowledge of the molecular events that mediate radiation response.
The advent of the functional genomics era has had a significant effect on clinical oncology ( 4). It is now fairly well established that genomic profiling can distinguish clinical prognostic groups in breast cancer ( 5– 8), lymphoma ( 9, 10), prostate cancer ( 11), and other malignancies ( 12– 16). Furthermore, it has been proposed that genomic profiling may lead to the development of classifiers that will predict prognosis based on genomic expression ( 17). Because classifiers correlate gene expression profiles with an end point, we hypothesized that a radiation sensitivity classifier or predictor could be developed based on cellular gene expression profiles derived from DNA microarrays. This hypothesis is based partly on the fact that the three main mechanisms, which have been proposed to play a role in clinical failure after radiation therapy (hypoxia, intrinsic radiosensitivity, and cellular proliferation), are known to induce genetic change ( 18– 24). Furthermore, we reasoned that the development of a genomic-based radiation sensitivity classifier could potentially lead to the identification of novel genes or targets involved in radiation response. If validated, the clinical potential of such an analytic approach would be significant, as it would allow a genome-wide high-throughput analysis of potential targets involved in radiation response.
Materials and Methods
Cell lines. Cell lines were obtained directly from the National Cancer Institute (NCI). Cells were cultured as recommended by the NCI in RPMI 1640 supplemented with glutamine (2 mmol/L), antibiotics (penicillin/streptomycin, 10 units/mL), and heat-inactivated fetal bovine serum (10%) at 37°C with an atmosphere of 5% CO2.
Radiation survival assays (survival fraction at 2 Gy). The survival fraction at 2 Gy (SF2) of cell lines used in the classifier was obtained from the literature in 23 of the 35 cell lines in our analysis. For cell lines obtained from the literature, we used articles that reported on clonogenic assays that had been done without the use of any substrate (i.e., agar) and that required cells to be in log phase at the time of irradiation. We also required cell lines to have at least two reported values in the literature by different laboratories. We then determined the mean SF2 reported for each cell line and used this value for the generation of the classifier. The remaining 12 cell lines were done in our lab (MCF-7, MDA-MB-435, KM-12, HOP62, H23, BT549, MDA-MB-231, HCT-116, HT29, H460, OVCAR5, and PC3). Clonogenic survival assays after 2 Gy of radiation were done as previously described ( 19). Plating efficiency for each cell line was determined before SF2 determination. Cells were plated so that 50 to 100 colonies would form per plate and incubated overnight at 37°C to allow for adherence. Cells were then radiated with 2 Gy using a Cesium Irradiator (J.L. Sheperd, Model I 68A, San Fernando, CA). Exposure time was adjusted for decay every 3 months. After irradiation, cells were incubated for 10 to 14 days at 37°C before being stained with crystal violet. Only colonies with at least 50 cells were counted. SF2 was determined by the following formula:
Microarrays. Gene expression profiles were from Affymetrix HU-6800 chips (7,129 genes) from a previously published study ( 25). These are publicly available as supplementary data to the published study. The gene expression data had been previously preprocessed using the Affymetrix MAS 4.0 algorithm in average difference units. Negative expression values were set to zero and the chips were normalized to the same mean intensity.
RNA isolation, cDNA synthesis, and quantitative real-time PCR. RNA isolation was done as previously described ( 26). Total RNA was extracted using TRIzol reagent (Life Technologies, Gaithersburg, MD) as recommended by the manufacturer. Briefly, cell lines were resuspended in 1 mL of TRIzol and incubated for 5 minutes at room temperature. After adding 0.25 mL of chloroform, samples were centrifuged at 12,000 × g for 15 minutes at 4°C. The aqueous phase was recovered, and RNA was precipitated with isopropanol (0.5 mL/mL of TRIzol used). The RNA pellet was washed with ice-cold 70% ethanol once and resuspended in diethylpyrocarbonate-treated water.
First-strand cDNA synthesis was done using the Omniscript RT kit (Qiagen, Valencia, CA) as recommended by the manufacturer. In brief, 2 μg of RNA were incubated with 1 μg of Oligo-dT primers in the presence of 500 nmol/L deoxynucleotide triphosphate, 10 units of RNase inhibitor (Qiagen), and 10 units of Omniscript RT in 1× reverse transcription buffer (Qiagen).
Sequencing primers and probes were designed using the PrimerSelect software (Applied Biosystems, Foster City, CA). Each primer set consisted of standard PCR primers designed to span gene exons to exclude any possible genomic DNA contamination. Detection and quantitation of each gene was accomplished using an amplicon-specific fluorescent oligonucleotide probe with a 5′-reporter dye (FAM) and a downstream 3′-quencher dye (BHQ-1). For the construction of standard curves, a cDNA serial dilution was used. Quantitative real-time PCR analyses were done using the ABI PRISM 7700 Sequence Detection System (Applied Biosystems). All samples were tested in 96-well optical reaction plates. PCR was carried out with the Platinum Quantitative PCR SuperMix-UDG (Invitrogen, Carlsbad, CA) using 2 μL of cDNA, 1× primers, and probe in a 25-μL final reaction mixture. After a 2-minute incubation at 50°C, PlatinumTaq was activated by a 10-minute incubation at 95°C followed by 40 PCR cycles consisting of 15 seconds of denaturation at 95°C and hybridization of probe and primers for 1 minute at 60°C. Data were analyzed using SDS software version 2.1 and exported into an Excel spreadsheet. Assays were done in triplicate and the housekeeping gene GAPDH was used to normalize mRNA values.
The sequences for each primer and probe were as follows: ribose 5-phosphate isomerase A (NM_144563)—forward primer: 5′-GAACCTCGTCTGTATTCCCACTTC-3′, reverse primer: 5′-GGGTCAAGCCATACTGCAGG-3′, probe: 5′-FAM/CGTGTCCGGGAGGGCATCAA/BHQ-3′; G-protein signaling regulator 19 (NM_005873)—forward primer: 5′-CAAGGAGGTGAGCCTGGACTC-3′, reverse primer: 5′-GACGGCTCCTGCATCTTCTT-3′, probe: 5′-FAM/CGTGTCCGGGAGGGCATCAA/BHQ-3′; and retinoblastoma binding protein 4 (RbAp48; NM_005610)—forward primer: 5′-TCAAGATCAACCATGAAGGAGAAGT-3′, reverse primer: 5′-TTTGTTGCGATGATACAAGGGT-3′, probe: 5′-FAM/AACAGGGCCCGTTATATGCCCCAG/BHQ-3′.
Western blots. Assays were done as previously described by our group ( 27). Protein was isolated using lysis buffer which included radioimmunoprecipitation assay buffer [50 mmol/L Tris (pH 7.5), 150 mmol/L NaCl, 0.5% deoxycholate, 1% NP40, 2 mmol/L sodium vanadate], protease inhibitors (leupeptin, antipain, and aprotinin), and phenylmethylsulfonyl fluoride (1 mmol/L). Protein was quantitated using the Bradford assay. Ten micrograms of protein were loaded onto an SDS-PAGE gel, transferred onto nitrocellulose, and probed with the antibody of interest. Anti-RbAp48 was obtained from Upstate (Lake Placid, NY). Total Akt and P-Akt (serine-473) were obtained from Cell Signaling Technology (Beverly, MA). Membranes were then washed, reprobed with a horseradish peroxidase-conjugated secondary antibody (Amersham Biosciences, Piscataway, NJ), and developed with enhanced chemiluminescence (Amersham).
Cloning and transfection of RbAp48. PCR amplification of the cDNA was done using forward primer 5′-atggccgacaaggaagcagc-3′ and reverse primer 5′-ctaggacccttgtccttctggat-3′. These primers generated a product of 1,278 bp. The PCR fragment was gel purified using Gel Extraction Kit (Qiagen) and cloned into pcDNA 3.1/V5-His-TOPO vector containing a neomycin-resistant gene (Invitrogen). Ultracompetent cells were transformed with the RbAp48 plasmid and an empty vector (mock). Selection was done with ampicillin. Clones were sequenced to determine that the desired sequences were present and in-frame. Stable RbAp48 transfectants were generated using the Superfect Reagent (Qiagen) following the recommendations of the manufacturer. G418 was used to select transfected colonies. Western blots were done to show overexpression of RbAp48.
Cell cycle analysis. Cells (1 × 106) were fixed in 70% ethanol at −20°C overnight. Cells were then centrifuged at 1,000 rpm for 5 minutes and washed once in PBS. Cells were then resuspended in 1 mL of propidium iodide/Triton X-100 staining solution with RNase A (0.1% Triton X-100, 200 μg/mL DNase-free RNase A, 20 μg/mL propidium iodide) and stained for 30 minutes at room temperature. Analysis was done by flow cytometry at the Flow Cytometry Core at the H. Lee Moffitt Cancer Center and Research Institute.
Statistical analysis and classifier design. We first defined the classification problem by considering the end point variable. Radiation sensitivity in this study was defined as a continuous variable.
The classifier consisted of two distinct components: gene selection and a multivariate linear regression model. The gene selection step was done using the significance analysis of microarrays (SAM) method. SAM tests the correlation between a continuous target variable (radiation response) and the expression levels for a gene. The method involves testing each of the genes one at a time and then performing permutation tests to estimate the false discovery rate that arises from individual testing of 7,129 genes. We selected the number of significant genes corresponding to a target false discovery rate of 5%.
The second step in constructing the classifier is the multivariate linear regression model. Genes selected by the SAM method were used to construct a predictive model of the SF2 values in the training set. Linear regression was used, as implemented by the Weka data mining software ( 28).
Evaluation of the classifier was done using leave-one-out cross-validation. This process involves repeatedly holding out a single sample from the data set, constructing a classifier on the remaining data, and using the classifier to predict the SF2 level of the held-out sample. An estimate of overall accuracy can be obtained from this approach. Leave-one-out cross-validation can also be used as a method of estimating the overall usefulness of selected genes in the classification. We adopted the technique described by Dyrskjot et al. ( 29), in which genes that are selected in most of the classifiers constructed (across all 35 leave-one-out iterations) are most likely to be informative.
A correct prediction was defined as any value within ±10% of the measured or reported SF2. This was intended to account for the known variability associated with clonogenic assays, which had been previously reported by Peters to be ±9% ( 2). However, in 3 of 12 cell lines measured in our laboratory (MDAMB-231, OVCAR5, and PC3), the range of measured SF2s was slightly higher, and thus for these three lines we accepted as correct any value within the actual measured range. Finally, the significance of the classifier results was assessed using a permutation test.
Development of a radiation classifier in the National Cancer Institute panel of 60 cell lines. The general scheme of our classifier is shown in Fig. 1 . In our approach, we used a subset of 35 cell lines from the NCI panel of 60 for which radiation sensitivity data as defined by survival fraction after 2 Gy (SF2) were either published in the literature (23 cell lines) or determined in our lab (12 cell lines). Radiation sensitivity (SF2) was defined as a continuous variable. The baseline genomic expression for each of these cell lines was obtained from the study by Stauton (Affymetrix HU-6800 microarrays, 7,129 genes). We used a leave-one-out cross-validation approach, in which the classifier was developed using 34 of 35 cell lines as a training set, leaving one cell line as a test set. The basal gene expression profiles and the radiation sensitivity of all 34 cell lines in the training set were used to identify genes that were correlated with radiation response. This was done using SAM analysis ( 30). To limit the role that chance could have in our gene selection process, we chose the number of genes such that an estimated false discovery rate of 5% was achieved. This was done by estimating the null distribution for the tests through the use of permutations of the existing data. Genes were selected as significantly different from this empirical null distribution.
Genes selected by SAM were then combined as “radiosensitivity predictors” during the construction of the classifier. The classifier is a multivariate linear regression model describing the contribution of each gene to the predictive process. Therefore, this process added an additional level of gene filtering to our analysis, as we required genes to continue to be “predictive” when the expression of other potential predictive genes was taken into consideration. The generated classifier was then tested by the test sample. The gene expression profile of the test sample was provided to the classifier, which then predicted the radiation sensitivity (SF2) of the test sample using the linear regression relationship developed during training. The predicted SF2 was deemed correct if it was within ±10% of the actual reported average SF2 (for those cell lines of which SF2 was obtained from the literature) or within the range of measured SF2s in our own experiments (occasionally more than 10%, in 3 of 12 cell lines measured in our lab). This process was then repeated an additional 34 times so that each cell line served as the test sample once.
Identification of gene signature predicting radiosensitivity. Table 1A and B shows the results of leave-one-out cross-validation on our classifier. The classifier correctly predicted SF2 values in 22 of 35 (62%) the cell lines. Thirteen samples were incorrectly classified. Of these, two samples missed their cutoff by <5%. These were cell lines KM-12 (predicted SF2 = 0.56, true SF2 = 0.42, cutoff = 0.52) and SF-539 (predicted SF2 = 0.677, true SF2 = 0.82, cutoff = 0.72). All three leukemia cell lines in our list (HL-60, CCRF-CEM, and MOLT-4) were incorrectly classified as well. However, for two of them (HL-60 and CCRF-CEM), the predicted SF2 was a negative number (−0.014 and −0.133, respectively). Therefore, the classifier judged these two cell lines to be in the radiosensitive side of the spectrum, which is indeed true. However, this failed to meet the criteria that we originally defined and thus they were classified as incorrect predictions.
We then analyzed the genes selected during the classification. Table 2 shows the list of genes that were selected by SAM analysis and the classifier to be predictive of radiation response. The four top genes were consistently selected by our analysis through all or most of the 35 rounds of training. The genes include ribose 5-phosphate isomerase A (selected in all 35 rounds of training), retinoblastoma binding protein 4, also known as RbAp48 (selected in 34 of 35 rounds of training), G-protein signaling regulator 19 (selected in 34 of 35 rounds of training), and an unknown gene (Affymetrix ID HG4236-HT4506, selected in 33 of 35 rounds of training). Figure 2A to C shows the linear correlation between the expression values of each of the top predictive genes and radiation sensitivity. In general, for each of the selected genes, a higher expression value was correlated with a more radiosensitive phenotype.
We then asked the question of what was the likelihood that chance alone would be able to predict 22 of 35 samples correctly. Although, the likelihood of any predicted SF2 being correct was 20% for most samples, not all possible SF2 values (0.01-1.0) had the same probability of being predicted by the classifier. Therefore, we did a statistical test in which 10,000 permutations of incorrectly paired predicted SF2 and true SF2 were generated. Only twice in the 10,000 permutations were 22 samples correctly classified by chance alone (P = 0.0002), further supporting the validity of our approach.
Quantitative PCR validates in silico genomic analysis. To validate our in silico genomic analysis, we determined the expression level of the three known genes selected by our analysis using quantitative real-time PCR. To perform these experiments of validation, for each candidate gene we selected cell lines that were on opposing ends of the radiation sensitivity and gene expression spectrum. As shown in Fig. 2A to C, there was excellent correlation between quantitative real-time PCR and microarrays for all three known genes selected by our analysis. A total of 26 cell lines were assayed (10 lines for RGS19, 8 lines for RbAp48, and 8 lines for R5PIA). Quantitative real-time PCR gene expression values for the gene of interest in 22 of the 26 cell lines fell within the expected range indicated by the Affymetrix HU-6800 chip.
Overexpression of RbAp48 induces radiosensitization. Because we were interested in exploring whether this analysis yielded biologically meaningful data, we decided to focus further experiments on RbAp48, one of the predictive genes identified by our classifier. RbAp48, a WD-40 protein, was first identified as a ubiquitous binding partner for the retinoblastoma protein ( 31). It has been proposed that RbAp48 is required for transcriptional repression of E2F-regulated genes ( 32). Furthermore, MSI1, an RbAp48 homologue in Saccharomyces cerevisiae, was shown to be a negative regulator of Ras ( 31). Because both proliferation and Ras have been implicated in pathways mediating radiation resistance, we hypothesized that RbAp48 could be mechanistically linked to radiation sensitivity.
We initially determined whether protein expression of RbAp48 mirrored its RNA expression. We did Western blots in the same group of cell lines that we had selected for quantitative real-time PCR. Figure 3 shows that RbAp48 protein expression correlated with its RNA expression. The radiosensitive cell lines (MOLT-4, CCRF-CEM, HCT-116, and KM-12) had a higher expression of the protein than radioresistant cell lines (MALME-3M, MDA-MB-231, H460, and HS-578T). To determine whether altering the expression level of RbAp48 would affect the radiophenotype of a cell line, we transfected RbAp48 into HS-578T, MDA-MB-231, and MALME-3M cells, two breast cancer and a melanoma cell line. As shown in Fig. 4A , overexpression of RbAp48 induced radiosensitization of all three cell lines when compared with an empty vector control. To further expand this observation, we did cell cycle analysis of HS-578T-RbAp48 overexpressors. As shown in Fig. 4B, a higher proportion (27%) of HS-578T-RbAp48 overexpressors were found in the G2-M, the most radiosensitive phase of the cell cycle, when compared with empty vector and wild-type controls (5% and 14%, respectively). Finally, because the Ras pathway has been proposed to play a role in radiation resistance through phosphorylation of Akt, we asked whether overexpression of RbAp48, a Ras antagonist, would alter the level of Akt phosphorylation. Figure 4C shows that overexpression of RbAp48 was correlated with dephosphorylation of Akt, suggesting that it achieves its radiosensitizing effect by opposing the Ras pathway. Taken together, these observations support a mechanistic link between RbAp48 and radiation sensitivity, thus validating the biological value of the classifier.
In this article, we describe the development of the first biologically validated radiation sensitivity classifier. In our approach, we developed a classifier that predicts the inherent radiosensitivity of tumor cell lines as measured by SF2, based on gene expression profiles. This was done in a subset of 35 cell lines from the NCI panel of 60 of which expression profiles (Affymetrix HU-6800) were available in the literature. The classifier correctly predicted the SF2 value in 22 of 35 cell lines from the NCI panel of 60, a result significantly different from chance (P = 0.0002). Furthermore, our gene selection step consistently chose three known genes and one unknown gene as correlated with radiosensitivity prediction. The expression of these selected predictive genes was validated by quantitative real-time PCR, in which excellent correlation between expression values, as measured by quantitative real-time PCR and microarrays, was shown. Importantly, overexpression of RbAp48, one of the predictive genes selected by the classifier, induced radiosensitization in three cancer cell lines (Hs-578T, MALME-3M, and MDA-MB-231), thus biologically validating our novel approach to genomic analysis.
The implications of our findings are significant. Although genomic profiling has been successfully used to distinguish prognostic groups in breast cancer ( 5– 8), lymphoma ( 9, 10), prostate cancer ( 11), and other malignancies ( 12– 16), the use of microarray technology in the process of gene discovery has been hindered by the complexities inherent to the analysis of microarray experiments. We chose to use several different methods for filtering the data, including selecting predictive genes identified by SAM analysis and building a classifier to measure their predictive value as gene expression profiles were correlated with radiation sensitivity. We then identified those genes most commonly selected by the accurate classifier models as candidate genes for further study.
In our algorithm, potential predictive genes are initially identified by SAM analysis, as gene expression profiles are correlated with radiation sensitivity. In this step, radiation sensitivity was defined as a continuous variable; therefore, the algorithm identified genes that correlated with radiation response throughout the radiation sensitivity spectrum. Other published classifiers in the literature have defined their end point in a binary fashion ( 25). The problem with this approach is that it requires two classes to be defined (i.e., radiosensitive versus radioresistant), and thus an artificial threshold bias is introduced in the algorithm.
It should be noted that different classifiers were constructed for each round of the leave-one-out approach because different iterations produced different genes and classifiers. Therefore, a “consensus classifier” could be constructed using the genes most commonly selected during each round of the leave-one-out cross-validation. Therefore, the knowledge developed from this approach (predictive genes, classifier design, and freely available data set) would allow others to test new data against the classifier.
Our approach to radiation sensitivity gene analysis consistently selected three known genes and one unknown gene as predictive of radiation sensitivity. These genes include ribose 5-phosphate isomerase A, an enzyme involved in the pentose-phosphate pathway ( 33), retinoblastoma binding protein 4, also known as RbAp48 ( 34), G-protein signaling regulator 19 ( 35), and an unknown gene that encodes a zinc finger protein. Importantly, we validated the genomic expression for each of the known genes derived from microarrays by performing quantitative RT-PCR. This was critical in our approach because the genomic expression was obtained from the literature and further argues for the validity of in silico experiments. However, the key observation that biologically validates the analytic approach presented in this work is the identification of RbAp48 as a novel gene involved in radiation response. We showed that overexpression of RbAp48 induced radiosensitization of three cell lines (HS-578T, MALME-3M, and MDA-MB-231) when compared with empty vector controls. We further expanded this observation by showing that HS-578T-RbAp48 overexpressors had a higher proportion of cells in the G2-M phase of the cell cycle than empty vector controls and that overexpression of RbAp48 was correlated with dephosphorylation of Akt. All these observations, taken together, strongly support the mechanistic involvement of RbAp48 in radiation response and validate biologically our genomic radiation sensitivity classifier.
The observation of RbAp48-induced radiosensitization raises several questions about the mechanism mediating this effect. One potential hypothesis is that it may be exerting its effect by antagonizing Ras, which has been proposed to be a central mediator of radiation resistance through the Akt pathway ( 19, 20). Therefore, our observation showing dephosphorylation of Akt in RbAp48 overexpressors is consistent with this hypothesis. Another potential mechanism may involve chromatin assembly and remodeling. RbAp48 is part of chromatin assembly factor 1 ( 36), histone deacetylase complex 1 ( 32), histone deacetylase complex 3 ( 37), and histone acetylase transferase 1 ( 38). Therefore, RbAp48 seems to play a role in nearly every stage of chromatin metabolism from assembly and remodeling to modification. There is a growing body of literature that supports the involvement of inhibitors of histone deacetylation as radiation sensitizers ( 39– 42) and, therefore, this is another potential mechanism by which RbAp48 could be exerting its effect.
This work has several clinical implications. We establish for the first time that radiation sensitivity is predictable based on gene expression profiles. However, how to translate this idea into the clinic is debatable. One of the reasons for working in cell lines initially was that we could address radiation sensitivity isolated from other patient factors. However, radiation sensitivity is only one of the variables that determine whether a tumor is radiocurable. Furthermore, the value of radiation sensitivity tests in vitro and its correlation with clinical outcome are controversial ( 43– 48). However, it is reasonable to propose the use of this technology as a genomic approach to the identification of novel molecular markers of radiation sensitivity. Additional markers besides those discussed in this work could be identified by increasing the total number of genes analyzed in the microarray from 7,129 (Affymetrix HU-6800) to 40,000 transcripts (Affymetrix U-133 Plus) or by analyzing genes up-regulated or down-regulated after treatment with radiation. These markers would need to be validated in clinical trials and could then be used in combination with other clinical factors to further improve our ability to define clinical prognostic groups.
In conclusion, we have developed the first biologically validated radiation sensitivity genomic classifier. We introduce a novel approach to radiation sensitivity predictive gene analysis that incorporates a multivariate predictive step and defines radiation sensitivity as a continuous variable. Overexpression of RbAp48, one of the genes selected by our approach, led to radiosensitization of three cell lines (HS-578T, MALME-3M, and MDA-MB-231) as predicted by the classifier. We propose the use of this technology as a genomic approach to the identification of novel molecular markers of radiation sensitivity.
Grant support: National Cancer Institute (Bethesda, MD) grants K08 CA108926-01 (J. Torres-Roca), R01-CA098522-01, K-24-CA85429-04, and U01-CA85052-02 (T. Yeatman), and in part by the Molecular Biology and Flow Cytometry Core Facility at the H. Lee Moffitt Cancer Center and Research Institute.
The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.
- Received February 25, 2005.
- Revision received April 29, 2005.
- Accepted June 9, 2005.
- ©2005 American Association for Cancer Research.